5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
84 EXTERN_C const struct regexp_engine my_reg_engine;
89 #include "dquote_static.c"
90 #include "charclass_invlists.h"
91 #include "inline_invlist.c"
92 #include "unicode_constants.h"
94 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
95 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
96 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
97 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
98 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
99 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
102 #define STATIC static
106 #define MIN(a,b) ((a) < (b) ? (a) : (b))
109 /* this is a chain of data about sub patterns we are processing that
110 need to be handled separately/specially in study_chunk. Its so
111 we can simulate recursion without losing state. */
113 typedef struct scan_frame {
114 regnode *last_regnode; /* last node to process in this frame */
115 regnode *next_regnode; /* next node to process when last is reached */
116 U32 prev_recursed_depth;
117 I32 stopparen; /* what stopparen do we use */
118 U32 is_top_frame; /* what flags do we use? */
120 struct scan_frame *this_prev_frame; /* this previous frame */
121 struct scan_frame *prev_frame; /* previous frame */
122 struct scan_frame *next_frame; /* next frame */
125 struct RExC_state_t {
126 U32 flags; /* RXf_* are we folding, multilining? */
127 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
128 char *precomp; /* uncompiled string. */
129 REGEXP *rx_sv; /* The SV that is the regexp. */
130 regexp *rx; /* perl core regexp structure */
131 regexp_internal *rxi; /* internal data for regexp object
133 char *start; /* Start of input for compile */
134 char *end; /* End of input for compile */
135 char *parse; /* Input-scan pointer. */
136 SSize_t whilem_seen; /* number of WHILEM in this expr */
137 regnode *emit_start; /* Start of emitted-code area */
138 regnode *emit_bound; /* First regnode outside of the
140 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
141 implies compiling, so don't emit */
142 regnode_ssc emit_dummy; /* placeholder for emit to point to;
143 large enough for the largest
144 non-EXACTish node, so can use it as
146 I32 naughty; /* How bad is this pattern? */
147 I32 sawback; /* Did we see \1, ...? */
149 SSize_t size; /* Code size. */
150 I32 npar; /* Capture buffer count, (OPEN) plus
151 one. ("par" 0 is the whole
153 I32 nestroot; /* root parens we are in - used by
157 regnode **open_parens; /* pointers to open parens */
158 regnode **close_parens; /* pointers to close parens */
159 regnode *opend; /* END node in program */
160 I32 utf8; /* whether the pattern is utf8 or not */
161 I32 orig_utf8; /* whether the pattern was originally in utf8 */
162 /* XXX use this for future optimisation of case
163 * where pattern must be upgraded to utf8. */
164 I32 uni_semantics; /* If a d charset modifier should use unicode
165 rules, even if the pattern is not in
167 HV *paren_names; /* Paren names */
169 regnode **recurse; /* Recurse regops */
170 I32 recurse_count; /* Number of recurse regops */
171 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
173 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
177 I32 override_recoding;
178 I32 in_multi_char_class;
179 struct reg_code_block *code_blocks; /* positions of literal (?{})
181 int num_code_blocks; /* size of code_blocks[] */
182 int code_index; /* next code_blocks[] slot */
183 SSize_t maxlen; /* mininum possible number of chars in string to match */
184 scan_frame *frame_head;
185 scan_frame *frame_last;
187 #ifdef ADD_TO_REGEXEC
188 char *starttry; /* -Dr: where regtry was called. */
189 #define RExC_starttry (pRExC_state->starttry)
191 SV *runtime_code_qr; /* qr with the runtime code blocks */
193 const char *lastparse;
195 AV *paren_name_list; /* idx -> name */
196 U32 study_chunk_recursed_count;
199 #define RExC_lastparse (pRExC_state->lastparse)
200 #define RExC_lastnum (pRExC_state->lastnum)
201 #define RExC_paren_name_list (pRExC_state->paren_name_list)
202 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
203 #define RExC_mysv (pRExC_state->mysv1)
204 #define RExC_mysv1 (pRExC_state->mysv1)
205 #define RExC_mysv2 (pRExC_state->mysv2)
210 #define RExC_flags (pRExC_state->flags)
211 #define RExC_pm_flags (pRExC_state->pm_flags)
212 #define RExC_precomp (pRExC_state->precomp)
213 #define RExC_rx_sv (pRExC_state->rx_sv)
214 #define RExC_rx (pRExC_state->rx)
215 #define RExC_rxi (pRExC_state->rxi)
216 #define RExC_start (pRExC_state->start)
217 #define RExC_end (pRExC_state->end)
218 #define RExC_parse (pRExC_state->parse)
219 #define RExC_whilem_seen (pRExC_state->whilem_seen)
220 #ifdef RE_TRACK_PATTERN_OFFSETS
221 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
224 #define RExC_emit (pRExC_state->emit)
225 #define RExC_emit_dummy (pRExC_state->emit_dummy)
226 #define RExC_emit_start (pRExC_state->emit_start)
227 #define RExC_emit_bound (pRExC_state->emit_bound)
228 #define RExC_sawback (pRExC_state->sawback)
229 #define RExC_seen (pRExC_state->seen)
230 #define RExC_size (pRExC_state->size)
231 #define RExC_maxlen (pRExC_state->maxlen)
232 #define RExC_npar (pRExC_state->npar)
233 #define RExC_nestroot (pRExC_state->nestroot)
234 #define RExC_extralen (pRExC_state->extralen)
235 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
236 #define RExC_utf8 (pRExC_state->utf8)
237 #define RExC_uni_semantics (pRExC_state->uni_semantics)
238 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
239 #define RExC_open_parens (pRExC_state->open_parens)
240 #define RExC_close_parens (pRExC_state->close_parens)
241 #define RExC_opend (pRExC_state->opend)
242 #define RExC_paren_names (pRExC_state->paren_names)
243 #define RExC_recurse (pRExC_state->recurse)
244 #define RExC_recurse_count (pRExC_state->recurse_count)
245 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
246 #define RExC_study_chunk_recursed_bytes \
247 (pRExC_state->study_chunk_recursed_bytes)
248 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
249 #define RExC_contains_locale (pRExC_state->contains_locale)
250 #define RExC_contains_i (pRExC_state->contains_i)
251 #define RExC_override_recoding (pRExC_state->override_recoding)
252 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
253 #define RExC_frame_head (pRExC_state->frame_head)
254 #define RExC_frame_last (pRExC_state->frame_last)
255 #define RExC_frame_count (pRExC_state->frame_count)
257 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
258 * a flag to disable back-off on the fixed/floating substrings - if it's
259 * a high complexity pattern we assume the benefit of avoiding a full match
260 * is worth the cost of checking for the substrings even if they rarely help.
262 #define RExC_naughty (pRExC_state->naughty)
263 #define TOO_NAUGHTY (10)
264 #define MARK_NAUGHTY(add) \
265 if (RExC_naughty < TOO_NAUGHTY) \
266 RExC_naughty += (add)
267 #define MARK_NAUGHTY_EXP(exp, add) \
268 if (RExC_naughty < TOO_NAUGHTY) \
269 RExC_naughty += RExC_naughty / (exp) + (add)
271 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
272 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
273 ((*s) == '{' && regcurly(s)))
276 * Flags to be passed up and down.
278 #define WORST 0 /* Worst case. */
279 #define HASWIDTH 0x01 /* Known to match non-null strings. */
281 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
282 * character. (There needs to be a case: in the switch statement in regexec.c
283 * for any node marked SIMPLE.) Note that this is not the same thing as
286 #define SPSTART 0x04 /* Starts with * or + */
287 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
288 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
289 #define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
291 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
293 /* whether trie related optimizations are enabled */
294 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
295 #define TRIE_STUDY_OPT
296 #define FULL_TRIE_STUDY
302 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
303 #define PBITVAL(paren) (1 << ((paren) & 7))
304 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
305 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
306 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
308 #define REQUIRE_UTF8 STMT_START { \
310 *flagp = RESTART_UTF8; \
315 /* This converts the named class defined in regcomp.h to its equivalent class
316 * number defined in handy.h. */
317 #define namedclass_to_classnum(class) ((int) ((class) / 2))
318 #define classnum_to_namedclass(classnum) ((classnum) * 2)
320 #define _invlist_union_complement_2nd(a, b, output) \
321 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
322 #define _invlist_intersection_complement_2nd(a, b, output) \
323 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
325 /* About scan_data_t.
327 During optimisation we recurse through the regexp program performing
328 various inplace (keyhole style) optimisations. In addition study_chunk
329 and scan_commit populate this data structure with information about
330 what strings MUST appear in the pattern. We look for the longest
331 string that must appear at a fixed location, and we look for the
332 longest string that may appear at a floating location. So for instance
337 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
338 strings (because they follow a .* construct). study_chunk will identify
339 both FOO and BAR as being the longest fixed and floating strings respectively.
341 The strings can be composites, for instance
345 will result in a composite fixed substring 'foo'.
347 For each string some basic information is maintained:
349 - offset or min_offset
350 This is the position the string must appear at, or not before.
351 It also implicitly (when combined with minlenp) tells us how many
352 characters must match before the string we are searching for.
353 Likewise when combined with minlenp and the length of the string it
354 tells us how many characters must appear after the string we have
358 Only used for floating strings. This is the rightmost point that
359 the string can appear at. If set to SSize_t_MAX it indicates that the
360 string can occur infinitely far to the right.
363 A pointer to the minimum number of characters of the pattern that the
364 string was found inside. This is important as in the case of positive
365 lookahead or positive lookbehind we can have multiple patterns
370 The minimum length of the pattern overall is 3, the minimum length
371 of the lookahead part is 3, but the minimum length of the part that
372 will actually match is 1. So 'FOO's minimum length is 3, but the
373 minimum length for the F is 1. This is important as the minimum length
374 is used to determine offsets in front of and behind the string being
375 looked for. Since strings can be composites this is the length of the
376 pattern at the time it was committed with a scan_commit. Note that
377 the length is calculated by study_chunk, so that the minimum lengths
378 are not known until the full pattern has been compiled, thus the
379 pointer to the value.
383 In the case of lookbehind the string being searched for can be
384 offset past the start point of the final matching string.
385 If this value was just blithely removed from the min_offset it would
386 invalidate some of the calculations for how many chars must match
387 before or after (as they are derived from min_offset and minlen and
388 the length of the string being searched for).
389 When the final pattern is compiled and the data is moved from the
390 scan_data_t structure into the regexp structure the information
391 about lookbehind is factored in, with the information that would
392 have been lost precalculated in the end_shift field for the
395 The fields pos_min and pos_delta are used to store the minimum offset
396 and the delta to the maximum offset at the current point in the pattern.
400 typedef struct scan_data_t {
401 /*I32 len_min; unused */
402 /*I32 len_delta; unused */
406 SSize_t last_end; /* min value, <0 unless valid. */
407 SSize_t last_start_min;
408 SSize_t last_start_max;
409 SV **longest; /* Either &l_fixed, or &l_float. */
410 SV *longest_fixed; /* longest fixed string found in pattern */
411 SSize_t offset_fixed; /* offset where it starts */
412 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
413 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
414 SV *longest_float; /* longest floating string found in pattern */
415 SSize_t offset_float_min; /* earliest point in string it can appear */
416 SSize_t offset_float_max; /* latest point in string it can appear */
417 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
418 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
421 SSize_t *last_closep;
422 regnode_ssc *start_class;
426 * Forward declarations for pregcomp()'s friends.
429 static const scan_data_t zero_scan_data =
430 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
432 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
433 #define SF_BEFORE_SEOL 0x0001
434 #define SF_BEFORE_MEOL 0x0002
435 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
436 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
438 #define SF_FIX_SHIFT_EOL (+2)
439 #define SF_FL_SHIFT_EOL (+4)
441 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
442 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
444 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
445 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
446 #define SF_IS_INF 0x0040
447 #define SF_HAS_PAR 0x0080
448 #define SF_IN_PAR 0x0100
449 #define SF_HAS_EVAL 0x0200
450 #define SCF_DO_SUBSTR 0x0400
451 #define SCF_DO_STCLASS_AND 0x0800
452 #define SCF_DO_STCLASS_OR 0x1000
453 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
454 #define SCF_WHILEM_VISITED_POS 0x2000
456 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
457 #define SCF_SEEN_ACCEPT 0x8000
458 #define SCF_TRIE_DOING_RESTUDY 0x10000
459 #define SCF_IN_DEFINE 0x20000
464 #define UTF cBOOL(RExC_utf8)
466 /* The enums for all these are ordered so things work out correctly */
467 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
468 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
469 == REGEX_DEPENDS_CHARSET)
470 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
471 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
472 >= REGEX_UNICODE_CHARSET)
473 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
474 == REGEX_ASCII_RESTRICTED_CHARSET)
475 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
476 >= REGEX_ASCII_RESTRICTED_CHARSET)
477 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
478 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
480 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
482 /* For programs that want to be strictly Unicode compatible by dying if any
483 * attempt is made to match a non-Unicode code point against a Unicode
485 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
487 #define OOB_NAMEDCLASS -1
489 /* There is no code point that is out-of-bounds, so this is problematic. But
490 * its only current use is to initialize a variable that is always set before
492 #define OOB_UNICODE 0xDEADBEEF
494 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
495 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
498 /* length of regex to show in messages that don't mark a position within */
499 #define RegexLengthToShowInErrorMessages 127
502 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
503 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
504 * op/pragma/warn/regcomp.
506 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
507 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
509 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
510 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
512 #define REPORT_LOCATION_ARGS(offset) \
513 UTF8fARG(UTF, offset, RExC_precomp), \
514 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
517 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
518 * arg. Show regex, up to a maximum length. If it's too long, chop and add
521 #define _FAIL(code) STMT_START { \
522 const char *ellipses = ""; \
523 IV len = RExC_end - RExC_precomp; \
526 SAVEFREESV(RExC_rx_sv); \
527 if (len > RegexLengthToShowInErrorMessages) { \
528 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
529 len = RegexLengthToShowInErrorMessages - 10; \
535 #define FAIL(msg) _FAIL( \
536 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
537 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
539 #define FAIL2(msg,arg) _FAIL( \
540 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
541 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
544 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
546 #define Simple_vFAIL(m) STMT_START { \
548 (RExC_parse > RExC_end ? RExC_end : RExC_parse) - RExC_precomp; \
549 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
550 m, REPORT_LOCATION_ARGS(offset)); \
554 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
556 #define vFAIL(m) STMT_START { \
558 SAVEFREESV(RExC_rx_sv); \
563 * Like Simple_vFAIL(), but accepts two arguments.
565 #define Simple_vFAIL2(m,a1) STMT_START { \
566 const IV offset = RExC_parse - RExC_precomp; \
567 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
568 REPORT_LOCATION_ARGS(offset)); \
572 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
574 #define vFAIL2(m,a1) STMT_START { \
576 SAVEFREESV(RExC_rx_sv); \
577 Simple_vFAIL2(m, a1); \
582 * Like Simple_vFAIL(), but accepts three arguments.
584 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
585 const IV offset = RExC_parse - RExC_precomp; \
586 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
587 REPORT_LOCATION_ARGS(offset)); \
591 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
593 #define vFAIL3(m,a1,a2) STMT_START { \
595 SAVEFREESV(RExC_rx_sv); \
596 Simple_vFAIL3(m, a1, a2); \
600 * Like Simple_vFAIL(), but accepts four arguments.
602 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
603 const IV offset = RExC_parse - RExC_precomp; \
604 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
605 REPORT_LOCATION_ARGS(offset)); \
608 #define vFAIL4(m,a1,a2,a3) STMT_START { \
610 SAVEFREESV(RExC_rx_sv); \
611 Simple_vFAIL4(m, a1, a2, a3); \
614 /* A specialized version of vFAIL2 that works with UTF8f */
615 #define vFAIL2utf8f(m, a1) STMT_START { \
616 const IV offset = RExC_parse - RExC_precomp; \
618 SAVEFREESV(RExC_rx_sv); \
619 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
620 REPORT_LOCATION_ARGS(offset)); \
623 /* These have asserts in them because of [perl #122671] Many warnings in
624 * regcomp.c can occur twice. If they get output in pass1 and later in that
625 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
626 * would get output again. So they should be output in pass2, and these
627 * asserts make sure new warnings follow that paradigm. */
629 /* m is not necessarily a "literal string", in this macro */
630 #define reg_warn_non_literal_string(loc, m) STMT_START { \
631 const IV offset = loc - RExC_precomp; \
632 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
633 m, REPORT_LOCATION_ARGS(offset)); \
636 #define ckWARNreg(loc,m) STMT_START { \
637 const IV offset = loc - RExC_precomp; \
638 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
639 REPORT_LOCATION_ARGS(offset)); \
642 #define vWARN_dep(loc, m) STMT_START { \
643 const IV offset = loc - RExC_precomp; \
644 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
645 REPORT_LOCATION_ARGS(offset)); \
648 #define ckWARNdep(loc,m) STMT_START { \
649 const IV offset = loc - RExC_precomp; \
650 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
652 REPORT_LOCATION_ARGS(offset)); \
655 #define ckWARNregdep(loc,m) STMT_START { \
656 const IV offset = loc - RExC_precomp; \
657 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
659 REPORT_LOCATION_ARGS(offset)); \
662 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
663 const IV offset = loc - RExC_precomp; \
664 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
666 a1, REPORT_LOCATION_ARGS(offset)); \
669 #define ckWARN2reg(loc, m, a1) STMT_START { \
670 const IV offset = loc - RExC_precomp; \
671 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
672 a1, REPORT_LOCATION_ARGS(offset)); \
675 #define vWARN3(loc, m, a1, a2) STMT_START { \
676 const IV offset = loc - RExC_precomp; \
677 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
678 a1, a2, REPORT_LOCATION_ARGS(offset)); \
681 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
682 const IV offset = loc - RExC_precomp; \
683 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
684 a1, a2, REPORT_LOCATION_ARGS(offset)); \
687 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
688 const IV offset = loc - RExC_precomp; \
689 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
690 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
693 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
694 const IV offset = loc - RExC_precomp; \
695 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
696 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
699 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
700 const IV offset = loc - RExC_precomp; \
701 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
702 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
705 /* Macros for recording node offsets. 20001227 mjd@plover.com
706 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
707 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
708 * Element 0 holds the number n.
709 * Position is 1 indexed.
711 #ifndef RE_TRACK_PATTERN_OFFSETS
712 #define Set_Node_Offset_To_R(node,byte)
713 #define Set_Node_Offset(node,byte)
714 #define Set_Cur_Node_Offset
715 #define Set_Node_Length_To_R(node,len)
716 #define Set_Node_Length(node,len)
717 #define Set_Node_Cur_Length(node,start)
718 #define Node_Offset(n)
719 #define Node_Length(n)
720 #define Set_Node_Offset_Length(node,offset,len)
721 #define ProgLen(ri) ri->u.proglen
722 #define SetProgLen(ri,x) ri->u.proglen = x
724 #define ProgLen(ri) ri->u.offsets[0]
725 #define SetProgLen(ri,x) ri->u.offsets[0] = x
726 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
728 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
729 __LINE__, (int)(node), (int)(byte))); \
731 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
734 RExC_offsets[2*(node)-1] = (byte); \
739 #define Set_Node_Offset(node,byte) \
740 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
741 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
743 #define Set_Node_Length_To_R(node,len) STMT_START { \
745 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
746 __LINE__, (int)(node), (int)(len))); \
748 Perl_croak(aTHX_ "value of node is %d in Length macro", \
751 RExC_offsets[2*(node)] = (len); \
756 #define Set_Node_Length(node,len) \
757 Set_Node_Length_To_R((node)-RExC_emit_start, len)
758 #define Set_Node_Cur_Length(node, start) \
759 Set_Node_Length(node, RExC_parse - start)
761 /* Get offsets and lengths */
762 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
763 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
765 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
766 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
767 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
771 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
772 #define EXPERIMENTAL_INPLACESCAN
773 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
775 #define DEBUG_RExC_seen() \
776 DEBUG_OPTIMISE_MORE_r({ \
777 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
779 if (RExC_seen & REG_ZERO_LEN_SEEN) \
780 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
782 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
783 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
785 if (RExC_seen & REG_GPOS_SEEN) \
786 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
788 if (RExC_seen & REG_CANY_SEEN) \
789 PerlIO_printf(Perl_debug_log,"REG_CANY_SEEN "); \
791 if (RExC_seen & REG_RECURSE_SEEN) \
792 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
794 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
795 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
797 if (RExC_seen & REG_VERBARG_SEEN) \
798 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
800 if (RExC_seen & REG_CUTGROUP_SEEN) \
801 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
803 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
804 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
806 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
807 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
809 if (RExC_seen & REG_GOSTART_SEEN) \
810 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
812 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
813 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
815 PerlIO_printf(Perl_debug_log,"\n"); \
818 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
819 if ((flags) & flag) PerlIO_printf(Perl_debug_log, "%s ", #flag)
821 #define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
823 PerlIO_printf(Perl_debug_log, "%s", open_str); \
824 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
825 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
826 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
827 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
828 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
829 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
830 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
831 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
832 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
833 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
834 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
835 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
836 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
837 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
838 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
839 PerlIO_printf(Perl_debug_log, "%s", close_str); \
843 #define DEBUG_STUDYDATA(str,data,depth) \
844 DEBUG_OPTIMISE_MORE_r(if(data){ \
845 PerlIO_printf(Perl_debug_log, \
846 "%*s" str "Pos:%"IVdf"/%"IVdf \
848 (int)(depth)*2, "", \
849 (IV)((data)->pos_min), \
850 (IV)((data)->pos_delta), \
851 (UV)((data)->flags) \
853 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
854 PerlIO_printf(Perl_debug_log, \
855 " Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
856 (IV)((data)->whilem_c), \
857 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
858 is_inf ? "INF " : "" \
860 if ((data)->last_found) \
861 PerlIO_printf(Perl_debug_log, \
862 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
863 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
864 SvPVX_const((data)->last_found), \
865 (IV)((data)->last_end), \
866 (IV)((data)->last_start_min), \
867 (IV)((data)->last_start_max), \
868 ((data)->longest && \
869 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
870 SvPVX_const((data)->longest_fixed), \
871 (IV)((data)->offset_fixed), \
872 ((data)->longest && \
873 (data)->longest==&((data)->longest_float)) ? "*" : "", \
874 SvPVX_const((data)->longest_float), \
875 (IV)((data)->offset_float_min), \
876 (IV)((data)->offset_float_max) \
878 PerlIO_printf(Perl_debug_log,"\n"); \
883 /* is c a control character for which we have a mnemonic? */
884 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
887 S_cntrl_to_mnemonic(const U8 c)
889 /* Returns the mnemonic string that represents character 'c', if one
890 * exists; NULL otherwise. The only ones that exist for the purposes of
891 * this routine are a few control characters */
894 case '\a': return "\\a";
895 case '\b': return "\\b";
896 case ESC_NATIVE: return "\\e";
897 case '\f': return "\\f";
898 case '\n': return "\\n";
899 case '\r': return "\\r";
900 case '\t': return "\\t";
908 /* Mark that we cannot extend a found fixed substring at this point.
909 Update the longest found anchored substring and the longest found
910 floating substrings if needed. */
913 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
914 SSize_t *minlenp, int is_inf)
916 const STRLEN l = CHR_SVLEN(data->last_found);
917 const STRLEN old_l = CHR_SVLEN(*data->longest);
918 GET_RE_DEBUG_FLAGS_DECL;
920 PERL_ARGS_ASSERT_SCAN_COMMIT;
922 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
923 SvSetMagicSV(*data->longest, data->last_found);
924 if (*data->longest == data->longest_fixed) {
925 data->offset_fixed = l ? data->last_start_min : data->pos_min;
926 if (data->flags & SF_BEFORE_EOL)
928 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
930 data->flags &= ~SF_FIX_BEFORE_EOL;
931 data->minlen_fixed=minlenp;
932 data->lookbehind_fixed=0;
934 else { /* *data->longest == data->longest_float */
935 data->offset_float_min = l ? data->last_start_min : data->pos_min;
936 data->offset_float_max = (l
937 ? data->last_start_max
938 : (data->pos_delta > SSize_t_MAX - data->pos_min
940 : data->pos_min + data->pos_delta));
942 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
943 data->offset_float_max = SSize_t_MAX;
944 if (data->flags & SF_BEFORE_EOL)
946 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
948 data->flags &= ~SF_FL_BEFORE_EOL;
949 data->minlen_float=minlenp;
950 data->lookbehind_float=0;
953 SvCUR_set(data->last_found, 0);
955 SV * const sv = data->last_found;
956 if (SvUTF8(sv) && SvMAGICAL(sv)) {
957 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
963 data->flags &= ~SF_BEFORE_EOL;
964 DEBUG_STUDYDATA("commit: ",data,0);
967 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
968 * list that describes which code points it matches */
971 S_ssc_anything(pTHX_ regnode_ssc *ssc)
973 /* Set the SSC 'ssc' to match an empty string or any code point */
975 PERL_ARGS_ASSERT_SSC_ANYTHING;
977 assert(is_ANYOF_SYNTHETIC(ssc));
979 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
980 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
981 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
985 S_ssc_is_anything(const regnode_ssc *ssc)
987 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
988 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
989 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
990 * in any way, so there's no point in using it */
995 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
997 assert(is_ANYOF_SYNTHETIC(ssc));
999 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1003 /* See if the list consists solely of the range 0 - Infinity */
1004 invlist_iterinit(ssc->invlist);
1005 ret = invlist_iternext(ssc->invlist, &start, &end)
1009 invlist_iterfinish(ssc->invlist);
1015 /* If e.g., both \w and \W are set, matches everything */
1016 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1018 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1019 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1029 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1031 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1032 * string, any code point, or any posix class under locale */
1034 PERL_ARGS_ASSERT_SSC_INIT;
1036 Zero(ssc, 1, regnode_ssc);
1037 set_ANYOF_SYNTHETIC(ssc);
1038 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1041 /* If any portion of the regex is to operate under locale rules that aren't
1042 * fully known at compile time, initialization includes it. The reason
1043 * this isn't done for all regexes is that the optimizer was written under
1044 * the assumption that locale was all-or-nothing. Given the complexity and
1045 * lack of documentation in the optimizer, and that there are inadequate
1046 * test cases for locale, many parts of it may not work properly, it is
1047 * safest to avoid locale unless necessary. */
1048 if (RExC_contains_locale) {
1049 ANYOF_POSIXL_SETALL(ssc);
1052 ANYOF_POSIXL_ZERO(ssc);
1057 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1058 const regnode_ssc *ssc)
1060 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1061 * to the list of code points matched, and locale posix classes; hence does
1062 * not check its flags) */
1067 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1069 assert(is_ANYOF_SYNTHETIC(ssc));
1071 invlist_iterinit(ssc->invlist);
1072 ret = invlist_iternext(ssc->invlist, &start, &end)
1076 invlist_iterfinish(ssc->invlist);
1082 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1090 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1091 const regnode_charclass* const node)
1093 /* Returns a mortal inversion list defining which code points are matched
1094 * by 'node', which is of type ANYOF. Handles complementing the result if
1095 * appropriate. If some code points aren't knowable at this time, the
1096 * returned list must, and will, contain every code point that is a
1099 SV* invlist = sv_2mortal(_new_invlist(0));
1100 SV* only_utf8_locale_invlist = NULL;
1102 const U32 n = ARG(node);
1103 bool new_node_has_latin1 = FALSE;
1105 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1107 /* Look at the data structure created by S_set_ANYOF_arg() */
1108 if (n != ANYOF_ONLY_HAS_BITMAP) {
1109 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1110 AV * const av = MUTABLE_AV(SvRV(rv));
1111 SV **const ary = AvARRAY(av);
1112 assert(RExC_rxi->data->what[n] == 's');
1114 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1115 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1117 else if (ary[0] && ary[0] != &PL_sv_undef) {
1119 /* Here, no compile-time swash, and there are things that won't be
1120 * known until runtime -- we have to assume it could be anything */
1121 return _add_range_to_invlist(invlist, 0, UV_MAX);
1123 else if (ary[3] && ary[3] != &PL_sv_undef) {
1125 /* Here no compile-time swash, and no run-time only data. Use the
1126 * node's inversion list */
1127 invlist = sv_2mortal(invlist_clone(ary[3]));
1130 /* Get the code points valid only under UTF-8 locales */
1131 if ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
1132 && ary[2] && ary[2] != &PL_sv_undef)
1134 only_utf8_locale_invlist = ary[2];
1138 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1139 * code points, and an inversion list for the others, but if there are code
1140 * points that should match only conditionally on the target string being
1141 * UTF-8, those are placed in the inversion list, and not the bitmap.
1142 * Since there are circumstances under which they could match, they are
1143 * included in the SSC. But if the ANYOF node is to be inverted, we have
1144 * to exclude them here, so that when we invert below, the end result
1145 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1146 * have to do this here before we add the unconditionally matched code
1148 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1149 _invlist_intersection_complement_2nd(invlist,
1154 /* Add in the points from the bit map */
1155 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1156 if (ANYOF_BITMAP_TEST(node, i)) {
1157 invlist = add_cp_to_invlist(invlist, i);
1158 new_node_has_latin1 = TRUE;
1162 /* If this can match all upper Latin1 code points, have to add them
1164 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_NON_UTF8_NON_ASCII) {
1165 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1168 /* Similarly for these */
1169 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1170 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1173 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1174 _invlist_invert(invlist);
1176 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1178 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1179 * locale. We can skip this if there are no 0-255 at all. */
1180 _invlist_union(invlist, PL_Latin1, &invlist);
1183 /* Similarly add the UTF-8 locale possible matches. These have to be
1184 * deferred until after the non-UTF-8 locale ones are taken care of just
1185 * above, or it leads to wrong results under ANYOF_INVERT */
1186 if (only_utf8_locale_invlist) {
1187 _invlist_union_maybe_complement_2nd(invlist,
1188 only_utf8_locale_invlist,
1189 ANYOF_FLAGS(node) & ANYOF_INVERT,
1196 /* These two functions currently do the exact same thing */
1197 #define ssc_init_zero ssc_init
1199 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1200 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1202 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1203 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1204 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1207 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1208 const regnode_charclass *and_with)
1210 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1211 * another SSC or a regular ANYOF class. Can create false positives. */
1216 PERL_ARGS_ASSERT_SSC_AND;
1218 assert(is_ANYOF_SYNTHETIC(ssc));
1220 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1221 * the code point inversion list and just the relevant flags */
1222 if (is_ANYOF_SYNTHETIC(and_with)) {
1223 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1224 anded_flags = ANYOF_FLAGS(and_with);
1226 /* XXX This is a kludge around what appears to be deficiencies in the
1227 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1228 * there are paths through the optimizer where it doesn't get weeded
1229 * out when it should. And if we don't make some extra provision for
1230 * it like the code just below, it doesn't get added when it should.
1231 * This solution is to add it only when AND'ing, which is here, and
1232 * only when what is being AND'ed is the pristine, original node
1233 * matching anything. Thus it is like adding it to ssc_anything() but
1234 * only when the result is to be AND'ed. Probably the same solution
1235 * could be adopted for the same problem we have with /l matching,
1236 * which is solved differently in S_ssc_init(), and that would lead to
1237 * fewer false positives than that solution has. But if this solution
1238 * creates bugs, the consequences are only that a warning isn't raised
1239 * that should be; while the consequences for having /l bugs is
1240 * incorrect matches */
1241 if (ssc_is_anything((regnode_ssc *)and_with)) {
1242 anded_flags |= ANYOF_WARN_SUPER;
1246 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1247 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1250 ANYOF_FLAGS(ssc) &= anded_flags;
1252 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1253 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1254 * 'and_with' may be inverted. When not inverted, we have the situation of
1256 * (C1 | P1) & (C2 | P2)
1257 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1258 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1259 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1260 * <= ((C1 & C2) | P1 | P2)
1261 * Alternatively, the last few steps could be:
1262 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1263 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1264 * <= (C1 | C2 | (P1 & P2))
1265 * We favor the second approach if either P1 or P2 is non-empty. This is
1266 * because these components are a barrier to doing optimizations, as what
1267 * they match cannot be known until the moment of matching as they are
1268 * dependent on the current locale, 'AND"ing them likely will reduce or
1270 * But we can do better if we know that C1,P1 are in their initial state (a
1271 * frequent occurrence), each matching everything:
1272 * (<everything>) & (C2 | P2) = C2 | P2
1273 * Similarly, if C2,P2 are in their initial state (again a frequent
1274 * occurrence), the result is a no-op
1275 * (C1 | P1) & (<everything>) = C1 | P1
1278 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1279 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1280 * <= (C1 & ~C2) | (P1 & ~P2)
1283 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1284 && ! is_ANYOF_SYNTHETIC(and_with))
1288 ssc_intersection(ssc,
1290 FALSE /* Has already been inverted */
1293 /* If either P1 or P2 is empty, the intersection will be also; can skip
1295 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1296 ANYOF_POSIXL_ZERO(ssc);
1298 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1300 /* Note that the Posix class component P from 'and_with' actually
1302 * P = Pa | Pb | ... | Pn
1303 * where each component is one posix class, such as in [\w\s].
1305 * ~P = ~(Pa | Pb | ... | Pn)
1306 * = ~Pa & ~Pb & ... & ~Pn
1307 * <= ~Pa | ~Pb | ... | ~Pn
1308 * The last is something we can easily calculate, but unfortunately
1309 * is likely to have many false positives. We could do better
1310 * in some (but certainly not all) instances if two classes in
1311 * P have known relationships. For example
1312 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1314 * :lower: & :print: = :lower:
1315 * And similarly for classes that must be disjoint. For example,
1316 * since \s and \w can have no elements in common based on rules in
1317 * the POSIX standard,
1318 * \w & ^\S = nothing
1319 * Unfortunately, some vendor locales do not meet the Posix
1320 * standard, in particular almost everything by Microsoft.
1321 * The loop below just changes e.g., \w into \W and vice versa */
1323 regnode_charclass_posixl temp;
1324 int add = 1; /* To calculate the index of the complement */
1326 ANYOF_POSIXL_ZERO(&temp);
1327 for (i = 0; i < ANYOF_MAX; i++) {
1329 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1330 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1332 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1333 ANYOF_POSIXL_SET(&temp, i + add);
1335 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1337 ANYOF_POSIXL_AND(&temp, ssc);
1339 } /* else ssc already has no posixes */
1340 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1341 in its initial state */
1342 else if (! is_ANYOF_SYNTHETIC(and_with)
1343 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1345 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1346 * copy it over 'ssc' */
1347 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1348 if (is_ANYOF_SYNTHETIC(and_with)) {
1349 StructCopy(and_with, ssc, regnode_ssc);
1352 ssc->invlist = anded_cp_list;
1353 ANYOF_POSIXL_ZERO(ssc);
1354 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1355 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1359 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1360 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1362 /* One or the other of P1, P2 is non-empty. */
1363 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1364 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1366 ssc_union(ssc, anded_cp_list, FALSE);
1368 else { /* P1 = P2 = empty */
1369 ssc_intersection(ssc, anded_cp_list, FALSE);
1375 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1376 const regnode_charclass *or_with)
1378 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1379 * another SSC or a regular ANYOF class. Can create false positives if
1380 * 'or_with' is to be inverted. */
1385 PERL_ARGS_ASSERT_SSC_OR;
1387 assert(is_ANYOF_SYNTHETIC(ssc));
1389 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1390 * the code point inversion list and just the relevant flags */
1391 if (is_ANYOF_SYNTHETIC(or_with)) {
1392 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1393 ored_flags = ANYOF_FLAGS(or_with);
1396 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1397 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1400 ANYOF_FLAGS(ssc) |= ored_flags;
1402 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1403 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1404 * 'or_with' may be inverted. When not inverted, we have the simple
1405 * situation of computing:
1406 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1407 * If P1|P2 yields a situation with both a class and its complement are
1408 * set, like having both \w and \W, this matches all code points, and we
1409 * can delete these from the P component of the ssc going forward. XXX We
1410 * might be able to delete all the P components, but I (khw) am not certain
1411 * about this, and it is better to be safe.
1414 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1415 * <= (C1 | P1) | ~C2
1416 * <= (C1 | ~C2) | P1
1417 * (which results in actually simpler code than the non-inverted case)
1420 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1421 && ! is_ANYOF_SYNTHETIC(or_with))
1423 /* We ignore P2, leaving P1 going forward */
1424 } /* else Not inverted */
1425 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1426 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1427 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1429 for (i = 0; i < ANYOF_MAX; i += 2) {
1430 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1432 ssc_match_all_cp(ssc);
1433 ANYOF_POSIXL_CLEAR(ssc, i);
1434 ANYOF_POSIXL_CLEAR(ssc, i+1);
1442 FALSE /* Already has been inverted */
1446 PERL_STATIC_INLINE void
1447 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1449 PERL_ARGS_ASSERT_SSC_UNION;
1451 assert(is_ANYOF_SYNTHETIC(ssc));
1453 _invlist_union_maybe_complement_2nd(ssc->invlist,
1459 PERL_STATIC_INLINE void
1460 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1462 const bool invert2nd)
1464 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1466 assert(is_ANYOF_SYNTHETIC(ssc));
1468 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1474 PERL_STATIC_INLINE void
1475 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1477 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1479 assert(is_ANYOF_SYNTHETIC(ssc));
1481 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1484 PERL_STATIC_INLINE void
1485 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1487 /* AND just the single code point 'cp' into the SSC 'ssc' */
1489 SV* cp_list = _new_invlist(2);
1491 PERL_ARGS_ASSERT_SSC_CP_AND;
1493 assert(is_ANYOF_SYNTHETIC(ssc));
1495 cp_list = add_cp_to_invlist(cp_list, cp);
1496 ssc_intersection(ssc, cp_list,
1497 FALSE /* Not inverted */
1499 SvREFCNT_dec_NN(cp_list);
1502 PERL_STATIC_INLINE void
1503 S_ssc_clear_locale(regnode_ssc *ssc)
1505 /* Set the SSC 'ssc' to not match any locale things */
1506 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1508 assert(is_ANYOF_SYNTHETIC(ssc));
1510 ANYOF_POSIXL_ZERO(ssc);
1511 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1514 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1517 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1519 /* The synthetic start class is used to hopefully quickly winnow down
1520 * places where a pattern could start a match in the target string. If it
1521 * doesn't really narrow things down that much, there isn't much point to
1522 * having the overhead of using it. This function uses some very crude
1523 * heuristics to decide if to use the ssc or not.
1525 * It returns TRUE if 'ssc' rules out more than half what it considers to
1526 * be the "likely" possible matches, but of course it doesn't know what the
1527 * actual things being matched are going to be; these are only guesses
1529 * For /l matches, it assumes that the only likely matches are going to be
1530 * in the 0-255 range, uniformly distributed, so half of that is 127
1531 * For /a and /d matches, it assumes that the likely matches will be just
1532 * the ASCII range, so half of that is 63
1533 * For /u and there isn't anything matching above the Latin1 range, it
1534 * assumes that that is the only range likely to be matched, and uses
1535 * half that as the cut-off: 127. If anything matches above Latin1,
1536 * it assumes that all of Unicode could match (uniformly), except for
1537 * non-Unicode code points and things in the General Category "Other"
1538 * (unassigned, private use, surrogates, controls and formats). This
1539 * is a much large number. */
1541 const U32 max_match = (LOC)
1545 : (invlist_highest(ssc->invlist) < 256)
1547 : ((NON_OTHER_COUNT + 1) / 2) - 1;
1548 U32 count = 0; /* Running total of number of code points matched by
1550 UV start, end; /* Start and end points of current range in inversion
1553 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1555 invlist_iterinit(ssc->invlist);
1556 while (invlist_iternext(ssc->invlist, &start, &end)) {
1558 /* /u is the only thing that we expect to match above 255; so if not /u
1559 * and even if there are matches above 255, ignore them. This catches
1560 * things like \d under /d which does match the digits above 255, but
1561 * since the pattern is /d, it is not likely to be expecting them */
1562 if (! UNI_SEMANTICS) {
1566 end = MIN(end, 255);
1568 count += end - start + 1;
1569 if (count > max_match) {
1570 invlist_iterfinish(ssc->invlist);
1580 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1582 /* The inversion list in the SSC is marked mortal; now we need a more
1583 * permanent copy, which is stored the same way that is done in a regular
1584 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1587 SV* invlist = invlist_clone(ssc->invlist);
1589 PERL_ARGS_ASSERT_SSC_FINALIZE;
1591 assert(is_ANYOF_SYNTHETIC(ssc));
1593 /* The code in this file assumes that all but these flags aren't relevant
1594 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1595 * by the time we reach here */
1596 assert(! (ANYOF_FLAGS(ssc) & ~ANYOF_COMMON_FLAGS));
1598 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1600 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1601 NULL, NULL, NULL, FALSE);
1603 /* Make sure is clone-safe */
1604 ssc->invlist = NULL;
1606 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1607 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1610 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1613 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1614 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1615 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1616 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1617 ? (TRIE_LIST_CUR( idx ) - 1) \
1623 dump_trie(trie,widecharmap,revcharmap)
1624 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1625 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1627 These routines dump out a trie in a somewhat readable format.
1628 The _interim_ variants are used for debugging the interim
1629 tables that are used to generate the final compressed
1630 representation which is what dump_trie expects.
1632 Part of the reason for their existence is to provide a form
1633 of documentation as to how the different representations function.
1638 Dumps the final compressed table form of the trie to Perl_debug_log.
1639 Used for debugging make_trie().
1643 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1644 AV *revcharmap, U32 depth)
1647 SV *sv=sv_newmortal();
1648 int colwidth= widecharmap ? 6 : 4;
1650 GET_RE_DEBUG_FLAGS_DECL;
1652 PERL_ARGS_ASSERT_DUMP_TRIE;
1654 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1655 (int)depth * 2 + 2,"",
1656 "Match","Base","Ofs" );
1658 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1659 SV ** const tmp = av_fetch( revcharmap, state, 0);
1661 PerlIO_printf( Perl_debug_log, "%*s",
1663 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1664 PL_colors[0], PL_colors[1],
1665 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1666 PERL_PV_ESCAPE_FIRSTCHAR
1671 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1672 (int)depth * 2 + 2,"");
1674 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1675 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1676 PerlIO_printf( Perl_debug_log, "\n");
1678 for( state = 1 ; state < trie->statecount ; state++ ) {
1679 const U32 base = trie->states[ state ].trans.base;
1681 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1682 (int)depth * 2 + 2,"", (UV)state);
1684 if ( trie->states[ state ].wordnum ) {
1685 PerlIO_printf( Perl_debug_log, " W%4X",
1686 trie->states[ state ].wordnum );
1688 PerlIO_printf( Perl_debug_log, "%6s", "" );
1691 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1696 while( ( base + ofs < trie->uniquecharcount ) ||
1697 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1698 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1702 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1704 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1705 if ( ( base + ofs >= trie->uniquecharcount )
1706 && ( base + ofs - trie->uniquecharcount
1708 && trie->trans[ base + ofs
1709 - trie->uniquecharcount ].check == state )
1711 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1713 (UV)trie->trans[ base + ofs
1714 - trie->uniquecharcount ].next );
1716 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1720 PerlIO_printf( Perl_debug_log, "]");
1723 PerlIO_printf( Perl_debug_log, "\n" );
1725 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1727 for (word=1; word <= trie->wordcount; word++) {
1728 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1729 (int)word, (int)(trie->wordinfo[word].prev),
1730 (int)(trie->wordinfo[word].len));
1732 PerlIO_printf(Perl_debug_log, "\n" );
1735 Dumps a fully constructed but uncompressed trie in list form.
1736 List tries normally only are used for construction when the number of
1737 possible chars (trie->uniquecharcount) is very high.
1738 Used for debugging make_trie().
1741 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1742 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1746 SV *sv=sv_newmortal();
1747 int colwidth= widecharmap ? 6 : 4;
1748 GET_RE_DEBUG_FLAGS_DECL;
1750 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1752 /* print out the table precompression. */
1753 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1754 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1755 "------:-----+-----------------\n" );
1757 for( state=1 ; state < next_alloc ; state ++ ) {
1760 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1761 (int)depth * 2 + 2,"", (UV)state );
1762 if ( ! trie->states[ state ].wordnum ) {
1763 PerlIO_printf( Perl_debug_log, "%5s| ","");
1765 PerlIO_printf( Perl_debug_log, "W%4x| ",
1766 trie->states[ state ].wordnum
1769 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1770 SV ** const tmp = av_fetch( revcharmap,
1771 TRIE_LIST_ITEM(state,charid).forid, 0);
1773 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1775 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1777 PL_colors[0], PL_colors[1],
1778 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1779 | PERL_PV_ESCAPE_FIRSTCHAR
1781 TRIE_LIST_ITEM(state,charid).forid,
1782 (UV)TRIE_LIST_ITEM(state,charid).newstate
1785 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1786 (int)((depth * 2) + 14), "");
1789 PerlIO_printf( Perl_debug_log, "\n");
1794 Dumps a fully constructed but uncompressed trie in table form.
1795 This is the normal DFA style state transition table, with a few
1796 twists to facilitate compression later.
1797 Used for debugging make_trie().
1800 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1801 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1806 SV *sv=sv_newmortal();
1807 int colwidth= widecharmap ? 6 : 4;
1808 GET_RE_DEBUG_FLAGS_DECL;
1810 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1813 print out the table precompression so that we can do a visual check
1814 that they are identical.
1817 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1819 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1820 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1822 PerlIO_printf( Perl_debug_log, "%*s",
1824 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1825 PL_colors[0], PL_colors[1],
1826 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1827 PERL_PV_ESCAPE_FIRSTCHAR
1833 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1835 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1836 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1839 PerlIO_printf( Perl_debug_log, "\n" );
1841 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1843 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1844 (int)depth * 2 + 2,"",
1845 (UV)TRIE_NODENUM( state ) );
1847 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1848 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1850 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1852 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1854 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1855 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1856 (UV)trie->trans[ state ].check );
1858 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1859 (UV)trie->trans[ state ].check,
1860 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1868 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1869 startbranch: the first branch in the whole branch sequence
1870 first : start branch of sequence of branch-exact nodes.
1871 May be the same as startbranch
1872 last : Thing following the last branch.
1873 May be the same as tail.
1874 tail : item following the branch sequence
1875 count : words in the sequence
1876 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
1877 depth : indent depth
1879 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1881 A trie is an N'ary tree where the branches are determined by digital
1882 decomposition of the key. IE, at the root node you look up the 1st character and
1883 follow that branch repeat until you find the end of the branches. Nodes can be
1884 marked as "accepting" meaning they represent a complete word. Eg:
1888 would convert into the following structure. Numbers represent states, letters
1889 following numbers represent valid transitions on the letter from that state, if
1890 the number is in square brackets it represents an accepting state, otherwise it
1891 will be in parenthesis.
1893 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1897 (1) +-i->(6)-+-s->[7]
1899 +-s->(3)-+-h->(4)-+-e->[5]
1901 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1903 This shows that when matching against the string 'hers' we will begin at state 1
1904 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1905 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1906 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1907 single traverse. We store a mapping from accepting to state to which word was
1908 matched, and then when we have multiple possibilities we try to complete the
1909 rest of the regex in the order in which they occured in the alternation.
1911 The only prior NFA like behaviour that would be changed by the TRIE support is
1912 the silent ignoring of duplicate alternations which are of the form:
1914 / (DUPE|DUPE) X? (?{ ... }) Y /x
1916 Thus EVAL blocks following a trie may be called a different number of times with
1917 and without the optimisation. With the optimisations dupes will be silently
1918 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1919 the following demonstrates:
1921 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1923 which prints out 'word' three times, but
1925 'words'=~/(word|word|word)(?{ print $1 })S/
1927 which doesnt print it out at all. This is due to other optimisations kicking in.
1929 Example of what happens on a structural level:
1931 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1933 1: CURLYM[1] {1,32767}(18)
1944 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1945 and should turn into:
1947 1: CURLYM[1] {1,32767}(18)
1949 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1957 Cases where tail != last would be like /(?foo|bar)baz/:
1967 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1968 and would end up looking like:
1971 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1978 d = uvchr_to_utf8_flags(d, uv, 0);
1980 is the recommended Unicode-aware way of saying
1985 #define TRIE_STORE_REVCHAR(val) \
1988 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1989 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1990 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
1991 SvCUR_set(zlopp, kapow - flrbbbbb); \
1994 av_push(revcharmap, zlopp); \
1996 char ooooff = (char)val; \
1997 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2001 /* This gets the next character from the input, folding it if not already
2003 #define TRIE_READ_CHAR STMT_START { \
2006 /* if it is UTF then it is either already folded, or does not need \
2008 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2010 else if (folder == PL_fold_latin1) { \
2011 /* This folder implies Unicode rules, which in the range expressible \
2012 * by not UTF is the lower case, with the two exceptions, one of \
2013 * which should have been taken care of before calling this */ \
2014 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2015 uvc = toLOWER_L1(*uc); \
2016 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2019 /* raw data, will be folded later if needed */ \
2027 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2028 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2029 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2030 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2032 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2033 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2034 TRIE_LIST_CUR( state )++; \
2037 #define TRIE_LIST_NEW(state) STMT_START { \
2038 Newxz( trie->states[ state ].trans.list, \
2039 4, reg_trie_trans_le ); \
2040 TRIE_LIST_CUR( state ) = 1; \
2041 TRIE_LIST_LEN( state ) = 4; \
2044 #define TRIE_HANDLE_WORD(state) STMT_START { \
2045 U16 dupe= trie->states[ state ].wordnum; \
2046 regnode * const noper_next = regnext( noper ); \
2049 /* store the word for dumping */ \
2051 if (OP(noper) != NOTHING) \
2052 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2054 tmp = newSVpvn_utf8( "", 0, UTF ); \
2055 av_push( trie_words, tmp ); \
2059 trie->wordinfo[curword].prev = 0; \
2060 trie->wordinfo[curword].len = wordlen; \
2061 trie->wordinfo[curword].accept = state; \
2063 if ( noper_next < tail ) { \
2065 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2067 trie->jump[curword] = (U16)(noper_next - convert); \
2069 jumper = noper_next; \
2071 nextbranch= regnext(cur); \
2075 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2076 /* chain, so that when the bits of chain are later */\
2077 /* linked together, the dups appear in the chain */\
2078 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2079 trie->wordinfo[dupe].prev = curword; \
2081 /* we haven't inserted this word yet. */ \
2082 trie->states[ state ].wordnum = curword; \
2087 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2088 ( ( base + charid >= ucharcount \
2089 && base + charid < ubound \
2090 && state == trie->trans[ base - ucharcount + charid ].check \
2091 && trie->trans[ base - ucharcount + charid ].next ) \
2092 ? trie->trans[ base - ucharcount + charid ].next \
2093 : ( state==1 ? special : 0 ) \
2097 #define MADE_JUMP_TRIE 2
2098 #define MADE_EXACT_TRIE 4
2101 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2102 regnode *first, regnode *last, regnode *tail,
2103 U32 word_count, U32 flags, U32 depth)
2105 /* first pass, loop through and scan words */
2106 reg_trie_data *trie;
2107 HV *widecharmap = NULL;
2108 AV *revcharmap = newAV();
2114 regnode *jumper = NULL;
2115 regnode *nextbranch = NULL;
2116 regnode *convert = NULL;
2117 U32 *prev_states; /* temp array mapping each state to previous one */
2118 /* we just use folder as a flag in utf8 */
2119 const U8 * folder = NULL;
2122 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2123 AV *trie_words = NULL;
2124 /* along with revcharmap, this only used during construction but both are
2125 * useful during debugging so we store them in the struct when debugging.
2128 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2129 STRLEN trie_charcount=0;
2131 SV *re_trie_maxbuff;
2132 GET_RE_DEBUG_FLAGS_DECL;
2134 PERL_ARGS_ASSERT_MAKE_TRIE;
2136 PERL_UNUSED_ARG(depth);
2140 case EXACT: case EXACTL: break;
2144 case EXACTFLU8: folder = PL_fold_latin1; break;
2145 case EXACTF: folder = PL_fold; break;
2146 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2149 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2151 trie->startstate = 1;
2152 trie->wordcount = word_count;
2153 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2154 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2155 if (flags == EXACT || flags == EXACTL)
2156 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2157 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2158 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2161 trie_words = newAV();
2164 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2165 assert(re_trie_maxbuff);
2166 if (!SvIOK(re_trie_maxbuff)) {
2167 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2169 DEBUG_TRIE_COMPILE_r({
2170 PerlIO_printf( Perl_debug_log,
2171 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2172 (int)depth * 2 + 2, "",
2173 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2174 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2177 /* Find the node we are going to overwrite */
2178 if ( first == startbranch && OP( last ) != BRANCH ) {
2179 /* whole branch chain */
2182 /* branch sub-chain */
2183 convert = NEXTOPER( first );
2186 /* -- First loop and Setup --
2188 We first traverse the branches and scan each word to determine if it
2189 contains widechars, and how many unique chars there are, this is
2190 important as we have to build a table with at least as many columns as we
2193 We use an array of integers to represent the character codes 0..255
2194 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2195 the native representation of the character value as the key and IV's for
2198 *TODO* If we keep track of how many times each character is used we can
2199 remap the columns so that the table compression later on is more
2200 efficient in terms of memory by ensuring the most common value is in the
2201 middle and the least common are on the outside. IMO this would be better
2202 than a most to least common mapping as theres a decent chance the most
2203 common letter will share a node with the least common, meaning the node
2204 will not be compressible. With a middle is most common approach the worst
2205 case is when we have the least common nodes twice.
2209 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2210 regnode *noper = NEXTOPER( cur );
2211 const U8 *uc = (U8*)STRING( noper );
2212 const U8 *e = uc + STR_LEN( noper );
2214 U32 wordlen = 0; /* required init */
2215 STRLEN minchars = 0;
2216 STRLEN maxchars = 0;
2217 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2220 if (OP(noper) == NOTHING) {
2221 regnode *noper_next= regnext(noper);
2222 if (noper_next != tail && OP(noper_next) == flags) {
2224 uc= (U8*)STRING(noper);
2225 e= uc + STR_LEN(noper);
2226 trie->minlen= STR_LEN(noper);
2233 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2234 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2235 regardless of encoding */
2236 if (OP( noper ) == EXACTFU_SS) {
2237 /* false positives are ok, so just set this */
2238 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2241 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2243 TRIE_CHARCOUNT(trie)++;
2246 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2247 * is in effect. Under /i, this character can match itself, or
2248 * anything that folds to it. If not under /i, it can match just
2249 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2250 * all fold to k, and all are single characters. But some folds
2251 * expand to more than one character, so for example LATIN SMALL
2252 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2253 * the string beginning at 'uc' is 'ffi', it could be matched by
2254 * three characters, or just by the one ligature character. (It
2255 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2256 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2257 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2258 * match.) The trie needs to know the minimum and maximum number
2259 * of characters that could match so that it can use size alone to
2260 * quickly reject many match attempts. The max is simple: it is
2261 * the number of folded characters in this branch (since a fold is
2262 * never shorter than what folds to it. */
2266 /* And the min is equal to the max if not under /i (indicated by
2267 * 'folder' being NULL), or there are no multi-character folds. If
2268 * there is a multi-character fold, the min is incremented just
2269 * once, for the character that folds to the sequence. Each
2270 * character in the sequence needs to be added to the list below of
2271 * characters in the trie, but we count only the first towards the
2272 * min number of characters needed. This is done through the
2273 * variable 'foldlen', which is returned by the macros that look
2274 * for these sequences as the number of bytes the sequence
2275 * occupies. Each time through the loop, we decrement 'foldlen' by
2276 * how many bytes the current char occupies. Only when it reaches
2277 * 0 do we increment 'minchars' or look for another multi-character
2279 if (folder == NULL) {
2282 else if (foldlen > 0) {
2283 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2288 /* See if *uc is the beginning of a multi-character fold. If
2289 * so, we decrement the length remaining to look at, to account
2290 * for the current character this iteration. (We can use 'uc'
2291 * instead of the fold returned by TRIE_READ_CHAR because for
2292 * non-UTF, the latin1_safe macro is smart enough to account
2293 * for all the unfolded characters, and because for UTF, the
2294 * string will already have been folded earlier in the
2295 * compilation process */
2297 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2298 foldlen -= UTF8SKIP(uc);
2301 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2306 /* The current character (and any potential folds) should be added
2307 * to the possible matching characters for this position in this
2311 U8 folded= folder[ (U8) uvc ];
2312 if ( !trie->charmap[ folded ] ) {
2313 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2314 TRIE_STORE_REVCHAR( folded );
2317 if ( !trie->charmap[ uvc ] ) {
2318 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2319 TRIE_STORE_REVCHAR( uvc );
2322 /* store the codepoint in the bitmap, and its folded
2324 TRIE_BITMAP_SET(trie, uvc);
2326 /* store the folded codepoint */
2327 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2330 /* store first byte of utf8 representation of
2331 variant codepoints */
2332 if (! UVCHR_IS_INVARIANT(uvc)) {
2333 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2336 set_bit = 0; /* We've done our bit :-) */
2340 /* XXX We could come up with the list of code points that fold
2341 * to this using PL_utf8_foldclosures, except not for
2342 * multi-char folds, as there may be multiple combinations
2343 * there that could work, which needs to wait until runtime to
2344 * resolve (The comment about LIGATURE FFI above is such an
2349 widecharmap = newHV();
2351 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2354 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2356 if ( !SvTRUE( *svpp ) ) {
2357 sv_setiv( *svpp, ++trie->uniquecharcount );
2358 TRIE_STORE_REVCHAR(uvc);
2361 } /* end loop through characters in this branch of the trie */
2363 /* We take the min and max for this branch and combine to find the min
2364 * and max for all branches processed so far */
2365 if( cur == first ) {
2366 trie->minlen = minchars;
2367 trie->maxlen = maxchars;
2368 } else if (minchars < trie->minlen) {
2369 trie->minlen = minchars;
2370 } else if (maxchars > trie->maxlen) {
2371 trie->maxlen = maxchars;
2373 } /* end first pass */
2374 DEBUG_TRIE_COMPILE_r(
2375 PerlIO_printf( Perl_debug_log,
2376 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2377 (int)depth * 2 + 2,"",
2378 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2379 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2380 (int)trie->minlen, (int)trie->maxlen )
2384 We now know what we are dealing with in terms of unique chars and
2385 string sizes so we can calculate how much memory a naive
2386 representation using a flat table will take. If it's over a reasonable
2387 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2388 conservative but potentially much slower representation using an array
2391 At the end we convert both representations into the same compressed
2392 form that will be used in regexec.c for matching with. The latter
2393 is a form that cannot be used to construct with but has memory
2394 properties similar to the list form and access properties similar
2395 to the table form making it both suitable for fast searches and
2396 small enough that its feasable to store for the duration of a program.
2398 See the comment in the code where the compressed table is produced
2399 inplace from the flat tabe representation for an explanation of how
2400 the compression works.
2405 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2408 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2409 > SvIV(re_trie_maxbuff) )
2412 Second Pass -- Array Of Lists Representation
2414 Each state will be represented by a list of charid:state records
2415 (reg_trie_trans_le) the first such element holds the CUR and LEN
2416 points of the allocated array. (See defines above).
2418 We build the initial structure using the lists, and then convert
2419 it into the compressed table form which allows faster lookups
2420 (but cant be modified once converted).
2423 STRLEN transcount = 1;
2425 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2426 "%*sCompiling trie using list compiler\n",
2427 (int)depth * 2 + 2, ""));
2429 trie->states = (reg_trie_state *)
2430 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2431 sizeof(reg_trie_state) );
2435 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2437 regnode *noper = NEXTOPER( cur );
2438 U8 *uc = (U8*)STRING( noper );
2439 const U8 *e = uc + STR_LEN( noper );
2440 U32 state = 1; /* required init */
2441 U16 charid = 0; /* sanity init */
2442 U32 wordlen = 0; /* required init */
2444 if (OP(noper) == NOTHING) {
2445 regnode *noper_next= regnext(noper);
2446 if (noper_next != tail && OP(noper_next) == flags) {
2448 uc= (U8*)STRING(noper);
2449 e= uc + STR_LEN(noper);
2453 if (OP(noper) != NOTHING) {
2454 for ( ; uc < e ; uc += len ) {
2459 charid = trie->charmap[ uvc ];
2461 SV** const svpp = hv_fetch( widecharmap,
2468 charid=(U16)SvIV( *svpp );
2471 /* charid is now 0 if we dont know the char read, or
2472 * nonzero if we do */
2479 if ( !trie->states[ state ].trans.list ) {
2480 TRIE_LIST_NEW( state );
2483 check <= TRIE_LIST_USED( state );
2486 if ( TRIE_LIST_ITEM( state, check ).forid
2489 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2494 newstate = next_alloc++;
2495 prev_states[newstate] = state;
2496 TRIE_LIST_PUSH( state, charid, newstate );
2501 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2505 TRIE_HANDLE_WORD(state);
2507 } /* end second pass */
2509 /* next alloc is the NEXT state to be allocated */
2510 trie->statecount = next_alloc;
2511 trie->states = (reg_trie_state *)
2512 PerlMemShared_realloc( trie->states,
2514 * sizeof(reg_trie_state) );
2516 /* and now dump it out before we compress it */
2517 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2518 revcharmap, next_alloc,
2522 trie->trans = (reg_trie_trans *)
2523 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2530 for( state=1 ; state < next_alloc ; state ++ ) {
2534 DEBUG_TRIE_COMPILE_MORE_r(
2535 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2539 if (trie->states[state].trans.list) {
2540 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2544 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2545 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2546 if ( forid < minid ) {
2548 } else if ( forid > maxid ) {
2552 if ( transcount < tp + maxid - minid + 1) {
2554 trie->trans = (reg_trie_trans *)
2555 PerlMemShared_realloc( trie->trans,
2557 * sizeof(reg_trie_trans) );
2558 Zero( trie->trans + (transcount / 2),
2562 base = trie->uniquecharcount + tp - minid;
2563 if ( maxid == minid ) {
2565 for ( ; zp < tp ; zp++ ) {
2566 if ( ! trie->trans[ zp ].next ) {
2567 base = trie->uniquecharcount + zp - minid;
2568 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2570 trie->trans[ zp ].check = state;
2576 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2578 trie->trans[ tp ].check = state;
2583 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2584 const U32 tid = base
2585 - trie->uniquecharcount
2586 + TRIE_LIST_ITEM( state, idx ).forid;
2587 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2589 trie->trans[ tid ].check = state;
2591 tp += ( maxid - minid + 1 );
2593 Safefree(trie->states[ state ].trans.list);
2596 DEBUG_TRIE_COMPILE_MORE_r(
2597 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2600 trie->states[ state ].trans.base=base;
2602 trie->lasttrans = tp + 1;
2606 Second Pass -- Flat Table Representation.
2608 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2609 each. We know that we will need Charcount+1 trans at most to store
2610 the data (one row per char at worst case) So we preallocate both
2611 structures assuming worst case.
2613 We then construct the trie using only the .next slots of the entry
2616 We use the .check field of the first entry of the node temporarily
2617 to make compression both faster and easier by keeping track of how
2618 many non zero fields are in the node.
2620 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2623 There are two terms at use here: state as a TRIE_NODEIDX() which is
2624 a number representing the first entry of the node, and state as a
2625 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2626 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2627 if there are 2 entrys per node. eg:
2635 The table is internally in the right hand, idx form. However as we
2636 also have to deal with the states array which is indexed by nodenum
2637 we have to use TRIE_NODENUM() to convert.
2640 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2641 "%*sCompiling trie using table compiler\n",
2642 (int)depth * 2 + 2, ""));
2644 trie->trans = (reg_trie_trans *)
2645 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2646 * trie->uniquecharcount + 1,
2647 sizeof(reg_trie_trans) );
2648 trie->states = (reg_trie_state *)
2649 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2650 sizeof(reg_trie_state) );
2651 next_alloc = trie->uniquecharcount + 1;
2654 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2656 regnode *noper = NEXTOPER( cur );
2657 const U8 *uc = (U8*)STRING( noper );
2658 const U8 *e = uc + STR_LEN( noper );
2660 U32 state = 1; /* required init */
2662 U16 charid = 0; /* sanity init */
2663 U32 accept_state = 0; /* sanity init */
2665 U32 wordlen = 0; /* required init */
2667 if (OP(noper) == NOTHING) {
2668 regnode *noper_next= regnext(noper);
2669 if (noper_next != tail && OP(noper_next) == flags) {
2671 uc= (U8*)STRING(noper);
2672 e= uc + STR_LEN(noper);
2676 if ( OP(noper) != NOTHING ) {
2677 for ( ; uc < e ; uc += len ) {
2682 charid = trie->charmap[ uvc ];
2684 SV* const * const svpp = hv_fetch( widecharmap,
2688 charid = svpp ? (U16)SvIV(*svpp) : 0;
2692 if ( !trie->trans[ state + charid ].next ) {
2693 trie->trans[ state + charid ].next = next_alloc;
2694 trie->trans[ state ].check++;
2695 prev_states[TRIE_NODENUM(next_alloc)]
2696 = TRIE_NODENUM(state);
2697 next_alloc += trie->uniquecharcount;
2699 state = trie->trans[ state + charid ].next;
2701 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2703 /* charid is now 0 if we dont know the char read, or
2704 * nonzero if we do */
2707 accept_state = TRIE_NODENUM( state );
2708 TRIE_HANDLE_WORD(accept_state);
2710 } /* end second pass */
2712 /* and now dump it out before we compress it */
2713 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2715 next_alloc, depth+1));
2719 * Inplace compress the table.*
2721 For sparse data sets the table constructed by the trie algorithm will
2722 be mostly 0/FAIL transitions or to put it another way mostly empty.
2723 (Note that leaf nodes will not contain any transitions.)
2725 This algorithm compresses the tables by eliminating most such
2726 transitions, at the cost of a modest bit of extra work during lookup:
2728 - Each states[] entry contains a .base field which indicates the
2729 index in the state[] array wheres its transition data is stored.
2731 - If .base is 0 there are no valid transitions from that node.
2733 - If .base is nonzero then charid is added to it to find an entry in
2736 -If trans[states[state].base+charid].check!=state then the
2737 transition is taken to be a 0/Fail transition. Thus if there are fail
2738 transitions at the front of the node then the .base offset will point
2739 somewhere inside the previous nodes data (or maybe even into a node
2740 even earlier), but the .check field determines if the transition is
2744 The following process inplace converts the table to the compressed
2745 table: We first do not compress the root node 1,and mark all its
2746 .check pointers as 1 and set its .base pointer as 1 as well. This
2747 allows us to do a DFA construction from the compressed table later,
2748 and ensures that any .base pointers we calculate later are greater
2751 - We set 'pos' to indicate the first entry of the second node.
2753 - We then iterate over the columns of the node, finding the first and
2754 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2755 and set the .check pointers accordingly, and advance pos
2756 appropriately and repreat for the next node. Note that when we copy
2757 the next pointers we have to convert them from the original
2758 NODEIDX form to NODENUM form as the former is not valid post
2761 - If a node has no transitions used we mark its base as 0 and do not
2762 advance the pos pointer.
2764 - If a node only has one transition we use a second pointer into the
2765 structure to fill in allocated fail transitions from other states.
2766 This pointer is independent of the main pointer and scans forward
2767 looking for null transitions that are allocated to a state. When it
2768 finds one it writes the single transition into the "hole". If the
2769 pointer doesnt find one the single transition is appended as normal.
2771 - Once compressed we can Renew/realloc the structures to release the
2774 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2775 specifically Fig 3.47 and the associated pseudocode.
2779 const U32 laststate = TRIE_NODENUM( next_alloc );
2782 trie->statecount = laststate;
2784 for ( state = 1 ; state < laststate ; state++ ) {
2786 const U32 stateidx = TRIE_NODEIDX( state );
2787 const U32 o_used = trie->trans[ stateidx ].check;
2788 U32 used = trie->trans[ stateidx ].check;
2789 trie->trans[ stateidx ].check = 0;
2792 used && charid < trie->uniquecharcount;
2795 if ( flag || trie->trans[ stateidx + charid ].next ) {
2796 if ( trie->trans[ stateidx + charid ].next ) {
2798 for ( ; zp < pos ; zp++ ) {
2799 if ( ! trie->trans[ zp ].next ) {
2803 trie->states[ state ].trans.base
2805 + trie->uniquecharcount
2807 trie->trans[ zp ].next
2808 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2810 trie->trans[ zp ].check = state;
2811 if ( ++zp > pos ) pos = zp;
2818 trie->states[ state ].trans.base
2819 = pos + trie->uniquecharcount - charid ;
2821 trie->trans[ pos ].next
2822 = SAFE_TRIE_NODENUM(
2823 trie->trans[ stateidx + charid ].next );
2824 trie->trans[ pos ].check = state;
2829 trie->lasttrans = pos + 1;
2830 trie->states = (reg_trie_state *)
2831 PerlMemShared_realloc( trie->states, laststate
2832 * sizeof(reg_trie_state) );
2833 DEBUG_TRIE_COMPILE_MORE_r(
2834 PerlIO_printf( Perl_debug_log,
2835 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2836 (int)depth * 2 + 2,"",
2837 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2841 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2844 } /* end table compress */
2846 DEBUG_TRIE_COMPILE_MORE_r(
2847 PerlIO_printf(Perl_debug_log,
2848 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2849 (int)depth * 2 + 2, "",
2850 (UV)trie->statecount,
2851 (UV)trie->lasttrans)
2853 /* resize the trans array to remove unused space */
2854 trie->trans = (reg_trie_trans *)
2855 PerlMemShared_realloc( trie->trans, trie->lasttrans
2856 * sizeof(reg_trie_trans) );
2858 { /* Modify the program and insert the new TRIE node */
2859 U8 nodetype =(U8)(flags & 0xFF);
2863 regnode *optimize = NULL;
2864 #ifdef RE_TRACK_PATTERN_OFFSETS
2867 U32 mjd_nodelen = 0;
2868 #endif /* RE_TRACK_PATTERN_OFFSETS */
2869 #endif /* DEBUGGING */
2871 This means we convert either the first branch or the first Exact,
2872 depending on whether the thing following (in 'last') is a branch
2873 or not and whther first is the startbranch (ie is it a sub part of
2874 the alternation or is it the whole thing.)
2875 Assuming its a sub part we convert the EXACT otherwise we convert
2876 the whole branch sequence, including the first.
2878 /* Find the node we are going to overwrite */
2879 if ( first != startbranch || OP( last ) == BRANCH ) {
2880 /* branch sub-chain */
2881 NEXT_OFF( first ) = (U16)(last - first);
2882 #ifdef RE_TRACK_PATTERN_OFFSETS
2884 mjd_offset= Node_Offset((convert));
2885 mjd_nodelen= Node_Length((convert));
2888 /* whole branch chain */
2890 #ifdef RE_TRACK_PATTERN_OFFSETS
2893 const regnode *nop = NEXTOPER( convert );
2894 mjd_offset= Node_Offset((nop));
2895 mjd_nodelen= Node_Length((nop));
2899 PerlIO_printf(Perl_debug_log,
2900 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2901 (int)depth * 2 + 2, "",
2902 (UV)mjd_offset, (UV)mjd_nodelen)
2905 /* But first we check to see if there is a common prefix we can
2906 split out as an EXACT and put in front of the TRIE node. */
2907 trie->startstate= 1;
2908 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2910 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2914 const U32 base = trie->states[ state ].trans.base;
2916 if ( trie->states[state].wordnum )
2919 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2920 if ( ( base + ofs >= trie->uniquecharcount ) &&
2921 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2922 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2924 if ( ++count > 1 ) {
2925 SV **tmp = av_fetch( revcharmap, ofs, 0);
2926 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2927 if ( state == 1 ) break;
2929 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2931 PerlIO_printf(Perl_debug_log,
2932 "%*sNew Start State=%"UVuf" Class: [",
2933 (int)depth * 2 + 2, "",
2936 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2937 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2939 TRIE_BITMAP_SET(trie,*ch);
2941 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2943 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2947 TRIE_BITMAP_SET(trie,*ch);
2949 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2950 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2956 SV **tmp = av_fetch( revcharmap, idx, 0);
2958 char *ch = SvPV( *tmp, len );
2960 SV *sv=sv_newmortal();
2961 PerlIO_printf( Perl_debug_log,
2962 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2963 (int)depth * 2 + 2, "",
2965 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2966 PL_colors[0], PL_colors[1],
2967 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2968 PERL_PV_ESCAPE_FIRSTCHAR
2973 OP( convert ) = nodetype;
2974 str=STRING(convert);
2977 STR_LEN(convert) += len;
2983 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2988 trie->prefixlen = (state-1);
2990 regnode *n = convert+NODE_SZ_STR(convert);
2991 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2992 trie->startstate = state;
2993 trie->minlen -= (state - 1);
2994 trie->maxlen -= (state - 1);
2996 /* At least the UNICOS C compiler choked on this
2997 * being argument to DEBUG_r(), so let's just have
3000 #ifdef PERL_EXT_RE_BUILD
3006 regnode *fix = convert;
3007 U32 word = trie->wordcount;
3009 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3010 while( ++fix < n ) {
3011 Set_Node_Offset_Length(fix, 0, 0);
3014 SV ** const tmp = av_fetch( trie_words, word, 0 );
3016 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3017 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3019 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3027 NEXT_OFF(convert) = (U16)(tail - convert);
3028 DEBUG_r(optimize= n);
3034 if ( trie->maxlen ) {
3035 NEXT_OFF( convert ) = (U16)(tail - convert);
3036 ARG_SET( convert, data_slot );
3037 /* Store the offset to the first unabsorbed branch in
3038 jump[0], which is otherwise unused by the jump logic.
3039 We use this when dumping a trie and during optimisation. */
3041 trie->jump[0] = (U16)(nextbranch - convert);
3043 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3044 * and there is a bitmap
3045 * and the first "jump target" node we found leaves enough room
3046 * then convert the TRIE node into a TRIEC node, with the bitmap
3047 * embedded inline in the opcode - this is hypothetically faster.
3049 if ( !trie->states[trie->startstate].wordnum
3051 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3053 OP( convert ) = TRIEC;
3054 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3055 PerlMemShared_free(trie->bitmap);
3058 OP( convert ) = TRIE;
3060 /* store the type in the flags */
3061 convert->flags = nodetype;
3065 + regarglen[ OP( convert ) ];
3067 /* XXX We really should free up the resource in trie now,
3068 as we won't use them - (which resources?) dmq */
3070 /* needed for dumping*/
3071 DEBUG_r(if (optimize) {
3072 regnode *opt = convert;
3074 while ( ++opt < optimize) {
3075 Set_Node_Offset_Length(opt,0,0);
3078 Try to clean up some of the debris left after the
3081 while( optimize < jumper ) {
3082 mjd_nodelen += Node_Length((optimize));
3083 OP( optimize ) = OPTIMIZED;
3084 Set_Node_Offset_Length(optimize,0,0);
3087 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3089 } /* end node insert */
3091 /* Finish populating the prev field of the wordinfo array. Walk back
3092 * from each accept state until we find another accept state, and if
3093 * so, point the first word's .prev field at the second word. If the
3094 * second already has a .prev field set, stop now. This will be the
3095 * case either if we've already processed that word's accept state,
3096 * or that state had multiple words, and the overspill words were
3097 * already linked up earlier.
3104 for (word=1; word <= trie->wordcount; word++) {
3106 if (trie->wordinfo[word].prev)
3108 state = trie->wordinfo[word].accept;
3110 state = prev_states[state];
3113 prev = trie->states[state].wordnum;
3117 trie->wordinfo[word].prev = prev;
3119 Safefree(prev_states);
3123 /* and now dump out the compressed format */
3124 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3126 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3128 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3129 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3131 SvREFCNT_dec_NN(revcharmap);
3135 : trie->startstate>1
3141 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3143 /* The Trie is constructed and compressed now so we can build a fail array if
3146 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3148 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3152 We find the fail state for each state in the trie, this state is the longest
3153 proper suffix of the current state's 'word' that is also a proper prefix of
3154 another word in our trie. State 1 represents the word '' and is thus the
3155 default fail state. This allows the DFA not to have to restart after its
3156 tried and failed a word at a given point, it simply continues as though it
3157 had been matching the other word in the first place.
3159 'abcdgu'=~/abcdefg|cdgu/
3160 When we get to 'd' we are still matching the first word, we would encounter
3161 'g' which would fail, which would bring us to the state representing 'd' in
3162 the second word where we would try 'g' and succeed, proceeding to match
3165 /* add a fail transition */
3166 const U32 trie_offset = ARG(source);
3167 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3169 const U32 ucharcount = trie->uniquecharcount;
3170 const U32 numstates = trie->statecount;
3171 const U32 ubound = trie->lasttrans + ucharcount;
3175 U32 base = trie->states[ 1 ].trans.base;
3178 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3180 GET_RE_DEBUG_FLAGS_DECL;
3182 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3183 PERL_UNUSED_CONTEXT;
3185 PERL_UNUSED_ARG(depth);
3188 if ( OP(source) == TRIE ) {
3189 struct regnode_1 *op = (struct regnode_1 *)
3190 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3191 StructCopy(source,op,struct regnode_1);
3192 stclass = (regnode *)op;
3194 struct regnode_charclass *op = (struct regnode_charclass *)
3195 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3196 StructCopy(source,op,struct regnode_charclass);
3197 stclass = (regnode *)op;
3199 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3201 ARG_SET( stclass, data_slot );
3202 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3203 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3204 aho->trie=trie_offset;
3205 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3206 Copy( trie->states, aho->states, numstates, reg_trie_state );
3207 Newxz( q, numstates, U32);
3208 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3211 /* initialize fail[0..1] to be 1 so that we always have
3212 a valid final fail state */
3213 fail[ 0 ] = fail[ 1 ] = 1;
3215 for ( charid = 0; charid < ucharcount ; charid++ ) {
3216 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3218 q[ q_write ] = newstate;
3219 /* set to point at the root */
3220 fail[ q[ q_write++ ] ]=1;
3223 while ( q_read < q_write) {
3224 const U32 cur = q[ q_read++ % numstates ];
3225 base = trie->states[ cur ].trans.base;
3227 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3228 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3230 U32 fail_state = cur;
3233 fail_state = fail[ fail_state ];
3234 fail_base = aho->states[ fail_state ].trans.base;
3235 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3237 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3238 fail[ ch_state ] = fail_state;
3239 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3241 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3243 q[ q_write++ % numstates] = ch_state;
3247 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3248 when we fail in state 1, this allows us to use the
3249 charclass scan to find a valid start char. This is based on the principle
3250 that theres a good chance the string being searched contains lots of stuff
3251 that cant be a start char.
3253 fail[ 0 ] = fail[ 1 ] = 0;
3254 DEBUG_TRIE_COMPILE_r({
3255 PerlIO_printf(Perl_debug_log,
3256 "%*sStclass Failtable (%"UVuf" states): 0",
3257 (int)(depth * 2), "", (UV)numstates
3259 for( q_read=1; q_read<numstates; q_read++ ) {
3260 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3262 PerlIO_printf(Perl_debug_log, "\n");
3265 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3270 #define DEBUG_PEEP(str,scan,depth) \
3271 DEBUG_OPTIMISE_r({if (scan){ \
3272 regnode *Next = regnext(scan); \
3273 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state); \
3274 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)", \
3275 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3276 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3277 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3278 PerlIO_printf(Perl_debug_log, "\n"); \
3281 /* The below joins as many adjacent EXACTish nodes as possible into a single
3282 * one. The regop may be changed if the node(s) contain certain sequences that
3283 * require special handling. The joining is only done if:
3284 * 1) there is room in the current conglomerated node to entirely contain the
3286 * 2) they are the exact same node type
3288 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3289 * these get optimized out
3291 * If a node is to match under /i (folded), the number of characters it matches
3292 * can be different than its character length if it contains a multi-character
3293 * fold. *min_subtract is set to the total delta number of characters of the
3296 * And *unfolded_multi_char is set to indicate whether or not the node contains
3297 * an unfolded multi-char fold. This happens when whether the fold is valid or
3298 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3299 * SMALL LETTER SHARP S, as only if the target string being matched against
3300 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3301 * folding rules depend on the locale in force at runtime. (Multi-char folds
3302 * whose components are all above the Latin1 range are not run-time locale
3303 * dependent, and have already been folded by the time this function is
3306 * This is as good a place as any to discuss the design of handling these
3307 * multi-character fold sequences. It's been wrong in Perl for a very long
3308 * time. There are three code points in Unicode whose multi-character folds
3309 * were long ago discovered to mess things up. The previous designs for
3310 * dealing with these involved assigning a special node for them. This
3311 * approach doesn't always work, as evidenced by this example:
3312 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3313 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3314 * would match just the \xDF, it won't be able to handle the case where a
3315 * successful match would have to cross the node's boundary. The new approach
3316 * that hopefully generally solves the problem generates an EXACTFU_SS node
3317 * that is "sss" in this case.
3319 * It turns out that there are problems with all multi-character folds, and not
3320 * just these three. Now the code is general, for all such cases. The
3321 * approach taken is:
3322 * 1) This routine examines each EXACTFish node that could contain multi-
3323 * character folded sequences. Since a single character can fold into
3324 * such a sequence, the minimum match length for this node is less than
3325 * the number of characters in the node. This routine returns in
3326 * *min_subtract how many characters to subtract from the the actual
3327 * length of the string to get a real minimum match length; it is 0 if
3328 * there are no multi-char foldeds. This delta is used by the caller to
3329 * adjust the min length of the match, and the delta between min and max,
3330 * so that the optimizer doesn't reject these possibilities based on size
3332 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3333 * is used for an EXACTFU node that contains at least one "ss" sequence in
3334 * it. For non-UTF-8 patterns and strings, this is the only case where
3335 * there is a possible fold length change. That means that a regular
3336 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3337 * with length changes, and so can be processed faster. regexec.c takes
3338 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3339 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3340 * known until runtime). This saves effort in regex matching. However,
3341 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3342 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3343 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3344 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3345 * possibilities for the non-UTF8 patterns are quite simple, except for
3346 * the sharp s. All the ones that don't involve a UTF-8 target string are
3347 * members of a fold-pair, and arrays are set up for all of them so that
3348 * the other member of the pair can be found quickly. Code elsewhere in
3349 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3350 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3351 * described in the next item.
3352 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3353 * validity of the fold won't be known until runtime, and so must remain
3354 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3355 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3356 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3357 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3358 * The reason this is a problem is that the optimizer part of regexec.c
3359 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3360 * that a character in the pattern corresponds to at most a single
3361 * character in the target string. (And I do mean character, and not byte
3362 * here, unlike other parts of the documentation that have never been
3363 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3364 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3365 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3366 * nodes, violate the assumption, and they are the only instances where it
3367 * is violated. I'm reluctant to try to change the assumption, as the
3368 * code involved is impenetrable to me (khw), so instead the code here
3369 * punts. This routine examines EXACTFL nodes, and (when the pattern
3370 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3371 * boolean indicating whether or not the node contains such a fold. When
3372 * it is true, the caller sets a flag that later causes the optimizer in
3373 * this file to not set values for the floating and fixed string lengths,
3374 * and thus avoids the optimizer code in regexec.c that makes the invalid
3375 * assumption. Thus, there is no optimization based on string lengths for
3376 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3377 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3378 * assumption is wrong only in these cases is that all other non-UTF-8
3379 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3380 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3381 * EXACTF nodes because we don't know at compile time if it actually
3382 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3383 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3384 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3385 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3386 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3387 * string would require the pattern to be forced into UTF-8, the overhead
3388 * of which we want to avoid. Similarly the unfolded multi-char folds in
3389 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3392 * Similarly, the code that generates tries doesn't currently handle
3393 * not-already-folded multi-char folds, and it looks like a pain to change
3394 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3395 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3396 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3397 * using /iaa matching will be doing so almost entirely with ASCII
3398 * strings, so this should rarely be encountered in practice */
3400 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3401 if (PL_regkind[OP(scan)] == EXACT) \
3402 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3405 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3406 UV *min_subtract, bool *unfolded_multi_char,
3407 U32 flags,regnode *val, U32 depth)
3409 /* Merge several consecutive EXACTish nodes into one. */
3410 regnode *n = regnext(scan);
3412 regnode *next = scan + NODE_SZ_STR(scan);
3416 regnode *stop = scan;
3417 GET_RE_DEBUG_FLAGS_DECL;
3419 PERL_UNUSED_ARG(depth);
3422 PERL_ARGS_ASSERT_JOIN_EXACT;
3423 #ifndef EXPERIMENTAL_INPLACESCAN
3424 PERL_UNUSED_ARG(flags);
3425 PERL_UNUSED_ARG(val);
3427 DEBUG_PEEP("join",scan,depth);
3429 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3430 * EXACT ones that are mergeable to the current one. */
3432 && (PL_regkind[OP(n)] == NOTHING
3433 || (stringok && OP(n) == OP(scan)))
3435 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3438 if (OP(n) == TAIL || n > next)
3440 if (PL_regkind[OP(n)] == NOTHING) {
3441 DEBUG_PEEP("skip:",n,depth);
3442 NEXT_OFF(scan) += NEXT_OFF(n);
3443 next = n + NODE_STEP_REGNODE;
3450 else if (stringok) {
3451 const unsigned int oldl = STR_LEN(scan);
3452 regnode * const nnext = regnext(n);
3454 /* XXX I (khw) kind of doubt that this works on platforms (should
3455 * Perl ever run on one) where U8_MAX is above 255 because of lots
3456 * of other assumptions */
3457 /* Don't join if the sum can't fit into a single node */
3458 if (oldl + STR_LEN(n) > U8_MAX)
3461 DEBUG_PEEP("merg",n,depth);
3464 NEXT_OFF(scan) += NEXT_OFF(n);
3465 STR_LEN(scan) += STR_LEN(n);
3466 next = n + NODE_SZ_STR(n);
3467 /* Now we can overwrite *n : */
3468 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3476 #ifdef EXPERIMENTAL_INPLACESCAN
3477 if (flags && !NEXT_OFF(n)) {
3478 DEBUG_PEEP("atch", val, depth);
3479 if (reg_off_by_arg[OP(n)]) {
3480 ARG_SET(n, val - n);
3483 NEXT_OFF(n) = val - n;
3491 *unfolded_multi_char = FALSE;
3493 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3494 * can now analyze for sequences of problematic code points. (Prior to
3495 * this final joining, sequences could have been split over boundaries, and
3496 * hence missed). The sequences only happen in folding, hence for any
3497 * non-EXACT EXACTish node */
3498 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3499 U8* s0 = (U8*) STRING(scan);
3501 U8* s_end = s0 + STR_LEN(scan);
3503 int total_count_delta = 0; /* Total delta number of characters that
3504 multi-char folds expand to */
3506 /* One pass is made over the node's string looking for all the
3507 * possibilities. To avoid some tests in the loop, there are two main
3508 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3513 if (OP(scan) == EXACTFL) {
3516 /* An EXACTFL node would already have been changed to another
3517 * node type unless there is at least one character in it that
3518 * is problematic; likely a character whose fold definition
3519 * won't be known until runtime, and so has yet to be folded.
3520 * For all but the UTF-8 locale, folds are 1-1 in length, but
3521 * to handle the UTF-8 case, we need to create a temporary
3522 * folded copy using UTF-8 locale rules in order to analyze it.
3523 * This is because our macros that look to see if a sequence is
3524 * a multi-char fold assume everything is folded (otherwise the
3525 * tests in those macros would be too complicated and slow).
3526 * Note that here, the non-problematic folds will have already
3527 * been done, so we can just copy such characters. We actually
3528 * don't completely fold the EXACTFL string. We skip the
3529 * unfolded multi-char folds, as that would just create work
3530 * below to figure out the size they already are */
3532 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3535 STRLEN s_len = UTF8SKIP(s);
3536 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3537 Copy(s, d, s_len, U8);
3540 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3541 *unfolded_multi_char = TRUE;
3542 Copy(s, d, s_len, U8);
3545 else if (isASCII(*s)) {
3546 *(d++) = toFOLD(*s);
3550 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3556 /* Point the remainder of the routine to look at our temporary
3560 } /* End of creating folded copy of EXACTFL string */
3562 /* Examine the string for a multi-character fold sequence. UTF-8
3563 * patterns have all characters pre-folded by the time this code is
3565 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3566 length sequence we are looking for is 2 */
3568 int count = 0; /* How many characters in a multi-char fold */
3569 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3570 if (! len) { /* Not a multi-char fold: get next char */
3575 /* Nodes with 'ss' require special handling, except for
3576 * EXACTFA-ish for which there is no multi-char fold to this */
3577 if (len == 2 && *s == 's' && *(s+1) == 's'
3578 && OP(scan) != EXACTFA
3579 && OP(scan) != EXACTFA_NO_TRIE)
3582 if (OP(scan) != EXACTFL) {
3583 OP(scan) = EXACTFU_SS;
3587 else { /* Here is a generic multi-char fold. */
3588 U8* multi_end = s + len;
3590 /* Count how many characters are in it. In the case of
3591 * /aa, no folds which contain ASCII code points are
3592 * allowed, so check for those, and skip if found. */
3593 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3594 count = utf8_length(s, multi_end);
3598 while (s < multi_end) {
3601 goto next_iteration;
3611 /* The delta is how long the sequence is minus 1 (1 is how long
3612 * the character that folds to the sequence is) */
3613 total_count_delta += count - 1;
3617 /* We created a temporary folded copy of the string in EXACTFL
3618 * nodes. Therefore we need to be sure it doesn't go below zero,
3619 * as the real string could be shorter */
3620 if (OP(scan) == EXACTFL) {
3621 int total_chars = utf8_length((U8*) STRING(scan),
3622 (U8*) STRING(scan) + STR_LEN(scan));
3623 if (total_count_delta > total_chars) {
3624 total_count_delta = total_chars;
3628 *min_subtract += total_count_delta;
3631 else if (OP(scan) == EXACTFA) {
3633 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3634 * fold to the ASCII range (and there are no existing ones in the
3635 * upper latin1 range). But, as outlined in the comments preceding
3636 * this function, we need to flag any occurrences of the sharp s.
3637 * This character forbids trie formation (because of added
3640 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3641 OP(scan) = EXACTFA_NO_TRIE;
3642 *unfolded_multi_char = TRUE;
3651 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3652 * folds that are all Latin1. As explained in the comments
3653 * preceding this function, we look also for the sharp s in EXACTF
3654 * and EXACTFL nodes; it can be in the final position. Otherwise
3655 * we can stop looking 1 byte earlier because have to find at least
3656 * two characters for a multi-fold */
3657 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3662 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3663 if (! len) { /* Not a multi-char fold. */
3664 if (*s == LATIN_SMALL_LETTER_SHARP_S
3665 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3667 *unfolded_multi_char = TRUE;
3674 && isALPHA_FOLD_EQ(*s, 's')
3675 && isALPHA_FOLD_EQ(*(s+1), 's'))
3678 /* EXACTF nodes need to know that the minimum length
3679 * changed so that a sharp s in the string can match this
3680 * ss in the pattern, but they remain EXACTF nodes, as they
3681 * won't match this unless the target string is is UTF-8,
3682 * which we don't know until runtime. EXACTFL nodes can't
3683 * transform into EXACTFU nodes */
3684 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
3685 OP(scan) = EXACTFU_SS;
3689 *min_subtract += len - 1;
3696 /* Allow dumping but overwriting the collection of skipped
3697 * ops and/or strings with fake optimized ops */
3698 n = scan + NODE_SZ_STR(scan);
3706 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3710 /* REx optimizer. Converts nodes into quicker variants "in place".
3711 Finds fixed substrings. */
3713 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3714 to the position after last scanned or to NULL. */
3716 #define INIT_AND_WITHP \
3717 assert(!and_withp); \
3718 Newx(and_withp,1, regnode_ssc); \
3719 SAVEFREEPV(and_withp)
3723 S_unwind_scan_frames(pTHX_ const void *p)
3725 scan_frame *f= (scan_frame *)p;
3727 scan_frame *n= f->next_frame;
3735 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3736 SSize_t *minlenp, SSize_t *deltap,
3741 regnode_ssc *and_withp,
3742 U32 flags, U32 depth)
3743 /* scanp: Start here (read-write). */
3744 /* deltap: Write maxlen-minlen here. */
3745 /* last: Stop before this one. */
3746 /* data: string data about the pattern */
3747 /* stopparen: treat close N as END */
3748 /* recursed: which subroutines have we recursed into */
3749 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3751 /* There must be at least this number of characters to match */
3754 regnode *scan = *scanp, *next;
3756 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3757 int is_inf_internal = 0; /* The studied chunk is infinite */
3758 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3759 scan_data_t data_fake;
3760 SV *re_trie_maxbuff = NULL;
3761 regnode *first_non_open = scan;
3762 SSize_t stopmin = SSize_t_MAX;
3763 scan_frame *frame = NULL;
3764 GET_RE_DEBUG_FLAGS_DECL;
3766 PERL_ARGS_ASSERT_STUDY_CHUNK;
3770 while (first_non_open && OP(first_non_open) == OPEN)
3771 first_non_open=regnext(first_non_open);
3777 RExC_study_chunk_recursed_count++;
3779 DEBUG_OPTIMISE_MORE_r(
3781 PerlIO_printf(Perl_debug_log,
3782 "%*sstudy_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
3783 (int)(depth*2), "", (long)stopparen,
3784 (unsigned long)RExC_study_chunk_recursed_count,
3785 (unsigned long)depth, (unsigned long)recursed_depth,
3788 if (recursed_depth) {
3791 for ( j = 0 ; j < recursed_depth ; j++ ) {
3792 for ( i = 0 ; i < (U32)RExC_npar ; i++ ) {
3794 PAREN_TEST(RExC_study_chunk_recursed +
3795 ( j * RExC_study_chunk_recursed_bytes), i )
3798 !PAREN_TEST(RExC_study_chunk_recursed +
3799 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
3802 PerlIO_printf(Perl_debug_log," %d",(int)i);
3806 if ( j + 1 < recursed_depth ) {
3807 PerlIO_printf(Perl_debug_log, ",");
3811 PerlIO_printf(Perl_debug_log,"\n");
3814 while ( scan && OP(scan) != END && scan < last ){
3815 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3816 node length to get a real minimum (because
3817 the folded version may be shorter) */
3818 bool unfolded_multi_char = FALSE;
3819 /* Peephole optimizer: */
3820 DEBUG_STUDYDATA("Peep:", data, depth);
3821 DEBUG_PEEP("Peep", scan, depth);
3824 /* The reason we do this here we need to deal with things like /(?:f)(?:o)(?:o)/
3825 * which cant be dealt with by the normal EXACT parsing code, as each (?:..) is handled
3826 * by a different invocation of reg() -- Yves
3828 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
3830 /* Follow the next-chain of the current node and optimize
3831 away all the NOTHINGs from it. */
3832 if (OP(scan) != CURLYX) {
3833 const int max = (reg_off_by_arg[OP(scan)]
3835 /* I32 may be smaller than U16 on CRAYs! */
3836 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3837 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3841 /* Skip NOTHING and LONGJMP. */
3842 while ((n = regnext(n))
3843 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3844 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3845 && off + noff < max)
3847 if (reg_off_by_arg[OP(scan)])
3850 NEXT_OFF(scan) = off;
3853 /* The principal pseudo-switch. Cannot be a switch, since we
3854 look into several different things. */
3855 if ( OP(scan) == DEFINEP ) {
3857 SSize_t deltanext = 0;
3858 SSize_t fake_last_close = 0;
3859 I32 f = SCF_IN_DEFINE;
3861 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3862 scan = regnext(scan);
3863 assert( OP(scan) == IFTHEN );
3864 DEBUG_PEEP("expect IFTHEN", scan, depth);
3866 data_fake.last_closep= &fake_last_close;
3868 next = regnext(scan);
3869 scan = NEXTOPER(NEXTOPER(scan));
3870 DEBUG_PEEP("scan", scan, depth);
3871 DEBUG_PEEP("next", next, depth);
3873 /* we suppose the run is continuous, last=next...
3874 * NOTE we dont use the return here! */
3875 (void)study_chunk(pRExC_state, &scan, &minlen,
3876 &deltanext, next, &data_fake, stopparen,
3877 recursed_depth, NULL, f, depth+1);
3882 OP(scan) == BRANCH ||
3883 OP(scan) == BRANCHJ ||
3886 next = regnext(scan);
3889 /* The op(next)==code check below is to see if we
3890 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
3891 * IFTHEN is special as it might not appear in pairs.
3892 * Not sure whether BRANCH-BRANCHJ is possible, regardless
3893 * we dont handle it cleanly. */
3894 if (OP(next) == code || code == IFTHEN) {
3895 /* NOTE - There is similar code to this block below for
3896 * handling TRIE nodes on a re-study. If you change stuff here
3897 * check there too. */
3898 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3900 regnode * const startbranch=scan;
3902 if (flags & SCF_DO_SUBSTR) {
3903 /* Cannot merge strings after this. */
3904 scan_commit(pRExC_state, data, minlenp, is_inf);
3907 if (flags & SCF_DO_STCLASS)
3908 ssc_init_zero(pRExC_state, &accum);
3910 while (OP(scan) == code) {
3911 SSize_t deltanext, minnext, fake;
3913 regnode_ssc this_class;
3915 DEBUG_PEEP("Branch", scan, depth);
3918 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3920 data_fake.whilem_c = data->whilem_c;
3921 data_fake.last_closep = data->last_closep;
3924 data_fake.last_closep = &fake;
3926 data_fake.pos_delta = delta;
3927 next = regnext(scan);
3929 scan = NEXTOPER(scan); /* everything */
3930 if (code != BRANCH) /* everything but BRANCH */
3931 scan = NEXTOPER(scan);
3933 if (flags & SCF_DO_STCLASS) {
3934 ssc_init(pRExC_state, &this_class);
3935 data_fake.start_class = &this_class;
3936 f = SCF_DO_STCLASS_AND;
3938 if (flags & SCF_WHILEM_VISITED_POS)
3939 f |= SCF_WHILEM_VISITED_POS;
3941 /* we suppose the run is continuous, last=next...*/
3942 minnext = study_chunk(pRExC_state, &scan, minlenp,
3943 &deltanext, next, &data_fake, stopparen,
3944 recursed_depth, NULL, f,depth+1);
3948 if (deltanext == SSize_t_MAX) {
3949 is_inf = is_inf_internal = 1;
3951 } else if (max1 < minnext + deltanext)
3952 max1 = minnext + deltanext;
3954 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3956 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3957 if ( stopmin > minnext)
3958 stopmin = min + min1;
3959 flags &= ~SCF_DO_SUBSTR;
3961 data->flags |= SCF_SEEN_ACCEPT;
3964 if (data_fake.flags & SF_HAS_EVAL)
3965 data->flags |= SF_HAS_EVAL;
3966 data->whilem_c = data_fake.whilem_c;
3968 if (flags & SCF_DO_STCLASS)
3969 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
3971 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3973 if (flags & SCF_DO_SUBSTR) {
3974 data->pos_min += min1;
3975 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3976 data->pos_delta = SSize_t_MAX;
3978 data->pos_delta += max1 - min1;
3979 if (max1 != min1 || is_inf)
3980 data->longest = &(data->longest_float);
3983 if (delta == SSize_t_MAX
3984 || SSize_t_MAX - delta - (max1 - min1) < 0)
3985 delta = SSize_t_MAX;
3987 delta += max1 - min1;
3988 if (flags & SCF_DO_STCLASS_OR) {
3989 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
3991 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
3992 flags &= ~SCF_DO_STCLASS;
3995 else if (flags & SCF_DO_STCLASS_AND) {
3997 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
3998 flags &= ~SCF_DO_STCLASS;
4001 /* Switch to OR mode: cache the old value of
4002 * data->start_class */
4004 StructCopy(data->start_class, and_withp, regnode_ssc);
4005 flags &= ~SCF_DO_STCLASS_AND;
4006 StructCopy(&accum, data->start_class, regnode_ssc);
4007 flags |= SCF_DO_STCLASS_OR;
4011 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4012 OP( startbranch ) == BRANCH )
4016 Assuming this was/is a branch we are dealing with: 'scan'
4017 now points at the item that follows the branch sequence,
4018 whatever it is. We now start at the beginning of the
4019 sequence and look for subsequences of
4025 which would be constructed from a pattern like
4028 If we can find such a subsequence we need to turn the first
4029 element into a trie and then add the subsequent branch exact
4030 strings to the trie.
4034 1. patterns where the whole set of branches can be
4037 2. patterns where only a subset can be converted.
4039 In case 1 we can replace the whole set with a single regop
4040 for the trie. In case 2 we need to keep the start and end
4043 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4044 becomes BRANCH TRIE; BRANCH X;
4046 There is an additional case, that being where there is a
4047 common prefix, which gets split out into an EXACT like node
4048 preceding the TRIE node.
4050 If x(1..n)==tail then we can do a simple trie, if not we make
4051 a "jump" trie, such that when we match the appropriate word
4052 we "jump" to the appropriate tail node. Essentially we turn
4053 a nested if into a case structure of sorts.
4058 if (!re_trie_maxbuff) {
4059 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4060 if (!SvIOK(re_trie_maxbuff))
4061 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4063 if ( SvIV(re_trie_maxbuff)>=0 ) {
4065 regnode *first = (regnode *)NULL;
4066 regnode *last = (regnode *)NULL;
4067 regnode *tail = scan;
4071 /* var tail is used because there may be a TAIL
4072 regop in the way. Ie, the exacts will point to the
4073 thing following the TAIL, but the last branch will
4074 point at the TAIL. So we advance tail. If we
4075 have nested (?:) we may have to move through several
4079 while ( OP( tail ) == TAIL ) {
4080 /* this is the TAIL generated by (?:) */
4081 tail = regnext( tail );
4085 DEBUG_TRIE_COMPILE_r({
4086 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4087 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
4088 (int)depth * 2 + 2, "",
4089 "Looking for TRIE'able sequences. Tail node is: ",
4090 SvPV_nolen_const( RExC_mysv )
4096 Step through the branches
4097 cur represents each branch,
4098 noper is the first thing to be matched as part
4100 noper_next is the regnext() of that node.
4102 We normally handle a case like this
4103 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4104 support building with NOJUMPTRIE, which restricts
4105 the trie logic to structures like /FOO|BAR/.
4107 If noper is a trieable nodetype then the branch is
4108 a possible optimization target. If we are building
4109 under NOJUMPTRIE then we require that noper_next is
4110 the same as scan (our current position in the regex
4113 Once we have two or more consecutive such branches
4114 we can create a trie of the EXACT's contents and
4115 stitch it in place into the program.
4117 If the sequence represents all of the branches in
4118 the alternation we replace the entire thing with a
4121 Otherwise when it is a subsequence we need to
4122 stitch it in place and replace only the relevant
4123 branches. This means the first branch has to remain
4124 as it is used by the alternation logic, and its
4125 next pointer, and needs to be repointed at the item
4126 on the branch chain following the last branch we
4127 have optimized away.
4129 This could be either a BRANCH, in which case the
4130 subsequence is internal, or it could be the item
4131 following the branch sequence in which case the
4132 subsequence is at the end (which does not
4133 necessarily mean the first node is the start of the
4136 TRIE_TYPE(X) is a define which maps the optype to a
4140 ----------------+-----------
4144 EXACTFU_SS | EXACTFU
4147 EXACTFLU8 | EXACTFLU8
4151 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4153 : ( EXACT == (X) ) \
4155 : ( EXACTFU == (X) || EXACTFU_SS == (X) ) \
4157 : ( EXACTFA == (X) ) \
4159 : ( EXACTL == (X) ) \
4161 : ( EXACTFLU8 == (X) ) \
4165 /* dont use tail as the end marker for this traverse */
4166 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4167 regnode * const noper = NEXTOPER( cur );
4168 U8 noper_type = OP( noper );
4169 U8 noper_trietype = TRIE_TYPE( noper_type );
4170 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4171 regnode * const noper_next = regnext( noper );
4172 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
4173 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
4176 DEBUG_TRIE_COMPILE_r({
4177 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4178 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
4179 (int)depth * 2 + 2,"", SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4181 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4182 PerlIO_printf( Perl_debug_log, " -> %s",
4183 SvPV_nolen_const(RExC_mysv));
4186 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4187 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
4188 SvPV_nolen_const(RExC_mysv));
4190 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
4191 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4192 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4196 /* Is noper a trieable nodetype that can be merged
4197 * with the current trie (if there is one)? */
4201 ( noper_trietype == NOTHING)
4202 || ( trietype == NOTHING )
4203 || ( trietype == noper_trietype )
4206 && noper_next == tail
4210 /* Handle mergable triable node Either we are
4211 * the first node in a new trieable sequence,
4212 * in which case we do some bookkeeping,
4213 * otherwise we update the end pointer. */
4216 if ( noper_trietype == NOTHING ) {
4217 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4218 regnode * const noper_next = regnext( noper );
4219 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
4220 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4223 if ( noper_next_trietype ) {
4224 trietype = noper_next_trietype;
4225 } else if (noper_next_type) {
4226 /* a NOTHING regop is 1 regop wide.
4227 * We need at least two for a trie
4228 * so we can't merge this in */
4232 trietype = noper_trietype;
4235 if ( trietype == NOTHING )
4236 trietype = noper_trietype;
4241 } /* end handle mergable triable node */
4243 /* handle unmergable node -
4244 * noper may either be a triable node which can
4245 * not be tried together with the current trie,
4246 * or a non triable node */
4248 /* If last is set and trietype is not
4249 * NOTHING then we have found at least two
4250 * triable branch sequences in a row of a
4251 * similar trietype so we can turn them
4252 * into a trie. If/when we allow NOTHING to
4253 * start a trie sequence this condition
4254 * will be required, and it isn't expensive
4255 * so we leave it in for now. */
4256 if ( trietype && trietype != NOTHING )
4257 make_trie( pRExC_state,
4258 startbranch, first, cur, tail,
4259 count, trietype, depth+1 );
4260 last = NULL; /* note: we clear/update
4261 first, trietype etc below,
4262 so we dont do it here */
4266 && noper_next == tail
4269 /* noper is triable, so we can start a new
4273 trietype = noper_trietype;
4275 /* if we already saw a first but the
4276 * current node is not triable then we have
4277 * to reset the first information. */
4282 } /* end handle unmergable node */
4283 } /* loop over branches */
4284 DEBUG_TRIE_COMPILE_r({
4285 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4286 PerlIO_printf( Perl_debug_log,
4287 "%*s- %s (%d) <SCAN FINISHED>\n",
4289 "", SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4292 if ( last && trietype ) {
4293 if ( trietype != NOTHING ) {
4294 /* the last branch of the sequence was part of
4295 * a trie, so we have to construct it here
4296 * outside of the loop */
4297 made= make_trie( pRExC_state, startbranch,
4298 first, scan, tail, count,
4299 trietype, depth+1 );
4300 #ifdef TRIE_STUDY_OPT
4301 if ( ((made == MADE_EXACT_TRIE &&
4302 startbranch == first)
4303 || ( first_non_open == first )) &&
4305 flags |= SCF_TRIE_RESTUDY;
4306 if ( startbranch == first
4309 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4314 /* at this point we know whatever we have is a
4315 * NOTHING sequence/branch AND if 'startbranch'
4316 * is 'first' then we can turn the whole thing
4319 if ( startbranch == first ) {
4321 /* the entire thing is a NOTHING sequence,
4322 * something like this: (?:|) So we can
4323 * turn it into a plain NOTHING op. */
4324 DEBUG_TRIE_COMPILE_r({
4325 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4326 PerlIO_printf( Perl_debug_log,
4327 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
4328 "", SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4331 OP(startbranch)= NOTHING;
4332 NEXT_OFF(startbranch)= tail - startbranch;
4333 for ( opt= startbranch + 1; opt < tail ; opt++ )
4337 } /* end if ( last) */
4338 } /* TRIE_MAXBUF is non zero */
4343 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4344 scan = NEXTOPER(NEXTOPER(scan));
4345 } else /* single branch is optimized. */
4346 scan = NEXTOPER(scan);
4348 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
4350 regnode *start = NULL;
4351 regnode *end = NULL;
4352 U32 my_recursed_depth= recursed_depth;
4355 if (OP(scan) != SUSPEND) { /* GOSUB/GOSTART */
4356 /* Do setup, note this code has side effects beyond
4357 * the rest of this block. Specifically setting
4358 * RExC_recurse[] must happen at least once during
4360 if (OP(scan) == GOSUB) {
4362 RExC_recurse[ARG2L(scan)] = scan;
4363 start = RExC_open_parens[paren-1];
4364 end = RExC_close_parens[paren-1];
4366 start = RExC_rxi->program + 1;
4369 /* NOTE we MUST always execute the above code, even
4370 * if we do nothing with a GOSUB/GOSTART */
4372 ( flags & SCF_IN_DEFINE )
4375 (is_inf_internal || is_inf || data->flags & SF_IS_INF)
4377 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
4380 /* no need to do anything here if we are in a define. */
4381 /* or we are after some kind of infinite construct
4382 * so we can skip recursing into this item.
4383 * Since it is infinite we will not change the maxlen
4384 * or delta, and if we miss something that might raise
4385 * the minlen it will merely pessimise a little.
4387 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
4388 * might result in a minlen of 1 and not of 4,
4389 * but this doesn't make us mismatch, just try a bit
4390 * harder than we should.
4392 scan= regnext(scan);
4399 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4401 /* it is quite possible that there are more efficient ways
4402 * to do this. We maintain a bitmap per level of recursion
4403 * of which patterns we have entered so we can detect if a
4404 * pattern creates a possible infinite loop. When we
4405 * recurse down a level we copy the previous levels bitmap
4406 * down. When we are at recursion level 0 we zero the top
4407 * level bitmap. It would be nice to implement a different
4408 * more efficient way of doing this. In particular the top
4409 * level bitmap may be unnecessary.
4411 if (!recursed_depth) {
4412 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4414 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4415 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4416 RExC_study_chunk_recursed_bytes, U8);
4418 /* we havent recursed into this paren yet, so recurse into it */
4419 DEBUG_STUDYDATA("set:", data,depth);
4420 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4421 my_recursed_depth= recursed_depth + 1;
4423 DEBUG_STUDYDATA("inf:", data,depth);
4424 /* some form of infinite recursion, assume infinite length
4426 if (flags & SCF_DO_SUBSTR) {
4427 scan_commit(pRExC_state, data, minlenp, is_inf);
4428 data->longest = &(data->longest_float);
4430 is_inf = is_inf_internal = 1;
4431 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4432 ssc_anything(data->start_class);
4433 flags &= ~SCF_DO_STCLASS;
4435 start= NULL; /* reset start so we dont recurse later on. */
4440 end = regnext(scan);
4443 scan_frame *newframe;
4445 if (!RExC_frame_last) {
4446 Newxz(newframe, 1, scan_frame);
4447 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
4448 RExC_frame_head= newframe;
4450 } else if (!RExC_frame_last->next_frame) {
4451 Newxz(newframe,1,scan_frame);
4452 RExC_frame_last->next_frame= newframe;
4453 newframe->prev_frame= RExC_frame_last;
4456 newframe= RExC_frame_last->next_frame;
4458 RExC_frame_last= newframe;
4460 newframe->next_regnode = regnext(scan);
4461 newframe->last_regnode = last;
4462 newframe->stopparen = stopparen;
4463 newframe->prev_recursed_depth = recursed_depth;
4464 newframe->this_prev_frame= frame;
4466 DEBUG_STUDYDATA("frame-new:",data,depth);
4467 DEBUG_PEEP("fnew", scan, depth);
4474 recursed_depth= my_recursed_depth;
4479 else if (OP(scan) == EXACT || OP(scan) == EXACTL) {
4480 SSize_t l = STR_LEN(scan);
4483 const U8 * const s = (U8*)STRING(scan);
4484 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4485 l = utf8_length(s, s + l);
4487 uc = *((U8*)STRING(scan));
4490 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4491 /* The code below prefers earlier match for fixed
4492 offset, later match for variable offset. */
4493 if (data->last_end == -1) { /* Update the start info. */
4494 data->last_start_min = data->pos_min;
4495 data->last_start_max = is_inf
4496 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4498 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4500 SvUTF8_on(data->last_found);
4502 SV * const sv = data->last_found;
4503 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4504 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4505 if (mg && mg->mg_len >= 0)
4506 mg->mg_len += utf8_length((U8*)STRING(scan),
4507 (U8*)STRING(scan)+STR_LEN(scan));
4509 data->last_end = data->pos_min + l;
4510 data->pos_min += l; /* As in the first entry. */
4511 data->flags &= ~SF_BEFORE_EOL;
4514 /* ANDing the code point leaves at most it, and not in locale, and
4515 * can't match null string */
4516 if (flags & SCF_DO_STCLASS_AND) {
4517 ssc_cp_and(data->start_class, uc);
4518 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4519 ssc_clear_locale(data->start_class);
4521 else if (flags & SCF_DO_STCLASS_OR) {
4522 ssc_add_cp(data->start_class, uc);
4523 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4525 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4526 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4528 flags &= ~SCF_DO_STCLASS;
4530 else if (PL_regkind[OP(scan)] == EXACT) {
4531 /* But OP != EXACT!, so is EXACTFish */
4532 SSize_t l = STR_LEN(scan);
4533 const U8 * s = (U8*)STRING(scan);
4535 /* Search for fixed substrings supports EXACT only. */
4536 if (flags & SCF_DO_SUBSTR) {
4538 scan_commit(pRExC_state, data, minlenp, is_inf);
4541 l = utf8_length(s, s + l);
4543 if (unfolded_multi_char) {
4544 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4546 min += l - min_subtract;
4548 delta += min_subtract;
4549 if (flags & SCF_DO_SUBSTR) {
4550 data->pos_min += l - min_subtract;
4551 if (data->pos_min < 0) {
4554 data->pos_delta += min_subtract;
4556 data->longest = &(data->longest_float);
4560 if (flags & SCF_DO_STCLASS) {
4561 SV* EXACTF_invlist = _make_exactf_invlist(pRExC_state, scan);
4563 assert(EXACTF_invlist);
4564 if (flags & SCF_DO_STCLASS_AND) {
4565 if (OP(scan) != EXACTFL)
4566 ssc_clear_locale(data->start_class);
4567 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4568 ANYOF_POSIXL_ZERO(data->start_class);
4569 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4571 else { /* SCF_DO_STCLASS_OR */
4572 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4573 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4575 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4576 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4578 flags &= ~SCF_DO_STCLASS;
4579 SvREFCNT_dec(EXACTF_invlist);
4582 else if (REGNODE_VARIES(OP(scan))) {
4583 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4584 I32 fl = 0, f = flags;
4585 regnode * const oscan = scan;
4586 regnode_ssc this_class;
4587 regnode_ssc *oclass = NULL;
4588 I32 next_is_eval = 0;
4590 switch (PL_regkind[OP(scan)]) {
4591 case WHILEM: /* End of (?:...)* . */
4592 scan = NEXTOPER(scan);
4595 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
4596 next = NEXTOPER(scan);
4597 if (OP(next) == EXACT
4598 || OP(next) == EXACTL
4599 || (flags & SCF_DO_STCLASS))
4602 maxcount = REG_INFTY;
4603 next = regnext(scan);
4604 scan = NEXTOPER(scan);
4608 if (flags & SCF_DO_SUBSTR)
4613 if (flags & SCF_DO_STCLASS) {
4615 maxcount = REG_INFTY;
4616 next = regnext(scan);
4617 scan = NEXTOPER(scan);
4620 if (flags & SCF_DO_SUBSTR) {
4621 scan_commit(pRExC_state, data, minlenp, is_inf);
4622 /* Cannot extend fixed substrings */
4623 data->longest = &(data->longest_float);
4625 is_inf = is_inf_internal = 1;
4626 scan = regnext(scan);
4627 goto optimize_curly_tail;
4629 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
4630 && (scan->flags == stopparen))
4635 mincount = ARG1(scan);
4636 maxcount = ARG2(scan);
4638 next = regnext(scan);
4639 if (OP(scan) == CURLYX) {
4640 I32 lp = (data ? *(data->last_closep) : 0);
4641 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
4643 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
4644 next_is_eval = (OP(scan) == EVAL);
4646 if (flags & SCF_DO_SUBSTR) {
4648 scan_commit(pRExC_state, data, minlenp, is_inf);
4649 /* Cannot extend fixed substrings */
4650 pos_before = data->pos_min;
4654 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
4656 data->flags |= SF_IS_INF;
4658 if (flags & SCF_DO_STCLASS) {
4659 ssc_init(pRExC_state, &this_class);
4660 oclass = data->start_class;
4661 data->start_class = &this_class;
4662 f |= SCF_DO_STCLASS_AND;
4663 f &= ~SCF_DO_STCLASS_OR;
4665 /* Exclude from super-linear cache processing any {n,m}
4666 regops for which the combination of input pos and regex
4667 pos is not enough information to determine if a match
4670 For example, in the regex /foo(bar\s*){4,8}baz/ with the
4671 regex pos at the \s*, the prospects for a match depend not
4672 only on the input position but also on how many (bar\s*)
4673 repeats into the {4,8} we are. */
4674 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
4675 f &= ~SCF_WHILEM_VISITED_POS;
4677 /* This will finish on WHILEM, setting scan, or on NULL: */
4678 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
4679 last, data, stopparen, recursed_depth, NULL,
4681 ? (f & ~SCF_DO_SUBSTR)
4685 if (flags & SCF_DO_STCLASS)
4686 data->start_class = oclass;
4687 if (mincount == 0 || minnext == 0) {
4688 if (flags & SCF_DO_STCLASS_OR) {
4689 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4691 else if (flags & SCF_DO_STCLASS_AND) {
4692 /* Switch to OR mode: cache the old value of
4693 * data->start_class */
4695 StructCopy(data->start_class, and_withp, regnode_ssc);
4696 flags &= ~SCF_DO_STCLASS_AND;
4697 StructCopy(&this_class, data->start_class, regnode_ssc);
4698 flags |= SCF_DO_STCLASS_OR;
4699 ANYOF_FLAGS(data->start_class)
4700 |= SSC_MATCHES_EMPTY_STRING;
4702 } else { /* Non-zero len */
4703 if (flags & SCF_DO_STCLASS_OR) {
4704 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4705 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4707 else if (flags & SCF_DO_STCLASS_AND)
4708 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4709 flags &= ~SCF_DO_STCLASS;
4711 if (!scan) /* It was not CURLYX, but CURLY. */
4713 if (!(flags & SCF_TRIE_DOING_RESTUDY)
4714 /* ? quantifier ok, except for (?{ ... }) */
4715 && (next_is_eval || !(mincount == 0 && maxcount == 1))
4716 && (minnext == 0) && (deltanext == 0)
4717 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
4718 && maxcount <= REG_INFTY/3) /* Complement check for big
4721 /* Fatal warnings may leak the regexp without this: */
4722 SAVEFREESV(RExC_rx_sv);
4723 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
4724 "Quantifier unexpected on zero-length expression "
4725 "in regex m/%"UTF8f"/",
4726 UTF8fARG(UTF, RExC_end - RExC_precomp,
4728 (void)ReREFCNT_inc(RExC_rx_sv);
4731 min += minnext * mincount;
4732 is_inf_internal |= deltanext == SSize_t_MAX
4733 || (maxcount == REG_INFTY && minnext + deltanext > 0);
4734 is_inf |= is_inf_internal;
4736 delta = SSize_t_MAX;
4738 delta += (minnext + deltanext) * maxcount
4739 - minnext * mincount;
4741 /* Try powerful optimization CURLYX => CURLYN. */
4742 if ( OP(oscan) == CURLYX && data
4743 && data->flags & SF_IN_PAR
4744 && !(data->flags & SF_HAS_EVAL)
4745 && !deltanext && minnext == 1 ) {
4746 /* Try to optimize to CURLYN. */
4747 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
4748 regnode * const nxt1 = nxt;
4755 if (!REGNODE_SIMPLE(OP(nxt))
4756 && !(PL_regkind[OP(nxt)] == EXACT
4757 && STR_LEN(nxt) == 1))
4763 if (OP(nxt) != CLOSE)
4765 if (RExC_open_parens) {
4766 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4767 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
4769 /* Now we know that nxt2 is the only contents: */
4770 oscan->flags = (U8)ARG(nxt);
4772 OP(nxt1) = NOTHING; /* was OPEN. */
4775 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4776 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
4777 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
4778 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4779 OP(nxt + 1) = OPTIMIZED; /* was count. */
4780 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
4785 /* Try optimization CURLYX => CURLYM. */
4786 if ( OP(oscan) == CURLYX && data
4787 && !(data->flags & SF_HAS_PAR)
4788 && !(data->flags & SF_HAS_EVAL)
4789 && !deltanext /* atom is fixed width */
4790 && minnext != 0 /* CURLYM can't handle zero width */
4792 /* Nor characters whose fold at run-time may be
4793 * multi-character */
4794 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
4796 /* XXXX How to optimize if data == 0? */
4797 /* Optimize to a simpler form. */
4798 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
4802 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
4803 && (OP(nxt2) != WHILEM))
4805 OP(nxt2) = SUCCEED; /* Whas WHILEM */
4806 /* Need to optimize away parenths. */
4807 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
4808 /* Set the parenth number. */
4809 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
4811 oscan->flags = (U8)ARG(nxt);
4812 if (RExC_open_parens) {
4813 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4814 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
4816 OP(nxt1) = OPTIMIZED; /* was OPEN. */
4817 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4820 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4821 OP(nxt + 1) = OPTIMIZED; /* was count. */
4822 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
4823 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
4826 while ( nxt1 && (OP(nxt1) != WHILEM)) {
4827 regnode *nnxt = regnext(nxt1);
4829 if (reg_off_by_arg[OP(nxt1)])
4830 ARG_SET(nxt1, nxt2 - nxt1);
4831 else if (nxt2 - nxt1 < U16_MAX)
4832 NEXT_OFF(nxt1) = nxt2 - nxt1;
4834 OP(nxt) = NOTHING; /* Cannot beautify */
4839 /* Optimize again: */
4840 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4841 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
4846 else if ((OP(oscan) == CURLYX)
4847 && (flags & SCF_WHILEM_VISITED_POS)
4848 /* See the comment on a similar expression above.
4849 However, this time it's not a subexpression
4850 we care about, but the expression itself. */
4851 && (maxcount == REG_INFTY)
4852 && data && ++data->whilem_c < 16) {
4853 /* This stays as CURLYX, we can put the count/of pair. */
4854 /* Find WHILEM (as in regexec.c) */
4855 regnode *nxt = oscan + NEXT_OFF(oscan);
4857 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4859 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4860 | (RExC_whilem_seen << 4)); /* On WHILEM */
4862 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4864 if (flags & SCF_DO_SUBSTR) {
4865 SV *last_str = NULL;
4866 STRLEN last_chrs = 0;
4867 int counted = mincount != 0;
4869 if (data->last_end > 0 && mincount != 0) { /* Ends with a
4871 SSize_t b = pos_before >= data->last_start_min
4872 ? pos_before : data->last_start_min;
4874 const char * const s = SvPV_const(data->last_found, l);
4875 SSize_t old = b - data->last_start_min;
4878 old = utf8_hop((U8*)s, old) - (U8*)s;
4880 /* Get the added string: */
4881 last_str = newSVpvn_utf8(s + old, l, UTF);
4882 last_chrs = UTF ? utf8_length((U8*)(s + old),
4883 (U8*)(s + old + l)) : l;
4884 if (deltanext == 0 && pos_before == b) {
4885 /* What was added is a constant string */
4888 SvGROW(last_str, (mincount * l) + 1);
4889 repeatcpy(SvPVX(last_str) + l,
4890 SvPVX_const(last_str), l,
4892 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4893 /* Add additional parts. */
4894 SvCUR_set(data->last_found,
4895 SvCUR(data->last_found) - l);
4896 sv_catsv(data->last_found, last_str);
4898 SV * sv = data->last_found;
4900 SvUTF8(sv) && SvMAGICAL(sv) ?
4901 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4902 if (mg && mg->mg_len >= 0)
4903 mg->mg_len += last_chrs * (mincount-1);
4905 last_chrs *= mincount;
4906 data->last_end += l * (mincount - 1);
4909 /* start offset must point into the last copy */
4910 data->last_start_min += minnext * (mincount - 1);
4911 data->last_start_max =
4914 : data->last_start_max +
4915 (maxcount - 1) * (minnext + data->pos_delta);
4918 /* It is counted once already... */
4919 data->pos_min += minnext * (mincount - counted);
4921 PerlIO_printf(Perl_debug_log, "counted=%"UVuf" deltanext=%"UVuf
4922 " SSize_t_MAX=%"UVuf" minnext=%"UVuf
4923 " maxcount=%"UVuf" mincount=%"UVuf"\n",
4924 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
4926 if (deltanext != SSize_t_MAX)
4927 PerlIO_printf(Perl_debug_log, "LHS=%"UVuf" RHS=%"UVuf"\n",
4928 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
4929 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
4931 if (deltanext == SSize_t_MAX
4932 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
4933 data->pos_delta = SSize_t_MAX;
4935 data->pos_delta += - counted * deltanext +
4936 (minnext + deltanext) * maxcount - minnext * mincount;
4937 if (mincount != maxcount) {
4938 /* Cannot extend fixed substrings found inside
4940 scan_commit(pRExC_state, data, minlenp, is_inf);
4941 if (mincount && last_str) {
4942 SV * const sv = data->last_found;
4943 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4944 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4948 sv_setsv(sv, last_str);
4949 data->last_end = data->pos_min;
4950 data->last_start_min = data->pos_min - last_chrs;
4951 data->last_start_max = is_inf
4953 : data->pos_min + data->pos_delta - last_chrs;
4955 data->longest = &(data->longest_float);
4957 SvREFCNT_dec(last_str);
4959 if (data && (fl & SF_HAS_EVAL))
4960 data->flags |= SF_HAS_EVAL;
4961 optimize_curly_tail:
4962 if (OP(oscan) != CURLYX) {
4963 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4965 NEXT_OFF(oscan) += NEXT_OFF(next);
4971 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
4976 if (flags & SCF_DO_SUBSTR) {
4977 /* Cannot expect anything... */
4978 scan_commit(pRExC_state, data, minlenp, is_inf);
4979 data->longest = &(data->longest_float);
4981 is_inf = is_inf_internal = 1;
4982 if (flags & SCF_DO_STCLASS_OR) {
4983 if (OP(scan) == CLUMP) {
4984 /* Actually is any start char, but very few code points
4985 * aren't start characters */
4986 ssc_match_all_cp(data->start_class);
4989 ssc_anything(data->start_class);
4992 flags &= ~SCF_DO_STCLASS;
4996 else if (OP(scan) == LNBREAK) {
4997 if (flags & SCF_DO_STCLASS) {
4998 if (flags & SCF_DO_STCLASS_AND) {
4999 ssc_intersection(data->start_class,
5000 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5001 ssc_clear_locale(data->start_class);
5002 ANYOF_FLAGS(data->start_class)
5003 &= ~SSC_MATCHES_EMPTY_STRING;
5005 else if (flags & SCF_DO_STCLASS_OR) {
5006 ssc_union(data->start_class,
5007 PL_XPosix_ptrs[_CC_VERTSPACE],
5009 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5011 /* See commit msg for
5012 * 749e076fceedeb708a624933726e7989f2302f6a */
5013 ANYOF_FLAGS(data->start_class)
5014 &= ~SSC_MATCHES_EMPTY_STRING;
5016 flags &= ~SCF_DO_STCLASS;
5019 if (delta != SSize_t_MAX)
5020 delta++; /* Because of the 2 char string cr-lf */
5021 if (flags & SCF_DO_SUBSTR) {
5022 /* Cannot expect anything... */
5023 scan_commit(pRExC_state, data, minlenp, is_inf);
5025 data->pos_delta += 1;
5026 data->longest = &(data->longest_float);
5029 else if (REGNODE_SIMPLE(OP(scan))) {
5031 if (flags & SCF_DO_SUBSTR) {
5032 scan_commit(pRExC_state, data, minlenp, is_inf);
5036 if (flags & SCF_DO_STCLASS) {
5038 SV* my_invlist = NULL;
5041 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5042 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5044 /* Some of the logic below assumes that switching
5045 locale on will only add false positives. */
5050 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5055 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5056 ssc_match_all_cp(data->start_class);
5061 SV* REG_ANY_invlist = _new_invlist(2);
5062 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5064 if (flags & SCF_DO_STCLASS_OR) {
5065 ssc_union(data->start_class,
5067 TRUE /* TRUE => invert, hence all but \n
5071 else if (flags & SCF_DO_STCLASS_AND) {
5072 ssc_intersection(data->start_class,
5074 TRUE /* TRUE => invert */
5076 ssc_clear_locale(data->start_class);
5078 SvREFCNT_dec_NN(REG_ANY_invlist);
5084 if (flags & SCF_DO_STCLASS_AND)
5085 ssc_and(pRExC_state, data->start_class,
5086 (regnode_charclass *) scan);
5088 ssc_or(pRExC_state, data->start_class,
5089 (regnode_charclass *) scan);
5097 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5098 if (flags & SCF_DO_STCLASS_AND) {
5099 bool was_there = cBOOL(
5100 ANYOF_POSIXL_TEST(data->start_class,
5102 ANYOF_POSIXL_ZERO(data->start_class);
5103 if (was_there) { /* Do an AND */
5104 ANYOF_POSIXL_SET(data->start_class, namedclass);
5106 /* No individual code points can now match */
5107 data->start_class->invlist
5108 = sv_2mortal(_new_invlist(0));
5111 int complement = namedclass + ((invert) ? -1 : 1);
5113 assert(flags & SCF_DO_STCLASS_OR);
5115 /* If the complement of this class was already there,
5116 * the result is that they match all code points,
5117 * (\d + \D == everything). Remove the classes from
5118 * future consideration. Locale is not relevant in
5120 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5121 ssc_match_all_cp(data->start_class);
5122 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5123 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5125 else { /* The usual case; just add this class to the
5127 ANYOF_POSIXL_SET(data->start_class, namedclass);
5132 case NPOSIXA: /* For these, we always know the exact set of
5137 if (FLAGS(scan) == _CC_ASCII) {
5138 my_invlist = invlist_clone(PL_XPosix_ptrs[_CC_ASCII]);
5141 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
5142 PL_XPosix_ptrs[_CC_ASCII],
5153 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
5155 /* NPOSIXD matches all upper Latin1 code points unless the
5156 * target string being matched is UTF-8, which is
5157 * unknowable until match time. Since we are going to
5158 * invert, we want to get rid of all of them so that the
5159 * inversion will match all */
5160 if (OP(scan) == NPOSIXD) {
5161 _invlist_subtract(my_invlist, PL_UpperLatin1,
5167 if (flags & SCF_DO_STCLASS_AND) {
5168 ssc_intersection(data->start_class, my_invlist, invert);
5169 ssc_clear_locale(data->start_class);
5172 assert(flags & SCF_DO_STCLASS_OR);
5173 ssc_union(data->start_class, my_invlist, invert);
5175 SvREFCNT_dec(my_invlist);
5177 if (flags & SCF_DO_STCLASS_OR)
5178 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5179 flags &= ~SCF_DO_STCLASS;
5182 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5183 data->flags |= (OP(scan) == MEOL
5186 scan_commit(pRExC_state, data, minlenp, is_inf);
5189 else if ( PL_regkind[OP(scan)] == BRANCHJ
5190 /* Lookbehind, or need to calculate parens/evals/stclass: */
5191 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5192 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5194 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5195 || OP(scan) == UNLESSM )
5197 /* Negative Lookahead/lookbehind
5198 In this case we can't do fixed string optimisation.
5201 SSize_t deltanext, minnext, fake = 0;
5206 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5208 data_fake.whilem_c = data->whilem_c;
5209 data_fake.last_closep = data->last_closep;
5212 data_fake.last_closep = &fake;
5213 data_fake.pos_delta = delta;
5214 if ( flags & SCF_DO_STCLASS && !scan->flags
5215 && OP(scan) == IFMATCH ) { /* Lookahead */
5216 ssc_init(pRExC_state, &intrnl);
5217 data_fake.start_class = &intrnl;
5218 f |= SCF_DO_STCLASS_AND;
5220 if (flags & SCF_WHILEM_VISITED_POS)
5221 f |= SCF_WHILEM_VISITED_POS;
5222 next = regnext(scan);
5223 nscan = NEXTOPER(NEXTOPER(scan));
5224 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5225 last, &data_fake, stopparen,
5226 recursed_depth, NULL, f, depth+1);
5229 FAIL("Variable length lookbehind not implemented");
5231 else if (minnext > (I32)U8_MAX) {
5232 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5235 scan->flags = (U8)minnext;
5238 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5240 if (data_fake.flags & SF_HAS_EVAL)
5241 data->flags |= SF_HAS_EVAL;
5242 data->whilem_c = data_fake.whilem_c;
5244 if (f & SCF_DO_STCLASS_AND) {
5245 if (flags & SCF_DO_STCLASS_OR) {
5246 /* OR before, AND after: ideally we would recurse with
5247 * data_fake to get the AND applied by study of the
5248 * remainder of the pattern, and then derecurse;
5249 * *** HACK *** for now just treat as "no information".
5250 * See [perl #56690].
5252 ssc_init(pRExC_state, data->start_class);
5254 /* AND before and after: combine and continue. These
5255 * assertions are zero-length, so can match an EMPTY
5257 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5258 ANYOF_FLAGS(data->start_class)
5259 |= SSC_MATCHES_EMPTY_STRING;
5263 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5265 /* Positive Lookahead/lookbehind
5266 In this case we can do fixed string optimisation,
5267 but we must be careful about it. Note in the case of
5268 lookbehind the positions will be offset by the minimum
5269 length of the pattern, something we won't know about
5270 until after the recurse.
5272 SSize_t deltanext, fake = 0;
5276 /* We use SAVEFREEPV so that when the full compile
5277 is finished perl will clean up the allocated
5278 minlens when it's all done. This way we don't
5279 have to worry about freeing them when we know
5280 they wont be used, which would be a pain.
5283 Newx( minnextp, 1, SSize_t );
5284 SAVEFREEPV(minnextp);
5287 StructCopy(data, &data_fake, scan_data_t);
5288 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5291 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5292 data_fake.last_found=newSVsv(data->last_found);
5296 data_fake.last_closep = &fake;
5297 data_fake.flags = 0;
5298 data_fake.pos_delta = delta;
5300 data_fake.flags |= SF_IS_INF;
5301 if ( flags & SCF_DO_STCLASS && !scan->flags
5302 && OP(scan) == IFMATCH ) { /* Lookahead */
5303 ssc_init(pRExC_state, &intrnl);
5304 data_fake.start_class = &intrnl;
5305 f |= SCF_DO_STCLASS_AND;
5307 if (flags & SCF_WHILEM_VISITED_POS)
5308 f |= SCF_WHILEM_VISITED_POS;
5309 next = regnext(scan);
5310 nscan = NEXTOPER(NEXTOPER(scan));
5312 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5313 &deltanext, last, &data_fake,
5314 stopparen, recursed_depth, NULL,
5318 FAIL("Variable length lookbehind not implemented");
5320 else if (*minnextp > (I32)U8_MAX) {
5321 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5324 scan->flags = (U8)*minnextp;
5329 if (f & SCF_DO_STCLASS_AND) {
5330 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5331 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
5334 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5336 if (data_fake.flags & SF_HAS_EVAL)
5337 data->flags |= SF_HAS_EVAL;
5338 data->whilem_c = data_fake.whilem_c;
5339 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5340 if (RExC_rx->minlen<*minnextp)
5341 RExC_rx->minlen=*minnextp;
5342 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5343 SvREFCNT_dec_NN(data_fake.last_found);
5345 if ( data_fake.minlen_fixed != minlenp )
5347 data->offset_fixed= data_fake.offset_fixed;
5348 data->minlen_fixed= data_fake.minlen_fixed;
5349 data->lookbehind_fixed+= scan->flags;
5351 if ( data_fake.minlen_float != minlenp )
5353 data->minlen_float= data_fake.minlen_float;
5354 data->offset_float_min=data_fake.offset_float_min;
5355 data->offset_float_max=data_fake.offset_float_max;
5356 data->lookbehind_float+= scan->flags;
5363 else if (OP(scan) == OPEN) {
5364 if (stopparen != (I32)ARG(scan))
5367 else if (OP(scan) == CLOSE) {
5368 if (stopparen == (I32)ARG(scan)) {
5371 if ((I32)ARG(scan) == is_par) {
5372 next = regnext(scan);
5374 if ( next && (OP(next) != WHILEM) && next < last)
5375 is_par = 0; /* Disable optimization */
5378 *(data->last_closep) = ARG(scan);
5380 else if (OP(scan) == EVAL) {
5382 data->flags |= SF_HAS_EVAL;
5384 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5385 if (flags & SCF_DO_SUBSTR) {
5386 scan_commit(pRExC_state, data, minlenp, is_inf);
5387 flags &= ~SCF_DO_SUBSTR;
5389 if (data && OP(scan)==ACCEPT) {
5390 data->flags |= SCF_SEEN_ACCEPT;
5395 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5397 if (flags & SCF_DO_SUBSTR) {
5398 scan_commit(pRExC_state, data, minlenp, is_inf);
5399 data->longest = &(data->longest_float);
5401 is_inf = is_inf_internal = 1;
5402 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5403 ssc_anything(data->start_class);
5404 flags &= ~SCF_DO_STCLASS;
5406 else if (OP(scan) == GPOS) {
5407 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5408 !(delta || is_inf || (data && data->pos_delta)))
5410 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5411 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5412 if (RExC_rx->gofs < (STRLEN)min)
5413 RExC_rx->gofs = min;
5415 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5419 #ifdef TRIE_STUDY_OPT
5420 #ifdef FULL_TRIE_STUDY
5421 else if (PL_regkind[OP(scan)] == TRIE) {
5422 /* NOTE - There is similar code to this block above for handling
5423 BRANCH nodes on the initial study. If you change stuff here
5425 regnode *trie_node= scan;
5426 regnode *tail= regnext(scan);
5427 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5428 SSize_t max1 = 0, min1 = SSize_t_MAX;
5431 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5432 /* Cannot merge strings after this. */
5433 scan_commit(pRExC_state, data, minlenp, is_inf);
5435 if (flags & SCF_DO_STCLASS)
5436 ssc_init_zero(pRExC_state, &accum);
5442 const regnode *nextbranch= NULL;
5445 for ( word=1 ; word <= trie->wordcount ; word++)
5447 SSize_t deltanext=0, minnext=0, f = 0, fake;
5448 regnode_ssc this_class;
5450 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5452 data_fake.whilem_c = data->whilem_c;
5453 data_fake.last_closep = data->last_closep;
5456 data_fake.last_closep = &fake;
5457 data_fake.pos_delta = delta;
5458 if (flags & SCF_DO_STCLASS) {
5459 ssc_init(pRExC_state, &this_class);
5460 data_fake.start_class = &this_class;
5461 f = SCF_DO_STCLASS_AND;
5463 if (flags & SCF_WHILEM_VISITED_POS)
5464 f |= SCF_WHILEM_VISITED_POS;
5466 if (trie->jump[word]) {
5468 nextbranch = trie_node + trie->jump[0];
5469 scan= trie_node + trie->jump[word];
5470 /* We go from the jump point to the branch that follows
5471 it. Note this means we need the vestigal unused
5472 branches even though they arent otherwise used. */
5473 minnext = study_chunk(pRExC_state, &scan, minlenp,
5474 &deltanext, (regnode *)nextbranch, &data_fake,
5475 stopparen, recursed_depth, NULL, f,depth+1);
5477 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5478 nextbranch= regnext((regnode*)nextbranch);
5480 if (min1 > (SSize_t)(minnext + trie->minlen))
5481 min1 = minnext + trie->minlen;
5482 if (deltanext == SSize_t_MAX) {
5483 is_inf = is_inf_internal = 1;
5485 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5486 max1 = minnext + deltanext + trie->maxlen;
5488 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5490 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5491 if ( stopmin > min + min1)
5492 stopmin = min + min1;
5493 flags &= ~SCF_DO_SUBSTR;
5495 data->flags |= SCF_SEEN_ACCEPT;
5498 if (data_fake.flags & SF_HAS_EVAL)
5499 data->flags |= SF_HAS_EVAL;
5500 data->whilem_c = data_fake.whilem_c;
5502 if (flags & SCF_DO_STCLASS)
5503 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5506 if (flags & SCF_DO_SUBSTR) {
5507 data->pos_min += min1;
5508 data->pos_delta += max1 - min1;
5509 if (max1 != min1 || is_inf)
5510 data->longest = &(data->longest_float);
5513 if (delta != SSize_t_MAX)
5514 delta += max1 - min1;
5515 if (flags & SCF_DO_STCLASS_OR) {
5516 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5518 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5519 flags &= ~SCF_DO_STCLASS;
5522 else if (flags & SCF_DO_STCLASS_AND) {
5524 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5525 flags &= ~SCF_DO_STCLASS;
5528 /* Switch to OR mode: cache the old value of
5529 * data->start_class */
5531 StructCopy(data->start_class, and_withp, regnode_ssc);
5532 flags &= ~SCF_DO_STCLASS_AND;
5533 StructCopy(&accum, data->start_class, regnode_ssc);
5534 flags |= SCF_DO_STCLASS_OR;
5541 else if (PL_regkind[OP(scan)] == TRIE) {
5542 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5545 min += trie->minlen;
5546 delta += (trie->maxlen - trie->minlen);
5547 flags &= ~SCF_DO_STCLASS; /* xxx */
5548 if (flags & SCF_DO_SUBSTR) {
5549 /* Cannot expect anything... */
5550 scan_commit(pRExC_state, data, minlenp, is_inf);
5551 data->pos_min += trie->minlen;
5552 data->pos_delta += (trie->maxlen - trie->minlen);
5553 if (trie->maxlen != trie->minlen)
5554 data->longest = &(data->longest_float);
5556 if (trie->jump) /* no more substrings -- for now /grr*/
5557 flags &= ~SCF_DO_SUBSTR;
5559 #endif /* old or new */
5560 #endif /* TRIE_STUDY_OPT */
5562 /* Else: zero-length, ignore. */
5563 scan = regnext(scan);
5565 /* If we are exiting a recursion we can unset its recursed bit
5566 * and allow ourselves to enter it again - no danger of an
5567 * infinite loop there.
5568 if (stopparen > -1 && recursed) {
5569 DEBUG_STUDYDATA("unset:", data,depth);
5570 PAREN_UNSET( recursed, stopparen);
5576 DEBUG_STUDYDATA("frame-end:",data,depth);
5577 DEBUG_PEEP("fend", scan, depth);
5579 /* restore previous context */
5580 last = frame->last_regnode;
5581 scan = frame->next_regnode;
5582 stopparen = frame->stopparen;
5583 recursed_depth = frame->prev_recursed_depth;
5585 RExC_frame_last = frame->prev_frame;
5586 frame = frame->this_prev_frame;
5587 goto fake_study_recurse;
5592 DEBUG_STUDYDATA("pre-fin:",data,depth);
5595 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5597 if (flags & SCF_DO_SUBSTR && is_inf)
5598 data->pos_delta = SSize_t_MAX - data->pos_min;
5599 if (is_par > (I32)U8_MAX)
5601 if (is_par && pars==1 && data) {
5602 data->flags |= SF_IN_PAR;
5603 data->flags &= ~SF_HAS_PAR;
5605 else if (pars && data) {
5606 data->flags |= SF_HAS_PAR;
5607 data->flags &= ~SF_IN_PAR;
5609 if (flags & SCF_DO_STCLASS_OR)
5610 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5611 if (flags & SCF_TRIE_RESTUDY)
5612 data->flags |= SCF_TRIE_RESTUDY;
5614 DEBUG_STUDYDATA("post-fin:",data,depth);
5617 SSize_t final_minlen= min < stopmin ? min : stopmin;
5619 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
5620 if (final_minlen > SSize_t_MAX - delta)
5621 RExC_maxlen = SSize_t_MAX;
5622 else if (RExC_maxlen < final_minlen + delta)
5623 RExC_maxlen = final_minlen + delta;
5625 return final_minlen;
5631 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
5633 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
5635 PERL_ARGS_ASSERT_ADD_DATA;
5637 Renewc(RExC_rxi->data,
5638 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
5639 char, struct reg_data);
5641 Renew(RExC_rxi->data->what, count + n, U8);
5643 Newx(RExC_rxi->data->what, n, U8);
5644 RExC_rxi->data->count = count + n;
5645 Copy(s, RExC_rxi->data->what + count, n, U8);
5649 /*XXX: todo make this not included in a non debugging perl, but appears to be
5650 * used anyway there, in 'use re' */
5651 #ifndef PERL_IN_XSUB_RE
5653 Perl_reginitcolors(pTHX)
5655 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
5657 char *t = savepv(s);
5661 t = strchr(t, '\t');
5667 PL_colors[i] = t = (char *)"";
5672 PL_colors[i++] = (char *)"";
5679 #ifdef TRIE_STUDY_OPT
5680 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
5683 (data.flags & SCF_TRIE_RESTUDY) \
5691 #define CHECK_RESTUDY_GOTO_butfirst
5695 * pregcomp - compile a regular expression into internal code
5697 * Decides which engine's compiler to call based on the hint currently in
5701 #ifndef PERL_IN_XSUB_RE
5703 /* return the currently in-scope regex engine (or the default if none) */
5705 regexp_engine const *
5706 Perl_current_re_engine(pTHX)
5708 if (IN_PERL_COMPILETIME) {
5709 HV * const table = GvHV(PL_hintgv);
5712 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
5713 return &PL_core_reg_engine;
5714 ptr = hv_fetchs(table, "regcomp", FALSE);
5715 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
5716 return &PL_core_reg_engine;
5717 return INT2PTR(regexp_engine*,SvIV(*ptr));
5721 if (!PL_curcop->cop_hints_hash)
5722 return &PL_core_reg_engine;
5723 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
5724 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
5725 return &PL_core_reg_engine;
5726 return INT2PTR(regexp_engine*,SvIV(ptr));
5732 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
5734 regexp_engine const *eng = current_re_engine();
5735 GET_RE_DEBUG_FLAGS_DECL;
5737 PERL_ARGS_ASSERT_PREGCOMP;
5739 /* Dispatch a request to compile a regexp to correct regexp engine. */
5741 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
5744 return CALLREGCOMP_ENG(eng, pattern, flags);
5748 /* public(ish) entry point for the perl core's own regex compiling code.
5749 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
5750 * pattern rather than a list of OPs, and uses the internal engine rather
5751 * than the current one */
5754 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
5756 SV *pat = pattern; /* defeat constness! */
5757 PERL_ARGS_ASSERT_RE_COMPILE;
5758 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
5759 #ifdef PERL_IN_XSUB_RE
5762 &PL_core_reg_engine,
5764 NULL, NULL, rx_flags, 0);
5768 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
5769 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
5770 * point to the realloced string and length.
5772 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
5776 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
5777 char **pat_p, STRLEN *plen_p, int num_code_blocks)
5779 U8 *const src = (U8*)*pat_p;
5784 GET_RE_DEBUG_FLAGS_DECL;
5786 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5787 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
5789 Newx(dst, *plen_p * 2 + 1, U8);
5792 while (s < *plen_p) {
5793 append_utf8_from_native_byte(src[s], &d);
5794 if (n < num_code_blocks) {
5795 if (!do_end && pRExC_state->code_blocks[n].start == s) {
5796 pRExC_state->code_blocks[n].start = d - dst - 1;
5797 assert(*(d - 1) == '(');
5800 else if (do_end && pRExC_state->code_blocks[n].end == s) {
5801 pRExC_state->code_blocks[n].end = d - dst - 1;
5802 assert(*(d - 1) == ')');
5811 *pat_p = (char*) dst;
5813 RExC_orig_utf8 = RExC_utf8 = 1;
5818 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
5819 * while recording any code block indices, and handling overloading,
5820 * nested qr// objects etc. If pat is null, it will allocate a new
5821 * string, or just return the first arg, if there's only one.
5823 * Returns the malloced/updated pat.
5824 * patternp and pat_count is the array of SVs to be concatted;
5825 * oplist is the optional list of ops that generated the SVs;
5826 * recompile_p is a pointer to a boolean that will be set if
5827 * the regex will need to be recompiled.
5828 * delim, if non-null is an SV that will be inserted between each element
5832 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
5833 SV *pat, SV ** const patternp, int pat_count,
5834 OP *oplist, bool *recompile_p, SV *delim)
5838 bool use_delim = FALSE;
5839 bool alloced = FALSE;
5841 /* if we know we have at least two args, create an empty string,
5842 * then concatenate args to that. For no args, return an empty string */
5843 if (!pat && pat_count != 1) {
5849 for (svp = patternp; svp < patternp + pat_count; svp++) {
5852 STRLEN orig_patlen = 0;
5854 SV *msv = use_delim ? delim : *svp;
5855 if (!msv) msv = &PL_sv_undef;
5857 /* if we've got a delimiter, we go round the loop twice for each
5858 * svp slot (except the last), using the delimiter the second
5867 if (SvTYPE(msv) == SVt_PVAV) {
5868 /* we've encountered an interpolated array within
5869 * the pattern, e.g. /...@a..../. Expand the list of elements,
5870 * then recursively append elements.
5871 * The code in this block is based on S_pushav() */
5873 AV *const av = (AV*)msv;
5874 const SSize_t maxarg = AvFILL(av) + 1;
5878 assert(oplist->op_type == OP_PADAV
5879 || oplist->op_type == OP_RV2AV);
5880 oplist = OpSIBLING(oplist);
5883 if (SvRMAGICAL(av)) {
5886 Newx(array, maxarg, SV*);
5888 for (i=0; i < maxarg; i++) {
5889 SV ** const svp = av_fetch(av, i, FALSE);
5890 array[i] = svp ? *svp : &PL_sv_undef;
5894 array = AvARRAY(av);
5896 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5897 array, maxarg, NULL, recompile_p,
5899 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5905 /* we make the assumption here that each op in the list of
5906 * op_siblings maps to one SV pushed onto the stack,
5907 * except for code blocks, with have both an OP_NULL and
5909 * This allows us to match up the list of SVs against the
5910 * list of OPs to find the next code block.
5912 * Note that PUSHMARK PADSV PADSV ..
5914 * PADRANGE PADSV PADSV ..
5915 * so the alignment still works. */
5918 if (oplist->op_type == OP_NULL
5919 && (oplist->op_flags & OPf_SPECIAL))
5921 assert(n < pRExC_state->num_code_blocks);
5922 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5923 pRExC_state->code_blocks[n].block = oplist;
5924 pRExC_state->code_blocks[n].src_regex = NULL;
5927 oplist = OpSIBLING(oplist); /* skip CONST */
5930 oplist = OpSIBLING(oplist);;
5933 /* apply magic and QR overloading to arg */
5936 if (SvROK(msv) && SvAMAGIC(msv)) {
5937 SV *sv = AMG_CALLunary(msv, regexp_amg);
5941 if (SvTYPE(sv) != SVt_REGEXP)
5942 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5947 /* try concatenation overload ... */
5948 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5949 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5952 /* overloading involved: all bets are off over literal
5953 * code. Pretend we haven't seen it */
5954 pRExC_state->num_code_blocks -= n;
5958 /* ... or failing that, try "" overload */
5959 while (SvAMAGIC(msv)
5960 && (sv = AMG_CALLunary(msv, string_amg))
5964 && SvRV(msv) == SvRV(sv))
5969 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5973 /* this is a partially unrolled
5974 * sv_catsv_nomg(pat, msv);
5975 * that allows us to adjust code block indices if
5978 char *dst = SvPV_force_nomg(pat, dlen);
5980 if (SvUTF8(msv) && !SvUTF8(pat)) {
5981 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
5982 sv_setpvn(pat, dst, dlen);
5985 sv_catsv_nomg(pat, msv);
5992 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5995 /* extract any code blocks within any embedded qr//'s */
5996 if (rx && SvTYPE(rx) == SVt_REGEXP
5997 && RX_ENGINE((REGEXP*)rx)->op_comp)
6000 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6001 if (ri->num_code_blocks) {
6003 /* the presence of an embedded qr// with code means
6004 * we should always recompile: the text of the
6005 * qr// may not have changed, but it may be a
6006 * different closure than last time */
6008 Renew(pRExC_state->code_blocks,
6009 pRExC_state->num_code_blocks + ri->num_code_blocks,
6010 struct reg_code_block);
6011 pRExC_state->num_code_blocks += ri->num_code_blocks;
6013 for (i=0; i < ri->num_code_blocks; i++) {
6014 struct reg_code_block *src, *dst;
6015 STRLEN offset = orig_patlen
6016 + ReANY((REGEXP *)rx)->pre_prefix;
6017 assert(n < pRExC_state->num_code_blocks);
6018 src = &ri->code_blocks[i];
6019 dst = &pRExC_state->code_blocks[n];
6020 dst->start = src->start + offset;
6021 dst->end = src->end + offset;
6022 dst->block = src->block;
6023 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6032 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6041 /* see if there are any run-time code blocks in the pattern.
6042 * False positives are allowed */
6045 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6046 char *pat, STRLEN plen)
6051 PERL_UNUSED_CONTEXT;
6053 for (s = 0; s < plen; s++) {
6054 if (n < pRExC_state->num_code_blocks
6055 && s == pRExC_state->code_blocks[n].start)
6057 s = pRExC_state->code_blocks[n].end;
6061 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6063 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6065 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6072 /* Handle run-time code blocks. We will already have compiled any direct
6073 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6074 * copy of it, but with any literal code blocks blanked out and
6075 * appropriate chars escaped; then feed it into
6077 * eval "qr'modified_pattern'"
6081 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6085 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6087 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6088 * and merge them with any code blocks of the original regexp.
6090 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6091 * instead, just save the qr and return FALSE; this tells our caller that
6092 * the original pattern needs upgrading to utf8.
6096 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6097 char *pat, STRLEN plen)
6101 GET_RE_DEBUG_FLAGS_DECL;
6103 if (pRExC_state->runtime_code_qr) {
6104 /* this is the second time we've been called; this should
6105 * only happen if the main pattern got upgraded to utf8
6106 * during compilation; re-use the qr we compiled first time
6107 * round (which should be utf8 too)
6109 qr = pRExC_state->runtime_code_qr;
6110 pRExC_state->runtime_code_qr = NULL;
6111 assert(RExC_utf8 && SvUTF8(qr));
6117 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
6121 /* determine how many extra chars we need for ' and \ escaping */
6122 for (s = 0; s < plen; s++) {
6123 if (pat[s] == '\'' || pat[s] == '\\')
6127 Newx(newpat, newlen, char);
6129 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6131 for (s = 0; s < plen; s++) {
6132 if (n < pRExC_state->num_code_blocks
6133 && s == pRExC_state->code_blocks[n].start)
6135 /* blank out literal code block */
6136 assert(pat[s] == '(');
6137 while (s <= pRExC_state->code_blocks[n].end) {
6145 if (pat[s] == '\'' || pat[s] == '\\')
6150 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
6154 PerlIO_printf(Perl_debug_log,
6155 "%sre-parsing pattern for runtime code:%s %s\n",
6156 PL_colors[4],PL_colors[5],newpat);
6159 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
6164 PUSHSTACKi(PERLSI_REQUIRE);
6165 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
6166 * parsing qr''; normally only q'' does this. It also alters
6168 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
6169 SvREFCNT_dec_NN(sv);
6174 SV * const errsv = ERRSV;
6175 if (SvTRUE_NN(errsv))
6177 Safefree(pRExC_state->code_blocks);
6178 /* use croak_sv ? */
6179 Perl_croak_nocontext("%"SVf, SVfARG(errsv));
6182 assert(SvROK(qr_ref));
6184 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
6185 /* the leaving below frees the tmp qr_ref.
6186 * Give qr a life of its own */
6194 if (!RExC_utf8 && SvUTF8(qr)) {
6195 /* first time through; the pattern got upgraded; save the
6196 * qr for the next time through */
6197 assert(!pRExC_state->runtime_code_qr);
6198 pRExC_state->runtime_code_qr = qr;
6203 /* extract any code blocks within the returned qr// */
6206 /* merge the main (r1) and run-time (r2) code blocks into one */
6208 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6209 struct reg_code_block *new_block, *dst;
6210 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6213 if (!r2->num_code_blocks) /* we guessed wrong */
6215 SvREFCNT_dec_NN(qr);
6220 r1->num_code_blocks + r2->num_code_blocks,
6221 struct reg_code_block);
6224 while ( i1 < r1->num_code_blocks
6225 || i2 < r2->num_code_blocks)
6227 struct reg_code_block *src;
6230 if (i1 == r1->num_code_blocks) {
6231 src = &r2->code_blocks[i2++];
6234 else if (i2 == r2->num_code_blocks)
6235 src = &r1->code_blocks[i1++];
6236 else if ( r1->code_blocks[i1].start
6237 < r2->code_blocks[i2].start)
6239 src = &r1->code_blocks[i1++];
6240 assert(src->end < r2->code_blocks[i2].start);
6243 assert( r1->code_blocks[i1].start
6244 > r2->code_blocks[i2].start);
6245 src = &r2->code_blocks[i2++];
6247 assert(src->end < r1->code_blocks[i1].start);
6250 assert(pat[src->start] == '(');
6251 assert(pat[src->end] == ')');
6252 dst->start = src->start;
6253 dst->end = src->end;
6254 dst->block = src->block;
6255 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6259 r1->num_code_blocks += r2->num_code_blocks;
6260 Safefree(r1->code_blocks);
6261 r1->code_blocks = new_block;
6264 SvREFCNT_dec_NN(qr);
6270 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6271 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6272 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6273 STRLEN longest_length, bool eol, bool meol)
6275 /* This is the common code for setting up the floating and fixed length
6276 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6277 * as to whether succeeded or not */
6282 if (! (longest_length
6283 || (eol /* Can't have SEOL and MULTI */
6284 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6286 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6287 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6292 /* copy the information about the longest from the reg_scan_data
6293 over to the program. */
6294 if (SvUTF8(sv_longest)) {
6295 *rx_utf8 = sv_longest;
6298 *rx_substr = sv_longest;
6301 /* end_shift is how many chars that must be matched that
6302 follow this item. We calculate it ahead of time as once the
6303 lookbehind offset is added in we lose the ability to correctly
6305 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6306 *rx_end_shift = ml - offset
6307 - longest_length + (SvTAIL(sv_longest) != 0)
6310 t = (eol/* Can't have SEOL and MULTI */
6311 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6312 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6318 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6319 * regular expression into internal code.
6320 * The pattern may be passed either as:
6321 * a list of SVs (patternp plus pat_count)
6322 * a list of OPs (expr)
6323 * If both are passed, the SV list is used, but the OP list indicates
6324 * which SVs are actually pre-compiled code blocks
6326 * The SVs in the list have magic and qr overloading applied to them (and
6327 * the list may be modified in-place with replacement SVs in the latter
6330 * If the pattern hasn't changed from old_re, then old_re will be
6333 * eng is the current engine. If that engine has an op_comp method, then
6334 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6335 * do the initial concatenation of arguments and pass on to the external
6338 * If is_bare_re is not null, set it to a boolean indicating whether the
6339 * arg list reduced (after overloading) to a single bare regex which has
6340 * been returned (i.e. /$qr/).
6342 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6344 * pm_flags contains the PMf_* flags, typically based on those from the
6345 * pm_flags field of the related PMOP. Currently we're only interested in
6346 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6348 * We can't allocate space until we know how big the compiled form will be,
6349 * but we can't compile it (and thus know how big it is) until we've got a
6350 * place to put the code. So we cheat: we compile it twice, once with code
6351 * generation turned off and size counting turned on, and once "for real".
6352 * This also means that we don't allocate space until we are sure that the
6353 * thing really will compile successfully, and we never have to move the
6354 * code and thus invalidate pointers into it. (Note that it has to be in
6355 * one piece because free() must be able to free it all.) [NB: not true in perl]
6357 * Beware that the optimization-preparation code in here knows about some
6358 * of the structure of the compiled regexp. [I'll say.]
6362 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6363 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6364 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6368 regexp_internal *ri;
6376 SV *code_blocksv = NULL;
6377 SV** new_patternp = patternp;
6379 /* these are all flags - maybe they should be turned
6380 * into a single int with different bit masks */
6381 I32 sawlookahead = 0;
6386 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6388 bool runtime_code = 0;
6390 RExC_state_t RExC_state;
6391 RExC_state_t * const pRExC_state = &RExC_state;
6392 #ifdef TRIE_STUDY_OPT
6394 RExC_state_t copyRExC_state;
6396 GET_RE_DEBUG_FLAGS_DECL;
6398 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6400 DEBUG_r(if (!PL_colorset) reginitcolors());
6402 #ifndef PERL_IN_XSUB_RE
6403 /* Initialize these here instead of as-needed, as is quick and avoids
6404 * having to test them each time otherwise */
6405 if (! PL_AboveLatin1) {
6406 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6407 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6408 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6409 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6410 PL_HasMultiCharFold =
6411 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6413 /* This is calculated here, because the Perl program that generates the
6414 * static global ones doesn't currently have access to
6415 * NUM_ANYOF_CODE_POINTS */
6416 PL_InBitmap = _new_invlist(2);
6417 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
6418 NUM_ANYOF_CODE_POINTS - 1);
6422 pRExC_state->code_blocks = NULL;
6423 pRExC_state->num_code_blocks = 0;
6426 *is_bare_re = FALSE;
6428 if (expr && (expr->op_type == OP_LIST ||
6429 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6430 /* allocate code_blocks if needed */
6434 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
6435 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6436 ncode++; /* count of DO blocks */
6438 pRExC_state->num_code_blocks = ncode;
6439 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
6444 /* compile-time pattern with just OP_CONSTs and DO blocks */
6449 /* find how many CONSTs there are */
6452 if (expr->op_type == OP_CONST)
6455 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6456 if (o->op_type == OP_CONST)
6460 /* fake up an SV array */
6462 assert(!new_patternp);
6463 Newx(new_patternp, n, SV*);
6464 SAVEFREEPV(new_patternp);
6468 if (expr->op_type == OP_CONST)
6469 new_patternp[n] = cSVOPx_sv(expr);
6471 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6472 if (o->op_type == OP_CONST)
6473 new_patternp[n++] = cSVOPo_sv;
6478 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6479 "Assembling pattern from %d elements%s\n", pat_count,
6480 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6482 /* set expr to the first arg op */
6484 if (pRExC_state->num_code_blocks
6485 && expr->op_type != OP_CONST)
6487 expr = cLISTOPx(expr)->op_first;
6488 assert( expr->op_type == OP_PUSHMARK
6489 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6490 || expr->op_type == OP_PADRANGE);
6491 expr = OpSIBLING(expr);
6494 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6495 expr, &recompile, NULL);
6497 /* handle bare (possibly after overloading) regex: foo =~ $re */
6502 if (SvTYPE(re) == SVt_REGEXP) {
6506 Safefree(pRExC_state->code_blocks);
6507 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6508 "Precompiled pattern%s\n",
6509 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6515 exp = SvPV_nomg(pat, plen);
6517 if (!eng->op_comp) {
6518 if ((SvUTF8(pat) && IN_BYTES)
6519 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6521 /* make a temporary copy; either to convert to bytes,
6522 * or to avoid repeating get-magic / overloaded stringify */
6523 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6524 (IN_BYTES ? 0 : SvUTF8(pat)));
6526 Safefree(pRExC_state->code_blocks);
6527 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6530 /* ignore the utf8ness if the pattern is 0 length */
6531 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6532 RExC_uni_semantics = 0;
6533 RExC_contains_locale = 0;
6534 RExC_contains_i = 0;
6535 pRExC_state->runtime_code_qr = NULL;
6536 RExC_frame_head= NULL;
6537 RExC_frame_last= NULL;
6538 RExC_frame_count= 0;
6541 RExC_mysv1= sv_newmortal();
6542 RExC_mysv2= sv_newmortal();
6545 SV *dsv= sv_newmortal();
6546 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6547 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
6548 PL_colors[4],PL_colors[5],s);
6552 /* we jump here if we upgrade the pattern to utf8 and have to
6555 if ((pm_flags & PMf_USE_RE_EVAL)
6556 /* this second condition covers the non-regex literal case,
6557 * i.e. $foo =~ '(?{})'. */
6558 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
6560 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
6562 /* return old regex if pattern hasn't changed */
6563 /* XXX: note in the below we have to check the flags as well as the
6566 * Things get a touch tricky as we have to compare the utf8 flag
6567 * independently from the compile flags. */
6571 && !!RX_UTF8(old_re) == !!RExC_utf8
6572 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
6573 && RX_PRECOMP(old_re)
6574 && RX_PRELEN(old_re) == plen
6575 && memEQ(RX_PRECOMP(old_re), exp, plen)
6576 && !runtime_code /* with runtime code, always recompile */ )
6578 Safefree(pRExC_state->code_blocks);
6582 rx_flags = orig_rx_flags;
6584 if (rx_flags & PMf_FOLD) {
6585 RExC_contains_i = 1;
6587 if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
6589 /* Set to use unicode semantics if the pattern is in utf8 and has the
6590 * 'depends' charset specified, as it means unicode when utf8 */
6591 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6595 RExC_flags = rx_flags;
6596 RExC_pm_flags = pm_flags;
6599 if (TAINTING_get && TAINT_get)
6600 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
6602 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
6603 /* whoops, we have a non-utf8 pattern, whilst run-time code
6604 * got compiled as utf8. Try again with a utf8 pattern */
6605 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6606 pRExC_state->num_code_blocks);
6607 goto redo_first_pass;
6610 assert(!pRExC_state->runtime_code_qr);
6616 RExC_in_lookbehind = 0;
6617 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
6619 RExC_override_recoding = 0;
6620 RExC_in_multi_char_class = 0;
6622 /* First pass: determine size, legality. */
6625 RExC_end = exp + plen;
6630 RExC_emit = (regnode *) &RExC_emit_dummy;
6631 RExC_whilem_seen = 0;
6632 RExC_open_parens = NULL;
6633 RExC_close_parens = NULL;
6635 RExC_paren_names = NULL;
6637 RExC_paren_name_list = NULL;
6639 RExC_recurse = NULL;
6640 RExC_study_chunk_recursed = NULL;
6641 RExC_study_chunk_recursed_bytes= 0;
6642 RExC_recurse_count = 0;
6643 pRExC_state->code_index = 0;
6646 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
6648 RExC_lastparse=NULL;
6650 /* reg may croak on us, not giving us a chance to free
6651 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
6652 need it to survive as long as the regexp (qr/(?{})/).
6653 We must check that code_blocksv is not already set, because we may
6654 have jumped back to restart the sizing pass. */
6655 if (pRExC_state->code_blocks && !code_blocksv) {
6656 code_blocksv = newSV_type(SVt_PV);
6657 SAVEFREESV(code_blocksv);
6658 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
6659 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
6661 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6662 /* It's possible to write a regexp in ascii that represents Unicode
6663 codepoints outside of the byte range, such as via \x{100}. If we
6664 detect such a sequence we have to convert the entire pattern to utf8
6665 and then recompile, as our sizing calculation will have been based
6666 on 1 byte == 1 character, but we will need to use utf8 to encode
6667 at least some part of the pattern, and therefore must convert the whole
6670 if (flags & RESTART_UTF8) {
6671 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6672 pRExC_state->num_code_blocks);
6673 goto redo_first_pass;
6675 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
6678 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
6681 PerlIO_printf(Perl_debug_log,
6682 "Required size %"IVdf" nodes\n"
6683 "Starting second pass (creation)\n",
6686 RExC_lastparse=NULL;
6689 /* The first pass could have found things that force Unicode semantics */
6690 if ((RExC_utf8 || RExC_uni_semantics)
6691 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
6693 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6696 /* Small enough for pointer-storage convention?
6697 If extralen==0, this means that we will not need long jumps. */
6698 if (RExC_size >= 0x10000L && RExC_extralen)
6699 RExC_size += RExC_extralen;
6702 if (RExC_whilem_seen > 15)
6703 RExC_whilem_seen = 15;
6705 /* Allocate space and zero-initialize. Note, the two step process
6706 of zeroing when in debug mode, thus anything assigned has to
6707 happen after that */
6708 rx = (REGEXP*) newSV_type(SVt_REGEXP);
6710 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6711 char, regexp_internal);
6712 if ( r == NULL || ri == NULL )
6713 FAIL("Regexp out of space");
6715 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
6716 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6719 /* bulk initialize base fields with 0. */
6720 Zero(ri, sizeof(regexp_internal), char);
6723 /* non-zero initialization begins here */
6726 r->extflags = rx_flags;
6727 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
6729 if (pm_flags & PMf_IS_QR) {
6730 ri->code_blocks = pRExC_state->code_blocks;
6731 ri->num_code_blocks = pRExC_state->num_code_blocks;
6736 for (n = 0; n < pRExC_state->num_code_blocks; n++)
6737 if (pRExC_state->code_blocks[n].src_regex)
6738 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
6739 SAVEFREEPV(pRExC_state->code_blocks);
6743 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
6744 bool has_charset = (get_regex_charset(r->extflags)
6745 != REGEX_DEPENDS_CHARSET);
6747 /* The caret is output if there are any defaults: if not all the STD
6748 * flags are set, or if no character set specifier is needed */
6750 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
6752 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
6753 == REG_RUN_ON_COMMENT_SEEN);
6754 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
6755 >> RXf_PMf_STD_PMMOD_SHIFT);
6756 const char *fptr = STD_PAT_MODS; /*"msixn"*/
6758 /* Allocate for the worst case, which is all the std flags are turned
6759 * on. If more precision is desired, we could do a population count of
6760 * the flags set. This could be done with a small lookup table, or by
6761 * shifting, masking and adding, or even, when available, assembly
6762 * language for a machine-language population count.
6763 * We never output a minus, as all those are defaults, so are
6764 * covered by the caret */
6765 const STRLEN wraplen = plen + has_p + has_runon
6766 + has_default /* If needs a caret */
6768 /* If needs a character set specifier */
6769 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
6770 + (sizeof(STD_PAT_MODS) - 1)
6771 + (sizeof("(?:)") - 1);
6773 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
6774 r->xpv_len_u.xpvlenu_pv = p;
6776 SvFLAGS(rx) |= SVf_UTF8;
6779 /* If a default, cover it using the caret */
6781 *p++= DEFAULT_PAT_MOD;
6785 const char* const name = get_regex_charset_name(r->extflags, &len);
6786 Copy(name, p, len, char);
6790 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
6793 while((ch = *fptr++)) {
6801 Copy(RExC_precomp, p, plen, char);
6802 assert ((RX_WRAPPED(rx) - p) < 16);
6803 r->pre_prefix = p - RX_WRAPPED(rx);
6809 SvCUR_set(rx, p - RX_WRAPPED(rx));
6813 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
6815 /* setup various meta data about recursion, this all requires
6816 * RExC_npar to be correctly set, and a bit later on we clear it */
6817 if (RExC_seen & REG_RECURSE_SEEN) {
6818 Newxz(RExC_open_parens, RExC_npar,regnode *);
6819 SAVEFREEPV(RExC_open_parens);
6820 Newxz(RExC_close_parens,RExC_npar,regnode *);
6821 SAVEFREEPV(RExC_close_parens);
6823 if (RExC_seen & (REG_RECURSE_SEEN | REG_GOSTART_SEEN)) {
6824 /* Note, RExC_npar is 1 + the number of parens in a pattern.
6825 * So its 1 if there are no parens. */
6826 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
6827 ((RExC_npar & 0x07) != 0);
6828 Newx(RExC_study_chunk_recursed,
6829 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6830 SAVEFREEPV(RExC_study_chunk_recursed);
6833 /* Useful during FAIL. */
6834 #ifdef RE_TRACK_PATTERN_OFFSETS
6835 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
6836 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
6837 "%s %"UVuf" bytes for offset annotations.\n",
6838 ri->u.offsets ? "Got" : "Couldn't get",
6839 (UV)((2*RExC_size+1) * sizeof(U32))));
6841 SetProgLen(ri,RExC_size);
6846 /* Second pass: emit code. */
6847 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
6848 RExC_pm_flags = pm_flags;
6850 RExC_end = exp + plen;
6853 RExC_emit_start = ri->program;
6854 RExC_emit = ri->program;
6855 RExC_emit_bound = ri->program + RExC_size + 1;
6856 pRExC_state->code_index = 0;
6858 *((char*) RExC_emit++) = (char) REG_MAGIC;
6859 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6861 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#"UVxf"", (UV) flags);
6863 /* XXXX To minimize changes to RE engine we always allocate
6864 3-units-long substrs field. */
6865 Newx(r->substrs, 1, struct reg_substr_data);
6866 if (RExC_recurse_count) {
6867 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6868 SAVEFREEPV(RExC_recurse);
6872 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
6874 RExC_study_chunk_recursed_count= 0;
6876 Zero(r->substrs, 1, struct reg_substr_data);
6877 if (RExC_study_chunk_recursed) {
6878 Zero(RExC_study_chunk_recursed,
6879 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6883 #ifdef TRIE_STUDY_OPT
6885 StructCopy(&zero_scan_data, &data, scan_data_t);
6886 copyRExC_state = RExC_state;
6889 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6891 RExC_state = copyRExC_state;
6892 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
6893 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
6895 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
6896 StructCopy(&zero_scan_data, &data, scan_data_t);
6899 StructCopy(&zero_scan_data, &data, scan_data_t);
6902 /* Dig out information for optimizations. */
6903 r->extflags = RExC_flags; /* was pm_op */
6904 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6907 SvUTF8_on(rx); /* Unicode in it? */
6908 ri->regstclass = NULL;
6909 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
6910 r->intflags |= PREGf_NAUGHTY;
6911 scan = ri->program + 1; /* First BRANCH. */
6913 /* testing for BRANCH here tells us whether there is "must appear"
6914 data in the pattern. If there is then we can use it for optimisations */
6915 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
6918 STRLEN longest_float_length, longest_fixed_length;
6919 regnode_ssc ch_class; /* pointed to by data */
6921 SSize_t last_close = 0; /* pointed to by data */
6922 regnode *first= scan;
6923 regnode *first_next= regnext(first);
6925 * Skip introductions and multiplicators >= 1
6926 * so that we can extract the 'meat' of the pattern that must
6927 * match in the large if() sequence following.
6928 * NOTE that EXACT is NOT covered here, as it is normally
6929 * picked up by the optimiser separately.
6931 * This is unfortunate as the optimiser isnt handling lookahead
6932 * properly currently.
6935 while ((OP(first) == OPEN && (sawopen = 1)) ||
6936 /* An OR of *one* alternative - should not happen now. */
6937 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
6938 /* for now we can't handle lookbehind IFMATCH*/
6939 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
6940 (OP(first) == PLUS) ||
6941 (OP(first) == MINMOD) ||
6942 /* An {n,m} with n>0 */
6943 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
6944 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6947 * the only op that could be a regnode is PLUS, all the rest
6948 * will be regnode_1 or regnode_2.
6950 * (yves doesn't think this is true)
6952 if (OP(first) == PLUS)
6955 if (OP(first) == MINMOD)
6957 first += regarglen[OP(first)];
6959 first = NEXTOPER(first);
6960 first_next= regnext(first);
6963 /* Starting-point info. */
6965 DEBUG_PEEP("first:",first,0);
6966 /* Ignore EXACT as we deal with it later. */
6967 if (PL_regkind[OP(first)] == EXACT) {
6968 if (OP(first) == EXACT || OP(first) == EXACTL)
6969 NOOP; /* Empty, get anchored substr later. */
6971 ri->regstclass = first;
6974 else if (PL_regkind[OP(first)] == TRIE &&
6975 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6977 /* this can happen only on restudy */
6978 ri->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
6981 else if (REGNODE_SIMPLE(OP(first)))
6982 ri->regstclass = first;
6983 else if (PL_regkind[OP(first)] == BOUND ||
6984 PL_regkind[OP(first)] == NBOUND)
6985 ri->regstclass = first;
6986 else if (PL_regkind[OP(first)] == BOL) {
6987 r->intflags |= (OP(first) == MBOL
6990 first = NEXTOPER(first);
6993 else if (OP(first) == GPOS) {
6994 r->intflags |= PREGf_ANCH_GPOS;
6995 first = NEXTOPER(first);
6998 else if ((!sawopen || !RExC_sawback) &&
7000 (OP(first) == STAR &&
7001 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
7002 !(r->intflags & PREGf_ANCH) && !pRExC_state->num_code_blocks)
7004 /* turn .* into ^.* with an implied $*=1 */
7006 (OP(NEXTOPER(first)) == REG_ANY)
7009 r->intflags |= (type | PREGf_IMPLICIT);
7010 first = NEXTOPER(first);
7013 if (sawplus && !sawminmod && !sawlookahead
7014 && (!sawopen || !RExC_sawback)
7015 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
7016 /* x+ must match at the 1st pos of run of x's */
7017 r->intflags |= PREGf_SKIP;
7019 /* Scan is after the zeroth branch, first is atomic matcher. */
7020 #ifdef TRIE_STUDY_OPT
7023 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
7024 (IV)(first - scan + 1))
7028 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
7029 (IV)(first - scan + 1))
7035 * If there's something expensive in the r.e., find the
7036 * longest literal string that must appear and make it the
7037 * regmust. Resolve ties in favor of later strings, since
7038 * the regstart check works with the beginning of the r.e.
7039 * and avoiding duplication strengthens checking. Not a
7040 * strong reason, but sufficient in the absence of others.
7041 * [Now we resolve ties in favor of the earlier string if
7042 * it happens that c_offset_min has been invalidated, since the
7043 * earlier string may buy us something the later one won't.]
7046 data.longest_fixed = newSVpvs("");
7047 data.longest_float = newSVpvs("");
7048 data.last_found = newSVpvs("");
7049 data.longest = &(data.longest_fixed);
7050 ENTER_with_name("study_chunk");
7051 SAVEFREESV(data.longest_fixed);
7052 SAVEFREESV(data.longest_float);
7053 SAVEFREESV(data.last_found);
7055 if (!ri->regstclass) {
7056 ssc_init(pRExC_state, &ch_class);
7057 data.start_class = &ch_class;
7058 stclass_flag = SCF_DO_STCLASS_AND;
7059 } else /* XXXX Check for BOUND? */
7061 data.last_closep = &last_close;
7064 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
7065 scan + RExC_size, /* Up to end */
7067 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
7068 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
7072 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
7075 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
7076 && data.last_start_min == 0 && data.last_end > 0
7077 && !RExC_seen_zerolen
7078 && !(RExC_seen & REG_VERBARG_SEEN)
7079 && !(RExC_seen & REG_GPOS_SEEN)
7081 r->extflags |= RXf_CHECK_ALL;
7083 scan_commit(pRExC_state, &data,&minlen,0);
7085 longest_float_length = CHR_SVLEN(data.longest_float);
7087 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
7088 && data.offset_fixed == data.offset_float_min
7089 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
7090 && S_setup_longest (aTHX_ pRExC_state,
7094 &(r->float_end_shift),
7095 data.lookbehind_float,
7096 data.offset_float_min,
7098 longest_float_length,
7099 cBOOL(data.flags & SF_FL_BEFORE_EOL),
7100 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
7102 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
7103 r->float_max_offset = data.offset_float_max;
7104 if (data.offset_float_max < SSize_t_MAX) /* Don't offset infinity */
7105 r->float_max_offset -= data.lookbehind_float;
7106 SvREFCNT_inc_simple_void_NN(data.longest_float);
7109 r->float_substr = r->float_utf8 = NULL;
7110 longest_float_length = 0;
7113 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
7115 if (S_setup_longest (aTHX_ pRExC_state,
7117 &(r->anchored_utf8),
7118 &(r->anchored_substr),
7119 &(r->anchored_end_shift),
7120 data.lookbehind_fixed,
7123 longest_fixed_length,
7124 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
7125 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
7127 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
7128 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
7131 r->anchored_substr = r->anchored_utf8 = NULL;
7132 longest_fixed_length = 0;
7134 LEAVE_with_name("study_chunk");
7137 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
7138 ri->regstclass = NULL;
7140 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
7142 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7143 && is_ssc_worth_it(pRExC_state, data.start_class))
7145 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7147 ssc_finalize(pRExC_state, data.start_class);
7149 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7150 StructCopy(data.start_class,
7151 (regnode_ssc*)RExC_rxi->data->data[n],
7153 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7154 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7155 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
7156 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7157 PerlIO_printf(Perl_debug_log,
7158 "synthetic stclass \"%s\".\n",
7159 SvPVX_const(sv));});
7160 data.start_class = NULL;
7163 /* A temporary algorithm prefers floated substr to fixed one to dig
7165 if (longest_fixed_length > longest_float_length) {
7166 r->substrs->check_ix = 0;
7167 r->check_end_shift = r->anchored_end_shift;
7168 r->check_substr = r->anchored_substr;
7169 r->check_utf8 = r->anchored_utf8;
7170 r->check_offset_min = r->check_offset_max = r->anchored_offset;
7171 if (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS))
7172 r->intflags |= PREGf_NOSCAN;
7175 r->substrs->check_ix = 1;
7176 r->check_end_shift = r->float_end_shift;
7177 r->check_substr = r->float_substr;
7178 r->check_utf8 = r->float_utf8;
7179 r->check_offset_min = r->float_min_offset;
7180 r->check_offset_max = r->float_max_offset;
7182 if ((r->check_substr || r->check_utf8) ) {
7183 r->extflags |= RXf_USE_INTUIT;
7184 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
7185 r->extflags |= RXf_INTUIT_TAIL;
7187 r->substrs->data[0].max_offset = r->substrs->data[0].min_offset;
7189 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
7190 if ( (STRLEN)minlen < longest_float_length )
7191 minlen= longest_float_length;
7192 if ( (STRLEN)minlen < longest_fixed_length )
7193 minlen= longest_fixed_length;
7197 /* Several toplevels. Best we can is to set minlen. */
7199 regnode_ssc ch_class;
7200 SSize_t last_close = 0;
7202 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
7204 scan = ri->program + 1;
7205 ssc_init(pRExC_state, &ch_class);
7206 data.start_class = &ch_class;
7207 data.last_closep = &last_close;
7210 minlen = study_chunk(pRExC_state,
7211 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7212 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7213 ? SCF_TRIE_DOING_RESTUDY
7217 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7219 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
7220 = r->float_substr = r->float_utf8 = NULL;
7222 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7223 && is_ssc_worth_it(pRExC_state, data.start_class))
7225 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7227 ssc_finalize(pRExC_state, data.start_class);
7229 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7230 StructCopy(data.start_class,
7231 (regnode_ssc*)RExC_rxi->data->data[n],
7233 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7234 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7235 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7236 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7237 PerlIO_printf(Perl_debug_log,
7238 "synthetic stclass \"%s\".\n",
7239 SvPVX_const(sv));});
7240 data.start_class = NULL;
7244 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7245 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7246 r->maxlen = REG_INFTY;
7249 r->maxlen = RExC_maxlen;
7252 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7253 the "real" pattern. */
7255 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf" maxlen:%"IVdf"\n",
7256 (IV)minlen, (IV)r->minlen, (IV)RExC_maxlen);
7258 r->minlenret = minlen;
7259 if (r->minlen < minlen)
7262 if (RExC_seen & REG_GPOS_SEEN)
7263 r->intflags |= PREGf_GPOS_SEEN;
7264 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7265 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7267 if (pRExC_state->num_code_blocks)
7268 r->extflags |= RXf_EVAL_SEEN;
7269 if (RExC_seen & REG_CANY_SEEN)
7270 r->intflags |= PREGf_CANY_SEEN;
7271 if (RExC_seen & REG_VERBARG_SEEN)
7273 r->intflags |= PREGf_VERBARG_SEEN;
7274 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7276 if (RExC_seen & REG_CUTGROUP_SEEN)
7277 r->intflags |= PREGf_CUTGROUP_SEEN;
7278 if (pm_flags & PMf_USE_RE_EVAL)
7279 r->intflags |= PREGf_USE_RE_EVAL;
7280 if (RExC_paren_names)
7281 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7283 RXp_PAREN_NAMES(r) = NULL;
7285 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7286 * so it can be used in pp.c */
7287 if (r->intflags & PREGf_ANCH)
7288 r->extflags |= RXf_IS_ANCHORED;
7292 /* this is used to identify "special" patterns that might result
7293 * in Perl NOT calling the regex engine and instead doing the match "itself",
7294 * particularly special cases in split//. By having the regex compiler
7295 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7296 * we avoid weird issues with equivalent patterns resulting in different behavior,
7297 * AND we allow non Perl engines to get the same optimizations by the setting the
7298 * flags appropriately - Yves */
7299 regnode *first = ri->program + 1;
7301 regnode *next = NEXTOPER(first);
7304 if (PL_regkind[fop] == NOTHING && nop == END)
7305 r->extflags |= RXf_NULL;
7306 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
7307 /* when fop is SBOL first->flags will be true only when it was
7308 * produced by parsing /\A/, and not when parsing /^/. This is
7309 * very important for the split code as there we want to
7310 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
7311 * See rt #122761 for more details. -- Yves */
7312 r->extflags |= RXf_START_ONLY;
7313 else if (fop == PLUS
7314 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7315 && OP(regnext(first)) == END)
7316 r->extflags |= RXf_WHITE;
7317 else if ( r->extflags & RXf_SPLIT
7318 && (fop == EXACT || fop == EXACTL)
7319 && STR_LEN(first) == 1
7320 && *(STRING(first)) == ' '
7321 && OP(regnext(first)) == END )
7322 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7326 if (RExC_contains_locale) {
7327 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7331 if (RExC_paren_names) {
7332 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7333 ri->data->data[ri->name_list_idx]
7334 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7337 ri->name_list_idx = 0;
7339 if (RExC_recurse_count) {
7340 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
7341 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
7342 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
7345 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7346 /* assume we don't need to swap parens around before we match */
7348 PerlIO_printf(Perl_debug_log,"study_chunk_recursed_count: %lu\n",
7349 (unsigned long)RExC_study_chunk_recursed_count);
7353 PerlIO_printf(Perl_debug_log,"Final program:\n");
7356 #ifdef RE_TRACK_PATTERN_OFFSETS
7357 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7358 const STRLEN len = ri->u.offsets[0];
7360 GET_RE_DEBUG_FLAGS_DECL;
7361 PerlIO_printf(Perl_debug_log,
7362 "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
7363 for (i = 1; i <= len; i++) {
7364 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7365 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
7366 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7368 PerlIO_printf(Perl_debug_log, "\n");
7373 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7374 * by setting the regexp SV to readonly-only instead. If the
7375 * pattern's been recompiled, the USEDness should remain. */
7376 if (old_re && SvREADONLY(old_re))
7384 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7387 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7389 PERL_UNUSED_ARG(value);
7391 if (flags & RXapif_FETCH) {
7392 return reg_named_buff_fetch(rx, key, flags);
7393 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7394 Perl_croak_no_modify();
7396 } else if (flags & RXapif_EXISTS) {
7397 return reg_named_buff_exists(rx, key, flags)
7400 } else if (flags & RXapif_REGNAMES) {
7401 return reg_named_buff_all(rx, flags);
7402 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7403 return reg_named_buff_scalar(rx, flags);
7405 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7411 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7414 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7415 PERL_UNUSED_ARG(lastkey);
7417 if (flags & RXapif_FIRSTKEY)
7418 return reg_named_buff_firstkey(rx, flags);
7419 else if (flags & RXapif_NEXTKEY)
7420 return reg_named_buff_nextkey(rx, flags);
7422 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7429 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7432 AV *retarray = NULL;
7434 struct regexp *const rx = ReANY(r);
7436 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7438 if (flags & RXapif_ALL)
7441 if (rx && RXp_PAREN_NAMES(rx)) {
7442 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7445 SV* sv_dat=HeVAL(he_str);
7446 I32 *nums=(I32*)SvPVX(sv_dat);
7447 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7448 if ((I32)(rx->nparens) >= nums[i]
7449 && rx->offs[nums[i]].start != -1
7450 && rx->offs[nums[i]].end != -1)
7453 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7458 ret = newSVsv(&PL_sv_undef);
7461 av_push(retarray, ret);
7464 return newRV_noinc(MUTABLE_SV(retarray));
7471 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7474 struct regexp *const rx = ReANY(r);
7476 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7478 if (rx && RXp_PAREN_NAMES(rx)) {
7479 if (flags & RXapif_ALL) {
7480 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7482 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7484 SvREFCNT_dec_NN(sv);
7496 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7498 struct regexp *const rx = ReANY(r);
7500 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7502 if ( rx && RXp_PAREN_NAMES(rx) ) {
7503 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7505 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7512 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7514 struct regexp *const rx = ReANY(r);
7515 GET_RE_DEBUG_FLAGS_DECL;
7517 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7519 if (rx && RXp_PAREN_NAMES(rx)) {
7520 HV *hv = RXp_PAREN_NAMES(rx);
7522 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7525 SV* sv_dat = HeVAL(temphe);
7526 I32 *nums = (I32*)SvPVX(sv_dat);
7527 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7528 if ((I32)(rx->lastparen) >= nums[i] &&
7529 rx->offs[nums[i]].start != -1 &&
7530 rx->offs[nums[i]].end != -1)
7536 if (parno || flags & RXapif_ALL) {
7537 return newSVhek(HeKEY_hek(temphe));
7545 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
7550 struct regexp *const rx = ReANY(r);
7552 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
7554 if (rx && RXp_PAREN_NAMES(rx)) {
7555 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
7556 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
7557 } else if (flags & RXapif_ONE) {
7558 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
7559 av = MUTABLE_AV(SvRV(ret));
7560 length = av_tindex(av);
7561 SvREFCNT_dec_NN(ret);
7562 return newSViv(length + 1);
7564 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
7569 return &PL_sv_undef;
7573 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
7575 struct regexp *const rx = ReANY(r);
7578 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
7580 if (rx && RXp_PAREN_NAMES(rx)) {
7581 HV *hv= RXp_PAREN_NAMES(rx);
7583 (void)hv_iterinit(hv);
7584 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7587 SV* sv_dat = HeVAL(temphe);
7588 I32 *nums = (I32*)SvPVX(sv_dat);
7589 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7590 if ((I32)(rx->lastparen) >= nums[i] &&
7591 rx->offs[nums[i]].start != -1 &&
7592 rx->offs[nums[i]].end != -1)
7598 if (parno || flags & RXapif_ALL) {
7599 av_push(av, newSVhek(HeKEY_hek(temphe)));
7604 return newRV_noinc(MUTABLE_SV(av));
7608 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
7611 struct regexp *const rx = ReANY(r);
7617 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
7619 if ( n == RX_BUFF_IDX_CARET_PREMATCH
7620 || n == RX_BUFF_IDX_CARET_FULLMATCH
7621 || n == RX_BUFF_IDX_CARET_POSTMATCH
7624 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7626 /* on something like
7629 * the KEEPCOPY is set on the PMOP rather than the regex */
7630 if (PL_curpm && r == PM_GETRE(PL_curpm))
7631 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7640 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
7641 /* no need to distinguish between them any more */
7642 n = RX_BUFF_IDX_FULLMATCH;
7644 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
7645 && rx->offs[0].start != -1)
7647 /* $`, ${^PREMATCH} */
7648 i = rx->offs[0].start;
7652 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
7653 && rx->offs[0].end != -1)
7655 /* $', ${^POSTMATCH} */
7656 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
7657 i = rx->sublen + rx->suboffset - rx->offs[0].end;
7660 if ( 0 <= n && n <= (I32)rx->nparens &&
7661 (s1 = rx->offs[n].start) != -1 &&
7662 (t1 = rx->offs[n].end) != -1)
7664 /* $&, ${^MATCH}, $1 ... */
7666 s = rx->subbeg + s1 - rx->suboffset;
7671 assert(s >= rx->subbeg);
7672 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
7674 #ifdef NO_TAINT_SUPPORT
7675 sv_setpvn(sv, s, i);
7677 const int oldtainted = TAINT_get;
7679 sv_setpvn(sv, s, i);
7680 TAINT_set(oldtainted);
7682 if ( (rx->intflags & PREGf_CANY_SEEN)
7683 ? (RXp_MATCH_UTF8(rx)
7684 && (!i || is_utf8_string((U8*)s, i)))
7685 : (RXp_MATCH_UTF8(rx)) )
7692 if (RXp_MATCH_TAINTED(rx)) {
7693 if (SvTYPE(sv) >= SVt_PVMG) {
7694 MAGIC* const mg = SvMAGIC(sv);
7697 SvMAGIC_set(sv, mg->mg_moremagic);
7699 if ((mgt = SvMAGIC(sv))) {
7700 mg->mg_moremagic = mgt;
7701 SvMAGIC_set(sv, mg);
7712 sv_setsv(sv,&PL_sv_undef);
7718 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
7719 SV const * const value)
7721 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
7723 PERL_UNUSED_ARG(rx);
7724 PERL_UNUSED_ARG(paren);
7725 PERL_UNUSED_ARG(value);
7728 Perl_croak_no_modify();
7732 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
7735 struct regexp *const rx = ReANY(r);
7739 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
7741 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
7742 || paren == RX_BUFF_IDX_CARET_FULLMATCH
7743 || paren == RX_BUFF_IDX_CARET_POSTMATCH
7746 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7748 /* on something like
7751 * the KEEPCOPY is set on the PMOP rather than the regex */
7752 if (PL_curpm && r == PM_GETRE(PL_curpm))
7753 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7759 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
7761 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
7762 case RX_BUFF_IDX_PREMATCH: /* $` */
7763 if (rx->offs[0].start != -1) {
7764 i = rx->offs[0].start;
7773 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
7774 case RX_BUFF_IDX_POSTMATCH: /* $' */
7775 if (rx->offs[0].end != -1) {
7776 i = rx->sublen - rx->offs[0].end;
7778 s1 = rx->offs[0].end;
7785 default: /* $& / ${^MATCH}, $1, $2, ... */
7786 if (paren <= (I32)rx->nparens &&
7787 (s1 = rx->offs[paren].start) != -1 &&
7788 (t1 = rx->offs[paren].end) != -1)
7794 if (ckWARN(WARN_UNINITIALIZED))
7795 report_uninit((const SV *)sv);
7800 if (i > 0 && RXp_MATCH_UTF8(rx)) {
7801 const char * const s = rx->subbeg - rx->suboffset + s1;
7806 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
7813 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
7815 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
7816 PERL_UNUSED_ARG(rx);
7820 return newSVpvs("Regexp");
7823 /* Scans the name of a named buffer from the pattern.
7824 * If flags is REG_RSN_RETURN_NULL returns null.
7825 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
7826 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
7827 * to the parsed name as looked up in the RExC_paren_names hash.
7828 * If there is an error throws a vFAIL().. type exception.
7831 #define REG_RSN_RETURN_NULL 0
7832 #define REG_RSN_RETURN_NAME 1
7833 #define REG_RSN_RETURN_DATA 2
7836 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
7838 char *name_start = RExC_parse;
7840 PERL_ARGS_ASSERT_REG_SCAN_NAME;
7842 assert (RExC_parse <= RExC_end);
7843 if (RExC_parse == RExC_end) NOOP;
7844 else if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
7845 /* skip IDFIRST by using do...while */
7848 RExC_parse += UTF8SKIP(RExC_parse);
7849 } while (isWORDCHAR_utf8((U8*)RExC_parse));
7853 } while (isWORDCHAR(*RExC_parse));
7855 RExC_parse++; /* so the <- from the vFAIL is after the offending
7857 vFAIL("Group name must start with a non-digit word character");
7861 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
7862 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
7863 if ( flags == REG_RSN_RETURN_NAME)
7865 else if (flags==REG_RSN_RETURN_DATA) {
7868 if ( ! sv_name ) /* should not happen*/
7869 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
7870 if (RExC_paren_names)
7871 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
7873 sv_dat = HeVAL(he_str);
7875 vFAIL("Reference to nonexistent named group");
7879 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
7880 (unsigned long) flags);
7882 NOT_REACHED; /* NOT REACHED */
7887 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
7889 if (RExC_lastparse!=RExC_parse) { \
7890 PerlIO_printf(Perl_debug_log, "%s", \
7891 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
7892 RExC_end - RExC_parse, 16, \
7894 PERL_PV_ESCAPE_UNI_DETECT | \
7895 PERL_PV_PRETTY_ELLIPSES | \
7896 PERL_PV_PRETTY_LTGT | \
7897 PERL_PV_ESCAPE_RE | \
7898 PERL_PV_PRETTY_EXACTSIZE \
7902 PerlIO_printf(Perl_debug_log,"%16s",""); \
7905 num = RExC_size + 1; \
7907 num=REG_NODE_NUM(RExC_emit); \
7908 if (RExC_lastnum!=num) \
7909 PerlIO_printf(Perl_debug_log,"|%4d",num); \
7911 PerlIO_printf(Perl_debug_log,"|%4s",""); \
7912 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
7913 (int)((depth*2)), "", \
7917 RExC_lastparse=RExC_parse; \
7922 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
7923 DEBUG_PARSE_MSG((funcname)); \
7924 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
7926 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
7927 DEBUG_PARSE_MSG((funcname)); \
7928 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
7931 /* This section of code defines the inversion list object and its methods. The
7932 * interfaces are highly subject to change, so as much as possible is static to
7933 * this file. An inversion list is here implemented as a malloc'd C UV array
7934 * as an SVt_INVLIST scalar.
7936 * An inversion list for Unicode is an array of code points, sorted by ordinal
7937 * number. The zeroth element is the first code point in the list. The 1th
7938 * element is the first element beyond that not in the list. In other words,
7939 * the first range is
7940 * invlist[0]..(invlist[1]-1)
7941 * The other ranges follow. Thus every element whose index is divisible by two
7942 * marks the beginning of a range that is in the list, and every element not
7943 * divisible by two marks the beginning of a range not in the list. A single
7944 * element inversion list that contains the single code point N generally
7945 * consists of two elements
7948 * (The exception is when N is the highest representable value on the
7949 * machine, in which case the list containing just it would be a single
7950 * element, itself. By extension, if the last range in the list extends to
7951 * infinity, then the first element of that range will be in the inversion list
7952 * at a position that is divisible by two, and is the final element in the
7954 * Taking the complement (inverting) an inversion list is quite simple, if the
7955 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7956 * This implementation reserves an element at the beginning of each inversion
7957 * list to always contain 0; there is an additional flag in the header which
7958 * indicates if the list begins at the 0, or is offset to begin at the next
7961 * More about inversion lists can be found in "Unicode Demystified"
7962 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7963 * More will be coming when functionality is added later.
7965 * The inversion list data structure is currently implemented as an SV pointing
7966 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7967 * array of UV whose memory management is automatically handled by the existing
7968 * facilities for SV's.
7970 * Some of the methods should always be private to the implementation, and some
7971 * should eventually be made public */
7973 /* The header definitions are in F<inline_invlist.c> */
7975 PERL_STATIC_INLINE UV*
7976 S__invlist_array_init(SV* const invlist, const bool will_have_0)
7978 /* Returns a pointer to the first element in the inversion list's array.
7979 * This is called upon initialization of an inversion list. Where the
7980 * array begins depends on whether the list has the code point U+0000 in it
7981 * or not. The other parameter tells it whether the code that follows this
7982 * call is about to put a 0 in the inversion list or not. The first
7983 * element is either the element reserved for 0, if TRUE, or the element
7984 * after it, if FALSE */
7986 bool* offset = get_invlist_offset_addr(invlist);
7987 UV* zero_addr = (UV *) SvPVX(invlist);
7989 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
7992 assert(! _invlist_len(invlist));
7996 /* 1^1 = 0; 1^0 = 1 */
7997 *offset = 1 ^ will_have_0;
7998 return zero_addr + *offset;
8001 PERL_STATIC_INLINE UV*
8002 S_invlist_array(SV* const invlist)
8004 /* Returns the pointer to the inversion list's array. Every time the
8005 * length changes, this needs to be called in case malloc or realloc moved
8008 PERL_ARGS_ASSERT_INVLIST_ARRAY;
8010 /* Must not be empty. If these fail, you probably didn't check for <len>
8011 * being non-zero before trying to get the array */
8012 assert(_invlist_len(invlist));
8014 /* The very first element always contains zero, The array begins either
8015 * there, or if the inversion list is offset, at the element after it.
8016 * The offset header field determines which; it contains 0 or 1 to indicate
8017 * how much additionally to add */
8018 assert(0 == *(SvPVX(invlist)));
8019 return ((UV *) SvPVX(invlist) + *get_invlist_offset_addr(invlist));
8022 PERL_STATIC_INLINE void
8023 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8025 /* Sets the current number of elements stored in the inversion list.
8026 * Updates SvCUR correspondingly */
8027 PERL_UNUSED_CONTEXT;
8028 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8030 assert(SvTYPE(invlist) == SVt_INVLIST);
8035 : TO_INTERNAL_SIZE(len + offset));
8036 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
8039 #ifndef PERL_IN_XSUB_RE
8041 PERL_STATIC_INLINE IV*
8042 S_get_invlist_previous_index_addr(SV* invlist)
8044 /* Return the address of the IV that is reserved to hold the cached index
8046 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
8048 assert(SvTYPE(invlist) == SVt_INVLIST);
8050 return &(((XINVLIST*) SvANY(invlist))->prev_index);
8053 PERL_STATIC_INLINE IV
8054 S_invlist_previous_index(SV* const invlist)
8056 /* Returns cached index of previous search */
8058 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
8060 return *get_invlist_previous_index_addr(invlist);
8063 PERL_STATIC_INLINE void
8064 S_invlist_set_previous_index(SV* const invlist, const IV index)
8066 /* Caches <index> for later retrieval */
8068 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
8070 assert(index == 0 || index < (int) _invlist_len(invlist));
8072 *get_invlist_previous_index_addr(invlist) = index;
8075 PERL_STATIC_INLINE void
8076 S_invlist_trim(SV* const invlist)
8078 PERL_ARGS_ASSERT_INVLIST_TRIM;
8080 assert(SvTYPE(invlist) == SVt_INVLIST);
8082 /* Change the length of the inversion list to how many entries it currently
8084 SvPV_shrink_to_cur((SV *) invlist);
8087 PERL_STATIC_INLINE bool
8088 S_invlist_is_iterating(SV* const invlist)
8090 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8092 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8095 #endif /* ifndef PERL_IN_XSUB_RE */
8097 PERL_STATIC_INLINE UV
8098 S_invlist_max(SV* const invlist)
8100 /* Returns the maximum number of elements storable in the inversion list's
8101 * array, without having to realloc() */
8103 PERL_ARGS_ASSERT_INVLIST_MAX;
8105 assert(SvTYPE(invlist) == SVt_INVLIST);
8107 /* Assumes worst case, in which the 0 element is not counted in the
8108 * inversion list, so subtracts 1 for that */
8109 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
8110 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
8111 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
8114 #ifndef PERL_IN_XSUB_RE
8116 Perl__new_invlist(pTHX_ IV initial_size)
8119 /* Return a pointer to a newly constructed inversion list, with enough
8120 * space to store 'initial_size' elements. If that number is negative, a
8121 * system default is used instead */
8125 if (initial_size < 0) {
8129 /* Allocate the initial space */
8130 new_list = newSV_type(SVt_INVLIST);
8132 /* First 1 is in case the zero element isn't in the list; second 1 is for
8134 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
8135 invlist_set_len(new_list, 0, 0);
8137 /* Force iterinit() to be used to get iteration to work */
8138 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
8140 *get_invlist_previous_index_addr(new_list) = 0;
8146 Perl__new_invlist_C_array(pTHX_ const UV* const list)
8148 /* Return a pointer to a newly constructed inversion list, initialized to
8149 * point to <list>, which has to be in the exact correct inversion list
8150 * form, including internal fields. Thus this is a dangerous routine that
8151 * should not be used in the wrong hands. The passed in 'list' contains
8152 * several header fields at the beginning that are not part of the
8153 * inversion list body proper */
8155 const STRLEN length = (STRLEN) list[0];
8156 const UV version_id = list[1];
8157 const bool offset = cBOOL(list[2]);
8158 #define HEADER_LENGTH 3
8159 /* If any of the above changes in any way, you must change HEADER_LENGTH
8160 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
8161 * perl -E 'say int(rand 2**31-1)'
8163 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
8164 data structure type, so that one being
8165 passed in can be validated to be an
8166 inversion list of the correct vintage.
8169 SV* invlist = newSV_type(SVt_INVLIST);
8171 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8173 if (version_id != INVLIST_VERSION_ID) {
8174 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8177 /* The generated array passed in includes header elements that aren't part
8178 * of the list proper, so start it just after them */
8179 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8181 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8182 shouldn't touch it */
8184 *(get_invlist_offset_addr(invlist)) = offset;
8186 /* The 'length' passed to us is the physical number of elements in the
8187 * inversion list. But if there is an offset the logical number is one
8189 invlist_set_len(invlist, length - offset, offset);
8191 invlist_set_previous_index(invlist, 0);
8193 /* Initialize the iteration pointer. */
8194 invlist_iterfinish(invlist);
8196 SvREADONLY_on(invlist);
8200 #endif /* ifndef PERL_IN_XSUB_RE */
8203 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8205 /* Grow the maximum size of an inversion list */
8207 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8209 assert(SvTYPE(invlist) == SVt_INVLIST);
8211 /* Add one to account for the zero element at the beginning which may not
8212 * be counted by the calling parameters */
8213 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8217 S__append_range_to_invlist(pTHX_ SV* const invlist,
8218 const UV start, const UV end)
8220 /* Subject to change or removal. Append the range from 'start' to 'end' at
8221 * the end of the inversion list. The range must be above any existing
8225 UV max = invlist_max(invlist);
8226 UV len = _invlist_len(invlist);
8229 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8231 if (len == 0) { /* Empty lists must be initialized */
8232 offset = start != 0;
8233 array = _invlist_array_init(invlist, ! offset);
8236 /* Here, the existing list is non-empty. The current max entry in the
8237 * list is generally the first value not in the set, except when the
8238 * set extends to the end of permissible values, in which case it is
8239 * the first entry in that final set, and so this call is an attempt to
8240 * append out-of-order */
8242 UV final_element = len - 1;
8243 array = invlist_array(invlist);
8244 if (array[final_element] > start
8245 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8247 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",
8248 array[final_element], start,
8249 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8252 /* Here, it is a legal append. If the new range begins with the first
8253 * value not in the set, it is extending the set, so the new first
8254 * value not in the set is one greater than the newly extended range.
8256 offset = *get_invlist_offset_addr(invlist);
8257 if (array[final_element] == start) {
8258 if (end != UV_MAX) {
8259 array[final_element] = end + 1;
8262 /* But if the end is the maximum representable on the machine,
8263 * just let the range that this would extend to have no end */
8264 invlist_set_len(invlist, len - 1, offset);
8270 /* Here the new range doesn't extend any existing set. Add it */
8272 len += 2; /* Includes an element each for the start and end of range */
8274 /* If wll overflow the existing space, extend, which may cause the array to
8277 invlist_extend(invlist, len);
8279 /* Have to set len here to avoid assert failure in invlist_array() */
8280 invlist_set_len(invlist, len, offset);
8282 array = invlist_array(invlist);
8285 invlist_set_len(invlist, len, offset);
8288 /* The next item on the list starts the range, the one after that is
8289 * one past the new range. */
8290 array[len - 2] = start;
8291 if (end != UV_MAX) {
8292 array[len - 1] = end + 1;
8295 /* But if the end is the maximum representable on the machine, just let
8296 * the range have no end */
8297 invlist_set_len(invlist, len - 1, offset);
8301 #ifndef PERL_IN_XSUB_RE
8304 Perl__invlist_search(SV* const invlist, const UV cp)
8306 /* Searches the inversion list for the entry that contains the input code
8307 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8308 * return value is the index into the list's array of the range that
8313 IV high = _invlist_len(invlist);
8314 const IV highest_element = high - 1;
8317 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8319 /* If list is empty, return failure. */
8324 /* (We can't get the array unless we know the list is non-empty) */
8325 array = invlist_array(invlist);
8327 mid = invlist_previous_index(invlist);
8328 assert(mid >=0 && mid <= highest_element);
8330 /* <mid> contains the cache of the result of the previous call to this
8331 * function (0 the first time). See if this call is for the same result,
8332 * or if it is for mid-1. This is under the theory that calls to this
8333 * function will often be for related code points that are near each other.
8334 * And benchmarks show that caching gives better results. We also test
8335 * here if the code point is within the bounds of the list. These tests
8336 * replace others that would have had to be made anyway to make sure that
8337 * the array bounds were not exceeded, and these give us extra information
8338 * at the same time */
8339 if (cp >= array[mid]) {
8340 if (cp >= array[highest_element]) {
8341 return highest_element;
8344 /* Here, array[mid] <= cp < array[highest_element]. This means that
8345 * the final element is not the answer, so can exclude it; it also
8346 * means that <mid> is not the final element, so can refer to 'mid + 1'
8348 if (cp < array[mid + 1]) {
8354 else { /* cp < aray[mid] */
8355 if (cp < array[0]) { /* Fail if outside the array */
8359 if (cp >= array[mid - 1]) {
8364 /* Binary search. What we are looking for is <i> such that
8365 * array[i] <= cp < array[i+1]
8366 * The loop below converges on the i+1. Note that there may not be an
8367 * (i+1)th element in the array, and things work nonetheless */
8368 while (low < high) {
8369 mid = (low + high) / 2;
8370 assert(mid <= highest_element);
8371 if (array[mid] <= cp) { /* cp >= array[mid] */
8374 /* We could do this extra test to exit the loop early.
8375 if (cp < array[low]) {
8380 else { /* cp < array[mid] */
8387 invlist_set_previous_index(invlist, high);
8392 Perl__invlist_populate_swatch(SV* const invlist,
8393 const UV start, const UV end, U8* swatch)
8395 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8396 * but is used when the swash has an inversion list. This makes this much
8397 * faster, as it uses a binary search instead of a linear one. This is
8398 * intimately tied to that function, and perhaps should be in utf8.c,
8399 * except it is intimately tied to inversion lists as well. It assumes
8400 * that <swatch> is all 0's on input */
8403 const IV len = _invlist_len(invlist);
8407 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8409 if (len == 0) { /* Empty inversion list */
8413 array = invlist_array(invlist);
8415 /* Find which element it is */
8416 i = _invlist_search(invlist, start);
8418 /* We populate from <start> to <end> */
8419 while (current < end) {
8422 /* The inversion list gives the results for every possible code point
8423 * after the first one in the list. Only those ranges whose index is
8424 * even are ones that the inversion list matches. For the odd ones,
8425 * and if the initial code point is not in the list, we have to skip
8426 * forward to the next element */
8427 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8429 if (i >= len) { /* Finished if beyond the end of the array */
8433 if (current >= end) { /* Finished if beyond the end of what we
8435 if (LIKELY(end < UV_MAX)) {
8439 /* We get here when the upper bound is the maximum
8440 * representable on the machine, and we are looking for just
8441 * that code point. Have to special case it */
8443 goto join_end_of_list;
8446 assert(current >= start);
8448 /* The current range ends one below the next one, except don't go past
8451 upper = (i < len && array[i] < end) ? array[i] : end;
8453 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8454 * for each code point in it */
8455 for (; current < upper; current++) {
8456 const STRLEN offset = (STRLEN)(current - start);
8457 swatch[offset >> 3] |= 1 << (offset & 7);
8462 /* Quit if at the end of the list */
8465 /* But first, have to deal with the highest possible code point on
8466 * the platform. The previous code assumes that <end> is one
8467 * beyond where we want to populate, but that is impossible at the
8468 * platform's infinity, so have to handle it specially */
8469 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8471 const STRLEN offset = (STRLEN)(end - start);
8472 swatch[offset >> 3] |= 1 << (offset & 7);
8477 /* Advance to the next range, which will be for code points not in the
8486 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8487 const bool complement_b, SV** output)
8489 /* Take the union of two inversion lists and point <output> to it. *output
8490 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8491 * the reference count to that list will be decremented if not already a
8492 * temporary (mortal); otherwise *output will be made correspondingly
8493 * mortal. The first list, <a>, may be NULL, in which case a copy of the
8494 * second list is returned. If <complement_b> is TRUE, the union is taken
8495 * of the complement (inversion) of <b> instead of b itself.
8497 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8498 * Richard Gillam, published by Addison-Wesley, and explained at some
8499 * length there. The preface says to incorporate its examples into your
8500 * code at your own risk.
8502 * The algorithm is like a merge sort.
8504 * XXX A potential performance improvement is to keep track as we go along
8505 * if only one of the inputs contributes to the result, meaning the other
8506 * is a subset of that one. In that case, we can skip the final copy and
8507 * return the larger of the input lists, but then outside code might need
8508 * to keep track of whether to free the input list or not */
8510 const UV* array_a; /* a's array */
8512 UV len_a; /* length of a's array */
8515 SV* u; /* the resulting union */
8519 UV i_a = 0; /* current index into a's array */
8523 /* running count, as explained in the algorithm source book; items are
8524 * stopped accumulating and are output when the count changes to/from 0.
8525 * The count is incremented when we start a range that's in the set, and
8526 * decremented when we start a range that's not in the set. So its range
8527 * is 0 to 2. Only when the count is zero is something not in the set.
8531 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
8534 /* If either one is empty, the union is the other one */
8535 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
8536 bool make_temp = FALSE; /* Should we mortalize the result? */
8540 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8546 *output = invlist_clone(b);
8548 _invlist_invert(*output);
8550 } /* else *output already = b; */
8553 sv_2mortal(*output);
8557 else if ((len_b = _invlist_len(b)) == 0) {
8558 bool make_temp = FALSE;
8560 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8565 /* The complement of an empty list is a list that has everything in it,
8566 * so the union with <a> includes everything too */
8569 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8573 *output = _new_invlist(1);
8574 _append_range_to_invlist(*output, 0, UV_MAX);
8576 else if (*output != a) {
8577 *output = invlist_clone(a);
8579 /* else *output already = a; */
8582 sv_2mortal(*output);
8587 /* Here both lists exist and are non-empty */
8588 array_a = invlist_array(a);
8589 array_b = invlist_array(b);
8591 /* If are to take the union of 'a' with the complement of b, set it
8592 * up so are looking at b's complement. */
8595 /* To complement, we invert: if the first element is 0, remove it. To
8596 * do this, we just pretend the array starts one later */
8597 if (array_b[0] == 0) {
8603 /* But if the first element is not zero, we pretend the list starts
8604 * at the 0 that is always stored immediately before the array. */
8610 /* Size the union for the worst case: that the sets are completely
8612 u = _new_invlist(len_a + len_b);
8614 /* Will contain U+0000 if either component does */
8615 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
8616 || (len_b > 0 && array_b[0] == 0));
8618 /* Go through each list item by item, stopping when exhausted one of
8620 while (i_a < len_a && i_b < len_b) {
8621 UV cp; /* The element to potentially add to the union's array */
8622 bool cp_in_set; /* is it in the the input list's set or not */
8624 /* We need to take one or the other of the two inputs for the union.
8625 * Since we are merging two sorted lists, we take the smaller of the
8626 * next items. In case of a tie, we take the one that is in its set
8627 * first. If we took one not in the set first, it would decrement the
8628 * count, possibly to 0 which would cause it to be output as ending the
8629 * range, and the next time through we would take the same number, and
8630 * output it again as beginning the next range. By doing it the
8631 * opposite way, there is no possibility that the count will be
8632 * momentarily decremented to 0, and thus the two adjoining ranges will
8633 * be seamlessly merged. (In a tie and both are in the set or both not
8634 * in the set, it doesn't matter which we take first.) */
8635 if (array_a[i_a] < array_b[i_b]
8636 || (array_a[i_a] == array_b[i_b]
8637 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8639 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8643 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8644 cp = array_b[i_b++];
8647 /* Here, have chosen which of the two inputs to look at. Only output
8648 * if the running count changes to/from 0, which marks the
8649 * beginning/end of a range in that's in the set */
8652 array_u[i_u++] = cp;
8659 array_u[i_u++] = cp;
8664 /* Here, we are finished going through at least one of the lists, which
8665 * means there is something remaining in at most one. We check if the list
8666 * that hasn't been exhausted is positioned such that we are in the middle
8667 * of a range in its set or not. (i_a and i_b point to the element beyond
8668 * the one we care about.) If in the set, we decrement 'count'; if 0, there
8669 * is potentially more to output.
8670 * There are four cases:
8671 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
8672 * in the union is entirely from the non-exhausted set.
8673 * 2) Both were in their sets, count is 2. Nothing further should
8674 * be output, as everything that remains will be in the exhausted
8675 * list's set, hence in the union; decrementing to 1 but not 0 insures
8677 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
8678 * Nothing further should be output because the union includes
8679 * everything from the exhausted set. Not decrementing ensures that.
8680 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
8681 * decrementing to 0 insures that we look at the remainder of the
8682 * non-exhausted set */
8683 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8684 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8689 /* The final length is what we've output so far, plus what else is about to
8690 * be output. (If 'count' is non-zero, then the input list we exhausted
8691 * has everything remaining up to the machine's limit in its set, and hence
8692 * in the union, so there will be no further output. */
8695 /* At most one of the subexpressions will be non-zero */
8696 len_u += (len_a - i_a) + (len_b - i_b);
8699 /* Set result to final length, which can change the pointer to array_u, so
8701 if (len_u != _invlist_len(u)) {
8702 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
8704 array_u = invlist_array(u);
8707 /* When 'count' is 0, the list that was exhausted (if one was shorter than
8708 * the other) ended with everything above it not in its set. That means
8709 * that the remaining part of the union is precisely the same as the
8710 * non-exhausted list, so can just copy it unchanged. (If both list were
8711 * exhausted at the same time, then the operations below will be both 0.)
8714 IV copy_count; /* At most one will have a non-zero copy count */
8715 if ((copy_count = len_a - i_a) > 0) {
8716 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
8718 else if ((copy_count = len_b - i_b) > 0) {
8719 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
8723 /* We may be removing a reference to one of the inputs. If so, the output
8724 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8725 * count decremented) */
8726 if (a == *output || b == *output) {
8727 assert(! invlist_is_iterating(*output));
8728 if ((SvTEMP(*output))) {
8732 SvREFCNT_dec_NN(*output);
8742 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8743 const bool complement_b, SV** i)
8745 /* Take the intersection of two inversion lists and point <i> to it. *i
8746 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8747 * the reference count to that list will be decremented if not already a
8748 * temporary (mortal); otherwise *i will be made correspondingly mortal.
8749 * The first list, <a>, may be NULL, in which case an empty list is
8750 * returned. If <complement_b> is TRUE, the result will be the
8751 * intersection of <a> and the complement (or inversion) of <b> instead of
8754 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8755 * Richard Gillam, published by Addison-Wesley, and explained at some
8756 * length there. The preface says to incorporate its examples into your
8757 * code at your own risk. In fact, it had bugs
8759 * The algorithm is like a merge sort, and is essentially the same as the
8763 const UV* array_a; /* a's array */
8765 UV len_a; /* length of a's array */
8768 SV* r; /* the resulting intersection */
8772 UV i_a = 0; /* current index into a's array */
8776 /* running count, as explained in the algorithm source book; items are
8777 * stopped accumulating and are output when the count changes to/from 2.
8778 * The count is incremented when we start a range that's in the set, and
8779 * decremented when we start a range that's not in the set. So its range
8780 * is 0 to 2. Only when the count is 2 is something in the intersection.
8784 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
8787 /* Special case if either one is empty */
8788 len_a = (a == NULL) ? 0 : _invlist_len(a);
8789 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
8790 bool make_temp = FALSE;
8792 if (len_a != 0 && complement_b) {
8794 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
8795 * be empty. Here, also we are using 'b's complement, which hence
8796 * must be every possible code point. Thus the intersection is
8800 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8805 *i = invlist_clone(a);
8807 /* else *i is already 'a' */
8815 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
8816 * intersection must be empty */
8818 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8823 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8827 *i = _new_invlist(0);
8835 /* Here both lists exist and are non-empty */
8836 array_a = invlist_array(a);
8837 array_b = invlist_array(b);
8839 /* If are to take the intersection of 'a' with the complement of b, set it
8840 * up so are looking at b's complement. */
8843 /* To complement, we invert: if the first element is 0, remove it. To
8844 * do this, we just pretend the array starts one later */
8845 if (array_b[0] == 0) {
8851 /* But if the first element is not zero, we pretend the list starts
8852 * at the 0 that is always stored immediately before the array. */
8858 /* Size the intersection for the worst case: that the intersection ends up
8859 * fragmenting everything to be completely disjoint */
8860 r= _new_invlist(len_a + len_b);
8862 /* Will contain U+0000 iff both components do */
8863 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
8864 && len_b > 0 && array_b[0] == 0);
8866 /* Go through each list item by item, stopping when exhausted one of
8868 while (i_a < len_a && i_b < len_b) {
8869 UV cp; /* The element to potentially add to the intersection's
8871 bool cp_in_set; /* Is it in the input list's set or not */
8873 /* We need to take one or the other of the two inputs for the
8874 * intersection. Since we are merging two sorted lists, we take the
8875 * smaller of the next items. In case of a tie, we take the one that
8876 * is not in its set first (a difference from the union algorithm). If
8877 * we took one in the set first, it would increment the count, possibly
8878 * to 2 which would cause it to be output as starting a range in the
8879 * intersection, and the next time through we would take that same
8880 * number, and output it again as ending the set. By doing it the
8881 * opposite of this, there is no possibility that the count will be
8882 * momentarily incremented to 2. (In a tie and both are in the set or
8883 * both not in the set, it doesn't matter which we take first.) */
8884 if (array_a[i_a] < array_b[i_b]
8885 || (array_a[i_a] == array_b[i_b]
8886 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8888 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8892 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8896 /* Here, have chosen which of the two inputs to look at. Only output
8897 * if the running count changes to/from 2, which marks the
8898 * beginning/end of a range that's in the intersection */
8902 array_r[i_r++] = cp;
8907 array_r[i_r++] = cp;
8913 /* Here, we are finished going through at least one of the lists, which
8914 * means there is something remaining in at most one. We check if the list
8915 * that has been exhausted is positioned such that we are in the middle
8916 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
8917 * the ones we care about.) There are four cases:
8918 * 1) Both weren't in their sets, count is 0, and remains 0. There's
8919 * nothing left in the intersection.
8920 * 2) Both were in their sets, count is 2 and perhaps is incremented to
8921 * above 2. What should be output is exactly that which is in the
8922 * non-exhausted set, as everything it has is also in the intersection
8923 * set, and everything it doesn't have can't be in the intersection
8924 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
8925 * gets incremented to 2. Like the previous case, the intersection is
8926 * everything that remains in the non-exhausted set.
8927 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
8928 * remains 1. And the intersection has nothing more. */
8929 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8930 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8935 /* The final length is what we've output so far plus what else is in the
8936 * intersection. At most one of the subexpressions below will be non-zero
8940 len_r += (len_a - i_a) + (len_b - i_b);
8943 /* Set result to final length, which can change the pointer to array_r, so
8945 if (len_r != _invlist_len(r)) {
8946 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
8948 array_r = invlist_array(r);
8951 /* Finish outputting any remaining */
8952 if (count >= 2) { /* At most one will have a non-zero copy count */
8954 if ((copy_count = len_a - i_a) > 0) {
8955 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
8957 else if ((copy_count = len_b - i_b) > 0) {
8958 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
8962 /* We may be removing a reference to one of the inputs. If so, the output
8963 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8964 * count decremented) */
8965 if (a == *i || b == *i) {
8966 assert(! invlist_is_iterating(*i));
8971 SvREFCNT_dec_NN(*i);
8981 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
8983 /* Add the range from 'start' to 'end' inclusive to the inversion list's
8984 * set. A pointer to the inversion list is returned. This may actually be
8985 * a new list, in which case the passed in one has been destroyed. The
8986 * passed-in inversion list can be NULL, in which case a new one is created
8987 * with just the one range in it */
8992 if (invlist == NULL) {
8993 invlist = _new_invlist(2);
8997 len = _invlist_len(invlist);
9000 /* If comes after the final entry actually in the list, can just append it
9003 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
9004 && start >= invlist_array(invlist)[len - 1]))
9006 _append_range_to_invlist(invlist, start, end);
9010 /* Here, can't just append things, create and return a new inversion list
9011 * which is the union of this range and the existing inversion list */
9012 range_invlist = _new_invlist(2);
9013 _append_range_to_invlist(range_invlist, start, end);
9015 _invlist_union(invlist, range_invlist, &invlist);
9017 /* The temporary can be freed */
9018 SvREFCNT_dec_NN(range_invlist);
9024 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9025 UV** other_elements_ptr)
9027 /* Create and return an inversion list whose contents are to be populated
9028 * by the caller. The caller gives the number of elements (in 'size') and
9029 * the very first element ('element0'). This function will set
9030 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9033 * Obviously there is some trust involved that the caller will properly
9034 * fill in the other elements of the array.
9036 * (The first element needs to be passed in, as the underlying code does
9037 * things differently depending on whether it is zero or non-zero) */
9039 SV* invlist = _new_invlist(size);
9042 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9044 _append_range_to_invlist(invlist, element0, element0);
9045 offset = *get_invlist_offset_addr(invlist);
9047 invlist_set_len(invlist, size, offset);
9048 *other_elements_ptr = invlist_array(invlist) + 1;
9054 PERL_STATIC_INLINE SV*
9055 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9056 return _add_range_to_invlist(invlist, cp, cp);
9059 #ifndef PERL_IN_XSUB_RE
9061 Perl__invlist_invert(pTHX_ SV* const invlist)
9063 /* Complement the input inversion list. This adds a 0 if the list didn't
9064 * have a zero; removes it otherwise. As described above, the data
9065 * structure is set up so that this is very efficient */
9067 PERL_ARGS_ASSERT__INVLIST_INVERT;
9069 assert(! invlist_is_iterating(invlist));
9071 /* The inverse of matching nothing is matching everything */
9072 if (_invlist_len(invlist) == 0) {
9073 _append_range_to_invlist(invlist, 0, UV_MAX);
9077 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9082 PERL_STATIC_INLINE SV*
9083 S_invlist_clone(pTHX_ SV* const invlist)
9086 /* Return a new inversion list that is a copy of the input one, which is
9087 * unchanged. The new list will not be mortal even if the old one was. */
9089 /* Need to allocate extra space to accommodate Perl's addition of a
9090 * trailing NUL to SvPV's, since it thinks they are always strings */
9091 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9092 STRLEN physical_length = SvCUR(invlist);
9093 bool offset = *(get_invlist_offset_addr(invlist));
9095 PERL_ARGS_ASSERT_INVLIST_CLONE;
9097 *(get_invlist_offset_addr(new_invlist)) = offset;
9098 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9099 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9104 PERL_STATIC_INLINE STRLEN*
9105 S_get_invlist_iter_addr(SV* invlist)
9107 /* Return the address of the UV that contains the current iteration
9110 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9112 assert(SvTYPE(invlist) == SVt_INVLIST);
9114 return &(((XINVLIST*) SvANY(invlist))->iterator);
9117 PERL_STATIC_INLINE void
9118 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9120 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9122 *get_invlist_iter_addr(invlist) = 0;
9125 PERL_STATIC_INLINE void
9126 S_invlist_iterfinish(SV* invlist)
9128 /* Terminate iterator for invlist. This is to catch development errors.
9129 * Any iteration that is interrupted before completed should call this
9130 * function. Functions that add code points anywhere else but to the end
9131 * of an inversion list assert that they are not in the middle of an
9132 * iteration. If they were, the addition would make the iteration
9133 * problematical: if the iteration hadn't reached the place where things
9134 * were being added, it would be ok */
9136 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
9138 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
9142 S_invlist_iternext(SV* invlist, UV* start, UV* end)
9144 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
9145 * This call sets in <*start> and <*end>, the next range in <invlist>.
9146 * Returns <TRUE> if successful and the next call will return the next
9147 * range; <FALSE> if was already at the end of the list. If the latter,
9148 * <*start> and <*end> are unchanged, and the next call to this function
9149 * will start over at the beginning of the list */
9151 STRLEN* pos = get_invlist_iter_addr(invlist);
9152 UV len = _invlist_len(invlist);
9155 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9158 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9162 array = invlist_array(invlist);
9164 *start = array[(*pos)++];
9170 *end = array[(*pos)++] - 1;
9176 PERL_STATIC_INLINE UV
9177 S_invlist_highest(SV* const invlist)
9179 /* Returns the highest code point that matches an inversion list. This API
9180 * has an ambiguity, as it returns 0 under either the highest is actually
9181 * 0, or if the list is empty. If this distinction matters to you, check
9182 * for emptiness before calling this function */
9184 UV len = _invlist_len(invlist);
9187 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
9193 array = invlist_array(invlist);
9195 /* The last element in the array in the inversion list always starts a
9196 * range that goes to infinity. That range may be for code points that are
9197 * matched in the inversion list, or it may be for ones that aren't
9198 * matched. In the latter case, the highest code point in the set is one
9199 * less than the beginning of this range; otherwise it is the final element
9200 * of this range: infinity */
9201 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
9203 : array[len - 1] - 1;
9206 #ifndef PERL_IN_XSUB_RE
9208 Perl__invlist_contents(pTHX_ SV* const invlist)
9210 /* Get the contents of an inversion list into a string SV so that they can
9211 * be printed out. It uses the format traditionally done for debug tracing
9215 SV* output = newSVpvs("\n");
9217 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
9219 assert(! invlist_is_iterating(invlist));
9221 invlist_iterinit(invlist);
9222 while (invlist_iternext(invlist, &start, &end)) {
9223 if (end == UV_MAX) {
9224 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
9226 else if (end != start) {
9227 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
9231 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
9239 #ifndef PERL_IN_XSUB_RE
9241 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
9242 const char * const indent, SV* const invlist)
9244 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
9245 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
9246 * the string 'indent'. The output looks like this:
9247 [0] 0x000A .. 0x000D
9249 [4] 0x2028 .. 0x2029
9250 [6] 0x3104 .. INFINITY
9251 * This means that the first range of code points matched by the list are
9252 * 0xA through 0xD; the second range contains only the single code point
9253 * 0x85, etc. An inversion list is an array of UVs. Two array elements
9254 * are used to define each range (except if the final range extends to
9255 * infinity, only a single element is needed). The array index of the
9256 * first element for the corresponding range is given in brackets. */
9261 PERL_ARGS_ASSERT__INVLIST_DUMP;
9263 if (invlist_is_iterating(invlist)) {
9264 Perl_dump_indent(aTHX_ level, file,
9265 "%sCan't dump inversion list because is in middle of iterating\n",
9270 invlist_iterinit(invlist);
9271 while (invlist_iternext(invlist, &start, &end)) {
9272 if (end == UV_MAX) {
9273 Perl_dump_indent(aTHX_ level, file,
9274 "%s[%"UVuf"] 0x%04"UVXf" .. INFINITY\n",
9275 indent, (UV)count, start);
9277 else if (end != start) {
9278 Perl_dump_indent(aTHX_ level, file,
9279 "%s[%"UVuf"] 0x%04"UVXf" .. 0x%04"UVXf"\n",
9280 indent, (UV)count, start, end);
9283 Perl_dump_indent(aTHX_ level, file, "%s[%"UVuf"] 0x%04"UVXf"\n",
9284 indent, (UV)count, start);
9291 Perl__load_PL_utf8_foldclosures (pTHX)
9293 assert(! PL_utf8_foldclosures);
9295 /* If the folds haven't been read in, call a fold function
9297 if (! PL_utf8_tofold) {
9298 U8 dummy[UTF8_MAXBYTES_CASE+1];
9300 /* This string is just a short named one above \xff */
9301 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
9302 assert(PL_utf8_tofold); /* Verify that worked */
9304 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
9308 #ifdef PERL_ARGS_ASSERT__INVLISTEQ
9310 S__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
9312 /* Return a boolean as to if the two passed in inversion lists are
9313 * identical. The final argument, if TRUE, says to take the complement of
9314 * the second inversion list before doing the comparison */
9316 const UV* array_a = invlist_array(a);
9317 const UV* array_b = invlist_array(b);
9318 UV len_a = _invlist_len(a);
9319 UV len_b = _invlist_len(b);
9321 UV i = 0; /* current index into the arrays */
9322 bool retval = TRUE; /* Assume are identical until proven otherwise */
9324 PERL_ARGS_ASSERT__INVLISTEQ;
9326 /* If are to compare 'a' with the complement of b, set it
9327 * up so are looking at b's complement. */
9330 /* The complement of nothing is everything, so <a> would have to have
9331 * just one element, starting at zero (ending at infinity) */
9333 return (len_a == 1 && array_a[0] == 0);
9335 else if (array_b[0] == 0) {
9337 /* Otherwise, to complement, we invert. Here, the first element is
9338 * 0, just remove it. To do this, we just pretend the array starts
9346 /* But if the first element is not zero, we pretend the list starts
9347 * at the 0 that is always stored immediately before the array. */
9353 /* Make sure that the lengths are the same, as well as the final element
9354 * before looping through the remainder. (Thus we test the length, final,
9355 * and first elements right off the bat) */
9356 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
9359 else for (i = 0; i < len_a - 1; i++) {
9360 if (array_a[i] != array_b[i]) {
9371 * As best we can, determine the characters that can match the start of
9372 * the given EXACTF-ish node.
9374 * Returns the invlist as a new SV*; it is the caller's responsibility to
9375 * call SvREFCNT_dec() when done with it.
9378 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
9380 const U8 * s = (U8*)STRING(node);
9381 SSize_t bytelen = STR_LEN(node);
9383 /* Start out big enough for 2 separate code points */
9384 SV* invlist = _new_invlist(4);
9386 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
9391 /* We punt and assume can match anything if the node begins
9392 * with a multi-character fold. Things are complicated. For
9393 * example, /ffi/i could match any of:
9394 * "\N{LATIN SMALL LIGATURE FFI}"
9395 * "\N{LATIN SMALL LIGATURE FF}I"
9396 * "F\N{LATIN SMALL LIGATURE FI}"
9397 * plus several other things; and making sure we have all the
9398 * possibilities is hard. */
9399 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
9400 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
9403 /* Any Latin1 range character can potentially match any
9404 * other depending on the locale */
9405 if (OP(node) == EXACTFL) {
9406 _invlist_union(invlist, PL_Latin1, &invlist);
9409 /* But otherwise, it matches at least itself. We can
9410 * quickly tell if it has a distinct fold, and if so,
9411 * it matches that as well */
9412 invlist = add_cp_to_invlist(invlist, uc);
9413 if (IS_IN_SOME_FOLD_L1(uc))
9414 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
9417 /* Some characters match above-Latin1 ones under /i. This
9418 * is true of EXACTFL ones when the locale is UTF-8 */
9419 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
9420 && (! isASCII(uc) || (OP(node) != EXACTFA
9421 && OP(node) != EXACTFA_NO_TRIE)))
9423 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
9427 else { /* Pattern is UTF-8 */
9428 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
9429 STRLEN foldlen = UTF8SKIP(s);
9430 const U8* e = s + bytelen;
9433 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
9435 /* The only code points that aren't folded in a UTF EXACTFish
9436 * node are are the problematic ones in EXACTFL nodes */
9437 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
9438 /* We need to check for the possibility that this EXACTFL
9439 * node begins with a multi-char fold. Therefore we fold
9440 * the first few characters of it so that we can make that
9445 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
9447 *(d++) = (U8) toFOLD(*s);
9452 to_utf8_fold(s, d, &len);
9458 /* And set up so the code below that looks in this folded
9459 * buffer instead of the node's string */
9461 foldlen = UTF8SKIP(folded);
9465 /* When we reach here 's' points to the fold of the first
9466 * character(s) of the node; and 'e' points to far enough along
9467 * the folded string to be just past any possible multi-char
9468 * fold. 'foldlen' is the length in bytes of the first
9471 * Unlike the non-UTF-8 case, the macro for determining if a
9472 * string is a multi-char fold requires all the characters to
9473 * already be folded. This is because of all the complications
9474 * if not. Note that they are folded anyway, except in EXACTFL
9475 * nodes. Like the non-UTF case above, we punt if the node
9476 * begins with a multi-char fold */
9478 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
9479 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
9481 else { /* Single char fold */
9483 /* It matches all the things that fold to it, which are
9484 * found in PL_utf8_foldclosures (including itself) */
9485 invlist = add_cp_to_invlist(invlist, uc);
9486 if (! PL_utf8_foldclosures)
9487 _load_PL_utf8_foldclosures();
9488 if ((listp = hv_fetch(PL_utf8_foldclosures,
9489 (char *) s, foldlen, FALSE)))
9491 AV* list = (AV*) *listp;
9493 for (k = 0; k <= av_tindex(list); k++) {
9494 SV** c_p = av_fetch(list, k, FALSE);
9500 /* /aa doesn't allow folds between ASCII and non- */
9501 if ((OP(node) == EXACTFA || OP(node) == EXACTFA_NO_TRIE)
9502 && isASCII(c) != isASCII(uc))
9507 invlist = add_cp_to_invlist(invlist, c);
9516 #undef HEADER_LENGTH
9517 #undef TO_INTERNAL_SIZE
9518 #undef FROM_INTERNAL_SIZE
9519 #undef INVLIST_VERSION_ID
9521 /* End of inversion list object */
9524 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
9526 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
9527 * constructs, and updates RExC_flags with them. On input, RExC_parse
9528 * should point to the first flag; it is updated on output to point to the
9529 * final ')' or ':'. There needs to be at least one flag, or this will
9532 /* for (?g), (?gc), and (?o) warnings; warning
9533 about (?c) will warn about (?g) -- japhy */
9535 #define WASTED_O 0x01
9536 #define WASTED_G 0x02
9537 #define WASTED_C 0x04
9538 #define WASTED_GC (WASTED_G|WASTED_C)
9539 I32 wastedflags = 0x00;
9540 U32 posflags = 0, negflags = 0;
9541 U32 *flagsp = &posflags;
9542 char has_charset_modifier = '\0';
9544 bool has_use_defaults = FALSE;
9545 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
9546 int x_mod_count = 0;
9548 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
9550 /* '^' as an initial flag sets certain defaults */
9551 if (UCHARAT(RExC_parse) == '^') {
9553 has_use_defaults = TRUE;
9554 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
9555 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
9556 ? REGEX_UNICODE_CHARSET
9557 : REGEX_DEPENDS_CHARSET);
9560 cs = get_regex_charset(RExC_flags);
9561 if (cs == REGEX_DEPENDS_CHARSET
9562 && (RExC_utf8 || RExC_uni_semantics))
9564 cs = REGEX_UNICODE_CHARSET;
9567 while (*RExC_parse) {
9568 /* && strchr("iogcmsx", *RExC_parse) */
9569 /* (?g), (?gc) and (?o) are useless here
9570 and must be globally applied -- japhy */
9571 switch (*RExC_parse) {
9573 /* Code for the imsxn flags */
9574 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
9576 case LOCALE_PAT_MOD:
9577 if (has_charset_modifier) {
9578 goto excess_modifier;
9580 else if (flagsp == &negflags) {
9583 cs = REGEX_LOCALE_CHARSET;
9584 has_charset_modifier = LOCALE_PAT_MOD;
9586 case UNICODE_PAT_MOD:
9587 if (has_charset_modifier) {
9588 goto excess_modifier;
9590 else if (flagsp == &negflags) {
9593 cs = REGEX_UNICODE_CHARSET;
9594 has_charset_modifier = UNICODE_PAT_MOD;
9596 case ASCII_RESTRICT_PAT_MOD:
9597 if (flagsp == &negflags) {
9600 if (has_charset_modifier) {
9601 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
9602 goto excess_modifier;
9604 /* Doubled modifier implies more restricted */
9605 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
9608 cs = REGEX_ASCII_RESTRICTED_CHARSET;
9610 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
9612 case DEPENDS_PAT_MOD:
9613 if (has_use_defaults) {
9614 goto fail_modifiers;
9616 else if (flagsp == &negflags) {
9619 else if (has_charset_modifier) {
9620 goto excess_modifier;
9623 /* The dual charset means unicode semantics if the
9624 * pattern (or target, not known until runtime) are
9625 * utf8, or something in the pattern indicates unicode
9627 cs = (RExC_utf8 || RExC_uni_semantics)
9628 ? REGEX_UNICODE_CHARSET
9629 : REGEX_DEPENDS_CHARSET;
9630 has_charset_modifier = DEPENDS_PAT_MOD;
9634 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
9635 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
9637 else if (has_charset_modifier == *(RExC_parse - 1)) {
9638 vFAIL2("Regexp modifier \"%c\" may not appear twice",
9642 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
9644 NOT_REACHED; /*NOTREACHED*/
9647 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
9649 NOT_REACHED; /*NOTREACHED*/
9650 case ONCE_PAT_MOD: /* 'o' */
9651 case GLOBAL_PAT_MOD: /* 'g' */
9652 if (PASS2 && ckWARN(WARN_REGEXP)) {
9653 const I32 wflagbit = *RExC_parse == 'o'
9656 if (! (wastedflags & wflagbit) ) {
9657 wastedflags |= wflagbit;
9658 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9661 "Useless (%s%c) - %suse /%c modifier",
9662 flagsp == &negflags ? "?-" : "?",
9664 flagsp == &negflags ? "don't " : "",
9671 case CONTINUE_PAT_MOD: /* 'c' */
9672 if (PASS2 && ckWARN(WARN_REGEXP)) {
9673 if (! (wastedflags & WASTED_C) ) {
9674 wastedflags |= WASTED_GC;
9675 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9678 "Useless (%sc) - %suse /gc modifier",
9679 flagsp == &negflags ? "?-" : "?",
9680 flagsp == &negflags ? "don't " : ""
9685 case KEEPCOPY_PAT_MOD: /* 'p' */
9686 if (flagsp == &negflags) {
9688 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
9690 *flagsp |= RXf_PMf_KEEPCOPY;
9694 /* A flag is a default iff it is following a minus, so
9695 * if there is a minus, it means will be trying to
9696 * re-specify a default which is an error */
9697 if (has_use_defaults || flagsp == &negflags) {
9698 goto fail_modifiers;
9701 wastedflags = 0; /* reset so (?g-c) warns twice */
9705 RExC_flags |= posflags;
9706 RExC_flags &= ~negflags;
9707 set_regex_charset(&RExC_flags, cs);
9708 if (RExC_flags & RXf_PMf_FOLD) {
9709 RExC_contains_i = 1;
9712 STD_PMMOD_FLAGS_PARSE_X_WARN(x_mod_count);
9718 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9719 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9720 vFAIL2utf8f("Sequence (%"UTF8f"...) not recognized",
9721 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
9722 NOT_REACHED; /*NOTREACHED*/
9729 STD_PMMOD_FLAGS_PARSE_X_WARN(x_mod_count);
9734 - reg - regular expression, i.e. main body or parenthesized thing
9736 * Caller must absorb opening parenthesis.
9738 * Combining parenthesis handling with the base level of regular expression
9739 * is a trifle forced, but the need to tie the tails of the branches to what
9740 * follows makes it hard to avoid.
9742 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
9744 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
9746 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
9749 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
9750 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
9751 needs to be restarted.
9752 Otherwise would only return NULL if regbranch() returns NULL, which
9755 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
9756 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
9757 * 2 is like 1, but indicates that nextchar() has been called to advance
9758 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
9759 * this flag alerts us to the need to check for that */
9761 regnode *ret; /* Will be the head of the group. */
9764 regnode *ender = NULL;
9767 U32 oregflags = RExC_flags;
9768 bool have_branch = 0;
9770 I32 freeze_paren = 0;
9771 I32 after_freeze = 0;
9772 I32 num; /* numeric backreferences */
9774 char * parse_start = RExC_parse; /* MJD */
9775 char * const oregcomp_parse = RExC_parse;
9777 GET_RE_DEBUG_FLAGS_DECL;
9779 PERL_ARGS_ASSERT_REG;
9780 DEBUG_PARSE("reg ");
9782 *flagp = 0; /* Tentatively. */
9785 /* Make an OPEN node, if parenthesized. */
9788 /* Under /x, space and comments can be gobbled up between the '(' and
9789 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
9790 * intervening space, as the sequence is a token, and a token should be
9792 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
9794 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
9795 char *start_verb = RExC_parse;
9796 STRLEN verb_len = 0;
9797 char *start_arg = NULL;
9798 unsigned char op = 0;
9800 int internal_argval = 0; /* internal_argval is only useful if
9803 if (has_intervening_patws) {
9805 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
9807 while ( *RExC_parse && *RExC_parse != ')' ) {
9808 if ( *RExC_parse == ':' ) {
9809 start_arg = RExC_parse + 1;
9815 verb_len = RExC_parse - start_verb;
9818 while ( *RExC_parse && *RExC_parse != ')' )
9820 if ( *RExC_parse != ')' )
9821 vFAIL("Unterminated verb pattern argument");
9822 if ( RExC_parse == start_arg )
9825 if ( *RExC_parse != ')' )
9826 vFAIL("Unterminated verb pattern");
9829 switch ( *start_verb ) {
9830 case 'A': /* (*ACCEPT) */
9831 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
9833 internal_argval = RExC_nestroot;
9836 case 'C': /* (*COMMIT) */
9837 if ( memEQs(start_verb,verb_len,"COMMIT") )
9840 case 'F': /* (*FAIL) */
9841 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
9846 case ':': /* (*:NAME) */
9847 case 'M': /* (*MARK:NAME) */
9848 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
9853 case 'P': /* (*PRUNE) */
9854 if ( memEQs(start_verb,verb_len,"PRUNE") )
9857 case 'S': /* (*SKIP) */
9858 if ( memEQs(start_verb,verb_len,"SKIP") )
9861 case 'T': /* (*THEN) */
9862 /* [19:06] <TimToady> :: is then */
9863 if ( memEQs(start_verb,verb_len,"THEN") ) {
9865 RExC_seen |= REG_CUTGROUP_SEEN;
9870 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9872 "Unknown verb pattern '%"UTF8f"'",
9873 UTF8fARG(UTF, verb_len, start_verb));
9876 if ( start_arg && internal_argval ) {
9877 vFAIL3("Verb pattern '%.*s' may not have an argument",
9878 verb_len, start_verb);
9879 } else if ( argok < 0 && !start_arg ) {
9880 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
9881 verb_len, start_verb);
9883 ret = reganode(pRExC_state, op, internal_argval);
9884 if ( ! internal_argval && ! SIZE_ONLY ) {
9886 SV *sv = newSVpvn( start_arg,
9887 RExC_parse - start_arg);
9888 ARG(ret) = add_data( pRExC_state,
9890 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
9897 if (!internal_argval)
9898 RExC_seen |= REG_VERBARG_SEEN;
9899 } else if ( start_arg ) {
9900 vFAIL3("Verb pattern '%.*s' may not have an argument",
9901 verb_len, start_verb);
9903 ret = reg_node(pRExC_state, op);
9905 nextchar(pRExC_state);
9908 else if (*RExC_parse == '?') { /* (?...) */
9909 bool is_logical = 0;
9910 const char * const seqstart = RExC_parse;
9911 const char * endptr;
9912 if (has_intervening_patws) {
9914 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
9918 paren = *RExC_parse++;
9919 ret = NULL; /* For look-ahead/behind. */
9922 case 'P': /* (?P...) variants for those used to PCRE/Python */
9923 paren = *RExC_parse++;
9924 if ( paren == '<') /* (?P<...>) named capture */
9926 else if (paren == '>') { /* (?P>name) named recursion */
9927 goto named_recursion;
9929 else if (paren == '=') { /* (?P=...) named backref */
9930 /* this pretty much dupes the code for \k<NAME> in
9931 * regatom(), if you change this make sure you change that
9933 char* name_start = RExC_parse;
9935 SV *sv_dat = reg_scan_name(pRExC_state,
9936 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9937 if (RExC_parse == name_start || *RExC_parse != ')')
9938 /* diag_listed_as: Sequence ?P=... not terminated in regex; marked by <-- HERE in m/%s/ */
9939 vFAIL2("Sequence %.3s... not terminated",parse_start);
9942 num = add_data( pRExC_state, STR_WITH_LEN("S"));
9943 RExC_rxi->data->data[num]=(void*)sv_dat;
9944 SvREFCNT_inc_simple_void(sv_dat);
9947 ret = reganode(pRExC_state,
9950 : (ASCII_FOLD_RESTRICTED)
9952 : (AT_LEAST_UNI_SEMANTICS)
9960 Set_Node_Offset(ret, parse_start+1);
9961 Set_Node_Cur_Length(ret, parse_start);
9963 nextchar(pRExC_state);
9967 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9968 vFAIL3("Sequence (%.*s...) not recognized",
9969 RExC_parse-seqstart, seqstart);
9970 NOT_REACHED; /*NOTREACHED*/
9971 case '<': /* (?<...) */
9972 if (*RExC_parse == '!')
9974 else if (*RExC_parse != '=')
9980 case '\'': /* (?'...') */
9981 name_start= RExC_parse;
9982 svname = reg_scan_name(pRExC_state,
9983 SIZE_ONLY /* reverse test from the others */
9984 ? REG_RSN_RETURN_NAME
9985 : REG_RSN_RETURN_NULL);
9986 if (RExC_parse == name_start || *RExC_parse != paren)
9987 vFAIL2("Sequence (?%c... not terminated",
9988 paren=='>' ? '<' : paren);
9992 if (!svname) /* shouldn't happen */
9994 "panic: reg_scan_name returned NULL");
9995 if (!RExC_paren_names) {
9996 RExC_paren_names= newHV();
9997 sv_2mortal(MUTABLE_SV(RExC_paren_names));
9999 RExC_paren_name_list= newAV();
10000 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
10003 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
10005 sv_dat = HeVAL(he_str);
10007 /* croak baby croak */
10009 "panic: paren_name hash element allocation failed");
10010 } else if ( SvPOK(sv_dat) ) {
10011 /* (?|...) can mean we have dupes so scan to check
10012 its already been stored. Maybe a flag indicating
10013 we are inside such a construct would be useful,
10014 but the arrays are likely to be quite small, so
10015 for now we punt -- dmq */
10016 IV count = SvIV(sv_dat);
10017 I32 *pv = (I32*)SvPVX(sv_dat);
10019 for ( i = 0 ; i < count ; i++ ) {
10020 if ( pv[i] == RExC_npar ) {
10026 pv = (I32*)SvGROW(sv_dat,
10027 SvCUR(sv_dat) + sizeof(I32)+1);
10028 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
10029 pv[count] = RExC_npar;
10030 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
10033 (void)SvUPGRADE(sv_dat,SVt_PVNV);
10034 sv_setpvn(sv_dat, (char *)&(RExC_npar),
10037 SvIV_set(sv_dat, 1);
10040 /* Yes this does cause a memory leak in debugging Perls
10042 if (!av_store(RExC_paren_name_list,
10043 RExC_npar, SvREFCNT_inc(svname)))
10044 SvREFCNT_dec_NN(svname);
10047 /*sv_dump(sv_dat);*/
10049 nextchar(pRExC_state);
10051 goto capturing_parens;
10053 RExC_seen |= REG_LOOKBEHIND_SEEN;
10054 RExC_in_lookbehind++;
10057 case '=': /* (?=...) */
10058 RExC_seen_zerolen++;
10060 case '!': /* (?!...) */
10061 RExC_seen_zerolen++;
10062 /* check if we're really just a "FAIL" assertion */
10064 nextchar(pRExC_state);
10065 if (*RExC_parse == ')') {
10066 ret=reg_node(pRExC_state, OPFAIL);
10067 nextchar(pRExC_state);
10071 case '|': /* (?|...) */
10072 /* branch reset, behave like a (?:...) except that
10073 buffers in alternations share the same numbers */
10075 after_freeze = freeze_paren = RExC_npar;
10077 case ':': /* (?:...) */
10078 case '>': /* (?>...) */
10080 case '$': /* (?$...) */
10081 case '@': /* (?@...) */
10082 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
10084 case '0' : /* (?0) */
10085 case 'R' : /* (?R) */
10086 if (*RExC_parse != ')')
10087 FAIL("Sequence (?R) not terminated");
10088 ret = reg_node(pRExC_state, GOSTART);
10089 RExC_seen |= REG_GOSTART_SEEN;
10090 *flagp |= POSTPONED;
10091 nextchar(pRExC_state);
10094 /* named and numeric backreferences */
10095 case '&': /* (?&NAME) */
10096 parse_start = RExC_parse - 1;
10099 SV *sv_dat = reg_scan_name(pRExC_state,
10100 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10101 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10103 if (RExC_parse == RExC_end || *RExC_parse != ')')
10104 vFAIL("Sequence (?&... not terminated");
10105 goto gen_recurse_regop;
10108 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10110 vFAIL("Illegal pattern");
10112 goto parse_recursion;
10114 case '-': /* (?-1) */
10115 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10116 RExC_parse--; /* rewind to let it be handled later */
10120 case '1': case '2': case '3': case '4': /* (?1) */
10121 case '5': case '6': case '7': case '8': case '9':
10125 bool is_neg = FALSE;
10126 parse_start = RExC_parse - 1; /* MJD */
10127 if (*RExC_parse == '-') {
10131 num = grok_atou(RExC_parse, &endptr);
10133 RExC_parse = (char*)endptr;
10135 /* Some limit for num? */
10139 if (*RExC_parse!=')')
10140 vFAIL("Expecting close bracket");
10143 if ( paren == '-' ) {
10145 Diagram of capture buffer numbering.
10146 Top line is the normal capture buffer numbers
10147 Bottom line is the negative indexing as from
10151 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
10155 num = RExC_npar + num;
10158 vFAIL("Reference to nonexistent group");
10160 } else if ( paren == '+' ) {
10161 num = RExC_npar + num - 1;
10164 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
10166 if (num > (I32)RExC_rx->nparens) {
10168 vFAIL("Reference to nonexistent group");
10170 RExC_recurse_count++;
10171 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10172 "%*s%*s Recurse #%"UVuf" to %"IVdf"\n",
10173 22, "| |", (int)(depth * 2 + 1), "",
10174 (UV)ARG(ret), (IV)ARG2L(ret)));
10176 RExC_seen |= REG_RECURSE_SEEN;
10177 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
10178 Set_Node_Offset(ret, parse_start); /* MJD */
10180 *flagp |= POSTPONED;
10181 nextchar(pRExC_state);
10186 case '?': /* (??...) */
10188 if (*RExC_parse != '{') {
10190 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10192 "Sequence (%"UTF8f"...) not recognized",
10193 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10194 NOT_REACHED; /*NOTREACHED*/
10196 *flagp |= POSTPONED;
10197 paren = *RExC_parse++;
10199 case '{': /* (?{...}) */
10202 struct reg_code_block *cb;
10204 RExC_seen_zerolen++;
10206 if ( !pRExC_state->num_code_blocks
10207 || pRExC_state->code_index >= pRExC_state->num_code_blocks
10208 || pRExC_state->code_blocks[pRExC_state->code_index].start
10209 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
10212 if (RExC_pm_flags & PMf_USE_RE_EVAL)
10213 FAIL("panic: Sequence (?{...}): no code block found\n");
10214 FAIL("Eval-group not allowed at runtime, use re 'eval'");
10216 /* this is a pre-compiled code block (?{...}) */
10217 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
10218 RExC_parse = RExC_start + cb->end;
10221 if (cb->src_regex) {
10222 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
10223 RExC_rxi->data->data[n] =
10224 (void*)SvREFCNT_inc((SV*)cb->src_regex);
10225 RExC_rxi->data->data[n+1] = (void*)o;
10228 n = add_data(pRExC_state,
10229 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
10230 RExC_rxi->data->data[n] = (void*)o;
10233 pRExC_state->code_index++;
10234 nextchar(pRExC_state);
10238 ret = reg_node(pRExC_state, LOGICAL);
10240 eval = reg2Lanode(pRExC_state, EVAL,
10243 /* for later propagation into (??{})
10245 RExC_flags & RXf_PMf_COMPILETIME
10250 REGTAIL(pRExC_state, ret, eval);
10251 /* deal with the length of this later - MJD */
10254 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
10255 Set_Node_Length(ret, RExC_parse - parse_start + 1);
10256 Set_Node_Offset(ret, parse_start);
10259 case '(': /* (?(?{...})...) and (?(?=...)...) */
10262 const int DEFINE_len = sizeof("DEFINE") - 1;
10263 if (RExC_parse[0] == '?') { /* (?(?...)) */
10264 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
10265 || RExC_parse[1] == '<'
10266 || RExC_parse[1] == '{') { /* Lookahead or eval. */
10270 ret = reg_node(pRExC_state, LOGICAL);
10274 tail = reg(pRExC_state, 1, &flag, depth+1);
10275 if (flag & RESTART_UTF8) {
10276 *flagp = RESTART_UTF8;
10279 REGTAIL(pRExC_state, ret, tail);
10282 /* Fall through to ‘Unknown switch condition’ at the
10283 end of the if/else chain. */
10285 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
10286 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
10288 char ch = RExC_parse[0] == '<' ? '>' : '\'';
10289 char *name_start= RExC_parse++;
10291 SV *sv_dat=reg_scan_name(pRExC_state,
10292 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10293 if (RExC_parse == name_start || *RExC_parse != ch)
10294 vFAIL2("Sequence (?(%c... not terminated",
10295 (ch == '>' ? '<' : ch));
10298 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10299 RExC_rxi->data->data[num]=(void*)sv_dat;
10300 SvREFCNT_inc_simple_void(sv_dat);
10302 ret = reganode(pRExC_state,NGROUPP,num);
10303 goto insert_if_check_paren;
10305 else if (strnEQ(RExC_parse, "DEFINE",
10306 MIN(DEFINE_len, RExC_end - RExC_parse)))
10308 ret = reganode(pRExC_state,DEFINEP,0);
10309 RExC_parse += DEFINE_len;
10311 goto insert_if_check_paren;
10313 else if (RExC_parse[0] == 'R') {
10316 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
10317 parno = grok_atou(RExC_parse, &endptr);
10319 RExC_parse = (char*)endptr;
10320 } else if (RExC_parse[0] == '&') {
10323 sv_dat = reg_scan_name(pRExC_state,
10325 ? REG_RSN_RETURN_NULL
10326 : REG_RSN_RETURN_DATA);
10327 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10329 ret = reganode(pRExC_state,INSUBP,parno);
10330 goto insert_if_check_paren;
10332 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
10336 parno = grok_atou(RExC_parse, &endptr);
10338 RExC_parse = (char*)endptr;
10339 ret = reganode(pRExC_state, GROUPP, parno);
10341 insert_if_check_paren:
10342 if (*(tmp = nextchar(pRExC_state)) != ')') {
10343 /* nextchar also skips comments, so undo its work
10344 * and skip over the the next character.
10347 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10348 vFAIL("Switch condition not recognized");
10351 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
10352 br = regbranch(pRExC_state, &flags, 1,depth+1);
10354 if (flags & RESTART_UTF8) {
10355 *flagp = RESTART_UTF8;
10358 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
10361 REGTAIL(pRExC_state, br, reganode(pRExC_state,
10363 c = *nextchar(pRExC_state);
10364 if (flags&HASWIDTH)
10365 *flagp |= HASWIDTH;
10368 vFAIL("(?(DEFINE)....) does not allow branches");
10370 /* Fake one for optimizer. */
10371 lastbr = reganode(pRExC_state, IFTHEN, 0);
10373 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
10374 if (flags & RESTART_UTF8) {
10375 *flagp = RESTART_UTF8;
10378 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
10381 REGTAIL(pRExC_state, ret, lastbr);
10382 if (flags&HASWIDTH)
10383 *flagp |= HASWIDTH;
10384 c = *nextchar(pRExC_state);
10389 if (RExC_parse>RExC_end)
10390 vFAIL("Switch (?(condition)... not terminated");
10392 vFAIL("Switch (?(condition)... contains too many branches");
10394 ender = reg_node(pRExC_state, TAIL);
10395 REGTAIL(pRExC_state, br, ender);
10397 REGTAIL(pRExC_state, lastbr, ender);
10398 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10401 REGTAIL(pRExC_state, ret, ender);
10402 RExC_size++; /* XXX WHY do we need this?!!
10403 For large programs it seems to be required
10404 but I can't figure out why. -- dmq*/
10407 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10408 vFAIL("Unknown switch condition (?(...))");
10410 case '[': /* (?[ ... ]) */
10411 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
10414 RExC_parse--; /* for vFAIL to print correctly */
10415 vFAIL("Sequence (? incomplete");
10417 default: /* e.g., (?i) */
10420 parse_lparen_question_flags(pRExC_state);
10421 if (UCHARAT(RExC_parse) != ':') {
10422 nextchar(pRExC_state);
10427 nextchar(pRExC_state);
10432 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
10437 ret = reganode(pRExC_state, OPEN, parno);
10439 if (!RExC_nestroot)
10440 RExC_nestroot = parno;
10441 if (RExC_seen & REG_RECURSE_SEEN
10442 && !RExC_open_parens[parno-1])
10444 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10445 "%*s%*s Setting open paren #%"IVdf" to %d\n",
10446 22, "| |", (int)(depth * 2 + 1), "",
10447 (IV)parno, REG_NODE_NUM(ret)));
10448 RExC_open_parens[parno-1]= ret;
10451 Set_Node_Length(ret, 1); /* MJD */
10452 Set_Node_Offset(ret, RExC_parse); /* MJD */
10462 /* Pick up the branches, linking them together. */
10463 parse_start = RExC_parse; /* MJD */
10464 br = regbranch(pRExC_state, &flags, 1,depth+1);
10466 /* branch_len = (paren != 0); */
10469 if (flags & RESTART_UTF8) {
10470 *flagp = RESTART_UTF8;
10473 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10475 if (*RExC_parse == '|') {
10476 if (!SIZE_ONLY && RExC_extralen) {
10477 reginsert(pRExC_state, BRANCHJ, br, depth+1);
10480 reginsert(pRExC_state, BRANCH, br, depth+1);
10481 Set_Node_Length(br, paren != 0);
10482 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
10486 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
10488 else if (paren == ':') {
10489 *flagp |= flags&SIMPLE;
10491 if (is_open) { /* Starts with OPEN. */
10492 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
10494 else if (paren != '?') /* Not Conditional */
10496 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10498 while (*RExC_parse == '|') {
10499 if (!SIZE_ONLY && RExC_extralen) {
10500 ender = reganode(pRExC_state, LONGJMP,0);
10502 /* Append to the previous. */
10503 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10506 RExC_extralen += 2; /* Account for LONGJMP. */
10507 nextchar(pRExC_state);
10508 if (freeze_paren) {
10509 if (RExC_npar > after_freeze)
10510 after_freeze = RExC_npar;
10511 RExC_npar = freeze_paren;
10513 br = regbranch(pRExC_state, &flags, 0, depth+1);
10516 if (flags & RESTART_UTF8) {
10517 *flagp = RESTART_UTF8;
10520 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10522 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
10524 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10527 if (have_branch || paren != ':') {
10528 /* Make a closing node, and hook it on the end. */
10531 ender = reg_node(pRExC_state, TAIL);
10534 ender = reganode(pRExC_state, CLOSE, parno);
10535 if (!SIZE_ONLY && RExC_seen & REG_RECURSE_SEEN) {
10536 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10537 "%*s%*s Setting close paren #%"IVdf" to %d\n",
10538 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
10539 RExC_close_parens[parno-1]= ender;
10540 if (RExC_nestroot == parno)
10543 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
10544 Set_Node_Length(ender,1); /* MJD */
10550 *flagp &= ~HASWIDTH;
10553 ender = reg_node(pRExC_state, SUCCEED);
10556 ender = reg_node(pRExC_state, END);
10558 assert(!RExC_opend); /* there can only be one! */
10559 RExC_opend = ender;
10563 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10564 DEBUG_PARSE_MSG("lsbr");
10565 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
10566 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
10567 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10568 SvPV_nolen_const(RExC_mysv1),
10569 (IV)REG_NODE_NUM(lastbr),
10570 SvPV_nolen_const(RExC_mysv2),
10571 (IV)REG_NODE_NUM(ender),
10572 (IV)(ender - lastbr)
10575 REGTAIL(pRExC_state, lastbr, ender);
10577 if (have_branch && !SIZE_ONLY) {
10578 char is_nothing= 1;
10580 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
10582 /* Hook the tails of the branches to the closing node. */
10583 for (br = ret; br; br = regnext(br)) {
10584 const U8 op = PL_regkind[OP(br)];
10585 if (op == BRANCH) {
10586 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
10587 if ( OP(NEXTOPER(br)) != NOTHING
10588 || regnext(NEXTOPER(br)) != ender)
10591 else if (op == BRANCHJ) {
10592 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
10593 /* for now we always disable this optimisation * /
10594 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
10595 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
10601 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
10602 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10603 DEBUG_PARSE_MSG("NADA");
10604 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
10605 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
10606 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10607 SvPV_nolen_const(RExC_mysv1),
10608 (IV)REG_NODE_NUM(ret),
10609 SvPV_nolen_const(RExC_mysv2),
10610 (IV)REG_NODE_NUM(ender),
10615 if (OP(ender) == TAIL) {
10620 for ( opt= br + 1; opt < ender ; opt++ )
10621 OP(opt)= OPTIMIZED;
10622 NEXT_OFF(br)= ender - br;
10630 static const char parens[] = "=!<,>";
10632 if (paren && (p = strchr(parens, paren))) {
10633 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
10634 int flag = (p - parens) > 1;
10637 node = SUSPEND, flag = 0;
10638 reginsert(pRExC_state, node,ret, depth+1);
10639 Set_Node_Cur_Length(ret, parse_start);
10640 Set_Node_Offset(ret, parse_start + 1);
10642 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
10646 /* Check for proper termination. */
10648 /* restore original flags, but keep (?p) */
10649 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
10650 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
10651 RExC_parse = oregcomp_parse;
10652 vFAIL("Unmatched (");
10655 else if (!paren && RExC_parse < RExC_end) {
10656 if (*RExC_parse == ')') {
10658 vFAIL("Unmatched )");
10661 FAIL("Junk on end of regexp"); /* "Can't happen". */
10662 NOT_REACHED; /* NOTREACHED */
10665 if (RExC_in_lookbehind) {
10666 RExC_in_lookbehind--;
10668 if (after_freeze > RExC_npar)
10669 RExC_npar = after_freeze;
10674 - regbranch - one alternative of an | operator
10676 * Implements the concatenation operator.
10678 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10682 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
10685 regnode *chain = NULL;
10687 I32 flags = 0, c = 0;
10688 GET_RE_DEBUG_FLAGS_DECL;
10690 PERL_ARGS_ASSERT_REGBRANCH;
10692 DEBUG_PARSE("brnc");
10697 if (!SIZE_ONLY && RExC_extralen)
10698 ret = reganode(pRExC_state, BRANCHJ,0);
10700 ret = reg_node(pRExC_state, BRANCH);
10701 Set_Node_Length(ret, 1);
10705 if (!first && SIZE_ONLY)
10706 RExC_extralen += 1; /* BRANCHJ */
10708 *flagp = WORST; /* Tentatively. */
10711 nextchar(pRExC_state);
10712 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
10713 flags &= ~TRYAGAIN;
10714 latest = regpiece(pRExC_state, &flags,depth+1);
10715 if (latest == NULL) {
10716 if (flags & TRYAGAIN)
10718 if (flags & RESTART_UTF8) {
10719 *flagp = RESTART_UTF8;
10722 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
10724 else if (ret == NULL)
10726 *flagp |= flags&(HASWIDTH|POSTPONED);
10727 if (chain == NULL) /* First piece. */
10728 *flagp |= flags&SPSTART;
10730 /* FIXME adding one for every branch after the first is probably
10731 * excessive now we have TRIE support. (hv) */
10733 REGTAIL(pRExC_state, chain, latest);
10738 if (chain == NULL) { /* Loop ran zero times. */
10739 chain = reg_node(pRExC_state, NOTHING);
10744 *flagp |= flags&SIMPLE;
10751 - regpiece - something followed by possible [*+?]
10753 * Note that the branching code sequences used for ? and the general cases
10754 * of * and + are somewhat optimized: they use the same NOTHING node as
10755 * both the endmarker for their branch list and the body of the last branch.
10756 * It might seem that this node could be dispensed with entirely, but the
10757 * endmarker role is not redundant.
10759 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
10761 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10765 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10771 const char * const origparse = RExC_parse;
10773 I32 max = REG_INFTY;
10774 #ifdef RE_TRACK_PATTERN_OFFSETS
10777 const char *maxpos = NULL;
10779 /* Save the original in case we change the emitted regop to a FAIL. */
10780 regnode * const orig_emit = RExC_emit;
10782 GET_RE_DEBUG_FLAGS_DECL;
10784 PERL_ARGS_ASSERT_REGPIECE;
10786 DEBUG_PARSE("piec");
10788 ret = regatom(pRExC_state, &flags,depth+1);
10790 if (flags & (TRYAGAIN|RESTART_UTF8))
10791 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
10793 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
10799 if (op == '{' && regcurly(RExC_parse)) {
10801 #ifdef RE_TRACK_PATTERN_OFFSETS
10802 parse_start = RExC_parse; /* MJD */
10804 next = RExC_parse + 1;
10805 while (isDIGIT(*next) || *next == ',') {
10806 if (*next == ',') {
10814 if (*next == '}') { /* got one */
10815 const char* endptr;
10819 min = grok_atou(RExC_parse, &endptr);
10820 if (*maxpos == ',')
10823 maxpos = RExC_parse;
10824 max = grok_atou(maxpos, &endptr);
10825 if (!max && *maxpos != '0')
10826 max = REG_INFTY; /* meaning "infinity" */
10827 else if (max >= REG_INFTY)
10828 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
10830 nextchar(pRExC_state);
10831 if (max < min) { /* If can't match, warn and optimize to fail
10835 /* We can't back off the size because we have to reserve
10836 * enough space for all the things we are about to throw
10837 * away, but we can shrink it by the ammount we are about
10838 * to re-use here */
10839 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
10842 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
10843 RExC_emit = orig_emit;
10845 ret = reg_node(pRExC_state, OPFAIL);
10848 else if (min == max
10849 && RExC_parse < RExC_end
10850 && (*RExC_parse == '?' || *RExC_parse == '+'))
10853 ckWARN2reg(RExC_parse + 1,
10854 "Useless use of greediness modifier '%c'",
10857 /* Absorb the modifier, so later code doesn't see nor use
10859 nextchar(pRExC_state);
10863 if ((flags&SIMPLE)) {
10864 MARK_NAUGHTY_EXP(2, 2);
10865 reginsert(pRExC_state, CURLY, ret, depth+1);
10866 Set_Node_Offset(ret, parse_start+1); /* MJD */
10867 Set_Node_Cur_Length(ret, parse_start);
10870 regnode * const w = reg_node(pRExC_state, WHILEM);
10873 REGTAIL(pRExC_state, ret, w);
10874 if (!SIZE_ONLY && RExC_extralen) {
10875 reginsert(pRExC_state, LONGJMP,ret, depth+1);
10876 reginsert(pRExC_state, NOTHING,ret, depth+1);
10877 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
10879 reginsert(pRExC_state, CURLYX,ret, depth+1);
10881 Set_Node_Offset(ret, parse_start+1);
10882 Set_Node_Length(ret,
10883 op == '{' ? (RExC_parse - parse_start) : 1);
10885 if (!SIZE_ONLY && RExC_extralen)
10886 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
10887 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
10889 RExC_whilem_seen++, RExC_extralen += 3;
10890 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
10897 *flagp |= HASWIDTH;
10899 ARG1_SET(ret, (U16)min);
10900 ARG2_SET(ret, (U16)max);
10902 if (max == REG_INFTY)
10903 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10909 if (!ISMULT1(op)) {
10914 #if 0 /* Now runtime fix should be reliable. */
10916 /* if this is reinstated, don't forget to put this back into perldiag:
10918 =item Regexp *+ operand could be empty at {#} in regex m/%s/
10920 (F) The part of the regexp subject to either the * or + quantifier
10921 could match an empty string. The {#} shows in the regular
10922 expression about where the problem was discovered.
10926 if (!(flags&HASWIDTH) && op != '?')
10927 vFAIL("Regexp *+ operand could be empty");
10930 #ifdef RE_TRACK_PATTERN_OFFSETS
10931 parse_start = RExC_parse;
10933 nextchar(pRExC_state);
10935 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
10937 if (op == '*' && (flags&SIMPLE)) {
10938 reginsert(pRExC_state, STAR, ret, depth+1);
10941 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10943 else if (op == '*') {
10947 else if (op == '+' && (flags&SIMPLE)) {
10948 reginsert(pRExC_state, PLUS, ret, depth+1);
10951 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10953 else if (op == '+') {
10957 else if (op == '?') {
10962 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
10963 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
10964 ckWARN2reg(RExC_parse,
10965 "%"UTF8f" matches null string many times",
10966 UTF8fARG(UTF, (RExC_parse >= origparse
10967 ? RExC_parse - origparse
10970 (void)ReREFCNT_inc(RExC_rx_sv);
10973 if (RExC_parse < RExC_end && *RExC_parse == '?') {
10974 nextchar(pRExC_state);
10975 reginsert(pRExC_state, MINMOD, ret, depth+1);
10976 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
10979 if (RExC_parse < RExC_end && *RExC_parse == '+') {
10981 nextchar(pRExC_state);
10982 ender = reg_node(pRExC_state, SUCCEED);
10983 REGTAIL(pRExC_state, ret, ender);
10984 reginsert(pRExC_state, SUSPEND, ret, depth+1);
10986 ender = reg_node(pRExC_state, TAIL);
10987 REGTAIL(pRExC_state, ret, ender);
10990 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
10992 vFAIL("Nested quantifiers");
10999 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p,
11000 UV *valuep, I32 *flagp, U32 depth, SV** substitute_parse
11004 /* This is expected to be called by a parser routine that has recognized '\N'
11005 and needs to handle the rest. RExC_parse is expected to point at the first
11006 char following the N at the time of the call. On successful return,
11007 RExC_parse has been updated to point to just after the sequence identified
11008 by this routine, <*flagp> has been updated, and the non-NULL input pointers
11009 have been set appropriately.
11011 The typical case for this is \N{some character name}. This is usually
11012 called while parsing the input, filling in or ready to fill in an EXACTish
11013 node, and the code point for the character should be returned, so that it
11014 can be added to the node, and parsing continued with the next input
11015 character. But it may be that instead of a single character the \N{}
11016 expands to more than one, a named sequence. In this case any following
11017 quantifier applies to the whole sequence, and it is easier, given the code
11018 structure that calls this, to handle it from a different area of the code.
11019 For this reason, the input parameters can be set so that it returns valid
11020 only on one or the other of these cases.
11022 Another possibility is for the input to be an empty \N{}, which for
11023 backwards compatibility we accept, but generate a NOTHING node which should
11024 later get optimized out. This is handled from the area of code which can
11025 handle a named sequence, so if called with the parameters for the other, it
11028 Still another possibility is for the \N to mean [^\n], and not a single
11029 character or explicit sequence at all. This is determined by context.
11030 Again, this is handled from the area of code which can handle a named
11031 sequence, so if called with the parameters for the other, it also fails.
11033 And the final possibility is for the \N to be called from within a bracketed
11034 character class. In this case the [^\n] meaning makes no sense, and so is
11035 an error. Other anomalous situations are left to the calling code to handle.
11037 For non-single-quoted regexes, the tokenizer has attempted to decide which
11038 of the above applies, and in the case of a named sequence, has converted it
11039 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
11040 where c1... are the characters in the sequence. For single-quoted regexes,
11041 the tokenizer passes the \N sequence through unchanged; this code will not
11042 attempt to determine this nor expand those, instead raising a syntax error.
11043 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
11044 or there is no '}', it signals that this \N occurrence means to match a
11045 non-newline. (This mostly was done because of [perl #56444].)
11047 The API is somewhat convoluted due to historical and the above reasons.
11049 The function raises an error (via vFAIL), and doesn't return for various
11050 syntax errors. For other failures, it returns (STRLEN) -1. For successes,
11051 it returns a count of how many characters were accounted for by it. (This
11052 can be 0 for \N{}; 1 for it meaning [^\n]; and otherwise the number of code
11053 points in the sequence. It sets <node_p>, <valuep>, and/or
11054 <substitute_parse> on success.
11056 If <valuep> is non-null, it means the caller can accept an input sequence
11057 consisting of just a single code point; <*valuep> is set to the value of the
11058 only or first code point in the input.
11060 If <substitute_parse> is non-null, it means the caller can accept an input
11061 sequence consisting of one or more code points; <*substitute_parse> is a
11062 newly created mortal SV* in this case, containing \x{} escapes representing
11065 Both <valuep> and <substitute_parse> can be non-NULL.
11067 If <node_p> is non-null, <substitute_parse> must be NULL. This signifies
11068 that the caller can accept any legal sequence other than a single code
11069 point. To wit, <*node_p> is set as follows:
11070 1) \N means not-a-NL: points to a newly created REG_ANY node; return is 1
11071 2) \N{}: points to a new NOTHING node; return is 0
11072 3) otherwise: points to a new EXACT node containing the resolved
11073 string; return is the number of code points in the
11074 string. This will never be 1.
11075 Note that failure is returned for single code point sequences if <valuep> is
11076 null and <node_p> is not.
11079 char * endbrace; /* '}' following the name */
11081 char *endchar; /* Points to '.' or '}' ending cur char in the input
11083 bool has_multiple_chars; /* true if the input stream contains a sequence of
11084 more than one character */
11085 bool in_char_class = substitute_parse != NULL;
11086 STRLEN count = 0; /* Number of characters in this sequence */
11088 GET_RE_DEBUG_FLAGS_DECL;
11090 PERL_ARGS_ASSERT_GROK_BSLASH_N;
11092 GET_RE_DEBUG_FLAGS;
11094 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
11095 assert(! (node_p && substitute_parse)); /* At most 1 should be set */
11097 /* The [^\n] meaning of \N ignores spaces and comments under the /x
11098 * modifier. The other meaning does not, so use a temporary until we find
11099 * out which we are being called with */
11100 p = (RExC_flags & RXf_PMf_EXTENDED)
11101 ? regpatws(pRExC_state, RExC_parse,
11102 TRUE) /* means recognize comments */
11105 /* Disambiguate between \N meaning a named character versus \N meaning
11106 * [^\n]. The former is assumed when it can't be the latter. */
11107 if (*p != '{' || regcurly(p)) {
11110 /* no bare \N allowed in a charclass */
11111 if (in_char_class) {
11112 vFAIL("\\N in a character class must be a named character: \\N{...}");
11114 return (STRLEN) -1;
11116 RExC_parse--; /* Need to back off so nextchar() doesn't skip the
11118 nextchar(pRExC_state);
11119 *node_p = reg_node(pRExC_state, REG_ANY);
11120 *flagp |= HASWIDTH|SIMPLE;
11122 Set_Node_Length(*node_p, 1); /* MJD */
11126 /* Here, we have decided it should be a named character or sequence */
11128 /* The test above made sure that the next real character is a '{', but
11129 * under the /x modifier, it could be separated by space (or a comment and
11130 * \n) and this is not allowed (for consistency with \x{...} and the
11131 * tokenizer handling of \N{NAME}). */
11132 if (*RExC_parse != '{') {
11133 vFAIL("Missing braces on \\N{}");
11136 RExC_parse++; /* Skip past the '{' */
11138 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
11139 || ! (endbrace == RExC_parse /* nothing between the {} */
11140 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked... */
11141 && strnEQ(RExC_parse, "U+", 2)))) /* ... below for a better
11144 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
11145 vFAIL("\\N{NAME} must be resolved by the lexer");
11148 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
11150 if (endbrace == RExC_parse) { /* empty: \N{} */
11152 *node_p = reg_node(pRExC_state,NOTHING);
11154 else if (! in_char_class) {
11155 return (STRLEN) -1;
11157 nextchar(pRExC_state);
11161 RExC_parse += 2; /* Skip past the 'U+' */
11163 endchar = RExC_parse + strcspn(RExC_parse, ".}");
11165 /* Code points are separated by dots. If none, there is only one code
11166 * point, and is terminated by the brace */
11167 has_multiple_chars = (endchar < endbrace);
11169 /* We get the first code point if we want it, and either there is only one,
11170 * or we can accept both cases of one and there is more than one */
11171 if (valuep && (substitute_parse || ! has_multiple_chars)) {
11172 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
11173 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
11174 | PERL_SCAN_DISALLOW_PREFIX
11176 /* No errors in the first pass (See [perl
11177 * #122671].) We let the code below find the
11178 * errors when there are multiple chars. */
11179 | ((SIZE_ONLY || has_multiple_chars)
11180 ? PERL_SCAN_SILENT_ILLDIGIT
11183 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
11185 /* The tokenizer should have guaranteed validity, but it's possible to
11186 * bypass it by using single quoting, so check. Don't do the check
11187 * here when there are multiple chars; we do it below anyway. */
11188 if (! has_multiple_chars) {
11189 if (length_of_hex == 0
11190 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
11192 RExC_parse += length_of_hex; /* Includes all the valid */
11193 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
11194 ? UTF8SKIP(RExC_parse)
11196 /* Guard against malformed utf8 */
11197 if (RExC_parse >= endchar) {
11198 RExC_parse = endchar;
11200 vFAIL("Invalid hexadecimal number in \\N{U+...}");
11203 RExC_parse = endbrace + 1;
11208 /* Here, we should have already handled the case where a single character
11209 * is expected and found. So it is a failure if we aren't expecting
11210 * multiple chars and got them; or didn't get them but wanted them. We
11211 * fail without advancing the parse, so that the caller can try again with
11212 * different acceptance criteria */
11213 if ((! node_p && ! substitute_parse) || ! has_multiple_chars) {
11215 return (STRLEN) -1;
11219 /* What is done here is to convert this to a sub-pattern of the form
11220 * \x{char1}\x{char2}...
11221 * and then either return it in <*substitute_parse> if non-null; or
11222 * call reg recursively to parse it (enclosing in "(?: ... )" ). That
11223 * way, it retains its atomicness, while not having to worry about
11224 * special handling that some code points may have. toke.c has
11225 * converted the original Unicode values to native, so that we can just
11226 * pass on the hex values unchanged. We do have to set a flag to keep
11227 * recoding from happening in the recursion */
11231 char *orig_end = RExC_end;
11234 if (substitute_parse) {
11235 *substitute_parse = newSVpvs("");
11238 substitute_parse = &dummy;
11239 *substitute_parse = newSVpvs("?:");
11241 *substitute_parse = sv_2mortal(*substitute_parse);
11243 while (RExC_parse < endbrace) {
11245 /* Convert to notation the rest of the code understands */
11246 sv_catpv(*substitute_parse, "\\x{");
11247 sv_catpvn(*substitute_parse, RExC_parse, endchar - RExC_parse);
11248 sv_catpv(*substitute_parse, "}");
11250 /* Point to the beginning of the next character in the sequence. */
11251 RExC_parse = endchar + 1;
11252 endchar = RExC_parse + strcspn(RExC_parse, ".}");
11256 if (! in_char_class) {
11257 sv_catpv(*substitute_parse, ")");
11260 RExC_parse = SvPV(*substitute_parse, len);
11262 /* Don't allow empty number */
11263 if (len < (STRLEN) ((substitute_parse) ? 6 : 8)) {
11264 RExC_parse = endbrace;
11265 vFAIL("Invalid hexadecimal number in \\N{U+...}");
11267 RExC_end = RExC_parse + len;
11269 /* The values are Unicode, and therefore not subject to recoding */
11270 RExC_override_recoding = 1;
11273 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
11274 if (flags & RESTART_UTF8) {
11275 *flagp = RESTART_UTF8;
11276 return (STRLEN) -1;
11278 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
11281 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
11284 RExC_parse = endbrace;
11285 RExC_end = orig_end;
11286 RExC_override_recoding = 0;
11288 nextchar(pRExC_state);
11298 * It returns the code point in utf8 for the value in *encp.
11299 * value: a code value in the source encoding
11300 * encp: a pointer to an Encode object
11302 * If the result from Encode is not a single character,
11303 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
11306 S_reg_recode(pTHX_ const char value, SV **encp)
11309 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
11310 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
11311 const STRLEN newlen = SvCUR(sv);
11312 UV uv = UNICODE_REPLACEMENT;
11314 PERL_ARGS_ASSERT_REG_RECODE;
11318 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
11321 if (!newlen || numlen != newlen) {
11322 uv = UNICODE_REPLACEMENT;
11328 PERL_STATIC_INLINE U8
11329 S_compute_EXACTish(RExC_state_t *pRExC_state)
11333 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
11341 op = get_regex_charset(RExC_flags);
11342 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
11343 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
11344 been, so there is no hole */
11347 return op + EXACTF;
11350 PERL_STATIC_INLINE void
11351 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
11352 regnode *node, I32* flagp, STRLEN len, UV code_point,
11355 /* This knows the details about sizing an EXACTish node, setting flags for
11356 * it (by setting <*flagp>, and potentially populating it with a single
11359 * If <len> (the length in bytes) is non-zero, this function assumes that
11360 * the node has already been populated, and just does the sizing. In this
11361 * case <code_point> should be the final code point that has already been
11362 * placed into the node. This value will be ignored except that under some
11363 * circumstances <*flagp> is set based on it.
11365 * If <len> is zero, the function assumes that the node is to contain only
11366 * the single character given by <code_point> and calculates what <len>
11367 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
11368 * additionally will populate the node's STRING with <code_point> or its
11371 * In both cases <*flagp> is appropriately set
11373 * It knows that under FOLD, the Latin Sharp S and UTF characters above
11374 * 255, must be folded (the former only when the rules indicate it can
11377 * When it does the populating, it looks at the flag 'downgradable'. If
11378 * true with a node that folds, it checks if the single code point
11379 * participates in a fold, and if not downgrades the node to an EXACT.
11380 * This helps the optimizer */
11382 bool len_passed_in = cBOOL(len != 0);
11383 U8 character[UTF8_MAXBYTES_CASE+1];
11385 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
11387 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
11388 * sizing difference, and is extra work that is thrown away */
11389 if (downgradable && ! PASS2) {
11390 downgradable = FALSE;
11393 if (! len_passed_in) {
11395 if (UVCHR_IS_INVARIANT(code_point)) {
11396 if (LOC || ! FOLD) { /* /l defers folding until runtime */
11397 *character = (U8) code_point;
11399 else { /* Here is /i and not /l. (toFOLD() is defined on just
11400 ASCII, which isn't the same thing as INVARIANT on
11401 EBCDIC, but it works there, as the extra invariants
11402 fold to themselves) */
11403 *character = toFOLD((U8) code_point);
11405 /* We can downgrade to an EXACT node if this character
11406 * isn't a folding one. Note that this assumes that
11407 * nothing above Latin1 folds to some other invariant than
11408 * one of these alphabetics; otherwise we would also have
11410 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
11411 * || ASCII_FOLD_RESTRICTED))
11413 if (downgradable && PL_fold[code_point] == code_point) {
11419 else if (FOLD && (! LOC
11420 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
11421 { /* Folding, and ok to do so now */
11422 UV folded = _to_uni_fold_flags(
11426 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
11427 ? FOLD_FLAGS_NOMIX_ASCII
11430 && folded == code_point /* This quickly rules out many
11431 cases, avoiding the
11432 _invlist_contains_cp() overhead
11434 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
11441 else if (code_point <= MAX_UTF8_TWO_BYTE) {
11443 /* Not folding this cp, and can output it directly */
11444 *character = UTF8_TWO_BYTE_HI(code_point);
11445 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
11449 uvchr_to_utf8( character, code_point);
11450 len = UTF8SKIP(character);
11452 } /* Else pattern isn't UTF8. */
11454 *character = (U8) code_point;
11456 } /* Else is folded non-UTF8 */
11457 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
11459 /* We don't fold any non-UTF8 except possibly the Sharp s (see
11460 * comments at join_exact()); */
11461 *character = (U8) code_point;
11464 /* Can turn into an EXACT node if we know the fold at compile time,
11465 * and it folds to itself and doesn't particpate in other folds */
11468 && PL_fold_latin1[code_point] == code_point
11469 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
11470 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
11474 } /* else is Sharp s. May need to fold it */
11475 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
11477 *(character + 1) = 's';
11481 *character = LATIN_SMALL_LETTER_SHARP_S;
11487 RExC_size += STR_SZ(len);
11490 RExC_emit += STR_SZ(len);
11491 STR_LEN(node) = len;
11492 if (! len_passed_in) {
11493 Copy((char *) character, STRING(node), len, char);
11497 *flagp |= HASWIDTH;
11499 /* A single character node is SIMPLE, except for the special-cased SHARP S
11501 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
11502 && (code_point != LATIN_SMALL_LETTER_SHARP_S
11503 || ! FOLD || ! DEPENDS_SEMANTICS))
11508 /* The OP may not be well defined in PASS1 */
11509 if (PASS2 && OP(node) == EXACTFL) {
11510 RExC_contains_locale = 1;
11515 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
11516 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
11519 S_backref_value(char *p)
11521 const char* endptr;
11522 UV val = grok_atou(p, &endptr);
11523 if (endptr == p || endptr == NULL || val > I32_MAX)
11530 - regatom - the lowest level
11532 Try to identify anything special at the start of the pattern. If there
11533 is, then handle it as required. This may involve generating a single regop,
11534 such as for an assertion; or it may involve recursing, such as to
11535 handle a () structure.
11537 If the string doesn't start with something special then we gobble up
11538 as much literal text as we can.
11540 Once we have been able to handle whatever type of thing started the
11541 sequence, we return.
11543 Note: we have to be careful with escapes, as they can be both literal
11544 and special, and in the case of \10 and friends, context determines which.
11546 A summary of the code structure is:
11548 switch (first_byte) {
11549 cases for each special:
11550 handle this special;
11553 switch (2nd byte) {
11554 cases for each unambiguous special:
11555 handle this special;
11557 cases for each ambigous special/literal:
11559 if (special) handle here
11561 default: // unambiguously literal:
11564 default: // is a literal char
11567 create EXACTish node for literal;
11568 while (more input and node isn't full) {
11569 switch (input_byte) {
11570 cases for each special;
11571 make sure parse pointer is set so that the next call to
11572 regatom will see this special first
11573 goto loopdone; // EXACTish node terminated by prev. char
11575 append char to EXACTISH node;
11577 get next input byte;
11581 return the generated node;
11583 Specifically there are two separate switches for handling
11584 escape sequences, with the one for handling literal escapes requiring
11585 a dummy entry for all of the special escapes that are actually handled
11588 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
11590 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
11592 Otherwise does not return NULL.
11596 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11598 regnode *ret = NULL;
11600 char *parse_start = RExC_parse;
11605 GET_RE_DEBUG_FLAGS_DECL;
11607 *flagp = WORST; /* Tentatively. */
11609 DEBUG_PARSE("atom");
11611 PERL_ARGS_ASSERT_REGATOM;
11614 switch ((U8)*RExC_parse) {
11616 RExC_seen_zerolen++;
11617 nextchar(pRExC_state);
11618 if (RExC_flags & RXf_PMf_MULTILINE)
11619 ret = reg_node(pRExC_state, MBOL);
11621 ret = reg_node(pRExC_state, SBOL);
11622 Set_Node_Length(ret, 1); /* MJD */
11625 nextchar(pRExC_state);
11627 RExC_seen_zerolen++;
11628 if (RExC_flags & RXf_PMf_MULTILINE)
11629 ret = reg_node(pRExC_state, MEOL);
11631 ret = reg_node(pRExC_state, SEOL);
11632 Set_Node_Length(ret, 1); /* MJD */
11635 nextchar(pRExC_state);
11636 if (RExC_flags & RXf_PMf_SINGLELINE)
11637 ret = reg_node(pRExC_state, SANY);
11639 ret = reg_node(pRExC_state, REG_ANY);
11640 *flagp |= HASWIDTH|SIMPLE;
11642 Set_Node_Length(ret, 1); /* MJD */
11646 char * const oregcomp_parse = ++RExC_parse;
11647 ret = regclass(pRExC_state, flagp,depth+1,
11648 FALSE, /* means parse the whole char class */
11649 TRUE, /* allow multi-char folds */
11650 FALSE, /* don't silence non-portable warnings. */
11652 if (*RExC_parse != ']') {
11653 RExC_parse = oregcomp_parse;
11654 vFAIL("Unmatched [");
11657 if (*flagp & RESTART_UTF8)
11659 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
11662 nextchar(pRExC_state);
11663 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
11667 nextchar(pRExC_state);
11668 ret = reg(pRExC_state, 2, &flags,depth+1);
11670 if (flags & TRYAGAIN) {
11671 if (RExC_parse == RExC_end) {
11672 /* Make parent create an empty node if needed. */
11673 *flagp |= TRYAGAIN;
11678 if (flags & RESTART_UTF8) {
11679 *flagp = RESTART_UTF8;
11682 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"",
11685 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
11689 if (flags & TRYAGAIN) {
11690 *flagp |= TRYAGAIN;
11693 vFAIL("Internal urp");
11694 /* Supposed to be caught earlier. */
11700 vFAIL("Quantifier follows nothing");
11705 This switch handles escape sequences that resolve to some kind
11706 of special regop and not to literal text. Escape sequnces that
11707 resolve to literal text are handled below in the switch marked
11710 Every entry in this switch *must* have a corresponding entry
11711 in the literal escape switch. However, the opposite is not
11712 required, as the default for this switch is to jump to the
11713 literal text handling code.
11715 switch ((U8)*++RExC_parse) {
11716 /* Special Escapes */
11718 RExC_seen_zerolen++;
11719 ret = reg_node(pRExC_state, SBOL);
11720 /* SBOL is shared with /^/ so we set the flags so we can tell
11721 * /\A/ from /^/ in split. We check ret because first pass we
11722 * have no regop struct to set the flags on. */
11726 goto finish_meta_pat;
11728 ret = reg_node(pRExC_state, GPOS);
11729 RExC_seen |= REG_GPOS_SEEN;
11731 goto finish_meta_pat;
11733 RExC_seen_zerolen++;
11734 ret = reg_node(pRExC_state, KEEPS);
11736 /* XXX:dmq : disabling in-place substitution seems to
11737 * be necessary here to avoid cases of memory corruption, as
11738 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
11740 RExC_seen |= REG_LOOKBEHIND_SEEN;
11741 goto finish_meta_pat;
11743 ret = reg_node(pRExC_state, SEOL);
11745 RExC_seen_zerolen++; /* Do not optimize RE away */
11746 goto finish_meta_pat;
11748 ret = reg_node(pRExC_state, EOS);
11750 RExC_seen_zerolen++; /* Do not optimize RE away */
11751 goto finish_meta_pat;
11753 ret = reg_node(pRExC_state, CANY);
11754 RExC_seen |= REG_CANY_SEEN;
11755 *flagp |= HASWIDTH|SIMPLE;
11757 ckWARNdep(RExC_parse+1, "\\C is deprecated");
11759 goto finish_meta_pat;
11761 ret = reg_node(pRExC_state, CLUMP);
11762 *flagp |= HASWIDTH;
11763 goto finish_meta_pat;
11769 arg = ANYOF_WORDCHAR;
11773 RExC_seen_zerolen++;
11774 RExC_seen |= REG_LOOKBEHIND_SEEN;
11775 op = BOUND + get_regex_charset(RExC_flags);
11776 if (op > BOUNDA) { /* /aa is same as /a */
11779 else if (op == BOUNDL) {
11780 RExC_contains_locale = 1;
11782 ret = reg_node(pRExC_state, op);
11783 FLAGS(ret) = get_regex_charset(RExC_flags);
11785 if ((U8) *(RExC_parse + 1) == '{') {
11786 /* diag_listed_as: Use "%s" instead of "%s" */
11787 vFAIL("Use \"\\b\\{\" instead of \"\\b{\"");
11789 goto finish_meta_pat;
11791 RExC_seen_zerolen++;
11792 RExC_seen |= REG_LOOKBEHIND_SEEN;
11793 op = NBOUND + get_regex_charset(RExC_flags);
11794 if (op > NBOUNDA) { /* /aa is same as /a */
11797 else if (op == NBOUNDL) {
11798 RExC_contains_locale = 1;
11800 ret = reg_node(pRExC_state, op);
11801 FLAGS(ret) = get_regex_charset(RExC_flags);
11803 if ((U8) *(RExC_parse + 1) == '{') {
11804 /* diag_listed_as: Use "%s" instead of "%s" */
11805 vFAIL("Use \"\\B\\{\" instead of \"\\B{\"");
11807 goto finish_meta_pat;
11817 ret = reg_node(pRExC_state, LNBREAK);
11818 *flagp |= HASWIDTH|SIMPLE;
11819 goto finish_meta_pat;
11827 goto join_posix_op_known;
11833 arg = ANYOF_VERTWS;
11835 goto join_posix_op_known;
11845 op = POSIXD + get_regex_charset(RExC_flags);
11846 if (op > POSIXA) { /* /aa is same as /a */
11849 else if (op == POSIXL) {
11850 RExC_contains_locale = 1;
11853 join_posix_op_known:
11856 op += NPOSIXD - POSIXD;
11859 ret = reg_node(pRExC_state, op);
11861 FLAGS(ret) = namedclass_to_classnum(arg);
11864 *flagp |= HASWIDTH|SIMPLE;
11868 nextchar(pRExC_state);
11869 Set_Node_Length(ret, 2); /* MJD */
11875 char* parse_start = RExC_parse - 2;
11880 ret = regclass(pRExC_state, flagp,depth+1,
11881 TRUE, /* means just parse this element */
11882 FALSE, /* don't allow multi-char folds */
11883 FALSE, /* don't silence non-portable warnings.
11884 It would be a bug if these returned
11887 /* regclass() can only return RESTART_UTF8 if multi-char folds
11890 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
11895 Set_Node_Offset(ret, parse_start + 2);
11896 Set_Node_Cur_Length(ret, parse_start);
11897 nextchar(pRExC_state);
11901 /* Handle \N and \N{NAME} with multiple code points here and not
11902 * below because it can be multicharacter. join_exact() will join
11903 * them up later on. Also this makes sure that things like
11904 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
11905 * The options to the grok function call causes it to fail if the
11906 * sequence is just a single code point. We then go treat it as
11907 * just another character in the current EXACT node, and hence it
11908 * gets uniform treatment with all the other characters. The
11909 * special treatment for quantifiers is not needed for such single
11910 * character sequences */
11912 if ((STRLEN) -1 == grok_bslash_N(pRExC_state, &ret, NULL, flagp,
11915 if (*flagp & RESTART_UTF8)
11921 case 'k': /* Handle \k<NAME> and \k'NAME' */
11924 char ch= RExC_parse[1];
11925 if (ch != '<' && ch != '\'' && ch != '{') {
11927 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11928 vFAIL2("Sequence %.2s... not terminated",parse_start);
11930 /* this pretty much dupes the code for (?P=...) in reg(), if
11931 you change this make sure you change that */
11932 char* name_start = (RExC_parse += 2);
11934 SV *sv_dat = reg_scan_name(pRExC_state,
11935 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11936 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
11937 if (RExC_parse == name_start || *RExC_parse != ch)
11938 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11939 vFAIL2("Sequence %.3s... not terminated",parse_start);
11942 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11943 RExC_rxi->data->data[num]=(void*)sv_dat;
11944 SvREFCNT_inc_simple_void(sv_dat);
11948 ret = reganode(pRExC_state,
11951 : (ASCII_FOLD_RESTRICTED)
11953 : (AT_LEAST_UNI_SEMANTICS)
11959 *flagp |= HASWIDTH;
11961 /* override incorrect value set in reganode MJD */
11962 Set_Node_Offset(ret, parse_start+1);
11963 Set_Node_Cur_Length(ret, parse_start);
11964 nextchar(pRExC_state);
11970 case '1': case '2': case '3': case '4':
11971 case '5': case '6': case '7': case '8': case '9':
11976 if (*RExC_parse == 'g') {
11980 if (*RExC_parse == '{') {
11984 if (*RExC_parse == '-') {
11988 if (hasbrace && !isDIGIT(*RExC_parse)) {
11989 if (isrel) RExC_parse--;
11991 goto parse_named_seq;
11994 num = S_backref_value(RExC_parse);
11996 vFAIL("Reference to invalid group 0");
11997 else if (num == I32_MAX) {
11998 if (isDIGIT(*RExC_parse))
11999 vFAIL("Reference to nonexistent group");
12001 vFAIL("Unterminated \\g... pattern");
12005 num = RExC_npar - num;
12007 vFAIL("Reference to nonexistent or unclosed group");
12011 num = S_backref_value(RExC_parse);
12012 /* bare \NNN might be backref or octal - if it is larger than or equal
12013 * RExC_npar then it is assumed to be and octal escape.
12014 * Note RExC_npar is +1 from the actual number of parens*/
12015 if (num == I32_MAX || (num > 9 && num >= RExC_npar
12016 && *RExC_parse != '8' && *RExC_parse != '9'))
12018 /* Probably a character specified in octal, e.g. \35 */
12023 /* at this point RExC_parse definitely points to a backref
12026 #ifdef RE_TRACK_PATTERN_OFFSETS
12027 char * const parse_start = RExC_parse - 1; /* MJD */
12029 while (isDIGIT(*RExC_parse))
12032 if (*RExC_parse != '}')
12033 vFAIL("Unterminated \\g{...} pattern");
12037 if (num > (I32)RExC_rx->nparens)
12038 vFAIL("Reference to nonexistent group");
12041 ret = reganode(pRExC_state,
12044 : (ASCII_FOLD_RESTRICTED)
12046 : (AT_LEAST_UNI_SEMANTICS)
12052 *flagp |= HASWIDTH;
12054 /* override incorrect value set in reganode MJD */
12055 Set_Node_Offset(ret, parse_start+1);
12056 Set_Node_Cur_Length(ret, parse_start);
12058 nextchar(pRExC_state);
12063 if (RExC_parse >= RExC_end)
12064 FAIL("Trailing \\");
12067 /* Do not generate "unrecognized" warnings here, we fall
12068 back into the quick-grab loop below */
12075 if (RExC_flags & RXf_PMf_EXTENDED) {
12076 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
12077 if (RExC_parse < RExC_end)
12084 parse_start = RExC_parse - 1;
12093 #define MAX_NODE_STRING_SIZE 127
12094 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
12096 U8 upper_parse = MAX_NODE_STRING_SIZE;
12097 U8 node_type = compute_EXACTish(pRExC_state);
12098 bool next_is_quantifier;
12099 char * oldp = NULL;
12101 /* We can convert EXACTF nodes to EXACTFU if they contain only
12102 * characters that match identically regardless of the target
12103 * string's UTF8ness. The reason to do this is that EXACTF is not
12104 * trie-able, EXACTFU is.
12106 * Similarly, we can convert EXACTFL nodes to EXACTFU if they
12107 * contain only above-Latin1 characters (hence must be in UTF8),
12108 * which don't participate in folds with Latin1-range characters,
12109 * as the latter's folds aren't known until runtime. (We don't
12110 * need to figure this out until pass 2) */
12111 bool maybe_exactfu = PASS2
12112 && (node_type == EXACTF || node_type == EXACTFL);
12114 /* If a folding node contains only code points that don't
12115 * participate in folds, it can be changed into an EXACT node,
12116 * which allows the optimizer more things to look for */
12119 ret = reg_node(pRExC_state, node_type);
12121 /* In pass1, folded, we use a temporary buffer instead of the
12122 * actual node, as the node doesn't exist yet */
12123 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
12129 /* We do the EXACTFish to EXACT node only if folding. (And we
12130 * don't need to figure this out until pass 2) */
12131 maybe_exact = FOLD && PASS2;
12133 /* XXX The node can hold up to 255 bytes, yet this only goes to
12134 * 127. I (khw) do not know why. Keeping it somewhat less than
12135 * 255 allows us to not have to worry about overflow due to
12136 * converting to utf8 and fold expansion, but that value is
12137 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
12138 * split up by this limit into a single one using the real max of
12139 * 255. Even at 127, this breaks under rare circumstances. If
12140 * folding, we do not want to split a node at a character that is a
12141 * non-final in a multi-char fold, as an input string could just
12142 * happen to want to match across the node boundary. The join
12143 * would solve that problem if the join actually happens. But a
12144 * series of more than two nodes in a row each of 127 would cause
12145 * the first join to succeed to get to 254, but then there wouldn't
12146 * be room for the next one, which could at be one of those split
12147 * multi-char folds. I don't know of any fool-proof solution. One
12148 * could back off to end with only a code point that isn't such a
12149 * non-final, but it is possible for there not to be any in the
12151 for (p = RExC_parse - 1;
12152 len < upper_parse && p < RExC_end;
12157 if (RExC_flags & RXf_PMf_EXTENDED)
12158 p = regpatws(pRExC_state, p,
12159 TRUE); /* means recognize comments */
12170 /* Literal Escapes Switch
12172 This switch is meant to handle escape sequences that
12173 resolve to a literal character.
12175 Every escape sequence that represents something
12176 else, like an assertion or a char class, is handled
12177 in the switch marked 'Special Escapes' above in this
12178 routine, but also has an entry here as anything that
12179 isn't explicitly mentioned here will be treated as
12180 an unescaped equivalent literal.
12183 switch ((U8)*++p) {
12184 /* These are all the special escapes. */
12185 case 'A': /* Start assertion */
12186 case 'b': case 'B': /* Word-boundary assertion*/
12187 case 'C': /* Single char !DANGEROUS! */
12188 case 'd': case 'D': /* digit class */
12189 case 'g': case 'G': /* generic-backref, pos assertion */
12190 case 'h': case 'H': /* HORIZWS */
12191 case 'k': case 'K': /* named backref, keep marker */
12192 case 'p': case 'P': /* Unicode property */
12193 case 'R': /* LNBREAK */
12194 case 's': case 'S': /* space class */
12195 case 'v': case 'V': /* VERTWS */
12196 case 'w': case 'W': /* word class */
12197 case 'X': /* eXtended Unicode "combining
12198 character sequence" */
12199 case 'z': case 'Z': /* End of line/string assertion */
12203 /* Anything after here is an escape that resolves to a
12204 literal. (Except digits, which may or may not)
12210 case 'N': /* Handle a single-code point named character. */
12211 /* The options cause it to fail if a multiple code
12212 * point sequence. Handle those in the switch() above
12214 RExC_parse = p + 1;
12215 if ((STRLEN) -1 == grok_bslash_N(pRExC_state, NULL,
12221 if (*flagp & RESTART_UTF8)
12222 FAIL("panic: grok_bslash_N set RESTART_UTF8");
12223 RExC_parse = p = oldp;
12227 if (ender > 0xff) {
12244 ender = ESC_NATIVE;
12254 const char* error_msg;
12256 bool valid = grok_bslash_o(&p,
12259 PASS2, /* out warnings */
12260 FALSE, /* not strict */
12261 TRUE, /* Output warnings
12266 RExC_parse = p; /* going to die anyway; point
12267 to exact spot of failure */
12271 if (IN_ENCODING && ender < 0x100) {
12272 goto recode_encoding;
12274 if (ender > 0xff) {
12281 UV result = UV_MAX; /* initialize to erroneous
12283 const char* error_msg;
12285 bool valid = grok_bslash_x(&p,
12288 PASS2, /* out warnings */
12289 FALSE, /* not strict */
12290 TRUE, /* Output warnings
12295 RExC_parse = p; /* going to die anyway; point
12296 to exact spot of failure */
12301 if (IN_ENCODING && ender < 0x100) {
12302 goto recode_encoding;
12304 if (ender > 0xff) {
12311 ender = grok_bslash_c(*p++, PASS2);
12313 case '8': case '9': /* must be a backreference */
12316 case '1': case '2': case '3':case '4':
12317 case '5': case '6': case '7':
12318 /* When we parse backslash escapes there is ambiguity
12319 * between backreferences and octal escapes. Any escape
12320 * from \1 - \9 is a backreference, any multi-digit
12321 * escape which does not start with 0 and which when
12322 * evaluated as decimal could refer to an already
12323 * parsed capture buffer is a backslash. Anything else
12326 * Note this implies that \118 could be interpreted as
12327 * 118 OR as "\11" . "8" depending on whether there
12328 * were 118 capture buffers defined already in the
12331 /* NOTE, RExC_npar is 1 more than the actual number of
12332 * parens we have seen so far, hence the < RExC_npar below. */
12334 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
12335 { /* Not to be treated as an octal constant, go
12343 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
12345 ender = grok_oct(p, &numlen, &flags, NULL);
12346 if (ender > 0xff) {
12350 if (PASS2 /* like \08, \178 */
12353 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
12355 reg_warn_non_literal_string(
12357 form_short_octal_warning(p, numlen));
12360 if (IN_ENCODING && ender < 0x100)
12361 goto recode_encoding;
12364 if (! RExC_override_recoding) {
12365 SV* enc = _get_encoding();
12366 ender = reg_recode((const char)(U8)ender, &enc);
12368 ckWARNreg(p, "Invalid escape in the specified encoding");
12374 FAIL("Trailing \\");
12377 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
12378 /* Include any { following the alpha to emphasize
12379 * that it could be part of an escape at some point
12381 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
12382 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
12384 goto normal_default;
12385 } /* End of switch on '\' */
12388 /* Currently we don't warn when the lbrace is at the start
12389 * of a construct. This catches it in the middle of a
12390 * literal string, or when its the first thing after
12391 * something like "\b" */
12393 && (len || (p > RExC_start && isALPHA_A(*(p -1)))))
12395 ckWARNregdep(p + 1, "Unescaped left brace in regex is deprecated, passed through");
12398 default: /* A literal character */
12400 if (UTF8_IS_START(*p) && UTF) {
12402 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
12403 &numlen, UTF8_ALLOW_DEFAULT);
12409 } /* End of switch on the literal */
12411 /* Here, have looked at the literal character and <ender>
12412 * contains its ordinal, <p> points to the character after it
12415 if ( RExC_flags & RXf_PMf_EXTENDED)
12416 p = regpatws(pRExC_state, p,
12417 TRUE); /* means recognize comments */
12419 /* If the next thing is a quantifier, it applies to this
12420 * character only, which means that this character has to be in
12421 * its own node and can't just be appended to the string in an
12422 * existing node, so if there are already other characters in
12423 * the node, close the node with just them, and set up to do
12424 * this character again next time through, when it will be the
12425 * only thing in its new node */
12426 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
12432 if (! FOLD) { /* The simple case, just append the literal */
12434 /* In the sizing pass, we need only the size of the
12435 * character we are appending, hence we can delay getting
12436 * its representation until PASS2. */
12439 const STRLEN unilen = UNISKIP(ender);
12442 /* We have to subtract 1 just below (and again in
12443 * the corresponding PASS2 code) because the loop
12444 * increments <len> each time, as all but this path
12445 * (and one other) through it add a single byte to
12446 * the EXACTish node. But these paths would change
12447 * len to be the correct final value, so cancel out
12448 * the increment that follows */
12454 } else { /* PASS2 */
12457 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
12458 len += (char *) new_s - s - 1;
12459 s = (char *) new_s;
12462 *(s++) = (char) ender;
12466 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
12468 /* Here are folding under /l, and the code point is
12469 * problematic. First, we know we can't simplify things */
12470 maybe_exact = FALSE;
12471 maybe_exactfu = FALSE;
12473 /* A problematic code point in this context means that its
12474 * fold isn't known until runtime, so we can't fold it now.
12475 * (The non-problematic code points are the above-Latin1
12476 * ones that fold to also all above-Latin1. Their folds
12477 * don't vary no matter what the locale is.) But here we
12478 * have characters whose fold depends on the locale.
12479 * Unlike the non-folding case above, we have to keep track
12480 * of these in the sizing pass, so that we can make sure we
12481 * don't split too-long nodes in the middle of a potential
12482 * multi-char fold. And unlike the regular fold case
12483 * handled in the else clauses below, we don't actually
12484 * fold and don't have special cases to consider. What we
12485 * do for both passes is the PASS2 code for non-folding */
12486 goto not_fold_common;
12488 else /* A regular FOLD code point */
12490 /* See comments for join_exact() as to why we fold this
12491 * non-UTF at compile time */
12492 || (node_type == EXACTFU
12493 && ender == LATIN_SMALL_LETTER_SHARP_S)))
12495 /* Here, are folding and are not UTF-8 encoded; therefore
12496 * the character must be in the range 0-255, and is not /l
12497 * (Not /l because we already handled these under /l in
12498 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
12499 if (IS_IN_SOME_FOLD_L1(ender)) {
12500 maybe_exact = FALSE;
12502 /* See if the character's fold differs between /d and
12503 * /u. This includes the multi-char fold SHARP S to
12506 && (PL_fold[ender] != PL_fold_latin1[ender]
12507 || ender == LATIN_SMALL_LETTER_SHARP_S
12509 && isALPHA_FOLD_EQ(ender, 's')
12510 && isALPHA_FOLD_EQ(*(s-1), 's'))))
12512 maybe_exactfu = FALSE;
12516 /* Even when folding, we store just the input character, as
12517 * we have an array that finds its fold quickly */
12518 *(s++) = (char) ender;
12520 else { /* FOLD and UTF */
12521 /* Unlike the non-fold case, we do actually have to
12522 * calculate the results here in pass 1. This is for two
12523 * reasons, the folded length may be longer than the
12524 * unfolded, and we have to calculate how many EXACTish
12525 * nodes it will take; and we may run out of room in a node
12526 * in the middle of a potential multi-char fold, and have
12527 * to back off accordingly. */
12530 if (isASCII_uni(ender)) {
12531 folded = toFOLD(ender);
12532 *(s)++ = (U8) folded;
12537 folded = _to_uni_fold_flags(
12541 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12542 ? FOLD_FLAGS_NOMIX_ASCII
12546 /* The loop increments <len> each time, as all but this
12547 * path (and one other) through it add a single byte to
12548 * the EXACTish node. But this one has changed len to
12549 * be the correct final value, so subtract one to
12550 * cancel out the increment that follows */
12551 len += foldlen - 1;
12553 /* If this node only contains non-folding code points so
12554 * far, see if this new one is also non-folding */
12556 if (folded != ender) {
12557 maybe_exact = FALSE;
12560 /* Here the fold is the original; we have to check
12561 * further to see if anything folds to it */
12562 if (_invlist_contains_cp(PL_utf8_foldable,
12565 maybe_exact = FALSE;
12572 if (next_is_quantifier) {
12574 /* Here, the next input is a quantifier, and to get here,
12575 * the current character is the only one in the node.
12576 * Also, here <len> doesn't include the final byte for this
12582 } /* End of loop through literal characters */
12584 /* Here we have either exhausted the input or ran out of room in
12585 * the node. (If we encountered a character that can't be in the
12586 * node, transfer is made directly to <loopdone>, and so we
12587 * wouldn't have fallen off the end of the loop.) In the latter
12588 * case, we artificially have to split the node into two, because
12589 * we just don't have enough space to hold everything. This
12590 * creates a problem if the final character participates in a
12591 * multi-character fold in the non-final position, as a match that
12592 * should have occurred won't, due to the way nodes are matched,
12593 * and our artificial boundary. So back off until we find a non-
12594 * problematic character -- one that isn't at the beginning or
12595 * middle of such a fold. (Either it doesn't participate in any
12596 * folds, or appears only in the final position of all the folds it
12597 * does participate in.) A better solution with far fewer false
12598 * positives, and that would fill the nodes more completely, would
12599 * be to actually have available all the multi-character folds to
12600 * test against, and to back-off only far enough to be sure that
12601 * this node isn't ending with a partial one. <upper_parse> is set
12602 * further below (if we need to reparse the node) to include just
12603 * up through that final non-problematic character that this code
12604 * identifies, so when it is set to less than the full node, we can
12605 * skip the rest of this */
12606 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
12608 const STRLEN full_len = len;
12610 assert(len >= MAX_NODE_STRING_SIZE);
12612 /* Here, <s> points to the final byte of the final character.
12613 * Look backwards through the string until find a non-
12614 * problematic character */
12618 /* This has no multi-char folds to non-UTF characters */
12619 if (ASCII_FOLD_RESTRICTED) {
12623 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
12627 if (! PL_NonL1NonFinalFold) {
12628 PL_NonL1NonFinalFold = _new_invlist_C_array(
12629 NonL1_Perl_Non_Final_Folds_invlist);
12632 /* Point to the first byte of the final character */
12633 s = (char *) utf8_hop((U8 *) s, -1);
12635 while (s >= s0) { /* Search backwards until find
12636 non-problematic char */
12637 if (UTF8_IS_INVARIANT(*s)) {
12639 /* There are no ascii characters that participate
12640 * in multi-char folds under /aa. In EBCDIC, the
12641 * non-ascii invariants are all control characters,
12642 * so don't ever participate in any folds. */
12643 if (ASCII_FOLD_RESTRICTED
12644 || ! IS_NON_FINAL_FOLD(*s))
12649 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
12650 if (! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_NATIVE(
12656 else if (! _invlist_contains_cp(
12657 PL_NonL1NonFinalFold,
12658 valid_utf8_to_uvchr((U8 *) s, NULL)))
12663 /* Here, the current character is problematic in that
12664 * it does occur in the non-final position of some
12665 * fold, so try the character before it, but have to
12666 * special case the very first byte in the string, so
12667 * we don't read outside the string */
12668 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
12669 } /* End of loop backwards through the string */
12671 /* If there were only problematic characters in the string,
12672 * <s> will point to before s0, in which case the length
12673 * should be 0, otherwise include the length of the
12674 * non-problematic character just found */
12675 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
12678 /* Here, have found the final character, if any, that is
12679 * non-problematic as far as ending the node without splitting
12680 * it across a potential multi-char fold. <len> contains the
12681 * number of bytes in the node up-to and including that
12682 * character, or is 0 if there is no such character, meaning
12683 * the whole node contains only problematic characters. In
12684 * this case, give up and just take the node as-is. We can't
12689 /* If the node ends in an 's' we make sure it stays EXACTF,
12690 * as if it turns into an EXACTFU, it could later get
12691 * joined with another 's' that would then wrongly match
12693 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
12695 maybe_exactfu = FALSE;
12699 /* Here, the node does contain some characters that aren't
12700 * problematic. If one such is the final character in the
12701 * node, we are done */
12702 if (len == full_len) {
12705 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
12707 /* If the final character is problematic, but the
12708 * penultimate is not, back-off that last character to
12709 * later start a new node with it */
12714 /* Here, the final non-problematic character is earlier
12715 * in the input than the penultimate character. What we do
12716 * is reparse from the beginning, going up only as far as
12717 * this final ok one, thus guaranteeing that the node ends
12718 * in an acceptable character. The reason we reparse is
12719 * that we know how far in the character is, but we don't
12720 * know how to correlate its position with the input parse.
12721 * An alternate implementation would be to build that
12722 * correlation as we go along during the original parse,
12723 * but that would entail extra work for every node, whereas
12724 * this code gets executed only when the string is too
12725 * large for the node, and the final two characters are
12726 * problematic, an infrequent occurrence. Yet another
12727 * possible strategy would be to save the tail of the
12728 * string, and the next time regatom is called, initialize
12729 * with that. The problem with this is that unless you
12730 * back off one more character, you won't be guaranteed
12731 * regatom will get called again, unless regbranch,
12732 * regpiece ... are also changed. If you do back off that
12733 * extra character, so that there is input guaranteed to
12734 * force calling regatom, you can't handle the case where
12735 * just the first character in the node is acceptable. I
12736 * (khw) decided to try this method which doesn't have that
12737 * pitfall; if performance issues are found, we can do a
12738 * combination of the current approach plus that one */
12744 } /* End of verifying node ends with an appropriate char */
12746 loopdone: /* Jumped to when encounters something that shouldn't be in
12749 /* I (khw) don't know if you can get here with zero length, but the
12750 * old code handled this situation by creating a zero-length EXACT
12751 * node. Might as well be NOTHING instead */
12757 /* If 'maybe_exact' is still set here, means there are no
12758 * code points in the node that participate in folds;
12759 * similarly for 'maybe_exactfu' and code points that match
12760 * differently depending on UTF8ness of the target string
12761 * (for /u), or depending on locale for /l */
12767 else if (maybe_exactfu) {
12773 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
12774 FALSE /* Don't look to see if could
12775 be turned into an EXACT
12776 node, as we have already
12781 RExC_parse = p - 1;
12782 Set_Node_Cur_Length(ret, parse_start);
12783 nextchar(pRExC_state);
12785 /* len is STRLEN which is unsigned, need to copy to signed */
12788 vFAIL("Internal disaster");
12791 } /* End of label 'defchar:' */
12793 } /* End of giant switch on input character */
12799 S_regpatws(RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
12801 /* Returns the next non-pattern-white space, non-comment character (the
12802 * latter only if 'recognize_comment is true) in the string p, which is
12803 * ended by RExC_end. See also reg_skipcomment */
12804 const char *e = RExC_end;
12806 PERL_ARGS_ASSERT_REGPATWS;
12810 if ((len = is_PATWS_safe(p, e, UTF))) {
12813 else if (recognize_comment && *p == '#') {
12814 p = reg_skipcomment(pRExC_state, p);
12823 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
12825 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
12826 * sets up the bitmap and any flags, removing those code points from the
12827 * inversion list, setting it to NULL should it become completely empty */
12829 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
12830 assert(PL_regkind[OP(node)] == ANYOF);
12832 ANYOF_BITMAP_ZERO(node);
12833 if (*invlist_ptr) {
12835 /* This gets set if we actually need to modify things */
12836 bool change_invlist = FALSE;
12840 /* Start looking through *invlist_ptr */
12841 invlist_iterinit(*invlist_ptr);
12842 while (invlist_iternext(*invlist_ptr, &start, &end)) {
12846 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
12847 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
12849 else if (end >= NUM_ANYOF_CODE_POINTS) {
12850 ANYOF_FLAGS(node) |= ANYOF_HAS_UTF8_NONBITMAP_MATCHES;
12853 /* Quit if are above what we should change */
12854 if (start >= NUM_ANYOF_CODE_POINTS) {
12858 change_invlist = TRUE;
12860 /* Set all the bits in the range, up to the max that we are doing */
12861 high = (end < NUM_ANYOF_CODE_POINTS - 1)
12863 : NUM_ANYOF_CODE_POINTS - 1;
12864 for (i = start; i <= (int) high; i++) {
12865 if (! ANYOF_BITMAP_TEST(node, i)) {
12866 ANYOF_BITMAP_SET(node, i);
12870 invlist_iterfinish(*invlist_ptr);
12872 /* Done with loop; remove any code points that are in the bitmap from
12873 * *invlist_ptr; similarly for code points above the bitmap if we have
12874 * a flag to match all of them anyways */
12875 if (change_invlist) {
12876 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
12878 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
12879 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
12882 /* If have completely emptied it, remove it completely */
12883 if (_invlist_len(*invlist_ptr) == 0) {
12884 SvREFCNT_dec_NN(*invlist_ptr);
12885 *invlist_ptr = NULL;
12890 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
12891 Character classes ([:foo:]) can also be negated ([:^foo:]).
12892 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
12893 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
12894 but trigger failures because they are currently unimplemented. */
12896 #define POSIXCC_DONE(c) ((c) == ':')
12897 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
12898 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
12900 PERL_STATIC_INLINE I32
12901 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
12903 I32 namedclass = OOB_NAMEDCLASS;
12905 PERL_ARGS_ASSERT_REGPPOSIXCC;
12907 if (value == '[' && RExC_parse + 1 < RExC_end &&
12908 /* I smell either [: or [= or [. -- POSIX has been here, right? */
12909 POSIXCC(UCHARAT(RExC_parse)))
12911 const char c = UCHARAT(RExC_parse);
12912 char* const s = RExC_parse++;
12914 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
12916 if (RExC_parse == RExC_end) {
12919 /* Try to give a better location for the error (than the end of
12920 * the string) by looking for the matching ']' */
12922 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
12925 vFAIL2("Unmatched '%c' in POSIX class", c);
12927 /* Grandfather lone [:, [=, [. */
12931 const char* const t = RExC_parse++; /* skip over the c */
12934 if (UCHARAT(RExC_parse) == ']') {
12935 const char *posixcc = s + 1;
12936 RExC_parse++; /* skip over the ending ] */
12939 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
12940 const I32 skip = t - posixcc;
12942 /* Initially switch on the length of the name. */
12945 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
12946 this is the Perl \w
12948 namedclass = ANYOF_WORDCHAR;
12951 /* Names all of length 5. */
12952 /* alnum alpha ascii blank cntrl digit graph lower
12953 print punct space upper */
12954 /* Offset 4 gives the best switch position. */
12955 switch (posixcc[4]) {
12957 if (memEQ(posixcc, "alph", 4)) /* alpha */
12958 namedclass = ANYOF_ALPHA;
12961 if (memEQ(posixcc, "spac", 4)) /* space */
12962 namedclass = ANYOF_PSXSPC;
12965 if (memEQ(posixcc, "grap", 4)) /* graph */
12966 namedclass = ANYOF_GRAPH;
12969 if (memEQ(posixcc, "asci", 4)) /* ascii */
12970 namedclass = ANYOF_ASCII;
12973 if (memEQ(posixcc, "blan", 4)) /* blank */
12974 namedclass = ANYOF_BLANK;
12977 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
12978 namedclass = ANYOF_CNTRL;
12981 if (memEQ(posixcc, "alnu", 4)) /* alnum */
12982 namedclass = ANYOF_ALPHANUMERIC;
12985 if (memEQ(posixcc, "lowe", 4)) /* lower */
12986 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
12987 else if (memEQ(posixcc, "uppe", 4)) /* upper */
12988 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
12991 if (memEQ(posixcc, "digi", 4)) /* digit */
12992 namedclass = ANYOF_DIGIT;
12993 else if (memEQ(posixcc, "prin", 4)) /* print */
12994 namedclass = ANYOF_PRINT;
12995 else if (memEQ(posixcc, "punc", 4)) /* punct */
12996 namedclass = ANYOF_PUNCT;
13001 if (memEQ(posixcc, "xdigit", 6))
13002 namedclass = ANYOF_XDIGIT;
13006 if (namedclass == OOB_NAMEDCLASS)
13008 "POSIX class [:%"UTF8f":] unknown",
13009 UTF8fARG(UTF, t - s - 1, s + 1));
13011 /* The #defines are structured so each complement is +1 to
13012 * the normal one */
13016 assert (posixcc[skip] == ':');
13017 assert (posixcc[skip+1] == ']');
13018 } else if (!SIZE_ONLY) {
13019 /* [[=foo=]] and [[.foo.]] are still future. */
13021 /* adjust RExC_parse so the warning shows after
13022 the class closes */
13023 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
13025 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
13028 /* Maternal grandfather:
13029 * "[:" ending in ":" but not in ":]" */
13031 vFAIL("Unmatched '[' in POSIX class");
13034 /* Grandfather lone [:, [=, [. */
13044 S_could_it_be_a_POSIX_class(RExC_state_t *pRExC_state)
13046 /* This applies some heuristics at the current parse position (which should
13047 * be at a '[') to see if what follows might be intended to be a [:posix:]
13048 * class. It returns true if it really is a posix class, of course, but it
13049 * also can return true if it thinks that what was intended was a posix
13050 * class that didn't quite make it.
13052 * It will return true for
13054 * [:alphanumerics] (as long as the ] isn't followed immediately by a
13055 * ')' indicating the end of the (?[
13056 * [:any garbage including %^&$ punctuation:]
13058 * This is designed to be called only from S_handle_regex_sets; it could be
13059 * easily adapted to be called from the spot at the beginning of regclass()
13060 * that checks to see in a normal bracketed class if the surrounding []
13061 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
13062 * change long-standing behavior, so I (khw) didn't do that */
13063 char* p = RExC_parse + 1;
13064 char first_char = *p;
13066 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
13068 assert(*(p - 1) == '[');
13070 if (! POSIXCC(first_char)) {
13075 while (p < RExC_end && isWORDCHAR(*p)) p++;
13077 if (p >= RExC_end) {
13081 if (p - RExC_parse > 2 /* Got at least 1 word character */
13082 && (*p == first_char
13083 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
13088 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
13091 && p - RExC_parse > 2 /* [:] evaluates to colon;
13092 [::] is a bad posix class. */
13093 && first_char == *(p - 1));
13097 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
13098 I32 *flagp, U32 depth,
13099 char * const oregcomp_parse)
13101 /* Handle the (?[...]) construct to do set operations */
13104 UV start, end; /* End points of code point ranges */
13106 char *save_end, *save_parse;
13111 const bool save_fold = FOLD;
13113 GET_RE_DEBUG_FLAGS_DECL;
13115 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
13118 vFAIL("(?[...]) not valid in locale");
13120 RExC_uni_semantics = 1;
13122 /* This will return only an ANYOF regnode, or (unlikely) something smaller
13123 * (such as EXACT). Thus we can skip most everything if just sizing. We
13124 * call regclass to handle '[]' so as to not have to reinvent its parsing
13125 * rules here (throwing away the size it computes each time). And, we exit
13126 * upon an unescaped ']' that isn't one ending a regclass. To do both
13127 * these things, we need to realize that something preceded by a backslash
13128 * is escaped, so we have to keep track of backslashes */
13130 Perl_ck_warner_d(aTHX_
13131 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
13132 "The regex_sets feature is experimental" REPORT_LOCATION,
13133 UTF8fARG(UTF, (RExC_parse - RExC_precomp), RExC_precomp),
13135 RExC_end - RExC_start - (RExC_parse - RExC_precomp),
13136 RExC_precomp + (RExC_parse - RExC_precomp)));
13139 UV depth = 0; /* how many nested (?[...]) constructs */
13141 while (RExC_parse < RExC_end) {
13142 SV* current = NULL;
13143 RExC_parse = regpatws(pRExC_state, RExC_parse,
13144 TRUE); /* means recognize comments */
13145 switch (*RExC_parse) {
13147 if (RExC_parse[1] == '[') depth++, RExC_parse++;
13152 /* Skip the next byte (which could cause us to end up in
13153 * the middle of a UTF-8 character, but since none of those
13154 * are confusable with anything we currently handle in this
13155 * switch (invariants all), it's safe. We'll just hit the
13156 * default: case next time and keep on incrementing until
13157 * we find one of the invariants we do handle. */
13162 /* If this looks like it is a [:posix:] class, leave the
13163 * parse pointer at the '[' to fool regclass() into
13164 * thinking it is part of a '[[:posix:]]'. That function
13165 * will use strict checking to force a syntax error if it
13166 * doesn't work out to a legitimate class */
13167 bool is_posix_class
13168 = could_it_be_a_POSIX_class(pRExC_state);
13169 if (! is_posix_class) {
13173 /* regclass() can only return RESTART_UTF8 if multi-char
13174 folds are allowed. */
13175 if (!regclass(pRExC_state, flagp,depth+1,
13176 is_posix_class, /* parse the whole char
13177 class only if not a
13179 FALSE, /* don't allow multi-char folds */
13180 TRUE, /* silence non-portable warnings. */
13182 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
13185 /* function call leaves parse pointing to the ']', except
13186 * if we faked it */
13187 if (is_posix_class) {
13191 SvREFCNT_dec(current); /* In case it returned something */
13196 if (depth--) break;
13198 if (RExC_parse < RExC_end
13199 && *RExC_parse == ')')
13201 node = reganode(pRExC_state, ANYOF, 0);
13202 RExC_size += ANYOF_SKIP;
13203 nextchar(pRExC_state);
13204 Set_Node_Length(node,
13205 RExC_parse - oregcomp_parse + 1); /* MJD */
13214 FAIL("Syntax error in (?[...])");
13217 /* Pass 2 only after this. Everything in this construct is a
13218 * metacharacter. Operands begin with either a '\' (for an escape
13219 * sequence), or a '[' for a bracketed character class. Any other
13220 * character should be an operator, or parenthesis for grouping. Both
13221 * types of operands are handled by calling regclass() to parse them. It
13222 * is called with a parameter to indicate to return the computed inversion
13223 * list. The parsing here is implemented via a stack. Each entry on the
13224 * stack is a single character representing one of the operators, or the
13225 * '('; or else a pointer to an operand inversion list. */
13227 #define IS_OPERAND(a) (! SvIOK(a))
13229 /* The stack starts empty. It is a syntax error if the first thing parsed
13230 * is a binary operator; everything else is pushed on the stack. When an
13231 * operand is parsed, the top of the stack is examined. If it is a binary
13232 * operator, the item before it should be an operand, and both are replaced
13233 * by the result of doing that operation on the new operand and the one on
13234 * the stack. Thus a sequence of binary operands is reduced to a single
13235 * one before the next one is parsed.
13237 * A unary operator may immediately follow a binary in the input, for
13240 * When an operand is parsed and the top of the stack is a unary operator,
13241 * the operation is performed, and then the stack is rechecked to see if
13242 * this new operand is part of a binary operation; if so, it is handled as
13245 * A '(' is simply pushed on the stack; it is valid only if the stack is
13246 * empty, or the top element of the stack is an operator or another '('
13247 * (for which the parenthesized expression will become an operand). By the
13248 * time the corresponding ')' is parsed everything in between should have
13249 * been parsed and evaluated to a single operand (or else is a syntax
13250 * error), and is handled as a regular operand */
13252 sv_2mortal((SV *)(stack = newAV()));
13254 while (RExC_parse < RExC_end) {
13255 I32 top_index = av_tindex(stack);
13257 SV* current = NULL;
13259 /* Skip white space */
13260 RExC_parse = regpatws(pRExC_state, RExC_parse,
13261 TRUE /* means recognize comments */ );
13262 if (RExC_parse >= RExC_end) {
13263 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
13265 if ((curchar = UCHARAT(RExC_parse)) == ']') {
13272 if (av_tindex(stack) >= 0 /* This makes sure that we can
13273 safely subtract 1 from
13274 RExC_parse in the next clause.
13275 If we have something on the
13276 stack, we have parsed something
13278 && UCHARAT(RExC_parse - 1) == '('
13279 && RExC_parse < RExC_end)
13281 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
13282 * This happens when we have some thing like
13284 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
13286 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
13288 * Here we would be handling the interpolated
13289 * '$thai_or_lao'. We handle this by a recursive call to
13290 * ourselves which returns the inversion list the
13291 * interpolated expression evaluates to. We use the flags
13292 * from the interpolated pattern. */
13293 U32 save_flags = RExC_flags;
13294 const char * const save_parse = ++RExC_parse;
13296 parse_lparen_question_flags(pRExC_state);
13298 if (RExC_parse == save_parse /* Makes sure there was at
13299 least one flag (or this
13300 embedding wasn't compiled)
13302 || RExC_parse >= RExC_end - 4
13303 || UCHARAT(RExC_parse) != ':'
13304 || UCHARAT(++RExC_parse) != '('
13305 || UCHARAT(++RExC_parse) != '?'
13306 || UCHARAT(++RExC_parse) != '[')
13309 /* In combination with the above, this moves the
13310 * pointer to the point just after the first erroneous
13311 * character (or if there are no flags, to where they
13312 * should have been) */
13313 if (RExC_parse >= RExC_end - 4) {
13314 RExC_parse = RExC_end;
13316 else if (RExC_parse != save_parse) {
13317 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13319 vFAIL("Expecting '(?flags:(?[...'");
13322 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
13323 depth+1, oregcomp_parse);
13325 /* Here, 'current' contains the embedded expression's
13326 * inversion list, and RExC_parse points to the trailing
13327 * ']'; the next character should be the ')' which will be
13328 * paired with the '(' that has been put on the stack, so
13329 * the whole embedded expression reduces to '(operand)' */
13332 RExC_flags = save_flags;
13333 goto handle_operand;
13338 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13339 vFAIL("Unexpected character");
13342 /* regclass() can only return RESTART_UTF8 if multi-char
13343 folds are allowed. */
13344 if (!regclass(pRExC_state, flagp,depth+1,
13345 TRUE, /* means parse just the next thing */
13346 FALSE, /* don't allow multi-char folds */
13347 FALSE, /* don't silence non-portable warnings. */
13349 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
13351 /* regclass() will return with parsing just the \ sequence,
13352 * leaving the parse pointer at the next thing to parse */
13354 goto handle_operand;
13356 case '[': /* Is a bracketed character class */
13358 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
13360 if (! is_posix_class) {
13364 /* regclass() can only return RESTART_UTF8 if multi-char
13365 folds are allowed. */
13366 if(!regclass(pRExC_state, flagp,depth+1,
13367 is_posix_class, /* parse the whole char class
13368 only if not a posix class */
13369 FALSE, /* don't allow multi-char folds */
13370 FALSE, /* don't silence non-portable warnings. */
13372 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
13374 /* function call leaves parse pointing to the ']', except if we
13376 if (is_posix_class) {
13380 goto handle_operand;
13389 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
13390 || ! IS_OPERAND(*top_ptr))
13393 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
13395 av_push(stack, newSVuv(curchar));
13399 av_push(stack, newSVuv(curchar));
13403 if (top_index >= 0) {
13404 top_ptr = av_fetch(stack, top_index, FALSE);
13406 if (IS_OPERAND(*top_ptr)) {
13408 vFAIL("Unexpected '(' with no preceding operator");
13411 av_push(stack, newSVuv(curchar));
13418 || ! (current = av_pop(stack))
13419 || ! IS_OPERAND(current)
13420 || ! (lparen = av_pop(stack))
13421 || IS_OPERAND(lparen)
13422 || SvUV(lparen) != '(')
13424 SvREFCNT_dec(current);
13426 vFAIL("Unexpected ')'");
13429 SvREFCNT_dec_NN(lparen);
13436 /* Here, we have an operand to process, in 'current' */
13438 if (top_index < 0) { /* Just push if stack is empty */
13439 av_push(stack, current);
13442 SV* top = av_pop(stack);
13444 char current_operator;
13446 if (IS_OPERAND(top)) {
13447 SvREFCNT_dec_NN(top);
13448 SvREFCNT_dec_NN(current);
13449 vFAIL("Operand with no preceding operator");
13451 current_operator = (char) SvUV(top);
13452 switch (current_operator) {
13453 case '(': /* Push the '(' back on followed by the new
13455 av_push(stack, top);
13456 av_push(stack, current);
13457 SvREFCNT_inc(top); /* Counters the '_dec' done
13458 just after the 'break', so
13459 it doesn't get wrongly freed
13464 _invlist_invert(current);
13466 /* Unlike binary operators, the top of the stack,
13467 * now that this unary one has been popped off, may
13468 * legally be an operator, and we now have operand
13471 SvREFCNT_dec_NN(top);
13472 goto handle_operand;
13475 prev = av_pop(stack);
13476 _invlist_intersection(prev,
13479 av_push(stack, current);
13484 prev = av_pop(stack);
13485 _invlist_union(prev, current, ¤t);
13486 av_push(stack, current);
13490 prev = av_pop(stack);;
13491 _invlist_subtract(prev, current, ¤t);
13492 av_push(stack, current);
13495 case '^': /* The union minus the intersection */
13501 prev = av_pop(stack);
13502 _invlist_union(prev, current, &u);
13503 _invlist_intersection(prev, current, &i);
13504 /* _invlist_subtract will overwrite current
13505 without freeing what it already contains */
13507 _invlist_subtract(u, i, ¤t);
13508 av_push(stack, current);
13509 SvREFCNT_dec_NN(i);
13510 SvREFCNT_dec_NN(u);
13511 SvREFCNT_dec_NN(element);
13516 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
13518 SvREFCNT_dec_NN(top);
13519 SvREFCNT_dec(prev);
13523 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13526 if (av_tindex(stack) < 0 /* Was empty */
13527 || ((final = av_pop(stack)) == NULL)
13528 || ! IS_OPERAND(final)
13529 || av_tindex(stack) >= 0) /* More left on stack */
13531 vFAIL("Incomplete expression within '(?[ ])'");
13534 /* Here, 'final' is the resultant inversion list from evaluating the
13535 * expression. Return it if so requested */
13536 if (return_invlist) {
13537 *return_invlist = final;
13541 /* Otherwise generate a resultant node, based on 'final'. regclass() is
13542 * expecting a string of ranges and individual code points */
13543 invlist_iterinit(final);
13544 result_string = newSVpvs("");
13545 while (invlist_iternext(final, &start, &end)) {
13546 if (start == end) {
13547 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
13550 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
13555 save_parse = RExC_parse;
13556 RExC_parse = SvPV(result_string, len);
13557 save_end = RExC_end;
13558 RExC_end = RExC_parse + len;
13560 /* We turn off folding around the call, as the class we have constructed
13561 * already has all folding taken into consideration, and we don't want
13562 * regclass() to add to that */
13563 RExC_flags &= ~RXf_PMf_FOLD;
13564 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
13566 node = regclass(pRExC_state, flagp,depth+1,
13567 FALSE, /* means parse the whole char class */
13568 FALSE, /* don't allow multi-char folds */
13569 TRUE, /* silence non-portable warnings. The above may very
13570 well have generated non-portable code points, but
13571 they're valid on this machine */
13574 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
13577 RExC_flags |= RXf_PMf_FOLD;
13579 RExC_parse = save_parse + 1;
13580 RExC_end = save_end;
13581 SvREFCNT_dec_NN(final);
13582 SvREFCNT_dec_NN(result_string);
13584 nextchar(pRExC_state);
13585 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
13591 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
13593 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
13594 * innocent-looking character class, like /[ks]/i won't have to go out to
13595 * disk to find the possible matches.
13597 * This should be called only for a Latin1-range code points, cp, which is
13598 * known to be involved in a simple fold with other code points above
13599 * Latin1. It would give false results if /aa has been specified.
13600 * Multi-char folds are outside the scope of this, and must be handled
13603 * XXX It would be better to generate these via regen, in case a new
13604 * version of the Unicode standard adds new mappings, though that is not
13605 * really likely, and may be caught by the default: case of the switch
13608 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
13610 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
13616 add_cp_to_invlist(*invlist, KELVIN_SIGN);
13620 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
13623 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
13624 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
13626 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
13627 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
13628 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
13630 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
13631 *invlist = add_cp_to_invlist(*invlist,
13632 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
13634 case LATIN_SMALL_LETTER_SHARP_S:
13635 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
13638 /* Use deprecated warning to increase the chances of this being
13641 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
13648 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
13650 /* This adds the string scalar <multi_string> to the array
13651 * <multi_char_matches>. <multi_string> is known to have exactly
13652 * <cp_count> code points in it. This is used when constructing a
13653 * bracketed character class and we find something that needs to match more
13654 * than a single character.
13656 * <multi_char_matches> is actually an array of arrays. Each top-level
13657 * element is an array that contains all the strings known so far that are
13658 * the same length. And that length (in number of code points) is the same
13659 * as the index of the top-level array. Hence, the [2] element is an
13660 * array, each element thereof is a string containing TWO code points;
13661 * while element [3] is for strings of THREE characters, and so on. Since
13662 * this is for multi-char strings there can never be a [0] nor [1] element.
13664 * When we rewrite the character class below, we will do so such that the
13665 * longest strings are written first, so that it prefers the longest
13666 * matching strings first. This is done even if it turns out that any
13667 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
13668 * Christiansen has agreed that this is ok. This makes the test for the
13669 * ligature 'ffi' come before the test for 'ff', for example */
13672 AV** this_array_ptr;
13674 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
13676 if (! multi_char_matches) {
13677 multi_char_matches = newAV();
13680 if (av_exists(multi_char_matches, cp_count)) {
13681 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
13682 this_array = *this_array_ptr;
13685 this_array = newAV();
13686 av_store(multi_char_matches, cp_count,
13689 av_push(this_array, multi_string);
13691 return multi_char_matches;
13694 /* The names of properties whose definitions are not known at compile time are
13695 * stored in this SV, after a constant heading. So if the length has been
13696 * changed since initialization, then there is a run-time definition. */
13697 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
13698 (SvCUR(listsv) != initial_listsv_len)
13701 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
13702 const bool stop_at_1, /* Just parse the next thing, don't
13703 look for a full character class */
13704 bool allow_multi_folds,
13705 const bool silence_non_portable, /* Don't output warnings
13708 SV** ret_invlist) /* Return an inversion list, not a node */
13710 /* parse a bracketed class specification. Most of these will produce an
13711 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
13712 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
13713 * under /i with multi-character folds: it will be rewritten following the
13714 * paradigm of this example, where the <multi-fold>s are characters which
13715 * fold to multiple character sequences:
13716 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
13717 * gets effectively rewritten as:
13718 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
13719 * reg() gets called (recursively) on the rewritten version, and this
13720 * function will return what it constructs. (Actually the <multi-fold>s
13721 * aren't physically removed from the [abcdefghi], it's just that they are
13722 * ignored in the recursion by means of a flag:
13723 * <RExC_in_multi_char_class>.)
13725 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
13726 * characters, with the corresponding bit set if that character is in the
13727 * list. For characters above this, a range list or swash is used. There
13728 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
13729 * determinable at compile time
13731 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
13732 * to be restarted. This can only happen if ret_invlist is non-NULL.
13735 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
13737 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
13740 IV namedclass = OOB_NAMEDCLASS;
13741 char *rangebegin = NULL;
13742 bool need_class = 0;
13744 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
13745 than just initialized. */
13746 SV* properties = NULL; /* Code points that match \p{} \P{} */
13747 SV* posixes = NULL; /* Code points that match classes like [:word:],
13748 extended beyond the Latin1 range. These have to
13749 be kept separate from other code points for much
13750 of this function because their handling is
13751 different under /i, and for most classes under
13753 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
13754 separate for a while from the non-complemented
13755 versions because of complications with /d
13757 UV element_count = 0; /* Number of distinct elements in the class.
13758 Optimizations may be possible if this is tiny */
13759 AV * multi_char_matches = NULL; /* Code points that fold to more than one
13760 character; used under /i */
13762 char * stop_ptr = RExC_end; /* where to stop parsing */
13763 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
13765 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
13767 /* Unicode properties are stored in a swash; this holds the current one
13768 * being parsed. If this swash is the only above-latin1 component of the
13769 * character class, an optimization is to pass it directly on to the
13770 * execution engine. Otherwise, it is set to NULL to indicate that there
13771 * are other things in the class that have to be dealt with at execution
13773 SV* swash = NULL; /* Code points that match \p{} \P{} */
13775 /* Set if a component of this character class is user-defined; just passed
13776 * on to the engine */
13777 bool has_user_defined_property = FALSE;
13779 /* inversion list of code points this node matches only when the target
13780 * string is in UTF-8. (Because is under /d) */
13781 SV* depends_list = NULL;
13783 /* Inversion list of code points this node matches regardless of things
13784 * like locale, folding, utf8ness of the target string */
13785 SV* cp_list = NULL;
13787 /* Like cp_list, but code points on this list need to be checked for things
13788 * that fold to/from them under /i */
13789 SV* cp_foldable_list = NULL;
13791 /* Like cp_list, but code points on this list are valid only when the
13792 * runtime locale is UTF-8 */
13793 SV* only_utf8_locale_list = NULL;
13796 /* In a range, counts how many 0-2 of the ends of it came from literals,
13797 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
13798 UV literal_endpoint = 0;
13800 /* Is the range unicode? which means on a platform that isn't 1-1 native
13801 * to Unicode (i.e. non-ASCII), each code point in it should be considered
13802 * to be a Unicode value. */
13803 bool unicode_range = FALSE;
13805 bool invert = FALSE; /* Is this class to be complemented */
13807 bool warn_super = ALWAYS_WARN_SUPER;
13809 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
13810 case we need to change the emitted regop to an EXACT. */
13811 const char * orig_parse = RExC_parse;
13812 const SSize_t orig_size = RExC_size;
13813 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
13814 GET_RE_DEBUG_FLAGS_DECL;
13816 PERL_ARGS_ASSERT_REGCLASS;
13818 PERL_UNUSED_ARG(depth);
13821 DEBUG_PARSE("clas");
13823 /* Assume we are going to generate an ANYOF node. */
13824 ret = reganode(pRExC_state,
13831 RExC_size += ANYOF_SKIP;
13832 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
13835 ANYOF_FLAGS(ret) = 0;
13837 RExC_emit += ANYOF_SKIP;
13838 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
13839 initial_listsv_len = SvCUR(listsv);
13840 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
13844 RExC_parse = regpatws(pRExC_state, RExC_parse,
13845 FALSE /* means don't recognize comments */ );
13848 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
13851 allow_multi_folds = FALSE;
13854 RExC_parse = regpatws(pRExC_state, RExC_parse,
13855 FALSE /* means don't recognize comments */ );
13859 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
13860 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
13861 const char *s = RExC_parse;
13862 const char c = *s++;
13864 while (isWORDCHAR(*s))
13866 if (*s && c == *s && s[1] == ']') {
13867 SAVEFREESV(RExC_rx_sv);
13869 "POSIX syntax [%c %c] belongs inside character classes",
13871 (void)ReREFCNT_inc(RExC_rx_sv);
13875 /* If the caller wants us to just parse a single element, accomplish this
13876 * by faking the loop ending condition */
13877 if (stop_at_1 && RExC_end > RExC_parse) {
13878 stop_ptr = RExC_parse + 1;
13881 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
13882 if (UCHARAT(RExC_parse) == ']')
13883 goto charclassloop;
13886 if (RExC_parse >= stop_ptr) {
13891 RExC_parse = regpatws(pRExC_state, RExC_parse,
13892 FALSE /* means don't recognize comments */ );
13895 if (UCHARAT(RExC_parse) == ']') {
13901 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
13902 save_value = value;
13903 save_prevvalue = prevvalue;
13906 rangebegin = RExC_parse;
13909 literal_endpoint = 0;
13913 value = utf8n_to_uvchr((U8*)RExC_parse,
13914 RExC_end - RExC_parse,
13915 &numlen, UTF8_ALLOW_DEFAULT);
13916 RExC_parse += numlen;
13919 value = UCHARAT(RExC_parse++);
13922 && RExC_parse < RExC_end
13923 && POSIXCC(UCHARAT(RExC_parse)))
13925 namedclass = regpposixcc(pRExC_state, value, strict);
13927 else if (value != '\\') {
13929 literal_endpoint++;
13933 /* Is a backslash; get the code point of the char after it */
13934 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
13935 value = utf8n_to_uvchr((U8*)RExC_parse,
13936 RExC_end - RExC_parse,
13937 &numlen, UTF8_ALLOW_DEFAULT);
13938 RExC_parse += numlen;
13941 value = UCHARAT(RExC_parse++);
13943 /* Some compilers cannot handle switching on 64-bit integer
13944 * values, therefore value cannot be an UV. Yes, this will
13945 * be a problem later if we want switch on Unicode.
13946 * A similar issue a little bit later when switching on
13947 * namedclass. --jhi */
13949 /* If the \ is escaping white space when white space is being
13950 * skipped, it means that that white space is wanted literally, and
13951 * is already in 'value'. Otherwise, need to translate the escape
13952 * into what it signifies. */
13953 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
13955 case 'w': namedclass = ANYOF_WORDCHAR; break;
13956 case 'W': namedclass = ANYOF_NWORDCHAR; break;
13957 case 's': namedclass = ANYOF_SPACE; break;
13958 case 'S': namedclass = ANYOF_NSPACE; break;
13959 case 'd': namedclass = ANYOF_DIGIT; break;
13960 case 'D': namedclass = ANYOF_NDIGIT; break;
13961 case 'v': namedclass = ANYOF_VERTWS; break;
13962 case 'V': namedclass = ANYOF_NVERTWS; break;
13963 case 'h': namedclass = ANYOF_HORIZWS; break;
13964 case 'H': namedclass = ANYOF_NHORIZWS; break;
13965 case 'N': /* Handle \N{NAME} in class */
13968 STRLEN cp_count = grok_bslash_N(pRExC_state, NULL, &value,
13969 flagp, depth, &as_text);
13970 if (*flagp & RESTART_UTF8)
13971 FAIL("panic: grok_bslash_N set RESTART_UTF8");
13972 if (cp_count != 1) { /* The typical case drops through */
13973 assert(cp_count != (STRLEN) -1);
13974 if (cp_count == 0) {
13976 RExC_parse++; /* Position after the "}" */
13977 vFAIL("Zero length \\N{}");
13980 ckWARNreg(RExC_parse,
13981 "Ignoring zero length \\N{} in character class");
13984 else { /* cp_count > 1 */
13985 if (! RExC_in_multi_char_class) {
13986 if (invert || range || *RExC_parse == '-') {
13989 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
13992 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
13997 = add_multi_match(multi_char_matches,
14001 break; /* <value> contains the first code
14002 point. Drop out of the switch to
14005 } /* End of cp_count != 1 */
14007 /* This element should not be processed further in this
14010 value = save_value;
14011 prevvalue = save_prevvalue;
14012 continue; /* Back to top of loop to get next char */
14014 /* Here, is a single code point, and <value> contains it */
14016 /* We consider named characters to be literal characters,
14017 * and they are Unicode */
14018 literal_endpoint++;
14019 unicode_range = TRUE;
14028 /* We will handle any undefined properties ourselves */
14029 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
14030 /* And we actually would prefer to get
14031 * the straight inversion list of the
14032 * swash, since we will be accessing it
14033 * anyway, to save a little time */
14034 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
14036 if (RExC_parse >= RExC_end)
14037 vFAIL2("Empty \\%c{}", (U8)value);
14038 if (*RExC_parse == '{') {
14039 const U8 c = (U8)value;
14040 e = strchr(RExC_parse++, '}');
14042 vFAIL2("Missing right brace on \\%c{}", c);
14043 while (isSPACE(*RExC_parse))
14045 if (e == RExC_parse)
14046 vFAIL2("Empty \\%c{}", c);
14047 n = e - RExC_parse;
14048 while (isSPACE(*(RExC_parse + n - 1)))
14059 if (UCHARAT(RExC_parse) == '^') {
14062 /* toggle. (The rhs xor gets the single bit that
14063 * differs between P and p; the other xor inverts just
14065 value ^= 'P' ^ 'p';
14067 while (isSPACE(*RExC_parse)) {
14072 /* Try to get the definition of the property into
14073 * <invlist>. If /i is in effect, the effective property
14074 * will have its name be <__NAME_i>. The design is
14075 * discussed in commit
14076 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
14077 name = savepv(Perl_form(aTHX_
14079 (FOLD) ? "__" : "",
14085 /* Look up the property name, and get its swash and
14086 * inversion list, if the property is found */
14088 SvREFCNT_dec_NN(swash);
14090 swash = _core_swash_init("utf8", name, &PL_sv_undef,
14093 NULL, /* No inversion list */
14096 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
14097 HV* curpkg = (IN_PERL_COMPILETIME)
14099 : CopSTASH(PL_curcop);
14101 SvREFCNT_dec_NN(swash);
14105 /* Here didn't find it. It could be a user-defined
14106 * property that will be available at run-time. If we
14107 * accept only compile-time properties, is an error;
14108 * otherwise add it to the list for run-time look up */
14110 RExC_parse = e + 1;
14112 "Property '%"UTF8f"' is unknown",
14113 UTF8fARG(UTF, n, name));
14116 /* If the property name doesn't already have a package
14117 * name, add the current one to it so that it can be
14118 * referred to outside it. [perl #121777] */
14119 if (curpkg && ! instr(name, "::")) {
14120 char* pkgname = HvNAME(curpkg);
14121 if (strNE(pkgname, "main")) {
14122 char* full_name = Perl_form(aTHX_
14126 n = strlen(full_name);
14128 name = savepvn(full_name, n);
14131 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%"UTF8f"\n",
14132 (value == 'p' ? '+' : '!'),
14133 UTF8fARG(UTF, n, name));
14134 has_user_defined_property = TRUE;
14136 /* We don't know yet, so have to assume that the
14137 * property could match something in the Latin1 range,
14138 * hence something that isn't utf8. Note that this
14139 * would cause things in <depends_list> to match
14140 * inappropriately, except that any \p{}, including
14141 * this one forces Unicode semantics, which means there
14142 * is no <depends_list> */
14144 |= ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES;
14148 /* Here, did get the swash and its inversion list. If
14149 * the swash is from a user-defined property, then this
14150 * whole character class should be regarded as such */
14151 if (swash_init_flags
14152 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
14154 has_user_defined_property = TRUE;
14157 /* We warn on matching an above-Unicode code point
14158 * if the match would return true, except don't
14159 * warn for \p{All}, which has exactly one element
14161 (_invlist_contains_cp(invlist, 0x110000)
14162 && (! (_invlist_len(invlist) == 1
14163 && *invlist_array(invlist) == 0)))
14169 /* Invert if asking for the complement */
14170 if (value == 'P') {
14171 _invlist_union_complement_2nd(properties,
14175 /* The swash can't be used as-is, because we've
14176 * inverted things; delay removing it to here after
14177 * have copied its invlist above */
14178 SvREFCNT_dec_NN(swash);
14182 _invlist_union(properties, invlist, &properties);
14187 RExC_parse = e + 1;
14188 namedclass = ANYOF_UNIPROP; /* no official name, but it's
14191 /* \p means they want Unicode semantics */
14192 RExC_uni_semantics = 1;
14195 case 'n': value = '\n'; break;
14196 case 'r': value = '\r'; break;
14197 case 't': value = '\t'; break;
14198 case 'f': value = '\f'; break;
14199 case 'b': value = '\b'; break;
14200 case 'e': value = ESC_NATIVE; break;
14201 case 'a': value = '\a'; break;
14203 RExC_parse--; /* function expects to be pointed at the 'o' */
14205 const char* error_msg;
14206 bool valid = grok_bslash_o(&RExC_parse,
14209 PASS2, /* warnings only in
14212 silence_non_portable,
14218 if (IN_ENCODING && value < 0x100) {
14219 goto recode_encoding;
14223 RExC_parse--; /* function expects to be pointed at the 'x' */
14225 const char* error_msg;
14226 bool valid = grok_bslash_x(&RExC_parse,
14229 PASS2, /* Output warnings */
14231 silence_non_portable,
14237 if (IN_ENCODING && value < 0x100)
14238 goto recode_encoding;
14241 value = grok_bslash_c(*RExC_parse++, PASS2);
14243 case '0': case '1': case '2': case '3': case '4':
14244 case '5': case '6': case '7':
14246 /* Take 1-3 octal digits */
14247 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
14248 numlen = (strict) ? 4 : 3;
14249 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
14250 RExC_parse += numlen;
14253 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
14254 vFAIL("Need exactly 3 octal digits");
14256 else if (! SIZE_ONLY /* like \08, \178 */
14258 && RExC_parse < RExC_end
14259 && isDIGIT(*RExC_parse)
14260 && ckWARN(WARN_REGEXP))
14262 SAVEFREESV(RExC_rx_sv);
14263 reg_warn_non_literal_string(
14265 form_short_octal_warning(RExC_parse, numlen));
14266 (void)ReREFCNT_inc(RExC_rx_sv);
14269 if (IN_ENCODING && value < 0x100)
14270 goto recode_encoding;
14274 if (! RExC_override_recoding) {
14275 SV* enc = _get_encoding();
14276 value = reg_recode((const char)(U8)value, &enc);
14279 vFAIL("Invalid escape in the specified encoding");
14282 ckWARNreg(RExC_parse,
14283 "Invalid escape in the specified encoding");
14289 /* Allow \_ to not give an error */
14290 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
14292 vFAIL2("Unrecognized escape \\%c in character class",
14296 SAVEFREESV(RExC_rx_sv);
14297 ckWARN2reg(RExC_parse,
14298 "Unrecognized escape \\%c in character class passed through",
14300 (void)ReREFCNT_inc(RExC_rx_sv);
14304 } /* End of switch on char following backslash */
14305 } /* end of handling backslash escape sequences */
14307 /* Here, we have the current token in 'value' */
14309 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
14312 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
14313 * literal, as is the character that began the false range, i.e.
14314 * the 'a' in the examples */
14317 const int w = (RExC_parse >= rangebegin)
14318 ? RExC_parse - rangebegin
14322 "False [] range \"%"UTF8f"\"",
14323 UTF8fARG(UTF, w, rangebegin));
14326 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
14327 ckWARN2reg(RExC_parse,
14328 "False [] range \"%"UTF8f"\"",
14329 UTF8fARG(UTF, w, rangebegin));
14330 (void)ReREFCNT_inc(RExC_rx_sv);
14331 cp_list = add_cp_to_invlist(cp_list, '-');
14332 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
14337 range = 0; /* this was not a true range */
14338 element_count += 2; /* So counts for three values */
14341 classnum = namedclass_to_classnum(namedclass);
14343 if (LOC && namedclass < ANYOF_POSIXL_MAX
14344 #ifndef HAS_ISASCII
14345 && classnum != _CC_ASCII
14348 /* What the Posix classes (like \w, [:space:]) match in locale
14349 * isn't knowable under locale until actual match time. Room
14350 * must be reserved (one time per outer bracketed class) to
14351 * store such classes. The space will contain a bit for each
14352 * named class that is to be matched against. This isn't
14353 * needed for \p{} and pseudo-classes, as they are not affected
14354 * by locale, and hence are dealt with separately */
14355 if (! need_class) {
14358 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
14361 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
14363 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
14364 ANYOF_POSIXL_ZERO(ret);
14367 /* Coverity thinks it is possible for this to be negative; both
14368 * jhi and khw think it's not, but be safer */
14369 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
14370 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
14372 /* See if it already matches the complement of this POSIX
14374 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
14375 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
14379 posixl_matches_all = TRUE;
14380 break; /* No need to continue. Since it matches both
14381 e.g., \w and \W, it matches everything, and the
14382 bracketed class can be optimized into qr/./s */
14385 /* Add this class to those that should be checked at runtime */
14386 ANYOF_POSIXL_SET(ret, namedclass);
14388 /* The above-Latin1 characters are not subject to locale rules.
14389 * Just add them, in the second pass, to the
14390 * unconditionally-matched list */
14392 SV* scratch_list = NULL;
14394 /* Get the list of the above-Latin1 code points this
14396 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
14397 PL_XPosix_ptrs[classnum],
14399 /* Odd numbers are complements, like
14400 * NDIGIT, NASCII, ... */
14401 namedclass % 2 != 0,
14403 /* Checking if 'cp_list' is NULL first saves an extra
14404 * clone. Its reference count will be decremented at the
14405 * next union, etc, or if this is the only instance, at the
14406 * end of the routine */
14408 cp_list = scratch_list;
14411 _invlist_union(cp_list, scratch_list, &cp_list);
14412 SvREFCNT_dec_NN(scratch_list);
14414 continue; /* Go get next character */
14417 else if (! SIZE_ONLY) {
14419 /* Here, not in pass1 (in that pass we skip calculating the
14420 * contents of this class), and is /l, or is a POSIX class for
14421 * which /l doesn't matter (or is a Unicode property, which is
14422 * skipped here). */
14423 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
14424 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
14426 /* Here, should be \h, \H, \v, or \V. None of /d, /i
14427 * nor /l make a difference in what these match,
14428 * therefore we just add what they match to cp_list. */
14429 if (classnum != _CC_VERTSPACE) {
14430 assert( namedclass == ANYOF_HORIZWS
14431 || namedclass == ANYOF_NHORIZWS);
14433 /* It turns out that \h is just a synonym for
14435 classnum = _CC_BLANK;
14438 _invlist_union_maybe_complement_2nd(
14440 PL_XPosix_ptrs[classnum],
14441 namedclass % 2 != 0, /* Complement if odd
14442 (NHORIZWS, NVERTWS)
14447 else { /* Garden variety class. If is NASCII, NDIGIT, ...
14448 complement and use nposixes */
14449 SV** posixes_ptr = namedclass % 2 == 0
14452 SV** source_ptr = &PL_XPosix_ptrs[classnum];
14453 _invlist_union_maybe_complement_2nd(
14456 namedclass % 2 != 0,
14460 } /* end of namedclass \blah */
14463 RExC_parse = regpatws(pRExC_state, RExC_parse,
14464 FALSE /* means don't recognize comments */ );
14467 /* If 'range' is set, 'value' is the ending of a range--check its
14468 * validity. (If value isn't a single code point in the case of a
14469 * range, we should have figured that out above in the code that
14470 * catches false ranges). Later, we will handle each individual code
14471 * point in the range. If 'range' isn't set, this could be the
14472 * beginning of a range, so check for that by looking ahead to see if
14473 * the next real character to be processed is the range indicator--the
14478 /* For unicode ranges, we have to test that the Unicode as opposed
14479 * to the native values are not decreasing. (Above 255, and there
14480 * is no difference between native and Unicode) */
14481 if (unicode_range && prevvalue < 255 && value < 255) {
14482 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
14483 goto backwards_range;
14488 if (prevvalue > value) /* b-a */ {
14493 w = RExC_parse - rangebegin;
14495 "Invalid [] range \"%"UTF8f"\"",
14496 UTF8fARG(UTF, w, rangebegin));
14497 NOT_REACHED; /* NOT REACHED */
14501 prevvalue = value; /* save the beginning of the potential range */
14502 if (! stop_at_1 /* Can't be a range if parsing just one thing */
14503 && *RExC_parse == '-')
14505 char* next_char_ptr = RExC_parse + 1;
14506 if (skip_white) { /* Get the next real char after the '-' */
14507 next_char_ptr = regpatws(pRExC_state,
14509 FALSE); /* means don't recognize
14513 /* If the '-' is at the end of the class (just before the ']',
14514 * it is a literal minus; otherwise it is a range */
14515 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
14516 RExC_parse = next_char_ptr;
14518 /* a bad range like \w-, [:word:]- ? */
14519 if (namedclass > OOB_NAMEDCLASS) {
14520 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
14521 const int w = RExC_parse >= rangebegin
14522 ? RExC_parse - rangebegin
14525 vFAIL4("False [] range \"%*.*s\"",
14530 "False [] range \"%*.*s\"",
14535 cp_list = add_cp_to_invlist(cp_list, '-');
14539 range = 1; /* yeah, it's a range! */
14540 continue; /* but do it the next time */
14545 if (namedclass > OOB_NAMEDCLASS) {
14549 /* Here, we have a single value this time through the loop, and
14550 * <prevvalue> is the beginning of the range, if any; or <value> if
14553 /* non-Latin1 code point implies unicode semantics. Must be set in
14554 * pass1 so is there for the whole of pass 2 */
14556 RExC_uni_semantics = 1;
14559 /* Ready to process either the single value, or the completed range.
14560 * For single-valued non-inverted ranges, we consider the possibility
14561 * of multi-char folds. (We made a conscious decision to not do this
14562 * for the other cases because it can often lead to non-intuitive
14563 * results. For example, you have the peculiar case that:
14564 * "s s" =~ /^[^\xDF]+$/i => Y
14565 * "ss" =~ /^[^\xDF]+$/i => N
14567 * See [perl #89750] */
14568 if (FOLD && allow_multi_folds && value == prevvalue) {
14569 if (value == LATIN_SMALL_LETTER_SHARP_S
14570 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
14573 /* Here <value> is indeed a multi-char fold. Get what it is */
14575 U8 foldbuf[UTF8_MAXBYTES_CASE];
14578 UV folded = _to_uni_fold_flags(
14582 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
14583 ? FOLD_FLAGS_NOMIX_ASCII
14587 /* Here, <folded> should be the first character of the
14588 * multi-char fold of <value>, with <foldbuf> containing the
14589 * whole thing. But, if this fold is not allowed (because of
14590 * the flags), <fold> will be the same as <value>, and should
14591 * be processed like any other character, so skip the special
14593 if (folded != value) {
14595 /* Skip if we are recursed, currently parsing the class
14596 * again. Otherwise add this character to the list of
14597 * multi-char folds. */
14598 if (! RExC_in_multi_char_class) {
14599 STRLEN cp_count = utf8_length(foldbuf,
14600 foldbuf + foldlen);
14601 SV* multi_fold = sv_2mortal(newSVpvs(""));
14603 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
14606 = add_multi_match(multi_char_matches,
14612 /* This element should not be processed further in this
14615 value = save_value;
14616 prevvalue = save_prevvalue;
14622 /* Deal with this element of the class */
14625 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
14628 /* On non-ASCII platforms, for ranges that span all of 0..255, and
14629 * ones that don't require special handling, we can just add the
14630 * range like we do for ASCII platforms */
14631 if ((UNLIKELY(prevvalue == 0) && value >= 255)
14632 || ! (prevvalue < 256
14634 || (literal_endpoint == 2
14635 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
14636 || (isUPPER_A(prevvalue)
14637 && isUPPER_A(value)))))))
14639 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
14643 /* Here, requires special handling. This can be because it is
14644 * a range whose code points are considered to be Unicode, and
14645 * so must be individually translated into native, or because
14646 * its a subrange of 'A-Z' or 'a-z' which each aren't
14647 * contiguous in EBCDIC, but we have defined them to include
14648 * only the "expected" upper or lower case ASCII alphabetics.
14649 * Subranges above 255 are the same in native and Unicode, so
14650 * can be added as a range */
14651 U8 start = NATIVE_TO_LATIN1(prevvalue);
14653 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
14654 for (j = start; j <= end; j++) {
14655 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
14658 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
14665 range = 0; /* this range (if it was one) is done now */
14666 } /* End of loop through all the text within the brackets */
14668 /* If anything in the class expands to more than one character, we have to
14669 * deal with them by building up a substitute parse string, and recursively
14670 * calling reg() on it, instead of proceeding */
14671 if (multi_char_matches) {
14672 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
14675 char *save_end = RExC_end;
14676 char *save_parse = RExC_parse;
14677 bool first_time = TRUE; /* First multi-char occurrence doesn't get
14682 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
14683 because too confusing */
14685 sv_catpv(substitute_parse, "(?:");
14689 /* Look at the longest folds first */
14690 for (cp_count = av_tindex(multi_char_matches); cp_count > 0; cp_count--) {
14692 if (av_exists(multi_char_matches, cp_count)) {
14693 AV** this_array_ptr;
14696 this_array_ptr = (AV**) av_fetch(multi_char_matches,
14698 while ((this_sequence = av_pop(*this_array_ptr)) !=
14701 if (! first_time) {
14702 sv_catpv(substitute_parse, "|");
14704 first_time = FALSE;
14706 sv_catpv(substitute_parse, SvPVX(this_sequence));
14711 /* If the character class contains anything else besides these
14712 * multi-character folds, have to include it in recursive parsing */
14713 if (element_count) {
14714 sv_catpv(substitute_parse, "|[");
14715 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
14716 sv_catpv(substitute_parse, "]");
14719 sv_catpv(substitute_parse, ")");
14722 /* This is a way to get the parse to skip forward a whole named
14723 * sequence instead of matching the 2nd character when it fails the
14725 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
14729 RExC_parse = SvPV(substitute_parse, len);
14730 RExC_end = RExC_parse + len;
14731 RExC_in_multi_char_class = 1;
14732 RExC_override_recoding = 1;
14733 RExC_emit = (regnode *)orig_emit;
14735 ret = reg(pRExC_state, 1, ®_flags, depth+1);
14737 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
14739 RExC_parse = save_parse;
14740 RExC_end = save_end;
14741 RExC_in_multi_char_class = 0;
14742 RExC_override_recoding = 0;
14743 SvREFCNT_dec_NN(multi_char_matches);
14747 /* Here, we've gone through the entire class and dealt with multi-char
14748 * folds. We are now in a position that we can do some checks to see if we
14749 * can optimize this ANYOF node into a simpler one, even in Pass 1.
14750 * Currently we only do two checks:
14751 * 1) is in the unlikely event that the user has specified both, eg. \w and
14752 * \W under /l, then the class matches everything. (This optimization
14753 * is done only to make the optimizer code run later work.)
14754 * 2) if the character class contains only a single element (including a
14755 * single range), we see if there is an equivalent node for it.
14756 * Other checks are possible */
14757 if (! ret_invlist /* Can't optimize if returning the constructed
14759 && (UNLIKELY(posixl_matches_all) || element_count == 1))
14764 if (UNLIKELY(posixl_matches_all)) {
14767 else if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like
14768 \w or [:digit:] or \p{foo}
14771 /* All named classes are mapped into POSIXish nodes, with its FLAG
14772 * argument giving which class it is */
14773 switch ((I32)namedclass) {
14774 case ANYOF_UNIPROP:
14777 /* These don't depend on the charset modifiers. They always
14778 * match under /u rules */
14779 case ANYOF_NHORIZWS:
14780 case ANYOF_HORIZWS:
14781 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
14784 case ANYOF_NVERTWS:
14789 /* The actual POSIXish node for all the rest depends on the
14790 * charset modifier. The ones in the first set depend only on
14791 * ASCII or, if available on this platform, locale */
14795 op = (LOC) ? POSIXL : POSIXA;
14806 /* under /a could be alpha */
14808 if (ASCII_RESTRICTED) {
14809 namedclass = ANYOF_ALPHA + (namedclass % 2);
14817 /* The rest have more possibilities depending on the charset.
14818 * We take advantage of the enum ordering of the charset
14819 * modifiers to get the exact node type, */
14821 op = POSIXD + get_regex_charset(RExC_flags);
14822 if (op > POSIXA) { /* /aa is same as /a */
14827 /* The odd numbered ones are the complements of the
14828 * next-lower even number one */
14829 if (namedclass % 2 == 1) {
14833 arg = namedclass_to_classnum(namedclass);
14837 else if (value == prevvalue) {
14839 /* Here, the class consists of just a single code point */
14842 if (! LOC && value == '\n') {
14843 op = REG_ANY; /* Optimize [^\n] */
14844 *flagp |= HASWIDTH|SIMPLE;
14848 else if (value < 256 || UTF) {
14850 /* Optimize a single value into an EXACTish node, but not if it
14851 * would require converting the pattern to UTF-8. */
14852 op = compute_EXACTish(pRExC_state);
14854 } /* Otherwise is a range */
14855 else if (! LOC) { /* locale could vary these */
14856 if (prevvalue == '0') {
14857 if (value == '9') {
14862 else if (prevvalue == 'A') {
14865 && literal_endpoint == 2
14868 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
14872 else if (prevvalue == 'a') {
14875 && literal_endpoint == 2
14878 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
14884 /* Here, we have changed <op> away from its initial value iff we found
14885 * an optimization */
14888 /* Throw away this ANYOF regnode, and emit the calculated one,
14889 * which should correspond to the beginning, not current, state of
14891 const char * cur_parse = RExC_parse;
14892 RExC_parse = (char *)orig_parse;
14896 /* To get locale nodes to not use the full ANYOF size would
14897 * require moving the code above that writes the portions
14898 * of it that aren't in other nodes to after this point.
14899 * e.g. ANYOF_POSIXL_SET */
14900 RExC_size = orig_size;
14904 RExC_emit = (regnode *)orig_emit;
14905 if (PL_regkind[op] == POSIXD) {
14906 if (op == POSIXL) {
14907 RExC_contains_locale = 1;
14910 op += NPOSIXD - POSIXD;
14915 ret = reg_node(pRExC_state, op);
14917 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
14921 *flagp |= HASWIDTH|SIMPLE;
14923 else if (PL_regkind[op] == EXACT) {
14924 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
14925 TRUE /* downgradable to EXACT */
14929 RExC_parse = (char *) cur_parse;
14931 SvREFCNT_dec(posixes);
14932 SvREFCNT_dec(nposixes);
14933 SvREFCNT_dec(cp_list);
14934 SvREFCNT_dec(cp_foldable_list);
14941 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
14943 /* If folding, we calculate all characters that could fold to or from the
14944 * ones already on the list */
14945 if (cp_foldable_list) {
14947 UV start, end; /* End points of code point ranges */
14949 SV* fold_intersection = NULL;
14952 /* Our calculated list will be for Unicode rules. For locale
14953 * matching, we have to keep a separate list that is consulted at
14954 * runtime only when the locale indicates Unicode rules. For
14955 * non-locale, we just use to the general list */
14957 use_list = &only_utf8_locale_list;
14960 use_list = &cp_list;
14963 /* Only the characters in this class that participate in folds need
14964 * be checked. Get the intersection of this class and all the
14965 * possible characters that are foldable. This can quickly narrow
14966 * down a large class */
14967 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
14968 &fold_intersection);
14970 /* The folds for all the Latin1 characters are hard-coded into this
14971 * program, but we have to go out to disk to get the others. */
14972 if (invlist_highest(cp_foldable_list) >= 256) {
14974 /* This is a hash that for a particular fold gives all
14975 * characters that are involved in it */
14976 if (! PL_utf8_foldclosures) {
14977 _load_PL_utf8_foldclosures();
14981 /* Now look at the foldable characters in this class individually */
14982 invlist_iterinit(fold_intersection);
14983 while (invlist_iternext(fold_intersection, &start, &end)) {
14986 /* Look at every character in the range */
14987 for (j = start; j <= end; j++) {
14988 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
14994 if (IS_IN_SOME_FOLD_L1(j)) {
14996 /* ASCII is always matched; non-ASCII is matched
14997 * only under Unicode rules (which could happen
14998 * under /l if the locale is a UTF-8 one */
14999 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
15000 *use_list = add_cp_to_invlist(*use_list,
15001 PL_fold_latin1[j]);
15005 add_cp_to_invlist(depends_list,
15006 PL_fold_latin1[j]);
15010 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
15011 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
15013 add_above_Latin1_folds(pRExC_state,
15020 /* Here is an above Latin1 character. We don't have the
15021 * rules hard-coded for it. First, get its fold. This is
15022 * the simple fold, as the multi-character folds have been
15023 * handled earlier and separated out */
15024 _to_uni_fold_flags(j, foldbuf, &foldlen,
15025 (ASCII_FOLD_RESTRICTED)
15026 ? FOLD_FLAGS_NOMIX_ASCII
15029 /* Single character fold of above Latin1. Add everything in
15030 * its fold closure to the list that this node should match.
15031 * The fold closures data structure is a hash with the keys
15032 * being the UTF-8 of every character that is folded to, like
15033 * 'k', and the values each an array of all code points that
15034 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
15035 * Multi-character folds are not included */
15036 if ((listp = hv_fetch(PL_utf8_foldclosures,
15037 (char *) foldbuf, foldlen, FALSE)))
15039 AV* list = (AV*) *listp;
15041 for (k = 0; k <= av_tindex(list); k++) {
15042 SV** c_p = av_fetch(list, k, FALSE);
15048 /* /aa doesn't allow folds between ASCII and non- */
15049 if ((ASCII_FOLD_RESTRICTED
15050 && (isASCII(c) != isASCII(j))))
15055 /* Folds under /l which cross the 255/256 boundary
15056 * are added to a separate list. (These are valid
15057 * only when the locale is UTF-8.) */
15058 if (c < 256 && LOC) {
15059 *use_list = add_cp_to_invlist(*use_list, c);
15063 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
15065 cp_list = add_cp_to_invlist(cp_list, c);
15068 /* Similarly folds involving non-ascii Latin1
15069 * characters under /d are added to their list */
15070 depends_list = add_cp_to_invlist(depends_list,
15077 SvREFCNT_dec_NN(fold_intersection);
15080 /* Now that we have finished adding all the folds, there is no reason
15081 * to keep the foldable list separate */
15082 _invlist_union(cp_list, cp_foldable_list, &cp_list);
15083 SvREFCNT_dec_NN(cp_foldable_list);
15086 /* And combine the result (if any) with any inversion list from posix
15087 * classes. The lists are kept separate up to now because we don't want to
15088 * fold the classes (folding of those is automatically handled by the swash
15089 * fetching code) */
15090 if (posixes || nposixes) {
15091 if (posixes && AT_LEAST_ASCII_RESTRICTED) {
15092 /* Under /a and /aa, nothing above ASCII matches these */
15093 _invlist_intersection(posixes,
15094 PL_XPosix_ptrs[_CC_ASCII],
15098 if (DEPENDS_SEMANTICS) {
15099 /* Under /d, everything in the upper half of the Latin1 range
15100 * matches these complements */
15101 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_ALL_NON_UTF8_NON_ASCII;
15103 else if (AT_LEAST_ASCII_RESTRICTED) {
15104 /* Under /a and /aa, everything above ASCII matches these
15106 _invlist_union_complement_2nd(nposixes,
15107 PL_XPosix_ptrs[_CC_ASCII],
15111 _invlist_union(posixes, nposixes, &posixes);
15112 SvREFCNT_dec_NN(nposixes);
15115 posixes = nposixes;
15118 if (! DEPENDS_SEMANTICS) {
15120 _invlist_union(cp_list, posixes, &cp_list);
15121 SvREFCNT_dec_NN(posixes);
15128 /* Under /d, we put into a separate list the Latin1 things that
15129 * match only when the target string is utf8 */
15130 SV* nonascii_but_latin1_properties = NULL;
15131 _invlist_intersection(posixes, PL_UpperLatin1,
15132 &nonascii_but_latin1_properties);
15133 _invlist_subtract(posixes, nonascii_but_latin1_properties,
15136 _invlist_union(cp_list, posixes, &cp_list);
15137 SvREFCNT_dec_NN(posixes);
15143 if (depends_list) {
15144 _invlist_union(depends_list, nonascii_but_latin1_properties,
15146 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
15149 depends_list = nonascii_but_latin1_properties;
15154 /* And combine the result (if any) with any inversion list from properties.
15155 * The lists are kept separate up to now so that we can distinguish the two
15156 * in regards to matching above-Unicode. A run-time warning is generated
15157 * if a Unicode property is matched against a non-Unicode code point. But,
15158 * we allow user-defined properties to match anything, without any warning,
15159 * and we also suppress the warning if there is a portion of the character
15160 * class that isn't a Unicode property, and which matches above Unicode, \W
15161 * or [\x{110000}] for example.
15162 * (Note that in this case, unlike the Posix one above, there is no
15163 * <depends_list>, because having a Unicode property forces Unicode
15168 /* If it matters to the final outcome, see if a non-property
15169 * component of the class matches above Unicode. If so, the
15170 * warning gets suppressed. This is true even if just a single
15171 * such code point is specified, as though not strictly correct if
15172 * another such code point is matched against, the fact that they
15173 * are using above-Unicode code points indicates they should know
15174 * the issues involved */
15176 warn_super = ! (invert
15177 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
15180 _invlist_union(properties, cp_list, &cp_list);
15181 SvREFCNT_dec_NN(properties);
15184 cp_list = properties;
15188 ANYOF_FLAGS(ret) |= ANYOF_WARN_SUPER;
15192 /* Here, we have calculated what code points should be in the character
15195 * Now we can see about various optimizations. Fold calculation (which we
15196 * did above) needs to take place before inversion. Otherwise /[^k]/i
15197 * would invert to include K, which under /i would match k, which it
15198 * shouldn't. Therefore we can't invert folded locale now, as it won't be
15199 * folded until runtime */
15201 /* If we didn't do folding, it's because some information isn't available
15202 * until runtime; set the run-time fold flag for these. (We don't have to
15203 * worry about properties folding, as that is taken care of by the swash
15204 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
15205 * locales, or the class matches at least one 0-255 range code point */
15207 if (only_utf8_locale_list) {
15208 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
15210 else if (cp_list) { /* Look to see if there a 0-255 code point is in
15213 invlist_iterinit(cp_list);
15214 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
15215 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
15217 invlist_iterfinish(cp_list);
15221 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
15222 * at compile time. Besides not inverting folded locale now, we can't
15223 * invert if there are things such as \w, which aren't known until runtime
15227 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
15229 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
15231 _invlist_invert(cp_list);
15233 /* Any swash can't be used as-is, because we've inverted things */
15235 SvREFCNT_dec_NN(swash);
15239 /* Clear the invert flag since have just done it here */
15244 *ret_invlist = cp_list;
15245 SvREFCNT_dec(swash);
15247 /* Discard the generated node */
15249 RExC_size = orig_size;
15252 RExC_emit = orig_emit;
15257 /* Some character classes are equivalent to other nodes. Such nodes take
15258 * up less room and generally fewer operations to execute than ANYOF nodes.
15259 * Above, we checked for and optimized into some such equivalents for
15260 * certain common classes that are easy to test. Getting to this point in
15261 * the code means that the class didn't get optimized there. Since this
15262 * code is only executed in Pass 2, it is too late to save space--it has
15263 * been allocated in Pass 1, and currently isn't given back. But turning
15264 * things into an EXACTish node can allow the optimizer to join it to any
15265 * adjacent such nodes. And if the class is equivalent to things like /./,
15266 * expensive run-time swashes can be avoided. Now that we have more
15267 * complete information, we can find things necessarily missed by the
15268 * earlier code. I (khw) am not sure how much to look for here. It would
15269 * be easy, but perhaps too slow, to check any candidates against all the
15270 * node types they could possibly match using _invlistEQ(). */
15275 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
15276 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
15278 /* We don't optimize if we are supposed to make sure all non-Unicode
15279 * code points raise a warning, as only ANYOF nodes have this check.
15281 && ! ((ANYOF_FLAGS(ret) & ANYOF_WARN_SUPER) && ALWAYS_WARN_SUPER))
15284 U8 op = END; /* The optimzation node-type */
15285 const char * cur_parse= RExC_parse;
15287 invlist_iterinit(cp_list);
15288 if (! invlist_iternext(cp_list, &start, &end)) {
15290 /* Here, the list is empty. This happens, for example, when a
15291 * Unicode property is the only thing in the character class, and
15292 * it doesn't match anything. (perluniprops.pod notes such
15295 *flagp |= HASWIDTH|SIMPLE;
15297 else if (start == end) { /* The range is a single code point */
15298 if (! invlist_iternext(cp_list, &start, &end)
15300 /* Don't do this optimization if it would require changing
15301 * the pattern to UTF-8 */
15302 && (start < 256 || UTF))
15304 /* Here, the list contains a single code point. Can optimize
15305 * into an EXACTish node */
15316 /* A locale node under folding with one code point can be
15317 * an EXACTFL, as its fold won't be calculated until
15323 /* Here, we are generally folding, but there is only one
15324 * code point to match. If we have to, we use an EXACT
15325 * node, but it would be better for joining with adjacent
15326 * nodes in the optimization pass if we used the same
15327 * EXACTFish node that any such are likely to be. We can
15328 * do this iff the code point doesn't participate in any
15329 * folds. For example, an EXACTF of a colon is the same as
15330 * an EXACT one, since nothing folds to or from a colon. */
15332 if (IS_IN_SOME_FOLD_L1(value)) {
15337 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
15342 /* If we haven't found the node type, above, it means we
15343 * can use the prevailing one */
15345 op = compute_EXACTish(pRExC_state);
15350 else if (start == 0) {
15351 if (end == UV_MAX) {
15353 *flagp |= HASWIDTH|SIMPLE;
15356 else if (end == '\n' - 1
15357 && invlist_iternext(cp_list, &start, &end)
15358 && start == '\n' + 1 && end == UV_MAX)
15361 *flagp |= HASWIDTH|SIMPLE;
15365 invlist_iterfinish(cp_list);
15368 RExC_parse = (char *)orig_parse;
15369 RExC_emit = (regnode *)orig_emit;
15371 ret = reg_node(pRExC_state, op);
15373 RExC_parse = (char *)cur_parse;
15375 if (PL_regkind[op] == EXACT) {
15376 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
15377 TRUE /* downgradable to EXACT */
15381 SvREFCNT_dec_NN(cp_list);
15386 /* Here, <cp_list> contains all the code points we can determine at
15387 * compile time that match under all conditions. Go through it, and
15388 * for things that belong in the bitmap, put them there, and delete from
15389 * <cp_list>. While we are at it, see if everything above 255 is in the
15390 * list, and if so, set a flag to speed up execution */
15392 populate_ANYOF_from_invlist(ret, &cp_list);
15395 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
15398 /* Here, the bitmap has been populated with all the Latin1 code points that
15399 * always match. Can now add to the overall list those that match only
15400 * when the target string is UTF-8 (<depends_list>). */
15401 if (depends_list) {
15403 _invlist_union(cp_list, depends_list, &cp_list);
15404 SvREFCNT_dec_NN(depends_list);
15407 cp_list = depends_list;
15409 ANYOF_FLAGS(ret) |= ANYOF_HAS_UTF8_NONBITMAP_MATCHES;
15412 /* If there is a swash and more than one element, we can't use the swash in
15413 * the optimization below. */
15414 if (swash && element_count > 1) {
15415 SvREFCNT_dec_NN(swash);
15419 /* Note that the optimization of using 'swash' if it is the only thing in
15420 * the class doesn't have us change swash at all, so it can include things
15421 * that are also in the bitmap; otherwise we have purposely deleted that
15422 * duplicate information */
15423 set_ANYOF_arg(pRExC_state, ret, cp_list,
15424 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
15426 only_utf8_locale_list,
15427 swash, has_user_defined_property);
15429 *flagp |= HASWIDTH|SIMPLE;
15431 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
15432 RExC_contains_locale = 1;
15438 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
15441 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
15442 regnode* const node,
15444 SV* const runtime_defns,
15445 SV* const only_utf8_locale_list,
15447 const bool has_user_defined_property)
15449 /* Sets the arg field of an ANYOF-type node 'node', using information about
15450 * the node passed-in. If there is nothing outside the node's bitmap, the
15451 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
15452 * the count returned by add_data(), having allocated and stored an array,
15453 * av, that that count references, as follows:
15454 * av[0] stores the character class description in its textual form.
15455 * This is used later (regexec.c:Perl_regclass_swash()) to
15456 * initialize the appropriate swash, and is also useful for dumping
15457 * the regnode. This is set to &PL_sv_undef if the textual
15458 * description is not needed at run-time (as happens if the other
15459 * elements completely define the class)
15460 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
15461 * computed from av[0]. But if no further computation need be done,
15462 * the swash is stored here now (and av[0] is &PL_sv_undef).
15463 * av[2] stores the inversion list of code points that match only if the
15464 * current locale is UTF-8
15465 * av[3] stores the cp_list inversion list for use in addition or instead
15466 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
15467 * (Otherwise everything needed is already in av[0] and av[1])
15468 * av[4] is set if any component of the class is from a user-defined
15469 * property; used only if av[3] exists */
15473 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
15475 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
15476 assert(! (ANYOF_FLAGS(node)
15477 & (ANYOF_HAS_UTF8_NONBITMAP_MATCHES
15478 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES)));
15479 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
15482 AV * const av = newAV();
15485 assert(ANYOF_FLAGS(node)
15486 & (ANYOF_HAS_UTF8_NONBITMAP_MATCHES
15487 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES|ANYOF_LOC_FOLD));
15489 av_store(av, 0, (runtime_defns)
15490 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
15493 av_store(av, 1, swash);
15494 SvREFCNT_dec_NN(cp_list);
15497 av_store(av, 1, &PL_sv_undef);
15499 av_store(av, 3, cp_list);
15500 av_store(av, 4, newSVuv(has_user_defined_property));
15504 if (only_utf8_locale_list) {
15505 av_store(av, 2, only_utf8_locale_list);
15508 av_store(av, 2, &PL_sv_undef);
15511 rv = newRV_noinc(MUTABLE_SV(av));
15512 n = add_data(pRExC_state, STR_WITH_LEN("s"));
15513 RExC_rxi->data->data[n] = (void*)rv;
15518 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
15520 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
15521 const regnode* node,
15524 SV** only_utf8_locale_ptr,
15528 /* For internal core use only.
15529 * Returns the swash for the input 'node' in the regex 'prog'.
15530 * If <doinit> is 'true', will attempt to create the swash if not already
15532 * If <listsvp> is non-null, will return the printable contents of the
15533 * swash. This can be used to get debugging information even before the
15534 * swash exists, by calling this function with 'doinit' set to false, in
15535 * which case the components that will be used to eventually create the
15536 * swash are returned (in a printable form).
15537 * If <exclude_list> is not NULL, it is an inversion list of things to
15538 * exclude from what's returned in <listsvp>.
15539 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
15540 * that, in spite of this function's name, the swash it returns may include
15541 * the bitmap data as well */
15544 SV *si = NULL; /* Input swash initialization string */
15545 SV* invlist = NULL;
15547 RXi_GET_DECL(prog,progi);
15548 const struct reg_data * const data = prog ? progi->data : NULL;
15550 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
15552 assert(ANYOF_FLAGS(node)
15553 & (ANYOF_HAS_UTF8_NONBITMAP_MATCHES
15554 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES|ANYOF_LOC_FOLD));
15556 if (data && data->count) {
15557 const U32 n = ARG(node);
15559 if (data->what[n] == 's') {
15560 SV * const rv = MUTABLE_SV(data->data[n]);
15561 AV * const av = MUTABLE_AV(SvRV(rv));
15562 SV **const ary = AvARRAY(av);
15563 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
15565 si = *ary; /* ary[0] = the string to initialize the swash with */
15567 /* Elements 3 and 4 are either both present or both absent. [3] is
15568 * any inversion list generated at compile time; [4] indicates if
15569 * that inversion list has any user-defined properties in it. */
15570 if (av_tindex(av) >= 2) {
15571 if (only_utf8_locale_ptr
15573 && ary[2] != &PL_sv_undef)
15575 *only_utf8_locale_ptr = ary[2];
15578 assert(only_utf8_locale_ptr);
15579 *only_utf8_locale_ptr = NULL;
15582 if (av_tindex(av) >= 3) {
15584 if (SvUV(ary[4])) {
15585 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
15593 /* Element [1] is reserved for the set-up swash. If already there,
15594 * return it; if not, create it and store it there */
15595 if (ary[1] && SvROK(ary[1])) {
15598 else if (doinit && ((si && si != &PL_sv_undef)
15599 || (invlist && invlist != &PL_sv_undef))) {
15601 sw = _core_swash_init("utf8", /* the utf8 package */
15605 0, /* not from tr/// */
15607 &swash_init_flags);
15608 (void)av_store(av, 1, sw);
15613 /* If requested, return a printable version of what this swash matches */
15615 SV* matches_string = newSVpvs("");
15617 /* The swash should be used, if possible, to get the data, as it
15618 * contains the resolved data. But this function can be called at
15619 * compile-time, before everything gets resolved, in which case we
15620 * return the currently best available information, which is the string
15621 * that will eventually be used to do that resolving, 'si' */
15622 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
15623 && (si && si != &PL_sv_undef))
15625 sv_catsv(matches_string, si);
15628 /* Add the inversion list to whatever we have. This may have come from
15629 * the swash, or from an input parameter */
15631 if (exclude_list) {
15632 SV* clone = invlist_clone(invlist);
15633 _invlist_subtract(clone, exclude_list, &clone);
15634 sv_catsv(matches_string, _invlist_contents(clone));
15635 SvREFCNT_dec_NN(clone);
15638 sv_catsv(matches_string, _invlist_contents(invlist));
15641 *listsvp = matches_string;
15646 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
15648 /* reg_skipcomment()
15650 Absorbs an /x style # comment from the input stream,
15651 returning a pointer to the first character beyond the comment, or if the
15652 comment terminates the pattern without anything following it, this returns
15653 one past the final character of the pattern (in other words, RExC_end) and
15654 sets the REG_RUN_ON_COMMENT_SEEN flag.
15656 Note it's the callers responsibility to ensure that we are
15657 actually in /x mode
15661 PERL_STATIC_INLINE char*
15662 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
15664 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
15668 while (p < RExC_end) {
15669 if (*(++p) == '\n') {
15674 /* we ran off the end of the pattern without ending the comment, so we have
15675 * to add an \n when wrapping */
15676 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
15682 Advances the parse position, and optionally absorbs
15683 "whitespace" from the inputstream.
15685 Without /x "whitespace" means (?#...) style comments only,
15686 with /x this means (?#...) and # comments and whitespace proper.
15688 Returns the RExC_parse point from BEFORE the scan occurs.
15690 This is the /x friendly way of saying RExC_parse++.
15694 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
15696 char* const retval = RExC_parse++;
15698 PERL_ARGS_ASSERT_NEXTCHAR;
15701 if (RExC_end - RExC_parse >= 3
15702 && *RExC_parse == '('
15703 && RExC_parse[1] == '?'
15704 && RExC_parse[2] == '#')
15706 while (*RExC_parse != ')') {
15707 if (RExC_parse == RExC_end)
15708 FAIL("Sequence (?#... not terminated");
15714 if (RExC_flags & RXf_PMf_EXTENDED) {
15715 char * p = regpatws(pRExC_state, RExC_parse,
15716 TRUE); /* means recognize comments */
15717 if (p != RExC_parse) {
15727 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
15729 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
15730 * space. In pass1, it aligns and increments RExC_size; in pass2,
15733 regnode * const ret = RExC_emit;
15734 GET_RE_DEBUG_FLAGS_DECL;
15736 PERL_ARGS_ASSERT_REGNODE_GUTS;
15738 assert(extra_size >= regarglen[op]);
15741 SIZE_ALIGN(RExC_size);
15742 RExC_size += 1 + extra_size;
15745 if (RExC_emit >= RExC_emit_bound)
15746 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
15747 op, (void*)RExC_emit, (void*)RExC_emit_bound);
15749 NODE_ALIGN_FILL(ret);
15750 #ifndef RE_TRACK_PATTERN_OFFSETS
15751 PERL_UNUSED_ARG(name);
15753 if (RExC_offsets) { /* MJD */
15755 ("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
15758 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
15759 ? "Overwriting end of array!\n" : "OK",
15760 (UV)(RExC_emit - RExC_emit_start),
15761 (UV)(RExC_parse - RExC_start),
15762 (UV)RExC_offsets[0]));
15763 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
15770 - reg_node - emit a node
15772 STATIC regnode * /* Location. */
15773 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
15775 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
15777 PERL_ARGS_ASSERT_REG_NODE;
15779 assert(regarglen[op] == 0);
15782 regnode *ptr = ret;
15783 FILL_ADVANCE_NODE(ptr, op);
15790 - reganode - emit a node with an argument
15792 STATIC regnode * /* Location. */
15793 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
15795 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
15797 PERL_ARGS_ASSERT_REGANODE;
15799 assert(regarglen[op] == 1);
15802 regnode *ptr = ret;
15803 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
15810 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
15812 /* emit a node with U32 and I32 arguments */
15814 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
15816 PERL_ARGS_ASSERT_REG2LANODE;
15818 assert(regarglen[op] == 2);
15821 regnode *ptr = ret;
15822 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
15829 - reginsert - insert an operator in front of already-emitted operand
15831 * Means relocating the operand.
15834 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
15839 const int offset = regarglen[(U8)op];
15840 const int size = NODE_STEP_REGNODE + offset;
15841 GET_RE_DEBUG_FLAGS_DECL;
15843 PERL_ARGS_ASSERT_REGINSERT;
15844 PERL_UNUSED_CONTEXT;
15845 PERL_UNUSED_ARG(depth);
15846 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
15847 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
15856 if (RExC_open_parens) {
15858 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
15859 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
15860 if ( RExC_open_parens[paren] >= opnd ) {
15861 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
15862 RExC_open_parens[paren] += size;
15864 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
15866 if ( RExC_close_parens[paren] >= opnd ) {
15867 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
15868 RExC_close_parens[paren] += size;
15870 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
15875 while (src > opnd) {
15876 StructCopy(--src, --dst, regnode);
15877 #ifdef RE_TRACK_PATTERN_OFFSETS
15878 if (RExC_offsets) { /* MJD 20010112 */
15880 ("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
15884 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
15885 ? "Overwriting end of array!\n" : "OK",
15886 (UV)(src - RExC_emit_start),
15887 (UV)(dst - RExC_emit_start),
15888 (UV)RExC_offsets[0]));
15889 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
15890 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
15896 place = opnd; /* Op node, where operand used to be. */
15897 #ifdef RE_TRACK_PATTERN_OFFSETS
15898 if (RExC_offsets) { /* MJD */
15900 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
15904 (UV)(place - RExC_emit_start) > RExC_offsets[0]
15905 ? "Overwriting end of array!\n" : "OK",
15906 (UV)(place - RExC_emit_start),
15907 (UV)(RExC_parse - RExC_start),
15908 (UV)RExC_offsets[0]));
15909 Set_Node_Offset(place, RExC_parse);
15910 Set_Node_Length(place, 1);
15913 src = NEXTOPER(place);
15914 FILL_ADVANCE_NODE(place, op);
15915 Zero(src, offset, regnode);
15919 - regtail - set the next-pointer at the end of a node chain of p to val.
15920 - SEE ALSO: regtail_study
15922 /* TODO: All three parms should be const */
15924 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p,
15925 const regnode *val,U32 depth)
15928 GET_RE_DEBUG_FLAGS_DECL;
15930 PERL_ARGS_ASSERT_REGTAIL;
15932 PERL_UNUSED_ARG(depth);
15938 /* Find last node. */
15941 regnode * const temp = regnext(scan);
15943 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
15944 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
15945 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
15946 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
15947 (temp == NULL ? "->" : ""),
15948 (temp == NULL ? PL_reg_name[OP(val)] : "")
15956 if (reg_off_by_arg[OP(scan)]) {
15957 ARG_SET(scan, val - scan);
15960 NEXT_OFF(scan) = val - scan;
15966 - regtail_study - set the next-pointer at the end of a node chain of p to val.
15967 - Look for optimizable sequences at the same time.
15968 - currently only looks for EXACT chains.
15970 This is experimental code. The idea is to use this routine to perform
15971 in place optimizations on branches and groups as they are constructed,
15972 with the long term intention of removing optimization from study_chunk so
15973 that it is purely analytical.
15975 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
15976 to control which is which.
15979 /* TODO: All four parms should be const */
15982 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
15983 const regnode *val,U32 depth)
15987 #ifdef EXPERIMENTAL_INPLACESCAN
15990 GET_RE_DEBUG_FLAGS_DECL;
15992 PERL_ARGS_ASSERT_REGTAIL_STUDY;
15998 /* Find last node. */
16002 regnode * const temp = regnext(scan);
16003 #ifdef EXPERIMENTAL_INPLACESCAN
16004 if (PL_regkind[OP(scan)] == EXACT) {
16005 bool unfolded_multi_char; /* Unexamined in this routine */
16006 if (join_exact(pRExC_state, scan, &min,
16007 &unfolded_multi_char, 1, val, depth+1))
16012 switch (OP(scan)) {
16016 case EXACTFA_NO_TRIE:
16022 if( exact == PSEUDO )
16024 else if ( exact != OP(scan) )
16033 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
16034 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
16035 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
16036 SvPV_nolen_const(RExC_mysv),
16037 REG_NODE_NUM(scan),
16038 PL_reg_name[exact]);
16045 DEBUG_PARSE_MSG("");
16046 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
16047 PerlIO_printf(Perl_debug_log,
16048 "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
16049 SvPV_nolen_const(RExC_mysv),
16050 (IV)REG_NODE_NUM(val),
16054 if (reg_off_by_arg[OP(scan)]) {
16055 ARG_SET(scan, val - scan);
16058 NEXT_OFF(scan) = val - scan;
16066 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
16071 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
16076 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
16078 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
16079 if (flags & (1<<bit)) {
16080 if (!set++ && lead)
16081 PerlIO_printf(Perl_debug_log, "%s",lead);
16082 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_intflags_name[bit]);
16087 PerlIO_printf(Perl_debug_log, "\n");
16089 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
16094 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
16100 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
16102 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
16103 if (flags & (1<<bit)) {
16104 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
16107 if (!set++ && lead)
16108 PerlIO_printf(Perl_debug_log, "%s",lead);
16109 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
16112 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
16113 if (!set++ && lead) {
16114 PerlIO_printf(Perl_debug_log, "%s",lead);
16117 case REGEX_UNICODE_CHARSET:
16118 PerlIO_printf(Perl_debug_log, "UNICODE");
16120 case REGEX_LOCALE_CHARSET:
16121 PerlIO_printf(Perl_debug_log, "LOCALE");
16123 case REGEX_ASCII_RESTRICTED_CHARSET:
16124 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
16126 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
16127 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
16130 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
16136 PerlIO_printf(Perl_debug_log, "\n");
16138 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
16144 Perl_regdump(pTHX_ const regexp *r)
16147 SV * const sv = sv_newmortal();
16148 SV *dsv= sv_newmortal();
16149 RXi_GET_DECL(r,ri);
16150 GET_RE_DEBUG_FLAGS_DECL;
16152 PERL_ARGS_ASSERT_REGDUMP;
16154 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
16156 /* Header fields of interest. */
16157 if (r->anchored_substr) {
16158 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
16159 RE_SV_DUMPLEN(r->anchored_substr), 30);
16160 PerlIO_printf(Perl_debug_log,
16161 "anchored %s%s at %"IVdf" ",
16162 s, RE_SV_TAIL(r->anchored_substr),
16163 (IV)r->anchored_offset);
16164 } else if (r->anchored_utf8) {
16165 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
16166 RE_SV_DUMPLEN(r->anchored_utf8), 30);
16167 PerlIO_printf(Perl_debug_log,
16168 "anchored utf8 %s%s at %"IVdf" ",
16169 s, RE_SV_TAIL(r->anchored_utf8),
16170 (IV)r->anchored_offset);
16172 if (r->float_substr) {
16173 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
16174 RE_SV_DUMPLEN(r->float_substr), 30);
16175 PerlIO_printf(Perl_debug_log,
16176 "floating %s%s at %"IVdf"..%"UVuf" ",
16177 s, RE_SV_TAIL(r->float_substr),
16178 (IV)r->float_min_offset, (UV)r->float_max_offset);
16179 } else if (r->float_utf8) {
16180 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
16181 RE_SV_DUMPLEN(r->float_utf8), 30);
16182 PerlIO_printf(Perl_debug_log,
16183 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
16184 s, RE_SV_TAIL(r->float_utf8),
16185 (IV)r->float_min_offset, (UV)r->float_max_offset);
16187 if (r->check_substr || r->check_utf8)
16188 PerlIO_printf(Perl_debug_log,
16190 (r->check_substr == r->float_substr
16191 && r->check_utf8 == r->float_utf8
16192 ? "(checking floating" : "(checking anchored"));
16193 if (r->intflags & PREGf_NOSCAN)
16194 PerlIO_printf(Perl_debug_log, " noscan");
16195 if (r->extflags & RXf_CHECK_ALL)
16196 PerlIO_printf(Perl_debug_log, " isall");
16197 if (r->check_substr || r->check_utf8)
16198 PerlIO_printf(Perl_debug_log, ") ");
16200 if (ri->regstclass) {
16201 regprop(r, sv, ri->regstclass, NULL, NULL);
16202 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
16204 if (r->intflags & PREGf_ANCH) {
16205 PerlIO_printf(Perl_debug_log, "anchored");
16206 if (r->intflags & PREGf_ANCH_MBOL)
16207 PerlIO_printf(Perl_debug_log, "(MBOL)");
16208 if (r->intflags & PREGf_ANCH_SBOL)
16209 PerlIO_printf(Perl_debug_log, "(SBOL)");
16210 if (r->intflags & PREGf_ANCH_GPOS)
16211 PerlIO_printf(Perl_debug_log, "(GPOS)");
16212 PerlIO_putc(Perl_debug_log, ' ');
16214 if (r->intflags & PREGf_GPOS_SEEN)
16215 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
16216 if (r->intflags & PREGf_SKIP)
16217 PerlIO_printf(Perl_debug_log, "plus ");
16218 if (r->intflags & PREGf_IMPLICIT)
16219 PerlIO_printf(Perl_debug_log, "implicit ");
16220 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
16221 if (r->extflags & RXf_EVAL_SEEN)
16222 PerlIO_printf(Perl_debug_log, "with eval ");
16223 PerlIO_printf(Perl_debug_log, "\n");
16225 regdump_extflags("r->extflags: ",r->extflags);
16226 regdump_intflags("r->intflags: ",r->intflags);
16229 PERL_ARGS_ASSERT_REGDUMP;
16230 PERL_UNUSED_CONTEXT;
16231 PERL_UNUSED_ARG(r);
16232 #endif /* DEBUGGING */
16236 - regprop - printable representation of opcode, with run time support
16240 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
16245 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
16246 static const char * const anyofs[] = {
16247 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
16248 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
16249 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
16250 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
16251 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
16252 || _CC_VERTSPACE != 16
16253 #error Need to adjust order of anyofs[]
16290 RXi_GET_DECL(prog,progi);
16291 GET_RE_DEBUG_FLAGS_DECL;
16293 PERL_ARGS_ASSERT_REGPROP;
16295 sv_setpvn(sv, "", 0);
16297 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
16298 /* It would be nice to FAIL() here, but this may be called from
16299 regexec.c, and it would be hard to supply pRExC_state. */
16300 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
16301 (int)OP(o), (int)REGNODE_MAX);
16302 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
16304 k = PL_regkind[OP(o)];
16307 sv_catpvs(sv, " ");
16308 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
16309 * is a crude hack but it may be the best for now since
16310 * we have no flag "this EXACTish node was UTF-8"
16312 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
16313 PERL_PV_ESCAPE_UNI_DETECT |
16314 PERL_PV_ESCAPE_NONASCII |
16315 PERL_PV_PRETTY_ELLIPSES |
16316 PERL_PV_PRETTY_LTGT |
16317 PERL_PV_PRETTY_NOCLEAR
16319 } else if (k == TRIE) {
16320 /* print the details of the trie in dumpuntil instead, as
16321 * progi->data isn't available here */
16322 const char op = OP(o);
16323 const U32 n = ARG(o);
16324 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
16325 (reg_ac_data *)progi->data->data[n] :
16327 const reg_trie_data * const trie
16328 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
16330 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
16331 DEBUG_TRIE_COMPILE_r(
16332 Perl_sv_catpvf(aTHX_ sv,
16333 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
16334 (UV)trie->startstate,
16335 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
16336 (UV)trie->wordcount,
16339 (UV)TRIE_CHARCOUNT(trie),
16340 (UV)trie->uniquecharcount
16343 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
16344 sv_catpvs(sv, "[");
16345 (void) put_charclass_bitmap_innards(sv,
16346 (IS_ANYOF_TRIE(op))
16348 : TRIE_BITMAP(trie),
16350 sv_catpvs(sv, "]");
16353 } else if (k == CURLY) {
16354 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
16355 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
16356 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
16358 else if (k == WHILEM && o->flags) /* Ordinal/of */
16359 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
16360 else if (k == REF || k == OPEN || k == CLOSE
16361 || k == GROUPP || OP(o)==ACCEPT)
16363 AV *name_list= NULL;
16364 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
16365 if ( RXp_PAREN_NAMES(prog) ) {
16366 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
16367 } else if ( pRExC_state ) {
16368 name_list= RExC_paren_name_list;
16371 if ( k != REF || (OP(o) < NREF)) {
16372 SV **name= av_fetch(name_list, ARG(o), 0 );
16374 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
16377 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
16378 I32 *nums=(I32*)SvPVX(sv_dat);
16379 SV **name= av_fetch(name_list, nums[0], 0 );
16382 for ( n=0; n<SvIVX(sv_dat); n++ ) {
16383 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
16384 (n ? "," : ""), (IV)nums[n]);
16386 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
16390 if ( k == REF && reginfo) {
16391 U32 n = ARG(o); /* which paren pair */
16392 I32 ln = prog->offs[n].start;
16393 if (prog->lastparen < n || ln == -1)
16394 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
16395 else if (ln == prog->offs[n].end)
16396 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
16398 const char *s = reginfo->strbeg + ln;
16399 Perl_sv_catpvf(aTHX_ sv, ": ");
16400 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
16401 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
16404 } else if (k == GOSUB) {
16405 AV *name_list= NULL;
16406 if ( RXp_PAREN_NAMES(prog) ) {
16407 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
16408 } else if ( pRExC_state ) {
16409 name_list= RExC_paren_name_list;
16412 /* Paren and offset */
16413 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o));
16415 SV **name= av_fetch(name_list, ARG(o), 0 );
16417 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
16420 else if (k == VERB) {
16422 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
16423 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
16424 } else if (k == LOGICAL)
16425 /* 2: embedded, otherwise 1 */
16426 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
16427 else if (k == ANYOF) {
16428 const U8 flags = ANYOF_FLAGS(o);
16430 SV* bitmap_invlist; /* Will hold what the bit map contains */
16433 if (OP(o) == ANYOFL)
16434 sv_catpvs(sv, "{loc}");
16435 if (flags & ANYOF_LOC_FOLD)
16436 sv_catpvs(sv, "{i}");
16437 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
16438 if (flags & ANYOF_INVERT)
16439 sv_catpvs(sv, "^");
16441 /* output what the standard cp 0-NUM_ANYOF_CODE_POINTS-1 bitmap matches
16443 do_sep = put_charclass_bitmap_innards(sv, ANYOF_BITMAP(o),
16446 /* output any special charclass tests (used entirely under use
16448 if (ANYOF_POSIXL_TEST_ANY_SET(o)) {
16450 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
16451 if (ANYOF_POSIXL_TEST(o,i)) {
16452 sv_catpv(sv, anyofs[i]);
16458 if ((flags & (ANYOF_MATCHES_ALL_ABOVE_BITMAP
16459 |ANYOF_HAS_UTF8_NONBITMAP_MATCHES
16460 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES
16464 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
16465 if (flags & ANYOF_INVERT)
16466 /*make sure the invert info is in each */
16467 sv_catpvs(sv, "^");
16470 if (flags & ANYOF_MATCHES_ALL_NON_UTF8_NON_ASCII) {
16471 sv_catpvs(sv, "{non-utf8-latin1-all}");
16474 /* output information about the unicode matching */
16475 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP)
16476 sv_catpvs(sv, "{above_bitmap_all}");
16477 else if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
16478 SV *lv; /* Set if there is something outside the bit map. */
16479 bool byte_output = FALSE; /* If something in the bitmap has
16481 SV *only_utf8_locale;
16483 /* Get the stuff that wasn't in the bitmap. 'bitmap_invlist'
16484 * is used to guarantee that nothing in the bitmap gets
16486 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
16487 &lv, &only_utf8_locale,
16489 if (lv && lv != &PL_sv_undef) {
16490 char *s = savesvpv(lv);
16491 char * const origs = s;
16493 while (*s && *s != '\n')
16497 const char * const t = ++s;
16499 if (flags & ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES) {
16500 sv_catpvs(sv, "{outside bitmap}");
16503 sv_catpvs(sv, "{utf8}");
16507 sv_catpvs(sv, " ");
16513 /* Truncate very long output */
16514 if (s - origs > 256) {
16515 Perl_sv_catpvf(aTHX_ sv,
16517 (int) (s - origs - 1),
16523 else if (*s == '\t') {
16537 SvREFCNT_dec_NN(lv);
16540 if ((flags & ANYOF_LOC_FOLD)
16541 && only_utf8_locale
16542 && only_utf8_locale != &PL_sv_undef)
16545 int max_entries = 256;
16547 sv_catpvs(sv, "{utf8 locale}");
16548 invlist_iterinit(only_utf8_locale);
16549 while (invlist_iternext(only_utf8_locale,
16551 put_range(sv, start, end, FALSE);
16553 if (max_entries < 0) {
16554 sv_catpvs(sv, "...");
16558 invlist_iterfinish(only_utf8_locale);
16562 SvREFCNT_dec(bitmap_invlist);
16565 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
16567 else if (k == POSIXD || k == NPOSIXD) {
16568 U8 index = FLAGS(o) * 2;
16569 if (index < C_ARRAY_LENGTH(anyofs)) {
16570 if (*anyofs[index] != '[') {
16573 sv_catpv(sv, anyofs[index]);
16574 if (*anyofs[index] != '[') {
16579 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
16582 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
16583 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
16584 else if (OP(o) == SBOL)
16585 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
16587 PERL_UNUSED_CONTEXT;
16588 PERL_UNUSED_ARG(sv);
16589 PERL_UNUSED_ARG(o);
16590 PERL_UNUSED_ARG(prog);
16591 PERL_UNUSED_ARG(reginfo);
16592 PERL_UNUSED_ARG(pRExC_state);
16593 #endif /* DEBUGGING */
16599 Perl_re_intuit_string(pTHX_ REGEXP * const r)
16600 { /* Assume that RE_INTUIT is set */
16601 struct regexp *const prog = ReANY(r);
16602 GET_RE_DEBUG_FLAGS_DECL;
16604 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
16605 PERL_UNUSED_CONTEXT;
16609 const char * const s = SvPV_nolen_const(prog->check_substr
16610 ? prog->check_substr : prog->check_utf8);
16612 if (!PL_colorset) reginitcolors();
16613 PerlIO_printf(Perl_debug_log,
16614 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
16616 prog->check_substr ? "" : "utf8 ",
16617 PL_colors[5],PL_colors[0],
16620 (strlen(s) > 60 ? "..." : ""));
16623 return prog->check_substr ? prog->check_substr : prog->check_utf8;
16629 handles refcounting and freeing the perl core regexp structure. When
16630 it is necessary to actually free the structure the first thing it
16631 does is call the 'free' method of the regexp_engine associated to
16632 the regexp, allowing the handling of the void *pprivate; member
16633 first. (This routine is not overridable by extensions, which is why
16634 the extensions free is called first.)
16636 See regdupe and regdupe_internal if you change anything here.
16638 #ifndef PERL_IN_XSUB_RE
16640 Perl_pregfree(pTHX_ REGEXP *r)
16646 Perl_pregfree2(pTHX_ REGEXP *rx)
16648 struct regexp *const r = ReANY(rx);
16649 GET_RE_DEBUG_FLAGS_DECL;
16651 PERL_ARGS_ASSERT_PREGFREE2;
16653 if (r->mother_re) {
16654 ReREFCNT_dec(r->mother_re);
16656 CALLREGFREE_PVT(rx); /* free the private data */
16657 SvREFCNT_dec(RXp_PAREN_NAMES(r));
16658 Safefree(r->xpv_len_u.xpvlenu_pv);
16661 SvREFCNT_dec(r->anchored_substr);
16662 SvREFCNT_dec(r->anchored_utf8);
16663 SvREFCNT_dec(r->float_substr);
16664 SvREFCNT_dec(r->float_utf8);
16665 Safefree(r->substrs);
16667 RX_MATCH_COPY_FREE(rx);
16668 #ifdef PERL_ANY_COW
16669 SvREFCNT_dec(r->saved_copy);
16672 SvREFCNT_dec(r->qr_anoncv);
16673 rx->sv_u.svu_rx = 0;
16678 This is a hacky workaround to the structural issue of match results
16679 being stored in the regexp structure which is in turn stored in
16680 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
16681 could be PL_curpm in multiple contexts, and could require multiple
16682 result sets being associated with the pattern simultaneously, such
16683 as when doing a recursive match with (??{$qr})
16685 The solution is to make a lightweight copy of the regexp structure
16686 when a qr// is returned from the code executed by (??{$qr}) this
16687 lightweight copy doesn't actually own any of its data except for
16688 the starp/end and the actual regexp structure itself.
16694 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
16696 struct regexp *ret;
16697 struct regexp *const r = ReANY(rx);
16698 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
16700 PERL_ARGS_ASSERT_REG_TEMP_COPY;
16703 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
16705 SvOK_off((SV *)ret_x);
16707 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
16708 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
16709 made both spots point to the same regexp body.) */
16710 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
16711 assert(!SvPVX(ret_x));
16712 ret_x->sv_u.svu_rx = temp->sv_any;
16713 temp->sv_any = NULL;
16714 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
16715 SvREFCNT_dec_NN(temp);
16716 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
16717 ing below will not set it. */
16718 SvCUR_set(ret_x, SvCUR(rx));
16721 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
16722 sv_force_normal(sv) is called. */
16724 ret = ReANY(ret_x);
16726 SvFLAGS(ret_x) |= SvUTF8(rx);
16727 /* We share the same string buffer as the original regexp, on which we
16728 hold a reference count, incremented when mother_re is set below.
16729 The string pointer is copied here, being part of the regexp struct.
16731 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
16732 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
16734 const I32 npar = r->nparens+1;
16735 Newx(ret->offs, npar, regexp_paren_pair);
16736 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
16739 Newx(ret->substrs, 1, struct reg_substr_data);
16740 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
16742 SvREFCNT_inc_void(ret->anchored_substr);
16743 SvREFCNT_inc_void(ret->anchored_utf8);
16744 SvREFCNT_inc_void(ret->float_substr);
16745 SvREFCNT_inc_void(ret->float_utf8);
16747 /* check_substr and check_utf8, if non-NULL, point to either their
16748 anchored or float namesakes, and don't hold a second reference. */
16750 RX_MATCH_COPIED_off(ret_x);
16751 #ifdef PERL_ANY_COW
16752 ret->saved_copy = NULL;
16754 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
16755 SvREFCNT_inc_void(ret->qr_anoncv);
16761 /* regfree_internal()
16763 Free the private data in a regexp. This is overloadable by
16764 extensions. Perl takes care of the regexp structure in pregfree(),
16765 this covers the *pprivate pointer which technically perl doesn't
16766 know about, however of course we have to handle the
16767 regexp_internal structure when no extension is in use.
16769 Note this is called before freeing anything in the regexp
16774 Perl_regfree_internal(pTHX_ REGEXP * const rx)
16776 struct regexp *const r = ReANY(rx);
16777 RXi_GET_DECL(r,ri);
16778 GET_RE_DEBUG_FLAGS_DECL;
16780 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
16786 SV *dsv= sv_newmortal();
16787 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
16788 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
16789 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
16790 PL_colors[4],PL_colors[5],s);
16793 #ifdef RE_TRACK_PATTERN_OFFSETS
16795 Safefree(ri->u.offsets); /* 20010421 MJD */
16797 if (ri->code_blocks) {
16799 for (n = 0; n < ri->num_code_blocks; n++)
16800 SvREFCNT_dec(ri->code_blocks[n].src_regex);
16801 Safefree(ri->code_blocks);
16805 int n = ri->data->count;
16808 /* If you add a ->what type here, update the comment in regcomp.h */
16809 switch (ri->data->what[n]) {
16815 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
16818 Safefree(ri->data->data[n]);
16824 { /* Aho Corasick add-on structure for a trie node.
16825 Used in stclass optimization only */
16827 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
16828 #ifdef USE_ITHREADS
16832 refcount = --aho->refcount;
16835 PerlMemShared_free(aho->states);
16836 PerlMemShared_free(aho->fail);
16837 /* do this last!!!! */
16838 PerlMemShared_free(ri->data->data[n]);
16839 /* we should only ever get called once, so
16840 * assert as much, and also guard the free
16841 * which /might/ happen twice. At the least
16842 * it will make code anlyzers happy and it
16843 * doesn't cost much. - Yves */
16844 assert(ri->regstclass);
16845 if (ri->regstclass) {
16846 PerlMemShared_free(ri->regstclass);
16847 ri->regstclass = 0;
16854 /* trie structure. */
16856 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
16857 #ifdef USE_ITHREADS
16861 refcount = --trie->refcount;
16864 PerlMemShared_free(trie->charmap);
16865 PerlMemShared_free(trie->states);
16866 PerlMemShared_free(trie->trans);
16868 PerlMemShared_free(trie->bitmap);
16870 PerlMemShared_free(trie->jump);
16871 PerlMemShared_free(trie->wordinfo);
16872 /* do this last!!!! */
16873 PerlMemShared_free(ri->data->data[n]);
16878 Perl_croak(aTHX_ "panic: regfree data code '%c'",
16879 ri->data->what[n]);
16882 Safefree(ri->data->what);
16883 Safefree(ri->data);
16889 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
16890 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
16891 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
16894 re_dup - duplicate a regexp.
16896 This routine is expected to clone a given regexp structure. It is only
16897 compiled under USE_ITHREADS.
16899 After all of the core data stored in struct regexp is duplicated
16900 the regexp_engine.dupe method is used to copy any private data
16901 stored in the *pprivate pointer. This allows extensions to handle
16902 any duplication it needs to do.
16904 See pregfree() and regfree_internal() if you change anything here.
16906 #if defined(USE_ITHREADS)
16907 #ifndef PERL_IN_XSUB_RE
16909 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
16913 const struct regexp *r = ReANY(sstr);
16914 struct regexp *ret = ReANY(dstr);
16916 PERL_ARGS_ASSERT_RE_DUP_GUTS;
16918 npar = r->nparens+1;
16919 Newx(ret->offs, npar, regexp_paren_pair);
16920 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
16922 if (ret->substrs) {
16923 /* Do it this way to avoid reading from *r after the StructCopy().
16924 That way, if any of the sv_dup_inc()s dislodge *r from the L1
16925 cache, it doesn't matter. */
16926 const bool anchored = r->check_substr
16927 ? r->check_substr == r->anchored_substr
16928 : r->check_utf8 == r->anchored_utf8;
16929 Newx(ret->substrs, 1, struct reg_substr_data);
16930 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
16932 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
16933 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
16934 ret->float_substr = sv_dup_inc(ret->float_substr, param);
16935 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
16937 /* check_substr and check_utf8, if non-NULL, point to either their
16938 anchored or float namesakes, and don't hold a second reference. */
16940 if (ret->check_substr) {
16942 assert(r->check_utf8 == r->anchored_utf8);
16943 ret->check_substr = ret->anchored_substr;
16944 ret->check_utf8 = ret->anchored_utf8;
16946 assert(r->check_substr == r->float_substr);
16947 assert(r->check_utf8 == r->float_utf8);
16948 ret->check_substr = ret->float_substr;
16949 ret->check_utf8 = ret->float_utf8;
16951 } else if (ret->check_utf8) {
16953 ret->check_utf8 = ret->anchored_utf8;
16955 ret->check_utf8 = ret->float_utf8;
16960 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
16961 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
16964 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
16966 if (RX_MATCH_COPIED(dstr))
16967 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
16969 ret->subbeg = NULL;
16970 #ifdef PERL_ANY_COW
16971 ret->saved_copy = NULL;
16974 /* Whether mother_re be set or no, we need to copy the string. We
16975 cannot refrain from copying it when the storage points directly to
16976 our mother regexp, because that's
16977 1: a buffer in a different thread
16978 2: something we no longer hold a reference on
16979 so we need to copy it locally. */
16980 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
16981 ret->mother_re = NULL;
16983 #endif /* PERL_IN_XSUB_RE */
16988 This is the internal complement to regdupe() which is used to copy
16989 the structure pointed to by the *pprivate pointer in the regexp.
16990 This is the core version of the extension overridable cloning hook.
16991 The regexp structure being duplicated will be copied by perl prior
16992 to this and will be provided as the regexp *r argument, however
16993 with the /old/ structures pprivate pointer value. Thus this routine
16994 may override any copying normally done by perl.
16996 It returns a pointer to the new regexp_internal structure.
17000 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
17003 struct regexp *const r = ReANY(rx);
17004 regexp_internal *reti;
17006 RXi_GET_DECL(r,ri);
17008 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
17012 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
17013 char, regexp_internal);
17014 Copy(ri->program, reti->program, len+1, regnode);
17016 reti->num_code_blocks = ri->num_code_blocks;
17017 if (ri->code_blocks) {
17019 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
17020 struct reg_code_block);
17021 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
17022 struct reg_code_block);
17023 for (n = 0; n < ri->num_code_blocks; n++)
17024 reti->code_blocks[n].src_regex = (REGEXP*)
17025 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
17028 reti->code_blocks = NULL;
17030 reti->regstclass = NULL;
17033 struct reg_data *d;
17034 const int count = ri->data->count;
17037 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
17038 char, struct reg_data);
17039 Newx(d->what, count, U8);
17042 for (i = 0; i < count; i++) {
17043 d->what[i] = ri->data->what[i];
17044 switch (d->what[i]) {
17045 /* see also regcomp.h and regfree_internal() */
17046 case 'a': /* actually an AV, but the dup function is identical. */
17050 case 'u': /* actually an HV, but the dup function is identical. */
17051 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
17054 /* This is cheating. */
17055 Newx(d->data[i], 1, regnode_ssc);
17056 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
17057 reti->regstclass = (regnode*)d->data[i];
17060 /* Trie stclasses are readonly and can thus be shared
17061 * without duplication. We free the stclass in pregfree
17062 * when the corresponding reg_ac_data struct is freed.
17064 reti->regstclass= ri->regstclass;
17068 ((reg_trie_data*)ri->data->data[i])->refcount++;
17073 d->data[i] = ri->data->data[i];
17076 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'",
17077 ri->data->what[i]);
17086 reti->name_list_idx = ri->name_list_idx;
17088 #ifdef RE_TRACK_PATTERN_OFFSETS
17089 if (ri->u.offsets) {
17090 Newx(reti->u.offsets, 2*len+1, U32);
17091 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
17094 SetProgLen(reti,len);
17097 return (void*)reti;
17100 #endif /* USE_ITHREADS */
17102 #ifndef PERL_IN_XSUB_RE
17105 - regnext - dig the "next" pointer out of a node
17108 Perl_regnext(pTHX_ regnode *p)
17115 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
17116 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
17117 (int)OP(p), (int)REGNODE_MAX);
17120 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
17129 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
17132 STRLEN l1 = strlen(pat1);
17133 STRLEN l2 = strlen(pat2);
17136 const char *message;
17138 PERL_ARGS_ASSERT_RE_CROAK2;
17144 Copy(pat1, buf, l1 , char);
17145 Copy(pat2, buf + l1, l2 , char);
17146 buf[l1 + l2] = '\n';
17147 buf[l1 + l2 + 1] = '\0';
17148 va_start(args, pat2);
17149 msv = vmess(buf, &args);
17151 message = SvPV_const(msv,l1);
17154 Copy(message, buf, l1 , char);
17155 /* l1-1 to avoid \n */
17156 Perl_croak(aTHX_ "%"UTF8f, UTF8fARG(utf8, l1-1, buf));
17160 /* Certain characters are output as a sequence with the first being a
17162 #define isBACKSLASHED_PUNCT(c) \
17163 ((c) == '-' || (c) == ']' || (c) == '\\' || (c) == '^')
17166 S_put_code_point(pTHX_ SV *sv, UV c)
17168 PERL_ARGS_ASSERT_PUT_CODE_POINT;
17171 Perl_sv_catpvf(aTHX_ sv, "\\x{%04"UVXf"}", c);
17173 else if (isPRINT(c)) {
17174 const char string = (char) c;
17175 if (isBACKSLASHED_PUNCT(c))
17176 sv_catpvs(sv, "\\");
17177 sv_catpvn(sv, &string, 1);
17180 const char * const mnemonic = cntrl_to_mnemonic((char) c);
17182 Perl_sv_catpvf(aTHX_ sv, "%s", mnemonic);
17185 Perl_sv_catpvf(aTHX_ sv, "\\x{%02X}", (U8) c);
17190 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
17193 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
17195 /* Appends to 'sv' a displayable version of the range of code points from
17196 * 'start' to 'end'. It assumes that only ASCII printables are displayable
17197 * as-is (though some of these will be escaped by put_code_point()). */
17199 const unsigned int min_range_count = 3;
17201 assert(start <= end);
17203 PERL_ARGS_ASSERT_PUT_RANGE;
17205 while (start <= end) {
17207 const char * format;
17209 if (end - start < min_range_count) {
17211 /* Individual chars in short ranges */
17212 for (; start <= end; start++) {
17213 put_code_point(sv, start);
17218 /* If permitted by the input options, and there is a possibility that
17219 * this range contains a printable literal, look to see if there is
17221 if (allow_literals && start <= MAX_PRINT_A) {
17223 /* If the range begin isn't an ASCII printable, effectively split
17224 * the range into two parts:
17225 * 1) the portion before the first such printable,
17227 * and output them separately. */
17228 if (! isPRINT_A(start)) {
17229 UV temp_end = start + 1;
17231 /* There is no point looking beyond the final possible
17232 * printable, in MAX_PRINT_A */
17233 UV max = MIN(end, MAX_PRINT_A);
17235 while (temp_end <= max && ! isPRINT_A(temp_end)) {
17239 /* Here, temp_end points to one beyond the first printable if
17240 * found, or to one beyond 'max' if not. If none found, make
17241 * sure that we use the entire range */
17242 if (temp_end > MAX_PRINT_A) {
17243 temp_end = end + 1;
17246 /* Output the first part of the split range, the part that
17247 * doesn't have printables, with no looking for literals
17248 * (otherwise we would infinitely recurse) */
17249 put_range(sv, start, temp_end - 1, FALSE);
17251 /* The 2nd part of the range (if any) starts here. */
17254 /* We continue instead of dropping down because even if the 2nd
17255 * part is non-empty, it could be so short that we want to
17256 * output it specially, as tested for at the top of this loop.
17261 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
17262 * output a sub-range of just the digits or letters, then process
17263 * the remaining portion as usual. */
17264 if (isALPHANUMERIC_A(start)) {
17265 UV mask = (isDIGIT_A(start))
17270 UV temp_end = start + 1;
17272 /* Find the end of the sub-range that includes just the
17273 * characters in the same class as the first character in it */
17274 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
17279 /* For short ranges, don't duplicate the code above to output
17280 * them; just call recursively */
17281 if (temp_end - start < min_range_count) {
17282 put_range(sv, start, temp_end, FALSE);
17284 else { /* Output as a range */
17285 put_code_point(sv, start);
17286 sv_catpvs(sv, "-");
17287 put_code_point(sv, temp_end);
17289 start = temp_end + 1;
17293 /* We output any other printables as individual characters */
17294 if (isPUNCT_A(start) || isSPACE_A(start)) {
17295 while (start <= end && (isPUNCT_A(start)
17296 || isSPACE_A(start)))
17298 put_code_point(sv, start);
17303 } /* End of looking for literals */
17305 /* Here is not to output as a literal. Some control characters have
17306 * mnemonic names. Split off any of those at the beginning and end of
17307 * the range to print mnemonically. It isn't possible for many of
17308 * these to be in a row, so this won't overwhelm with output */
17309 while (isMNEMONIC_CNTRL(start) && start <= end) {
17310 put_code_point(sv, start);
17313 if (start < end && isMNEMONIC_CNTRL(end)) {
17315 /* Here, the final character in the range has a mnemonic name.
17316 * Work backwards from the end to find the final non-mnemonic */
17317 UV temp_end = end - 1;
17318 while (isMNEMONIC_CNTRL(temp_end)) {
17322 /* And separately output the range that doesn't have mnemonics */
17323 put_range(sv, start, temp_end, FALSE);
17325 /* Then output the mnemonic trailing controls */
17326 start = temp_end + 1;
17327 while (start <= end) {
17328 put_code_point(sv, start);
17334 /* As a final resort, output the range or subrange as hex. */
17336 this_end = (end < NUM_ANYOF_CODE_POINTS)
17338 : NUM_ANYOF_CODE_POINTS - 1;
17339 format = (this_end < 256)
17340 ? "\\x{%02"UVXf"}-\\x{%02"UVXf"}"
17341 : "\\x{%04"UVXf"}-\\x{%04"UVXf"}";
17342 GCC_DIAG_IGNORE(-Wformat-nonliteral);
17343 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
17350 S_put_charclass_bitmap_innards(pTHX_ SV *sv, char *bitmap, SV** bitmap_invlist)
17352 /* Appends to 'sv' a displayable version of the innards of the bracketed
17353 * character class whose bitmap is 'bitmap'; Returns 'TRUE' if it actually
17354 * output anything, and bitmap_invlist, if not NULL, will point to an
17355 * inversion list of what is in the bit map */
17359 unsigned int punct_count = 0;
17360 SV* invlist = NULL;
17361 SV** invlist_ptr; /* Temporary, in case bitmap_invlist is NULL */
17362 bool allow_literals = TRUE;
17364 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
17366 invlist_ptr = (bitmap_invlist) ? bitmap_invlist : &invlist;
17368 /* Worst case is exactly every-other code point is in the list */
17369 *invlist_ptr = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
17371 /* Convert the bit map to an inversion list, keeping track of how many
17372 * ASCII puncts are set, including an extra amount for the backslashed
17374 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
17375 if (BITMAP_TEST(bitmap, i)) {
17376 *invlist_ptr = add_cp_to_invlist(*invlist_ptr, i);
17377 if (isPUNCT_A(i)) {
17379 if isBACKSLASHED_PUNCT(i) {
17386 /* Nothing to output */
17387 if (_invlist_len(*invlist_ptr) == 0) {
17388 SvREFCNT_dec(invlist);
17392 /* Generally, it is more readable if printable characters are output as
17393 * literals, but if a range (nearly) spans all of them, it's best to output
17394 * it as a single range. This code will use a single range if all but 2
17395 * printables are in it */
17396 invlist_iterinit(*invlist_ptr);
17397 while (invlist_iternext(*invlist_ptr, &start, &end)) {
17399 /* If range starts beyond final printable, it doesn't have any in it */
17400 if (start > MAX_PRINT_A) {
17404 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
17405 * all but two, the range must start and end no later than 2 from
17407 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
17408 if (end > MAX_PRINT_A) {
17414 if (end - start >= MAX_PRINT_A - ' ' - 2) {
17415 allow_literals = FALSE;
17420 invlist_iterfinish(*invlist_ptr);
17422 /* The legibility of the output depends mostly on how many punctuation
17423 * characters are output. There are 32 possible ASCII ones, and some have
17424 * an additional backslash, bringing it to currently 36, so if any more
17425 * than 18 are to be output, we can instead output it as its complement,
17426 * yielding fewer puncts, and making it more legible. But give some weight
17427 * to the fact that outputting it as a complement is less legible than a
17428 * straight output, so don't complement unless we are somewhat over the 18
17430 if (allow_literals && punct_count > 22) {
17431 sv_catpvs(sv, "^");
17433 /* Add everything remaining to the list, so when we invert it just
17434 * below, it will be excluded */
17435 _invlist_union_complement_2nd(*invlist_ptr, PL_InBitmap, invlist_ptr);
17436 _invlist_invert(*invlist_ptr);
17439 /* Here we have figured things out. Output each range */
17440 invlist_iterinit(*invlist_ptr);
17441 while (invlist_iternext(*invlist_ptr, &start, &end)) {
17442 if (start >= NUM_ANYOF_CODE_POINTS) {
17445 put_range(sv, start, end, allow_literals);
17447 invlist_iterfinish(*invlist_ptr);
17452 #define CLEAR_OPTSTART \
17453 if (optstart) STMT_START { \
17454 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, \
17455 " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
17459 #define DUMPUNTIL(b,e) \
17461 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
17463 STATIC const regnode *
17464 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
17465 const regnode *last, const regnode *plast,
17466 SV* sv, I32 indent, U32 depth)
17468 U8 op = PSEUDO; /* Arbitrary non-END op. */
17469 const regnode *next;
17470 const regnode *optstart= NULL;
17472 RXi_GET_DECL(r,ri);
17473 GET_RE_DEBUG_FLAGS_DECL;
17475 PERL_ARGS_ASSERT_DUMPUNTIL;
17477 #ifdef DEBUG_DUMPUNTIL
17478 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
17479 last ? last-start : 0,plast ? plast-start : 0);
17482 if (plast && plast < last)
17485 while (PL_regkind[op] != END && (!last || node < last)) {
17487 /* While that wasn't END last time... */
17490 if (op == CLOSE || op == WHILEM)
17492 next = regnext((regnode *)node);
17495 if (OP(node) == OPTIMIZED) {
17496 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
17503 regprop(r, sv, node, NULL, NULL);
17504 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
17505 (int)(2*indent + 1), "", SvPVX_const(sv));
17507 if (OP(node) != OPTIMIZED) {
17508 if (next == NULL) /* Next ptr. */
17509 PerlIO_printf(Perl_debug_log, " (0)");
17510 else if (PL_regkind[(U8)op] == BRANCH
17511 && PL_regkind[OP(next)] != BRANCH )
17512 PerlIO_printf(Perl_debug_log, " (FAIL)");
17514 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
17515 (void)PerlIO_putc(Perl_debug_log, '\n');
17519 if (PL_regkind[(U8)op] == BRANCHJ) {
17522 const regnode *nnode = (OP(next) == LONGJMP
17523 ? regnext((regnode *)next)
17525 if (last && nnode > last)
17527 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
17530 else if (PL_regkind[(U8)op] == BRANCH) {
17532 DUMPUNTIL(NEXTOPER(node), next);
17534 else if ( PL_regkind[(U8)op] == TRIE ) {
17535 const regnode *this_trie = node;
17536 const char op = OP(node);
17537 const U32 n = ARG(node);
17538 const reg_ac_data * const ac = op>=AHOCORASICK ?
17539 (reg_ac_data *)ri->data->data[n] :
17541 const reg_trie_data * const trie =
17542 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
17544 AV *const trie_words
17545 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
17547 const regnode *nextbranch= NULL;
17550 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
17551 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
17553 PerlIO_printf(Perl_debug_log, "%*s%s ",
17554 (int)(2*(indent+3)), "",
17556 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
17557 SvCUR(*elem_ptr), 60,
17558 PL_colors[0], PL_colors[1],
17560 ? PERL_PV_ESCAPE_UNI
17562 | PERL_PV_PRETTY_ELLIPSES
17563 | PERL_PV_PRETTY_LTGT
17568 U16 dist= trie->jump[word_idx+1];
17569 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
17570 (UV)((dist ? this_trie + dist : next) - start));
17573 nextbranch= this_trie + trie->jump[0];
17574 DUMPUNTIL(this_trie + dist, nextbranch);
17576 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
17577 nextbranch= regnext((regnode *)nextbranch);
17579 PerlIO_printf(Perl_debug_log, "\n");
17582 if (last && next > last)
17587 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
17588 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
17589 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
17591 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
17593 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
17595 else if ( op == PLUS || op == STAR) {
17596 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
17598 else if (PL_regkind[(U8)op] == ANYOF) {
17599 /* arglen 1 + class block */
17600 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
17601 ? ANYOF_POSIXL_SKIP
17603 node = NEXTOPER(node);
17605 else if (PL_regkind[(U8)op] == EXACT) {
17606 /* Literal string, where present. */
17607 node += NODE_SZ_STR(node) - 1;
17608 node = NEXTOPER(node);
17611 node = NEXTOPER(node);
17612 node += regarglen[(U8)op];
17614 if (op == CURLYX || op == OPEN)
17618 #ifdef DEBUG_DUMPUNTIL
17619 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
17624 #endif /* DEBUGGING */
17628 * c-indentation-style: bsd
17629 * c-basic-offset: 4
17630 * indent-tabs-mode: nil
17633 * ex: set ts=8 sts=4 sw=4 et: