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_inline.h"
90 #include "invlist_inline.h"
91 #include "unicode_constants.h"
93 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
94 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
96 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
97 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
98 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
101 #define STATIC static
104 /* this is a chain of data about sub patterns we are processing that
105 need to be handled separately/specially in study_chunk. Its so
106 we can simulate recursion without losing state. */
108 typedef struct scan_frame {
109 regnode *last_regnode; /* last node to process in this frame */
110 regnode *next_regnode; /* next node to process when last is reached */
111 U32 prev_recursed_depth;
112 I32 stopparen; /* what stopparen do we use */
114 struct scan_frame *this_prev_frame; /* this previous frame */
115 struct scan_frame *prev_frame; /* previous frame */
116 struct scan_frame *next_frame; /* next frame */
119 /* Certain characters are output as a sequence with the first being a
121 #define isBACKSLASHED_PUNCT(c) strchr("-[]\\^", c)
124 struct RExC_state_t {
125 U32 flags; /* RXf_* are we folding, multilining? */
126 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
127 char *precomp; /* uncompiled string. */
128 char *precomp_end; /* pointer to end of 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 char *adjusted_start; /* 'start', adjusted. See code use */
137 STRLEN precomp_adj; /* an offset beyond precomp. See code use */
138 SSize_t whilem_seen; /* number of WHILEM in this expr */
139 regnode *emit_start; /* Start of emitted-code area */
140 regnode *emit_bound; /* First regnode outside of the
142 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
143 implies compiling, so don't emit */
144 regnode_ssc emit_dummy; /* placeholder for emit to point to;
145 large enough for the largest
146 non-EXACTish node, so can use it as
148 I32 naughty; /* How bad is this pattern? */
149 I32 sawback; /* Did we see \1, ...? */
151 SSize_t size; /* Code size. */
152 I32 npar; /* Capture buffer count, (OPEN) plus
153 one. ("par" 0 is the whole
155 I32 nestroot; /* root parens we are in - used by
159 regnode **open_parens; /* pointers to open parens */
160 regnode **close_parens; /* pointers to close parens */
161 regnode *end_op; /* END node in program */
162 I32 utf8; /* whether the pattern is utf8 or not */
163 I32 orig_utf8; /* whether the pattern was originally in utf8 */
164 /* XXX use this for future optimisation of case
165 * where pattern must be upgraded to utf8. */
166 I32 uni_semantics; /* If a d charset modifier should use unicode
167 rules, even if the pattern is not in
169 HV *paren_names; /* Paren names */
171 regnode **recurse; /* Recurse regops */
172 I32 recurse_count; /* Number of recurse regops we have generated */
173 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
175 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
178 I32 override_recoding;
180 I32 recode_x_to_native;
182 I32 in_multi_char_class;
183 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
185 int code_index; /* next code_blocks[] slot */
186 SSize_t maxlen; /* mininum possible number of chars in string to match */
187 scan_frame *frame_head;
188 scan_frame *frame_last;
191 #ifdef ADD_TO_REGEXEC
192 char *starttry; /* -Dr: where regtry was called. */
193 #define RExC_starttry (pRExC_state->starttry)
195 SV *runtime_code_qr; /* qr with the runtime code blocks */
197 const char *lastparse;
199 AV *paren_name_list; /* idx -> name */
200 U32 study_chunk_recursed_count;
203 #define RExC_lastparse (pRExC_state->lastparse)
204 #define RExC_lastnum (pRExC_state->lastnum)
205 #define RExC_paren_name_list (pRExC_state->paren_name_list)
206 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
207 #define RExC_mysv (pRExC_state->mysv1)
208 #define RExC_mysv1 (pRExC_state->mysv1)
209 #define RExC_mysv2 (pRExC_state->mysv2)
212 bool seen_unfolded_sharp_s;
217 #define RExC_flags (pRExC_state->flags)
218 #define RExC_pm_flags (pRExC_state->pm_flags)
219 #define RExC_precomp (pRExC_state->precomp)
220 #define RExC_precomp_adj (pRExC_state->precomp_adj)
221 #define RExC_adjusted_start (pRExC_state->adjusted_start)
222 #define RExC_precomp_end (pRExC_state->precomp_end)
223 #define RExC_rx_sv (pRExC_state->rx_sv)
224 #define RExC_rx (pRExC_state->rx)
225 #define RExC_rxi (pRExC_state->rxi)
226 #define RExC_start (pRExC_state->start)
227 #define RExC_end (pRExC_state->end)
228 #define RExC_parse (pRExC_state->parse)
229 #define RExC_whilem_seen (pRExC_state->whilem_seen)
231 /* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
232 * EXACTF node, hence was parsed under /di rules. If later in the parse,
233 * something forces the pattern into using /ui rules, the sharp s should be
234 * folded into the sequence 'ss', which takes up more space than previously
235 * calculated. This means that the sizing pass needs to be restarted. (The
236 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
237 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
238 * so there is no need to resize [perl #125990]. */
239 #define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
241 #ifdef RE_TRACK_PATTERN_OFFSETS
242 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
245 #define RExC_emit (pRExC_state->emit)
246 #define RExC_emit_dummy (pRExC_state->emit_dummy)
247 #define RExC_emit_start (pRExC_state->emit_start)
248 #define RExC_emit_bound (pRExC_state->emit_bound)
249 #define RExC_sawback (pRExC_state->sawback)
250 #define RExC_seen (pRExC_state->seen)
251 #define RExC_size (pRExC_state->size)
252 #define RExC_maxlen (pRExC_state->maxlen)
253 #define RExC_npar (pRExC_state->npar)
254 #define RExC_nestroot (pRExC_state->nestroot)
255 #define RExC_extralen (pRExC_state->extralen)
256 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
257 #define RExC_utf8 (pRExC_state->utf8)
258 #define RExC_uni_semantics (pRExC_state->uni_semantics)
259 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
260 #define RExC_open_parens (pRExC_state->open_parens)
261 #define RExC_close_parens (pRExC_state->close_parens)
262 #define RExC_end_op (pRExC_state->end_op)
263 #define RExC_paren_names (pRExC_state->paren_names)
264 #define RExC_recurse (pRExC_state->recurse)
265 #define RExC_recurse_count (pRExC_state->recurse_count)
266 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
267 #define RExC_study_chunk_recursed_bytes \
268 (pRExC_state->study_chunk_recursed_bytes)
269 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
270 #define RExC_contains_locale (pRExC_state->contains_locale)
272 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
274 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
275 #define RExC_frame_head (pRExC_state->frame_head)
276 #define RExC_frame_last (pRExC_state->frame_last)
277 #define RExC_frame_count (pRExC_state->frame_count)
278 #define RExC_strict (pRExC_state->strict)
279 #define RExC_study_started (pRExC_state->study_started)
280 #define RExC_warn_text (pRExC_state->warn_text)
282 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
283 * a flag to disable back-off on the fixed/floating substrings - if it's
284 * a high complexity pattern we assume the benefit of avoiding a full match
285 * is worth the cost of checking for the substrings even if they rarely help.
287 #define RExC_naughty (pRExC_state->naughty)
288 #define TOO_NAUGHTY (10)
289 #define MARK_NAUGHTY(add) \
290 if (RExC_naughty < TOO_NAUGHTY) \
291 RExC_naughty += (add)
292 #define MARK_NAUGHTY_EXP(exp, add) \
293 if (RExC_naughty < TOO_NAUGHTY) \
294 RExC_naughty += RExC_naughty / (exp) + (add)
296 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
297 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
298 ((*s) == '{' && regcurly(s)))
301 * Flags to be passed up and down.
303 #define WORST 0 /* Worst case. */
304 #define HASWIDTH 0x01 /* Known to match non-null strings. */
306 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
307 * character. (There needs to be a case: in the switch statement in regexec.c
308 * for any node marked SIMPLE.) Note that this is not the same thing as
311 #define SPSTART 0x04 /* Starts with * or + */
312 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
313 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
314 #define RESTART_PASS1 0x20 /* Need to restart sizing pass */
315 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
316 calcuate sizes as UTF-8 */
318 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
320 /* whether trie related optimizations are enabled */
321 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
322 #define TRIE_STUDY_OPT
323 #define FULL_TRIE_STUDY
329 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
330 #define PBITVAL(paren) (1 << ((paren) & 7))
331 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
332 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
333 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
335 #define REQUIRE_UTF8(flagp) STMT_START { \
338 *flagp = RESTART_PASS1|NEED_UTF8; \
343 /* Change from /d into /u rules, and restart the parse if we've already seen
344 * something whose size would increase as a result, by setting *flagp and
345 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
346 * we've change to /u during the parse. */
347 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
349 if (DEPENDS_SEMANTICS) { \
351 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
352 RExC_uni_semantics = 1; \
353 if (RExC_seen_unfolded_sharp_s) { \
354 *flagp |= RESTART_PASS1; \
355 return restart_retval; \
360 /* This converts the named class defined in regcomp.h to its equivalent class
361 * number defined in handy.h. */
362 #define namedclass_to_classnum(class) ((int) ((class) / 2))
363 #define classnum_to_namedclass(classnum) ((classnum) * 2)
365 #define _invlist_union_complement_2nd(a, b, output) \
366 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
367 #define _invlist_intersection_complement_2nd(a, b, output) \
368 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
370 /* About scan_data_t.
372 During optimisation we recurse through the regexp program performing
373 various inplace (keyhole style) optimisations. In addition study_chunk
374 and scan_commit populate this data structure with information about
375 what strings MUST appear in the pattern. We look for the longest
376 string that must appear at a fixed location, and we look for the
377 longest string that may appear at a floating location. So for instance
382 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
383 strings (because they follow a .* construct). study_chunk will identify
384 both FOO and BAR as being the longest fixed and floating strings respectively.
386 The strings can be composites, for instance
390 will result in a composite fixed substring 'foo'.
392 For each string some basic information is maintained:
395 This is the position the string must appear at, or not before.
396 It also implicitly (when combined with minlenp) tells us how many
397 characters must match before the string we are searching for.
398 Likewise when combined with minlenp and the length of the string it
399 tells us how many characters must appear after the string we have
403 Only used for floating strings. This is the rightmost point that
404 the string can appear at. If set to SSize_t_MAX it indicates that the
405 string can occur infinitely far to the right.
406 For fixed strings, it is equal to min_offset.
409 A pointer to the minimum number of characters of the pattern that the
410 string was found inside. This is important as in the case of positive
411 lookahead or positive lookbehind we can have multiple patterns
416 The minimum length of the pattern overall is 3, the minimum length
417 of the lookahead part is 3, but the minimum length of the part that
418 will actually match is 1. So 'FOO's minimum length is 3, but the
419 minimum length for the F is 1. This is important as the minimum length
420 is used to determine offsets in front of and behind the string being
421 looked for. Since strings can be composites this is the length of the
422 pattern at the time it was committed with a scan_commit. Note that
423 the length is calculated by study_chunk, so that the minimum lengths
424 are not known until the full pattern has been compiled, thus the
425 pointer to the value.
429 In the case of lookbehind the string being searched for can be
430 offset past the start point of the final matching string.
431 If this value was just blithely removed from the min_offset it would
432 invalidate some of the calculations for how many chars must match
433 before or after (as they are derived from min_offset and minlen and
434 the length of the string being searched for).
435 When the final pattern is compiled and the data is moved from the
436 scan_data_t structure into the regexp structure the information
437 about lookbehind is factored in, with the information that would
438 have been lost precalculated in the end_shift field for the
441 The fields pos_min and pos_delta are used to store the minimum offset
442 and the delta to the maximum offset at the current point in the pattern.
446 struct scan_data_substrs {
447 SV *str; /* longest substring found in pattern */
448 SSize_t min_offset; /* earliest point in string it can appear */
449 SSize_t max_offset; /* latest point in string it can appear */
450 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
451 SSize_t lookbehind; /* is the pos of the string modified by LB */
452 I32 flags; /* per substring SF_* and SCF_* flags */
455 typedef struct scan_data_t {
456 /*I32 len_min; unused */
457 /*I32 len_delta; unused */
461 SSize_t last_end; /* min value, <0 unless valid. */
462 SSize_t last_start_min;
463 SSize_t last_start_max;
464 U8 cur_is_floating; /* whether the last_* values should be set as
465 * the next fixed (0) or floating (1)
468 /* [0] is longest fixed substring so far, [1] is longest float so far */
469 struct scan_data_substrs substrs[2];
471 I32 flags; /* common SF_* and SCF_* flags */
473 SSize_t *last_closep;
474 regnode_ssc *start_class;
478 * Forward declarations for pregcomp()'s friends.
481 static const scan_data_t zero_scan_data = {
482 0, 0, NULL, 0, 0, 0, 0,
484 { NULL, 0, 0, 0, 0, 0 },
485 { NULL, 0, 0, 0, 0, 0 },
492 #define SF_BEFORE_SEOL 0x0001
493 #define SF_BEFORE_MEOL 0x0002
494 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
496 #define SF_IS_INF 0x0040
497 #define SF_HAS_PAR 0x0080
498 #define SF_IN_PAR 0x0100
499 #define SF_HAS_EVAL 0x0200
502 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
503 * longest substring in the pattern. When it is not set the optimiser keeps
504 * track of position, but does not keep track of the actual strings seen,
506 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
509 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
510 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
511 * turned off because of the alternation (BRANCH). */
512 #define SCF_DO_SUBSTR 0x0400
514 #define SCF_DO_STCLASS_AND 0x0800
515 #define SCF_DO_STCLASS_OR 0x1000
516 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
517 #define SCF_WHILEM_VISITED_POS 0x2000
519 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
520 #define SCF_SEEN_ACCEPT 0x8000
521 #define SCF_TRIE_DOING_RESTUDY 0x10000
522 #define SCF_IN_DEFINE 0x20000
527 #define UTF cBOOL(RExC_utf8)
529 /* The enums for all these are ordered so things work out correctly */
530 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
531 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
532 == REGEX_DEPENDS_CHARSET)
533 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
534 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
535 >= REGEX_UNICODE_CHARSET)
536 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
537 == REGEX_ASCII_RESTRICTED_CHARSET)
538 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
539 >= REGEX_ASCII_RESTRICTED_CHARSET)
540 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
541 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
543 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
545 /* For programs that want to be strictly Unicode compatible by dying if any
546 * attempt is made to match a non-Unicode code point against a Unicode
548 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
550 #define OOB_NAMEDCLASS -1
552 /* There is no code point that is out-of-bounds, so this is problematic. But
553 * its only current use is to initialize a variable that is always set before
555 #define OOB_UNICODE 0xDEADBEEF
557 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
560 /* length of regex to show in messages that don't mark a position within */
561 #define RegexLengthToShowInErrorMessages 127
564 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
565 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
566 * op/pragma/warn/regcomp.
568 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
569 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
571 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
572 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
574 /* The code in this file in places uses one level of recursion with parsing
575 * rebased to an alternate string constructed by us in memory. This can take
576 * the form of something that is completely different from the input, or
577 * something that uses the input as part of the alternate. In the first case,
578 * there should be no possibility of an error, as we are in complete control of
579 * the alternate string. But in the second case we don't control the input
580 * portion, so there may be errors in that. Here's an example:
582 * is handled specially because \x{df} folds to a sequence of more than one
583 * character, 'ss'. What is done is to create and parse an alternate string,
584 * which looks like this:
585 * /(?:\x{DF}|[abc\x{DF}def])/ui
586 * where it uses the input unchanged in the middle of something it constructs,
587 * which is a branch for the DF outside the character class, and clustering
588 * parens around the whole thing. (It knows enough to skip the DF inside the
589 * class while in this substitute parse.) 'abc' and 'def' may have errors that
590 * need to be reported. The general situation looks like this:
593 * Input: ----------------------------------------------------
594 * Constructed: ---------------------------------------------------
597 * The input string sI..eI is the input pattern. The string sC..EC is the
598 * constructed substitute parse string. The portions sC..tC and eC..EC are
599 * constructed by us. The portion tC..eC is an exact duplicate of the input
600 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
601 * while parsing, we find an error at xC. We want to display a message showing
602 * the real input string. Thus we need to find the point xI in it which
603 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
604 * been constructed by us, and so shouldn't have errors. We get:
606 * xI = sI + (tI - sI) + (xC - tC)
608 * and, the offset into sI is:
610 * (xI - sI) = (tI - sI) + (xC - tC)
612 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
613 * and we save tC as RExC_adjusted_start.
615 * During normal processing of the input pattern, everything points to that,
616 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
619 #define tI_sI RExC_precomp_adj
620 #define tC RExC_adjusted_start
621 #define sC RExC_precomp
622 #define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
623 #define xI(xC) (sC + xI_offset(xC))
624 #define eC RExC_precomp_end
626 #define REPORT_LOCATION_ARGS(xC) \
628 (xI(xC) > eC) /* Don't run off end */ \
629 ? eC - sC /* Length before the <--HERE */ \
630 : ( __ASSERT_(xI_offset(xC) >= 0) xI_offset(xC) ), \
631 sC), /* The input pattern printed up to the <--HERE */ \
633 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
634 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
636 /* Used to point after bad bytes for an error message, but avoid skipping
637 * past a nul byte. */
638 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
641 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
642 * arg. Show regex, up to a maximum length. If it's too long, chop and add
645 #define _FAIL(code) STMT_START { \
646 const char *ellipses = ""; \
647 IV len = RExC_precomp_end - RExC_precomp; \
650 SAVEFREESV(RExC_rx_sv); \
651 if (len > RegexLengthToShowInErrorMessages) { \
652 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
653 len = RegexLengthToShowInErrorMessages - 10; \
659 #define FAIL(msg) _FAIL( \
660 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
661 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
663 #define FAIL2(msg,arg) _FAIL( \
664 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
665 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
668 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
670 #define Simple_vFAIL(m) STMT_START { \
671 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
672 m, REPORT_LOCATION_ARGS(RExC_parse)); \
676 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
678 #define vFAIL(m) STMT_START { \
680 SAVEFREESV(RExC_rx_sv); \
685 * Like Simple_vFAIL(), but accepts two arguments.
687 #define Simple_vFAIL2(m,a1) STMT_START { \
688 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
689 REPORT_LOCATION_ARGS(RExC_parse)); \
693 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
695 #define vFAIL2(m,a1) STMT_START { \
697 SAVEFREESV(RExC_rx_sv); \
698 Simple_vFAIL2(m, a1); \
703 * Like Simple_vFAIL(), but accepts three arguments.
705 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
706 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
707 REPORT_LOCATION_ARGS(RExC_parse)); \
711 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
713 #define vFAIL3(m,a1,a2) STMT_START { \
715 SAVEFREESV(RExC_rx_sv); \
716 Simple_vFAIL3(m, a1, a2); \
720 * Like Simple_vFAIL(), but accepts four arguments.
722 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
723 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
724 REPORT_LOCATION_ARGS(RExC_parse)); \
727 #define vFAIL4(m,a1,a2,a3) STMT_START { \
729 SAVEFREESV(RExC_rx_sv); \
730 Simple_vFAIL4(m, a1, a2, a3); \
733 /* A specialized version of vFAIL2 that works with UTF8f */
734 #define vFAIL2utf8f(m, a1) STMT_START { \
736 SAVEFREESV(RExC_rx_sv); \
737 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
738 REPORT_LOCATION_ARGS(RExC_parse)); \
741 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
743 SAVEFREESV(RExC_rx_sv); \
744 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
745 REPORT_LOCATION_ARGS(RExC_parse)); \
748 /* These have asserts in them because of [perl #122671] Many warnings in
749 * regcomp.c can occur twice. If they get output in pass1 and later in that
750 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
751 * would get output again. So they should be output in pass2, and these
752 * asserts make sure new warnings follow that paradigm. */
754 /* m is not necessarily a "literal string", in this macro */
755 #define reg_warn_non_literal_string(loc, m) STMT_START { \
756 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
757 "%s" REPORT_LOCATION, \
758 m, REPORT_LOCATION_ARGS(loc)); \
761 #define ckWARNreg(loc,m) STMT_START { \
762 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
764 REPORT_LOCATION_ARGS(loc)); \
767 #define vWARN(loc, m) STMT_START { \
768 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
770 REPORT_LOCATION_ARGS(loc)); \
773 #define vWARN_dep(loc, m) STMT_START { \
774 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
776 REPORT_LOCATION_ARGS(loc)); \
779 #define ckWARNdep(loc,m) STMT_START { \
780 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
782 REPORT_LOCATION_ARGS(loc)); \
785 #define ckWARNregdep(loc,m) STMT_START { \
786 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
789 REPORT_LOCATION_ARGS(loc)); \
792 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
793 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
795 a1, REPORT_LOCATION_ARGS(loc)); \
798 #define ckWARN2reg(loc, m, a1) STMT_START { \
799 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
801 a1, REPORT_LOCATION_ARGS(loc)); \
804 #define vWARN3(loc, m, a1, a2) STMT_START { \
805 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
807 a1, a2, REPORT_LOCATION_ARGS(loc)); \
810 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
811 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
814 REPORT_LOCATION_ARGS(loc)); \
817 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
818 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
821 REPORT_LOCATION_ARGS(loc)); \
824 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
825 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
828 REPORT_LOCATION_ARGS(loc)); \
831 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
832 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
835 REPORT_LOCATION_ARGS(loc)); \
838 /* Macros for recording node offsets. 20001227 mjd@plover.com
839 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
840 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
841 * Element 0 holds the number n.
842 * Position is 1 indexed.
844 #ifndef RE_TRACK_PATTERN_OFFSETS
845 #define Set_Node_Offset_To_R(node,byte)
846 #define Set_Node_Offset(node,byte)
847 #define Set_Cur_Node_Offset
848 #define Set_Node_Length_To_R(node,len)
849 #define Set_Node_Length(node,len)
850 #define Set_Node_Cur_Length(node,start)
851 #define Node_Offset(n)
852 #define Node_Length(n)
853 #define Set_Node_Offset_Length(node,offset,len)
854 #define ProgLen(ri) ri->u.proglen
855 #define SetProgLen(ri,x) ri->u.proglen = x
857 #define ProgLen(ri) ri->u.offsets[0]
858 #define SetProgLen(ri,x) ri->u.offsets[0] = x
859 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
861 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
862 __LINE__, (int)(node), (int)(byte))); \
864 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
867 RExC_offsets[2*(node)-1] = (byte); \
872 #define Set_Node_Offset(node,byte) \
873 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
874 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
876 #define Set_Node_Length_To_R(node,len) STMT_START { \
878 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
879 __LINE__, (int)(node), (int)(len))); \
881 Perl_croak(aTHX_ "value of node is %d in Length macro", \
884 RExC_offsets[2*(node)] = (len); \
889 #define Set_Node_Length(node,len) \
890 Set_Node_Length_To_R((node)-RExC_emit_start, len)
891 #define Set_Node_Cur_Length(node, start) \
892 Set_Node_Length(node, RExC_parse - start)
894 /* Get offsets and lengths */
895 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
896 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
898 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
899 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
900 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
904 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
905 #define EXPERIMENTAL_INPLACESCAN
906 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
910 Perl_re_printf(pTHX_ const char *fmt, ...)
914 PerlIO *f= Perl_debug_log;
915 PERL_ARGS_ASSERT_RE_PRINTF;
917 result = PerlIO_vprintf(f, fmt, ap);
923 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
927 PerlIO *f= Perl_debug_log;
928 PERL_ARGS_ASSERT_RE_INDENTF;
930 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
931 result = PerlIO_vprintf(f, fmt, ap);
935 #endif /* DEBUGGING */
937 #define DEBUG_RExC_seen() \
938 DEBUG_OPTIMISE_MORE_r({ \
939 Perl_re_printf( aTHX_ "RExC_seen: "); \
941 if (RExC_seen & REG_ZERO_LEN_SEEN) \
942 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
944 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
945 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
947 if (RExC_seen & REG_GPOS_SEEN) \
948 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
950 if (RExC_seen & REG_RECURSE_SEEN) \
951 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
953 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
954 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
956 if (RExC_seen & REG_VERBARG_SEEN) \
957 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
959 if (RExC_seen & REG_CUTGROUP_SEEN) \
960 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
962 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
963 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
965 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
966 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
968 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
969 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
971 Perl_re_printf( aTHX_ "\n"); \
974 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
975 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
980 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
981 const char *close_str)
986 Perl_re_printf( aTHX_ "%s", open_str);
987 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
988 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
989 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
990 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
991 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
992 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
993 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
994 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
995 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
996 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
997 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
998 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
999 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1000 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1001 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1002 Perl_re_printf( aTHX_ "%s", close_str);
1007 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1008 U32 depth, int is_inf)
1010 GET_RE_DEBUG_FLAGS_DECL;
1012 DEBUG_OPTIMISE_MORE_r({
1015 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1019 (IV)data->pos_delta,
1023 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1025 Perl_re_printf( aTHX_
1026 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1028 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1029 is_inf ? "INF " : ""
1032 if (data->last_found) {
1034 Perl_re_printf(aTHX_
1035 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1036 SvPVX_const(data->last_found),
1038 (IV)data->last_start_min,
1039 (IV)data->last_start_max
1042 for (i = 0; i < 2; i++) {
1043 Perl_re_printf(aTHX_
1044 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1045 data->cur_is_floating == i ? "*" : "",
1046 i ? "Float" : "Fixed",
1047 SvPVX_const(data->substrs[i].str),
1048 (IV)data->substrs[i].min_offset,
1049 (IV)data->substrs[i].max_offset
1051 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1055 Perl_re_printf( aTHX_ "\n");
1061 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1062 regnode *scan, U32 depth, U32 flags)
1064 GET_RE_DEBUG_FLAGS_DECL;
1071 Next = regnext(scan);
1072 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1073 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1076 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1077 Next ? (REG_NODE_NUM(Next)) : 0 );
1078 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1079 Perl_re_printf( aTHX_ "\n");
1084 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1085 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1087 # define DEBUG_PEEP(str, scan, depth, flags) \
1088 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1091 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1092 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1096 /* =========================================================
1097 * BEGIN edit_distance stuff.
1099 * This calculates how many single character changes of any type are needed to
1100 * transform a string into another one. It is taken from version 3.1 of
1102 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1105 /* Our unsorted dictionary linked list. */
1106 /* Note we use UVs, not chars. */
1111 struct dictionary* next;
1113 typedef struct dictionary item;
1116 PERL_STATIC_INLINE item*
1117 push(UV key,item* curr)
1120 Newx(head, 1, item);
1128 PERL_STATIC_INLINE item*
1129 find(item* head, UV key)
1131 item* iterator = head;
1133 if (iterator->key == key){
1136 iterator = iterator->next;
1142 PERL_STATIC_INLINE item*
1143 uniquePush(item* head,UV key)
1145 item* iterator = head;
1148 if (iterator->key == key) {
1151 iterator = iterator->next;
1154 return push(key,head);
1157 PERL_STATIC_INLINE void
1158 dict_free(item* head)
1160 item* iterator = head;
1163 item* temp = iterator;
1164 iterator = iterator->next;
1171 /* End of Dictionary Stuff */
1173 /* All calculations/work are done here */
1175 S_edit_distance(const UV* src,
1177 const STRLEN x, /* length of src[] */
1178 const STRLEN y, /* length of tgt[] */
1179 const SSize_t maxDistance
1183 UV swapCount,swapScore,targetCharCount,i,j;
1185 UV score_ceil = x + y;
1187 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1189 /* intialize matrix start values */
1190 Newx(scores, ( (x + 2) * (y + 2)), UV);
1191 scores[0] = score_ceil;
1192 scores[1 * (y + 2) + 0] = score_ceil;
1193 scores[0 * (y + 2) + 1] = score_ceil;
1194 scores[1 * (y + 2) + 1] = 0;
1195 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1200 for (i=1;i<=x;i++) {
1202 head = uniquePush(head,src[i]);
1203 scores[(i+1) * (y + 2) + 1] = i;
1204 scores[(i+1) * (y + 2) + 0] = score_ceil;
1207 for (j=1;j<=y;j++) {
1210 head = uniquePush(head,tgt[j]);
1211 scores[1 * (y + 2) + (j + 1)] = j;
1212 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1215 targetCharCount = find(head,tgt[j-1])->value;
1216 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1218 if (src[i-1] != tgt[j-1]){
1219 scores[(i+1) * (y + 2) + (j + 1)] = MIN(swapScore,(MIN(scores[i * (y + 2) + j], MIN(scores[(i+1) * (y + 2) + j], scores[i * (y + 2) + (j + 1)])) + 1));
1223 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1227 find(head,src[i-1])->value = i;
1231 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1234 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1238 /* END of edit_distance() stuff
1239 * ========================================================= */
1241 /* is c a control character for which we have a mnemonic? */
1242 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1245 S_cntrl_to_mnemonic(const U8 c)
1247 /* Returns the mnemonic string that represents character 'c', if one
1248 * exists; NULL otherwise. The only ones that exist for the purposes of
1249 * this routine are a few control characters */
1252 case '\a': return "\\a";
1253 case '\b': return "\\b";
1254 case ESC_NATIVE: return "\\e";
1255 case '\f': return "\\f";
1256 case '\n': return "\\n";
1257 case '\r': return "\\r";
1258 case '\t': return "\\t";
1264 /* Mark that we cannot extend a found fixed substring at this point.
1265 Update the longest found anchored substring or the longest found
1266 floating substrings if needed. */
1269 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1270 SSize_t *minlenp, int is_inf)
1272 const STRLEN l = CHR_SVLEN(data->last_found);
1273 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1274 const STRLEN old_l = CHR_SVLEN(longest_sv);
1275 GET_RE_DEBUG_FLAGS_DECL;
1277 PERL_ARGS_ASSERT_SCAN_COMMIT;
1279 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1280 const U8 i = data->cur_is_floating;
1281 SvSetMagicSV(longest_sv, data->last_found);
1282 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1285 data->substrs[0].max_offset = data->substrs[0].min_offset;
1287 data->substrs[1].max_offset = (l
1288 ? data->last_start_max
1289 : (data->pos_delta > SSize_t_MAX - data->pos_min
1291 : data->pos_min + data->pos_delta));
1293 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1294 data->substrs[1].max_offset = SSize_t_MAX;
1297 if (data->flags & SF_BEFORE_EOL)
1298 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1300 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1301 data->substrs[i].minlenp = minlenp;
1302 data->substrs[i].lookbehind = 0;
1305 SvCUR_set(data->last_found, 0);
1307 SV * const sv = data->last_found;
1308 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1309 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1314 data->last_end = -1;
1315 data->flags &= ~SF_BEFORE_EOL;
1316 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1319 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1320 * list that describes which code points it matches */
1323 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1325 /* Set the SSC 'ssc' to match an empty string or any code point */
1327 PERL_ARGS_ASSERT_SSC_ANYTHING;
1329 assert(is_ANYOF_SYNTHETIC(ssc));
1331 /* mortalize so won't leak */
1332 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1333 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1337 S_ssc_is_anything(const regnode_ssc *ssc)
1339 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1340 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1341 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1342 * in any way, so there's no point in using it */
1347 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1349 assert(is_ANYOF_SYNTHETIC(ssc));
1351 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1355 /* See if the list consists solely of the range 0 - Infinity */
1356 invlist_iterinit(ssc->invlist);
1357 ret = invlist_iternext(ssc->invlist, &start, &end)
1361 invlist_iterfinish(ssc->invlist);
1367 /* If e.g., both \w and \W are set, matches everything */
1368 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1370 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1371 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1381 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1383 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1384 * string, any code point, or any posix class under locale */
1386 PERL_ARGS_ASSERT_SSC_INIT;
1388 Zero(ssc, 1, regnode_ssc);
1389 set_ANYOF_SYNTHETIC(ssc);
1390 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1393 /* If any portion of the regex is to operate under locale rules that aren't
1394 * fully known at compile time, initialization includes it. The reason
1395 * this isn't done for all regexes is that the optimizer was written under
1396 * the assumption that locale was all-or-nothing. Given the complexity and
1397 * lack of documentation in the optimizer, and that there are inadequate
1398 * test cases for locale, many parts of it may not work properly, it is
1399 * safest to avoid locale unless necessary. */
1400 if (RExC_contains_locale) {
1401 ANYOF_POSIXL_SETALL(ssc);
1404 ANYOF_POSIXL_ZERO(ssc);
1409 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1410 const regnode_ssc *ssc)
1412 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1413 * to the list of code points matched, and locale posix classes; hence does
1414 * not check its flags) */
1419 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1421 assert(is_ANYOF_SYNTHETIC(ssc));
1423 invlist_iterinit(ssc->invlist);
1424 ret = invlist_iternext(ssc->invlist, &start, &end)
1428 invlist_iterfinish(ssc->invlist);
1434 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1442 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1443 const regnode_charclass* const node)
1445 /* Returns a mortal inversion list defining which code points are matched
1446 * by 'node', which is of type ANYOF. Handles complementing the result if
1447 * appropriate. If some code points aren't knowable at this time, the
1448 * returned list must, and will, contain every code point that is a
1452 SV* only_utf8_locale_invlist = NULL;
1454 const U32 n = ARG(node);
1455 bool new_node_has_latin1 = FALSE;
1457 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1459 /* Look at the data structure created by S_set_ANYOF_arg() */
1460 if (n != ANYOF_ONLY_HAS_BITMAP) {
1461 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1462 AV * const av = MUTABLE_AV(SvRV(rv));
1463 SV **const ary = AvARRAY(av);
1464 assert(RExC_rxi->data->what[n] == 's');
1466 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1467 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1469 else if (ary[0] && ary[0] != &PL_sv_undef) {
1471 /* Here, no compile-time swash, and there are things that won't be
1472 * known until runtime -- we have to assume it could be anything */
1473 invlist = sv_2mortal(_new_invlist(1));
1474 return _add_range_to_invlist(invlist, 0, UV_MAX);
1476 else if (ary[3] && ary[3] != &PL_sv_undef) {
1478 /* Here no compile-time swash, and no run-time only data. Use the
1479 * node's inversion list */
1480 invlist = sv_2mortal(invlist_clone(ary[3]));
1483 /* Get the code points valid only under UTF-8 locales */
1484 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1485 && ary[2] && ary[2] != &PL_sv_undef)
1487 only_utf8_locale_invlist = ary[2];
1492 invlist = sv_2mortal(_new_invlist(0));
1495 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1496 * code points, and an inversion list for the others, but if there are code
1497 * points that should match only conditionally on the target string being
1498 * UTF-8, those are placed in the inversion list, and not the bitmap.
1499 * Since there are circumstances under which they could match, they are
1500 * included in the SSC. But if the ANYOF node is to be inverted, we have
1501 * to exclude them here, so that when we invert below, the end result
1502 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1503 * have to do this here before we add the unconditionally matched code
1505 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1506 _invlist_intersection_complement_2nd(invlist,
1511 /* Add in the points from the bit map */
1512 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1513 if (ANYOF_BITMAP_TEST(node, i)) {
1514 unsigned int start = i++;
1516 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1519 invlist = _add_range_to_invlist(invlist, start, i-1);
1520 new_node_has_latin1 = TRUE;
1524 /* If this can match all upper Latin1 code points, have to add them
1525 * as well. But don't add them if inverting, as when that gets done below,
1526 * it would exclude all these characters, including the ones it shouldn't
1527 * that were added just above */
1528 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1529 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1531 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1534 /* Similarly for these */
1535 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1536 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1539 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1540 _invlist_invert(invlist);
1542 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1544 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1545 * locale. We can skip this if there are no 0-255 at all. */
1546 _invlist_union(invlist, PL_Latin1, &invlist);
1549 /* Similarly add the UTF-8 locale possible matches. These have to be
1550 * deferred until after the non-UTF-8 locale ones are taken care of just
1551 * above, or it leads to wrong results under ANYOF_INVERT */
1552 if (only_utf8_locale_invlist) {
1553 _invlist_union_maybe_complement_2nd(invlist,
1554 only_utf8_locale_invlist,
1555 ANYOF_FLAGS(node) & ANYOF_INVERT,
1562 /* These two functions currently do the exact same thing */
1563 #define ssc_init_zero ssc_init
1565 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1566 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1568 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1569 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1570 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1573 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1574 const regnode_charclass *and_with)
1576 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1577 * another SSC or a regular ANYOF class. Can create false positives. */
1582 PERL_ARGS_ASSERT_SSC_AND;
1584 assert(is_ANYOF_SYNTHETIC(ssc));
1586 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1587 * the code point inversion list and just the relevant flags */
1588 if (is_ANYOF_SYNTHETIC(and_with)) {
1589 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1590 anded_flags = ANYOF_FLAGS(and_with);
1592 /* XXX This is a kludge around what appears to be deficiencies in the
1593 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1594 * there are paths through the optimizer where it doesn't get weeded
1595 * out when it should. And if we don't make some extra provision for
1596 * it like the code just below, it doesn't get added when it should.
1597 * This solution is to add it only when AND'ing, which is here, and
1598 * only when what is being AND'ed is the pristine, original node
1599 * matching anything. Thus it is like adding it to ssc_anything() but
1600 * only when the result is to be AND'ed. Probably the same solution
1601 * could be adopted for the same problem we have with /l matching,
1602 * which is solved differently in S_ssc_init(), and that would lead to
1603 * fewer false positives than that solution has. But if this solution
1604 * creates bugs, the consequences are only that a warning isn't raised
1605 * that should be; while the consequences for having /l bugs is
1606 * incorrect matches */
1607 if (ssc_is_anything((regnode_ssc *)and_with)) {
1608 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1612 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1613 if (OP(and_with) == ANYOFD) {
1614 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1617 anded_flags = ANYOF_FLAGS(and_with)
1618 &( ANYOF_COMMON_FLAGS
1619 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1620 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1621 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1623 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1628 ANYOF_FLAGS(ssc) &= anded_flags;
1630 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1631 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1632 * 'and_with' may be inverted. When not inverted, we have the situation of
1634 * (C1 | P1) & (C2 | P2)
1635 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1636 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1637 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1638 * <= ((C1 & C2) | P1 | P2)
1639 * Alternatively, the last few steps could be:
1640 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1641 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1642 * <= (C1 | C2 | (P1 & P2))
1643 * We favor the second approach if either P1 or P2 is non-empty. This is
1644 * because these components are a barrier to doing optimizations, as what
1645 * they match cannot be known until the moment of matching as they are
1646 * dependent on the current locale, 'AND"ing them likely will reduce or
1648 * But we can do better if we know that C1,P1 are in their initial state (a
1649 * frequent occurrence), each matching everything:
1650 * (<everything>) & (C2 | P2) = C2 | P2
1651 * Similarly, if C2,P2 are in their initial state (again a frequent
1652 * occurrence), the result is a no-op
1653 * (C1 | P1) & (<everything>) = C1 | P1
1656 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1657 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1658 * <= (C1 & ~C2) | (P1 & ~P2)
1661 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1662 && ! is_ANYOF_SYNTHETIC(and_with))
1666 ssc_intersection(ssc,
1668 FALSE /* Has already been inverted */
1671 /* If either P1 or P2 is empty, the intersection will be also; can skip
1673 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1674 ANYOF_POSIXL_ZERO(ssc);
1676 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1678 /* Note that the Posix class component P from 'and_with' actually
1680 * P = Pa | Pb | ... | Pn
1681 * where each component is one posix class, such as in [\w\s].
1683 * ~P = ~(Pa | Pb | ... | Pn)
1684 * = ~Pa & ~Pb & ... & ~Pn
1685 * <= ~Pa | ~Pb | ... | ~Pn
1686 * The last is something we can easily calculate, but unfortunately
1687 * is likely to have many false positives. We could do better
1688 * in some (but certainly not all) instances if two classes in
1689 * P have known relationships. For example
1690 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1692 * :lower: & :print: = :lower:
1693 * And similarly for classes that must be disjoint. For example,
1694 * since \s and \w can have no elements in common based on rules in
1695 * the POSIX standard,
1696 * \w & ^\S = nothing
1697 * Unfortunately, some vendor locales do not meet the Posix
1698 * standard, in particular almost everything by Microsoft.
1699 * The loop below just changes e.g., \w into \W and vice versa */
1701 regnode_charclass_posixl temp;
1702 int add = 1; /* To calculate the index of the complement */
1704 Zero(&temp, 1, regnode_charclass_posixl);
1705 ANYOF_POSIXL_ZERO(&temp);
1706 for (i = 0; i < ANYOF_MAX; i++) {
1708 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1709 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1711 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1712 ANYOF_POSIXL_SET(&temp, i + add);
1714 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1716 ANYOF_POSIXL_AND(&temp, ssc);
1718 } /* else ssc already has no posixes */
1719 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1720 in its initial state */
1721 else if (! is_ANYOF_SYNTHETIC(and_with)
1722 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1724 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1725 * copy it over 'ssc' */
1726 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1727 if (is_ANYOF_SYNTHETIC(and_with)) {
1728 StructCopy(and_with, ssc, regnode_ssc);
1731 ssc->invlist = anded_cp_list;
1732 ANYOF_POSIXL_ZERO(ssc);
1733 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1734 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1738 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1739 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1741 /* One or the other of P1, P2 is non-empty. */
1742 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1743 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1745 ssc_union(ssc, anded_cp_list, FALSE);
1747 else { /* P1 = P2 = empty */
1748 ssc_intersection(ssc, anded_cp_list, FALSE);
1754 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1755 const regnode_charclass *or_with)
1757 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1758 * another SSC or a regular ANYOF class. Can create false positives if
1759 * 'or_with' is to be inverted. */
1764 PERL_ARGS_ASSERT_SSC_OR;
1766 assert(is_ANYOF_SYNTHETIC(ssc));
1768 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1769 * the code point inversion list and just the relevant flags */
1770 if (is_ANYOF_SYNTHETIC(or_with)) {
1771 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1772 ored_flags = ANYOF_FLAGS(or_with);
1775 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1776 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1777 if (OP(or_with) != ANYOFD) {
1779 |= ANYOF_FLAGS(or_with)
1780 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1781 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1782 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1784 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1789 ANYOF_FLAGS(ssc) |= ored_flags;
1791 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1792 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1793 * 'or_with' may be inverted. When not inverted, we have the simple
1794 * situation of computing:
1795 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1796 * If P1|P2 yields a situation with both a class and its complement are
1797 * set, like having both \w and \W, this matches all code points, and we
1798 * can delete these from the P component of the ssc going forward. XXX We
1799 * might be able to delete all the P components, but I (khw) am not certain
1800 * about this, and it is better to be safe.
1803 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1804 * <= (C1 | P1) | ~C2
1805 * <= (C1 | ~C2) | P1
1806 * (which results in actually simpler code than the non-inverted case)
1809 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1810 && ! is_ANYOF_SYNTHETIC(or_with))
1812 /* We ignore P2, leaving P1 going forward */
1813 } /* else Not inverted */
1814 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1815 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1816 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1818 for (i = 0; i < ANYOF_MAX; i += 2) {
1819 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1821 ssc_match_all_cp(ssc);
1822 ANYOF_POSIXL_CLEAR(ssc, i);
1823 ANYOF_POSIXL_CLEAR(ssc, i+1);
1831 FALSE /* Already has been inverted */
1835 PERL_STATIC_INLINE void
1836 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1838 PERL_ARGS_ASSERT_SSC_UNION;
1840 assert(is_ANYOF_SYNTHETIC(ssc));
1842 _invlist_union_maybe_complement_2nd(ssc->invlist,
1848 PERL_STATIC_INLINE void
1849 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1851 const bool invert2nd)
1853 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1855 assert(is_ANYOF_SYNTHETIC(ssc));
1857 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1863 PERL_STATIC_INLINE void
1864 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1866 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1868 assert(is_ANYOF_SYNTHETIC(ssc));
1870 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1873 PERL_STATIC_INLINE void
1874 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1876 /* AND just the single code point 'cp' into the SSC 'ssc' */
1878 SV* cp_list = _new_invlist(2);
1880 PERL_ARGS_ASSERT_SSC_CP_AND;
1882 assert(is_ANYOF_SYNTHETIC(ssc));
1884 cp_list = add_cp_to_invlist(cp_list, cp);
1885 ssc_intersection(ssc, cp_list,
1886 FALSE /* Not inverted */
1888 SvREFCNT_dec_NN(cp_list);
1891 PERL_STATIC_INLINE void
1892 S_ssc_clear_locale(regnode_ssc *ssc)
1894 /* Set the SSC 'ssc' to not match any locale things */
1895 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1897 assert(is_ANYOF_SYNTHETIC(ssc));
1899 ANYOF_POSIXL_ZERO(ssc);
1900 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1903 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1906 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1908 /* The synthetic start class is used to hopefully quickly winnow down
1909 * places where a pattern could start a match in the target string. If it
1910 * doesn't really narrow things down that much, there isn't much point to
1911 * having the overhead of using it. This function uses some very crude
1912 * heuristics to decide if to use the ssc or not.
1914 * It returns TRUE if 'ssc' rules out more than half what it considers to
1915 * be the "likely" possible matches, but of course it doesn't know what the
1916 * actual things being matched are going to be; these are only guesses
1918 * For /l matches, it assumes that the only likely matches are going to be
1919 * in the 0-255 range, uniformly distributed, so half of that is 127
1920 * For /a and /d matches, it assumes that the likely matches will be just
1921 * the ASCII range, so half of that is 63
1922 * For /u and there isn't anything matching above the Latin1 range, it
1923 * assumes that that is the only range likely to be matched, and uses
1924 * half that as the cut-off: 127. If anything matches above Latin1,
1925 * it assumes that all of Unicode could match (uniformly), except for
1926 * non-Unicode code points and things in the General Category "Other"
1927 * (unassigned, private use, surrogates, controls and formats). This
1928 * is a much large number. */
1930 U32 count = 0; /* Running total of number of code points matched by
1932 UV start, end; /* Start and end points of current range in inversion
1934 const U32 max_code_points = (LOC)
1936 : (( ! UNI_SEMANTICS
1937 || invlist_highest(ssc->invlist) < 256)
1940 const U32 max_match = max_code_points / 2;
1942 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1944 invlist_iterinit(ssc->invlist);
1945 while (invlist_iternext(ssc->invlist, &start, &end)) {
1946 if (start >= max_code_points) {
1949 end = MIN(end, max_code_points - 1);
1950 count += end - start + 1;
1951 if (count >= max_match) {
1952 invlist_iterfinish(ssc->invlist);
1962 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1964 /* The inversion list in the SSC is marked mortal; now we need a more
1965 * permanent copy, which is stored the same way that is done in a regular
1966 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1969 SV* invlist = invlist_clone(ssc->invlist);
1971 PERL_ARGS_ASSERT_SSC_FINALIZE;
1973 assert(is_ANYOF_SYNTHETIC(ssc));
1975 /* The code in this file assumes that all but these flags aren't relevant
1976 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1977 * by the time we reach here */
1978 assert(! (ANYOF_FLAGS(ssc)
1979 & ~( ANYOF_COMMON_FLAGS
1980 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1981 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
1983 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1985 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1986 NULL, NULL, NULL, FALSE);
1988 /* Make sure is clone-safe */
1989 ssc->invlist = NULL;
1991 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1992 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1995 if (RExC_contains_locale) {
1999 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2002 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2003 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2004 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2005 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2006 ? (TRIE_LIST_CUR( idx ) - 1) \
2012 dump_trie(trie,widecharmap,revcharmap)
2013 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2014 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2016 These routines dump out a trie in a somewhat readable format.
2017 The _interim_ variants are used for debugging the interim
2018 tables that are used to generate the final compressed
2019 representation which is what dump_trie expects.
2021 Part of the reason for their existence is to provide a form
2022 of documentation as to how the different representations function.
2027 Dumps the final compressed table form of the trie to Perl_debug_log.
2028 Used for debugging make_trie().
2032 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2033 AV *revcharmap, U32 depth)
2036 SV *sv=sv_newmortal();
2037 int colwidth= widecharmap ? 6 : 4;
2039 GET_RE_DEBUG_FLAGS_DECL;
2041 PERL_ARGS_ASSERT_DUMP_TRIE;
2043 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2044 depth+1, "Match","Base","Ofs" );
2046 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2047 SV ** const tmp = av_fetch( revcharmap, state, 0);
2049 Perl_re_printf( aTHX_ "%*s",
2051 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2052 PL_colors[0], PL_colors[1],
2053 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2054 PERL_PV_ESCAPE_FIRSTCHAR
2059 Perl_re_printf( aTHX_ "\n");
2060 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2062 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2063 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2064 Perl_re_printf( aTHX_ "\n");
2066 for( state = 1 ; state < trie->statecount ; state++ ) {
2067 const U32 base = trie->states[ state ].trans.base;
2069 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2071 if ( trie->states[ state ].wordnum ) {
2072 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2074 Perl_re_printf( aTHX_ "%6s", "" );
2077 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2082 while( ( base + ofs < trie->uniquecharcount ) ||
2083 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2084 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2088 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2090 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2091 if ( ( base + ofs >= trie->uniquecharcount )
2092 && ( base + ofs - trie->uniquecharcount
2094 && trie->trans[ base + ofs
2095 - trie->uniquecharcount ].check == state )
2097 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2098 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2101 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
2105 Perl_re_printf( aTHX_ "]");
2108 Perl_re_printf( aTHX_ "\n" );
2110 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2112 for (word=1; word <= trie->wordcount; word++) {
2113 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2114 (int)word, (int)(trie->wordinfo[word].prev),
2115 (int)(trie->wordinfo[word].len));
2117 Perl_re_printf( aTHX_ "\n" );
2120 Dumps a fully constructed but uncompressed trie in list form.
2121 List tries normally only are used for construction when the number of
2122 possible chars (trie->uniquecharcount) is very high.
2123 Used for debugging make_trie().
2126 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2127 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2131 SV *sv=sv_newmortal();
2132 int colwidth= widecharmap ? 6 : 4;
2133 GET_RE_DEBUG_FLAGS_DECL;
2135 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2137 /* print out the table precompression. */
2138 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2140 Perl_re_indentf( aTHX_ "%s",
2141 depth+1, "------:-----+-----------------\n" );
2143 for( state=1 ; state < next_alloc ; state ++ ) {
2146 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2147 depth+1, (UV)state );
2148 if ( ! trie->states[ state ].wordnum ) {
2149 Perl_re_printf( aTHX_ "%5s| ","");
2151 Perl_re_printf( aTHX_ "W%4x| ",
2152 trie->states[ state ].wordnum
2155 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2156 SV ** const tmp = av_fetch( revcharmap,
2157 TRIE_LIST_ITEM(state,charid).forid, 0);
2159 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2161 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2163 PL_colors[0], PL_colors[1],
2164 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2165 | PERL_PV_ESCAPE_FIRSTCHAR
2167 TRIE_LIST_ITEM(state,charid).forid,
2168 (UV)TRIE_LIST_ITEM(state,charid).newstate
2171 Perl_re_printf( aTHX_ "\n%*s| ",
2172 (int)((depth * 2) + 14), "");
2175 Perl_re_printf( aTHX_ "\n");
2180 Dumps a fully constructed but uncompressed trie in table form.
2181 This is the normal DFA style state transition table, with a few
2182 twists to facilitate compression later.
2183 Used for debugging make_trie().
2186 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2187 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2192 SV *sv=sv_newmortal();
2193 int colwidth= widecharmap ? 6 : 4;
2194 GET_RE_DEBUG_FLAGS_DECL;
2196 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2199 print out the table precompression so that we can do a visual check
2200 that they are identical.
2203 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2205 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2206 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2208 Perl_re_printf( aTHX_ "%*s",
2210 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2211 PL_colors[0], PL_colors[1],
2212 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2213 PERL_PV_ESCAPE_FIRSTCHAR
2219 Perl_re_printf( aTHX_ "\n");
2220 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2222 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2223 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2226 Perl_re_printf( aTHX_ "\n" );
2228 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2230 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2232 (UV)TRIE_NODENUM( state ) );
2234 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2235 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2237 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2239 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2241 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2242 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2243 (UV)trie->trans[ state ].check );
2245 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2246 (UV)trie->trans[ state ].check,
2247 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2255 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2256 startbranch: the first branch in the whole branch sequence
2257 first : start branch of sequence of branch-exact nodes.
2258 May be the same as startbranch
2259 last : Thing following the last branch.
2260 May be the same as tail.
2261 tail : item following the branch sequence
2262 count : words in the sequence
2263 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2264 depth : indent depth
2266 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2268 A trie is an N'ary tree where the branches are determined by digital
2269 decomposition of the key. IE, at the root node you look up the 1st character and
2270 follow that branch repeat until you find the end of the branches. Nodes can be
2271 marked as "accepting" meaning they represent a complete word. Eg:
2275 would convert into the following structure. Numbers represent states, letters
2276 following numbers represent valid transitions on the letter from that state, if
2277 the number is in square brackets it represents an accepting state, otherwise it
2278 will be in parenthesis.
2280 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2284 (1) +-i->(6)-+-s->[7]
2286 +-s->(3)-+-h->(4)-+-e->[5]
2288 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2290 This shows that when matching against the string 'hers' we will begin at state 1
2291 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2292 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2293 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2294 single traverse. We store a mapping from accepting to state to which word was
2295 matched, and then when we have multiple possibilities we try to complete the
2296 rest of the regex in the order in which they occurred in the alternation.
2298 The only prior NFA like behaviour that would be changed by the TRIE support is
2299 the silent ignoring of duplicate alternations which are of the form:
2301 / (DUPE|DUPE) X? (?{ ... }) Y /x
2303 Thus EVAL blocks following a trie may be called a different number of times with
2304 and without the optimisation. With the optimisations dupes will be silently
2305 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2306 the following demonstrates:
2308 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2310 which prints out 'word' three times, but
2312 'words'=~/(word|word|word)(?{ print $1 })S/
2314 which doesnt print it out at all. This is due to other optimisations kicking in.
2316 Example of what happens on a structural level:
2318 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2320 1: CURLYM[1] {1,32767}(18)
2331 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2332 and should turn into:
2334 1: CURLYM[1] {1,32767}(18)
2336 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2344 Cases where tail != last would be like /(?foo|bar)baz/:
2354 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2355 and would end up looking like:
2358 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2365 d = uvchr_to_utf8_flags(d, uv, 0);
2367 is the recommended Unicode-aware way of saying
2372 #define TRIE_STORE_REVCHAR(val) \
2375 SV *zlopp = newSV(UTF8_MAXBYTES); \
2376 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2377 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2378 SvCUR_set(zlopp, kapow - flrbbbbb); \
2381 av_push(revcharmap, zlopp); \
2383 char ooooff = (char)val; \
2384 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2388 /* This gets the next character from the input, folding it if not already
2390 #define TRIE_READ_CHAR STMT_START { \
2393 /* if it is UTF then it is either already folded, or does not need \
2395 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2397 else if (folder == PL_fold_latin1) { \
2398 /* This folder implies Unicode rules, which in the range expressible \
2399 * by not UTF is the lower case, with the two exceptions, one of \
2400 * which should have been taken care of before calling this */ \
2401 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2402 uvc = toLOWER_L1(*uc); \
2403 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2406 /* raw data, will be folded later if needed */ \
2414 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2415 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2416 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2417 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2418 TRIE_LIST_LEN( state ) = ging; \
2420 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2421 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2422 TRIE_LIST_CUR( state )++; \
2425 #define TRIE_LIST_NEW(state) STMT_START { \
2426 Newx( trie->states[ state ].trans.list, \
2427 4, reg_trie_trans_le ); \
2428 TRIE_LIST_CUR( state ) = 1; \
2429 TRIE_LIST_LEN( state ) = 4; \
2432 #define TRIE_HANDLE_WORD(state) STMT_START { \
2433 U16 dupe= trie->states[ state ].wordnum; \
2434 regnode * const noper_next = regnext( noper ); \
2437 /* store the word for dumping */ \
2439 if (OP(noper) != NOTHING) \
2440 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2442 tmp = newSVpvn_utf8( "", 0, UTF ); \
2443 av_push( trie_words, tmp ); \
2447 trie->wordinfo[curword].prev = 0; \
2448 trie->wordinfo[curword].len = wordlen; \
2449 trie->wordinfo[curword].accept = state; \
2451 if ( noper_next < tail ) { \
2453 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2455 trie->jump[curword] = (U16)(noper_next - convert); \
2457 jumper = noper_next; \
2459 nextbranch= regnext(cur); \
2463 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2464 /* chain, so that when the bits of chain are later */\
2465 /* linked together, the dups appear in the chain */\
2466 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2467 trie->wordinfo[dupe].prev = curword; \
2469 /* we haven't inserted this word yet. */ \
2470 trie->states[ state ].wordnum = curword; \
2475 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2476 ( ( base + charid >= ucharcount \
2477 && base + charid < ubound \
2478 && state == trie->trans[ base - ucharcount + charid ].check \
2479 && trie->trans[ base - ucharcount + charid ].next ) \
2480 ? trie->trans[ base - ucharcount + charid ].next \
2481 : ( state==1 ? special : 0 ) \
2484 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2486 TRIE_BITMAP_SET(trie, uvc); \
2487 /* store the folded codepoint */ \
2489 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2492 /* store first byte of utf8 representation of */ \
2493 /* variant codepoints */ \
2494 if (! UVCHR_IS_INVARIANT(uvc)) { \
2495 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2500 #define MADE_JUMP_TRIE 2
2501 #define MADE_EXACT_TRIE 4
2504 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2505 regnode *first, regnode *last, regnode *tail,
2506 U32 word_count, U32 flags, U32 depth)
2508 /* first pass, loop through and scan words */
2509 reg_trie_data *trie;
2510 HV *widecharmap = NULL;
2511 AV *revcharmap = newAV();
2517 regnode *jumper = NULL;
2518 regnode *nextbranch = NULL;
2519 regnode *convert = NULL;
2520 U32 *prev_states; /* temp array mapping each state to previous one */
2521 /* we just use folder as a flag in utf8 */
2522 const U8 * folder = NULL;
2524 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2525 * which stands for one trie structure, one hash, optionally followed
2528 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2529 AV *trie_words = NULL;
2530 /* along with revcharmap, this only used during construction but both are
2531 * useful during debugging so we store them in the struct when debugging.
2534 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2535 STRLEN trie_charcount=0;
2537 SV *re_trie_maxbuff;
2538 GET_RE_DEBUG_FLAGS_DECL;
2540 PERL_ARGS_ASSERT_MAKE_TRIE;
2542 PERL_UNUSED_ARG(depth);
2546 case EXACT: case EXACTL: break;
2550 case EXACTFLU8: folder = PL_fold_latin1; break;
2551 case EXACTF: folder = PL_fold; break;
2552 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2555 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2557 trie->startstate = 1;
2558 trie->wordcount = word_count;
2559 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2560 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2561 if (flags == EXACT || flags == EXACTL)
2562 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2563 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2564 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2567 trie_words = newAV();
2570 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2571 assert(re_trie_maxbuff);
2572 if (!SvIOK(re_trie_maxbuff)) {
2573 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2575 DEBUG_TRIE_COMPILE_r({
2576 Perl_re_indentf( aTHX_
2577 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2579 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2580 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2583 /* Find the node we are going to overwrite */
2584 if ( first == startbranch && OP( last ) != BRANCH ) {
2585 /* whole branch chain */
2588 /* branch sub-chain */
2589 convert = NEXTOPER( first );
2592 /* -- First loop and Setup --
2594 We first traverse the branches and scan each word to determine if it
2595 contains widechars, and how many unique chars there are, this is
2596 important as we have to build a table with at least as many columns as we
2599 We use an array of integers to represent the character codes 0..255
2600 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2601 the native representation of the character value as the key and IV's for
2604 *TODO* If we keep track of how many times each character is used we can
2605 remap the columns so that the table compression later on is more
2606 efficient in terms of memory by ensuring the most common value is in the
2607 middle and the least common are on the outside. IMO this would be better
2608 than a most to least common mapping as theres a decent chance the most
2609 common letter will share a node with the least common, meaning the node
2610 will not be compressible. With a middle is most common approach the worst
2611 case is when we have the least common nodes twice.
2615 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2616 regnode *noper = NEXTOPER( cur );
2620 U32 wordlen = 0; /* required init */
2621 STRLEN minchars = 0;
2622 STRLEN maxchars = 0;
2623 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2626 if (OP(noper) == NOTHING) {
2627 /* skip past a NOTHING at the start of an alternation
2628 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2630 regnode *noper_next= regnext(noper);
2631 if (noper_next < tail)
2635 if ( noper < tail &&
2637 OP(noper) == flags ||
2640 OP(noper) == EXACTFU_SS
2644 uc= (U8*)STRING(noper);
2645 e= uc + STR_LEN(noper);
2652 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2653 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2654 regardless of encoding */
2655 if (OP( noper ) == EXACTFU_SS) {
2656 /* false positives are ok, so just set this */
2657 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2661 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2663 TRIE_CHARCOUNT(trie)++;
2666 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2667 * is in effect. Under /i, this character can match itself, or
2668 * anything that folds to it. If not under /i, it can match just
2669 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2670 * all fold to k, and all are single characters. But some folds
2671 * expand to more than one character, so for example LATIN SMALL
2672 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2673 * the string beginning at 'uc' is 'ffi', it could be matched by
2674 * three characters, or just by the one ligature character. (It
2675 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2676 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2677 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2678 * match.) The trie needs to know the minimum and maximum number
2679 * of characters that could match so that it can use size alone to
2680 * quickly reject many match attempts. The max is simple: it is
2681 * the number of folded characters in this branch (since a fold is
2682 * never shorter than what folds to it. */
2686 /* And the min is equal to the max if not under /i (indicated by
2687 * 'folder' being NULL), or there are no multi-character folds. If
2688 * there is a multi-character fold, the min is incremented just
2689 * once, for the character that folds to the sequence. Each
2690 * character in the sequence needs to be added to the list below of
2691 * characters in the trie, but we count only the first towards the
2692 * min number of characters needed. This is done through the
2693 * variable 'foldlen', which is returned by the macros that look
2694 * for these sequences as the number of bytes the sequence
2695 * occupies. Each time through the loop, we decrement 'foldlen' by
2696 * how many bytes the current char occupies. Only when it reaches
2697 * 0 do we increment 'minchars' or look for another multi-character
2699 if (folder == NULL) {
2702 else if (foldlen > 0) {
2703 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2708 /* See if *uc is the beginning of a multi-character fold. If
2709 * so, we decrement the length remaining to look at, to account
2710 * for the current character this iteration. (We can use 'uc'
2711 * instead of the fold returned by TRIE_READ_CHAR because for
2712 * non-UTF, the latin1_safe macro is smart enough to account
2713 * for all the unfolded characters, and because for UTF, the
2714 * string will already have been folded earlier in the
2715 * compilation process */
2717 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2718 foldlen -= UTF8SKIP(uc);
2721 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2726 /* The current character (and any potential folds) should be added
2727 * to the possible matching characters for this position in this
2731 U8 folded= folder[ (U8) uvc ];
2732 if ( !trie->charmap[ folded ] ) {
2733 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2734 TRIE_STORE_REVCHAR( folded );
2737 if ( !trie->charmap[ uvc ] ) {
2738 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2739 TRIE_STORE_REVCHAR( uvc );
2742 /* store the codepoint in the bitmap, and its folded
2744 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2745 set_bit = 0; /* We've done our bit :-) */
2749 /* XXX We could come up with the list of code points that fold
2750 * to this using PL_utf8_foldclosures, except not for
2751 * multi-char folds, as there may be multiple combinations
2752 * there that could work, which needs to wait until runtime to
2753 * resolve (The comment about LIGATURE FFI above is such an
2758 widecharmap = newHV();
2760 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2763 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2765 if ( !SvTRUE( *svpp ) ) {
2766 sv_setiv( *svpp, ++trie->uniquecharcount );
2767 TRIE_STORE_REVCHAR(uvc);
2770 } /* end loop through characters in this branch of the trie */
2772 /* We take the min and max for this branch and combine to find the min
2773 * and max for all branches processed so far */
2774 if( cur == first ) {
2775 trie->minlen = minchars;
2776 trie->maxlen = maxchars;
2777 } else if (minchars < trie->minlen) {
2778 trie->minlen = minchars;
2779 } else if (maxchars > trie->maxlen) {
2780 trie->maxlen = maxchars;
2782 } /* end first pass */
2783 DEBUG_TRIE_COMPILE_r(
2784 Perl_re_indentf( aTHX_
2785 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2787 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2788 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2789 (int)trie->minlen, (int)trie->maxlen )
2793 We now know what we are dealing with in terms of unique chars and
2794 string sizes so we can calculate how much memory a naive
2795 representation using a flat table will take. If it's over a reasonable
2796 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2797 conservative but potentially much slower representation using an array
2800 At the end we convert both representations into the same compressed
2801 form that will be used in regexec.c for matching with. The latter
2802 is a form that cannot be used to construct with but has memory
2803 properties similar to the list form and access properties similar
2804 to the table form making it both suitable for fast searches and
2805 small enough that its feasable to store for the duration of a program.
2807 See the comment in the code where the compressed table is produced
2808 inplace from the flat tabe representation for an explanation of how
2809 the compression works.
2814 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2817 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2818 > SvIV(re_trie_maxbuff) )
2821 Second Pass -- Array Of Lists Representation
2823 Each state will be represented by a list of charid:state records
2824 (reg_trie_trans_le) the first such element holds the CUR and LEN
2825 points of the allocated array. (See defines above).
2827 We build the initial structure using the lists, and then convert
2828 it into the compressed table form which allows faster lookups
2829 (but cant be modified once converted).
2832 STRLEN transcount = 1;
2834 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2837 trie->states = (reg_trie_state *)
2838 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2839 sizeof(reg_trie_state) );
2843 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2845 regnode *noper = NEXTOPER( cur );
2846 U32 state = 1; /* required init */
2847 U16 charid = 0; /* sanity init */
2848 U32 wordlen = 0; /* required init */
2850 if (OP(noper) == NOTHING) {
2851 regnode *noper_next= regnext(noper);
2852 if (noper_next < tail)
2856 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2857 const U8 *uc= (U8*)STRING(noper);
2858 const U8 *e= uc + STR_LEN(noper);
2860 for ( ; uc < e ; uc += len ) {
2865 charid = trie->charmap[ uvc ];
2867 SV** const svpp = hv_fetch( widecharmap,
2874 charid=(U16)SvIV( *svpp );
2877 /* charid is now 0 if we dont know the char read, or
2878 * nonzero if we do */
2885 if ( !trie->states[ state ].trans.list ) {
2886 TRIE_LIST_NEW( state );
2889 check <= TRIE_LIST_USED( state );
2892 if ( TRIE_LIST_ITEM( state, check ).forid
2895 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2900 newstate = next_alloc++;
2901 prev_states[newstate] = state;
2902 TRIE_LIST_PUSH( state, charid, newstate );
2907 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
2911 TRIE_HANDLE_WORD(state);
2913 } /* end second pass */
2915 /* next alloc is the NEXT state to be allocated */
2916 trie->statecount = next_alloc;
2917 trie->states = (reg_trie_state *)
2918 PerlMemShared_realloc( trie->states,
2920 * sizeof(reg_trie_state) );
2922 /* and now dump it out before we compress it */
2923 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2924 revcharmap, next_alloc,
2928 trie->trans = (reg_trie_trans *)
2929 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2936 for( state=1 ; state < next_alloc ; state ++ ) {
2940 DEBUG_TRIE_COMPILE_MORE_r(
2941 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
2945 if (trie->states[state].trans.list) {
2946 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2950 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2951 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2952 if ( forid < minid ) {
2954 } else if ( forid > maxid ) {
2958 if ( transcount < tp + maxid - minid + 1) {
2960 trie->trans = (reg_trie_trans *)
2961 PerlMemShared_realloc( trie->trans,
2963 * sizeof(reg_trie_trans) );
2964 Zero( trie->trans + (transcount / 2),
2968 base = trie->uniquecharcount + tp - minid;
2969 if ( maxid == minid ) {
2971 for ( ; zp < tp ; zp++ ) {
2972 if ( ! trie->trans[ zp ].next ) {
2973 base = trie->uniquecharcount + zp - minid;
2974 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2976 trie->trans[ zp ].check = state;
2982 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2984 trie->trans[ tp ].check = state;
2989 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2990 const U32 tid = base
2991 - trie->uniquecharcount
2992 + TRIE_LIST_ITEM( state, idx ).forid;
2993 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2995 trie->trans[ tid ].check = state;
2997 tp += ( maxid - minid + 1 );
2999 Safefree(trie->states[ state ].trans.list);
3002 DEBUG_TRIE_COMPILE_MORE_r(
3003 Perl_re_printf( aTHX_ " base: %d\n",base);
3006 trie->states[ state ].trans.base=base;
3008 trie->lasttrans = tp + 1;
3012 Second Pass -- Flat Table Representation.
3014 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3015 each. We know that we will need Charcount+1 trans at most to store
3016 the data (one row per char at worst case) So we preallocate both
3017 structures assuming worst case.
3019 We then construct the trie using only the .next slots of the entry
3022 We use the .check field of the first entry of the node temporarily
3023 to make compression both faster and easier by keeping track of how
3024 many non zero fields are in the node.
3026 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3029 There are two terms at use here: state as a TRIE_NODEIDX() which is
3030 a number representing the first entry of the node, and state as a
3031 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3032 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3033 if there are 2 entrys per node. eg:
3041 The table is internally in the right hand, idx form. However as we
3042 also have to deal with the states array which is indexed by nodenum
3043 we have to use TRIE_NODENUM() to convert.
3046 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3049 trie->trans = (reg_trie_trans *)
3050 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3051 * trie->uniquecharcount + 1,
3052 sizeof(reg_trie_trans) );
3053 trie->states = (reg_trie_state *)
3054 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3055 sizeof(reg_trie_state) );
3056 next_alloc = trie->uniquecharcount + 1;
3059 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3061 regnode *noper = NEXTOPER( cur );
3063 U32 state = 1; /* required init */
3065 U16 charid = 0; /* sanity init */
3066 U32 accept_state = 0; /* sanity init */
3068 U32 wordlen = 0; /* required init */
3070 if (OP(noper) == NOTHING) {
3071 regnode *noper_next= regnext(noper);
3072 if (noper_next < tail)
3076 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3077 const U8 *uc= (U8*)STRING(noper);
3078 const U8 *e= uc + STR_LEN(noper);
3080 for ( ; uc < e ; uc += len ) {
3085 charid = trie->charmap[ uvc ];
3087 SV* const * const svpp = hv_fetch( widecharmap,
3091 charid = svpp ? (U16)SvIV(*svpp) : 0;
3095 if ( !trie->trans[ state + charid ].next ) {
3096 trie->trans[ state + charid ].next = next_alloc;
3097 trie->trans[ state ].check++;
3098 prev_states[TRIE_NODENUM(next_alloc)]
3099 = TRIE_NODENUM(state);
3100 next_alloc += trie->uniquecharcount;
3102 state = trie->trans[ state + charid ].next;
3104 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3106 /* charid is now 0 if we dont know the char read, or
3107 * nonzero if we do */
3110 accept_state = TRIE_NODENUM( state );
3111 TRIE_HANDLE_WORD(accept_state);
3113 } /* end second pass */
3115 /* and now dump it out before we compress it */
3116 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3118 next_alloc, depth+1));
3122 * Inplace compress the table.*
3124 For sparse data sets the table constructed by the trie algorithm will
3125 be mostly 0/FAIL transitions or to put it another way mostly empty.
3126 (Note that leaf nodes will not contain any transitions.)
3128 This algorithm compresses the tables by eliminating most such
3129 transitions, at the cost of a modest bit of extra work during lookup:
3131 - Each states[] entry contains a .base field which indicates the
3132 index in the state[] array wheres its transition data is stored.
3134 - If .base is 0 there are no valid transitions from that node.
3136 - If .base is nonzero then charid is added to it to find an entry in
3139 -If trans[states[state].base+charid].check!=state then the
3140 transition is taken to be a 0/Fail transition. Thus if there are fail
3141 transitions at the front of the node then the .base offset will point
3142 somewhere inside the previous nodes data (or maybe even into a node
3143 even earlier), but the .check field determines if the transition is
3147 The following process inplace converts the table to the compressed
3148 table: We first do not compress the root node 1,and mark all its
3149 .check pointers as 1 and set its .base pointer as 1 as well. This
3150 allows us to do a DFA construction from the compressed table later,
3151 and ensures that any .base pointers we calculate later are greater
3154 - We set 'pos' to indicate the first entry of the second node.
3156 - We then iterate over the columns of the node, finding the first and
3157 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3158 and set the .check pointers accordingly, and advance pos
3159 appropriately and repreat for the next node. Note that when we copy
3160 the next pointers we have to convert them from the original
3161 NODEIDX form to NODENUM form as the former is not valid post
3164 - If a node has no transitions used we mark its base as 0 and do not
3165 advance the pos pointer.
3167 - If a node only has one transition we use a second pointer into the
3168 structure to fill in allocated fail transitions from other states.
3169 This pointer is independent of the main pointer and scans forward
3170 looking for null transitions that are allocated to a state. When it
3171 finds one it writes the single transition into the "hole". If the
3172 pointer doesnt find one the single transition is appended as normal.
3174 - Once compressed we can Renew/realloc the structures to release the
3177 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3178 specifically Fig 3.47 and the associated pseudocode.
3182 const U32 laststate = TRIE_NODENUM( next_alloc );
3185 trie->statecount = laststate;
3187 for ( state = 1 ; state < laststate ; state++ ) {
3189 const U32 stateidx = TRIE_NODEIDX( state );
3190 const U32 o_used = trie->trans[ stateidx ].check;
3191 U32 used = trie->trans[ stateidx ].check;
3192 trie->trans[ stateidx ].check = 0;
3195 used && charid < trie->uniquecharcount;
3198 if ( flag || trie->trans[ stateidx + charid ].next ) {
3199 if ( trie->trans[ stateidx + charid ].next ) {
3201 for ( ; zp < pos ; zp++ ) {
3202 if ( ! trie->trans[ zp ].next ) {
3206 trie->states[ state ].trans.base
3208 + trie->uniquecharcount
3210 trie->trans[ zp ].next
3211 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3213 trie->trans[ zp ].check = state;
3214 if ( ++zp > pos ) pos = zp;
3221 trie->states[ state ].trans.base
3222 = pos + trie->uniquecharcount - charid ;
3224 trie->trans[ pos ].next
3225 = SAFE_TRIE_NODENUM(
3226 trie->trans[ stateidx + charid ].next );
3227 trie->trans[ pos ].check = state;
3232 trie->lasttrans = pos + 1;
3233 trie->states = (reg_trie_state *)
3234 PerlMemShared_realloc( trie->states, laststate
3235 * sizeof(reg_trie_state) );
3236 DEBUG_TRIE_COMPILE_MORE_r(
3237 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3239 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3243 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3246 } /* end table compress */
3248 DEBUG_TRIE_COMPILE_MORE_r(
3249 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3251 (UV)trie->statecount,
3252 (UV)trie->lasttrans)
3254 /* resize the trans array to remove unused space */
3255 trie->trans = (reg_trie_trans *)
3256 PerlMemShared_realloc( trie->trans, trie->lasttrans
3257 * sizeof(reg_trie_trans) );
3259 { /* Modify the program and insert the new TRIE node */
3260 U8 nodetype =(U8)(flags & 0xFF);
3264 regnode *optimize = NULL;
3265 #ifdef RE_TRACK_PATTERN_OFFSETS
3268 U32 mjd_nodelen = 0;
3269 #endif /* RE_TRACK_PATTERN_OFFSETS */
3270 #endif /* DEBUGGING */
3272 This means we convert either the first branch or the first Exact,
3273 depending on whether the thing following (in 'last') is a branch
3274 or not and whther first is the startbranch (ie is it a sub part of
3275 the alternation or is it the whole thing.)
3276 Assuming its a sub part we convert the EXACT otherwise we convert
3277 the whole branch sequence, including the first.
3279 /* Find the node we are going to overwrite */
3280 if ( first != startbranch || OP( last ) == BRANCH ) {
3281 /* branch sub-chain */
3282 NEXT_OFF( first ) = (U16)(last - first);
3283 #ifdef RE_TRACK_PATTERN_OFFSETS
3285 mjd_offset= Node_Offset((convert));
3286 mjd_nodelen= Node_Length((convert));
3289 /* whole branch chain */
3291 #ifdef RE_TRACK_PATTERN_OFFSETS
3294 const regnode *nop = NEXTOPER( convert );
3295 mjd_offset= Node_Offset((nop));
3296 mjd_nodelen= Node_Length((nop));
3300 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3302 (UV)mjd_offset, (UV)mjd_nodelen)
3305 /* But first we check to see if there is a common prefix we can
3306 split out as an EXACT and put in front of the TRIE node. */
3307 trie->startstate= 1;
3308 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3309 /* we want to find the first state that has more than
3310 * one transition, if that state is not the first state
3311 * then we have a common prefix which we can remove.
3314 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3316 I32 first_ofs = -1; /* keeps track of the ofs of the first
3317 transition, -1 means none */
3319 const U32 base = trie->states[ state ].trans.base;
3321 /* does this state terminate an alternation? */
3322 if ( trie->states[state].wordnum )
3325 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3326 if ( ( base + ofs >= trie->uniquecharcount ) &&
3327 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3328 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3330 if ( ++count > 1 ) {
3331 /* we have more than one transition */
3334 /* if this is the first state there is no common prefix
3335 * to extract, so we can exit */
3336 if ( state == 1 ) break;
3337 tmp = av_fetch( revcharmap, ofs, 0);
3338 ch = (U8*)SvPV_nolen_const( *tmp );
3340 /* if we are on count 2 then we need to initialize the
3341 * bitmap, and store the previous char if there was one
3344 /* clear the bitmap */
3345 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3347 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3350 if (first_ofs >= 0) {
3351 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3352 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3354 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3356 Perl_re_printf( aTHX_ "%s", (char*)ch)
3360 /* store the current firstchar in the bitmap */
3361 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3362 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3368 /* This state has only one transition, its transition is part
3369 * of a common prefix - we need to concatenate the char it
3370 * represents to what we have so far. */
3371 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3373 char *ch = SvPV( *tmp, len );
3375 SV *sv=sv_newmortal();
3376 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3378 (UV)state, (UV)first_ofs,
3379 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3380 PL_colors[0], PL_colors[1],
3381 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3382 PERL_PV_ESCAPE_FIRSTCHAR
3387 OP( convert ) = nodetype;
3388 str=STRING(convert);
3391 STR_LEN(convert) += len;
3397 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3402 trie->prefixlen = (state-1);
3404 regnode *n = convert+NODE_SZ_STR(convert);
3405 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3406 trie->startstate = state;
3407 trie->minlen -= (state - 1);
3408 trie->maxlen -= (state - 1);
3410 /* At least the UNICOS C compiler choked on this
3411 * being argument to DEBUG_r(), so let's just have
3414 #ifdef PERL_EXT_RE_BUILD
3420 regnode *fix = convert;
3421 U32 word = trie->wordcount;
3423 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3424 while( ++fix < n ) {
3425 Set_Node_Offset_Length(fix, 0, 0);
3428 SV ** const tmp = av_fetch( trie_words, word, 0 );
3430 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3431 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3433 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3441 NEXT_OFF(convert) = (U16)(tail - convert);
3442 DEBUG_r(optimize= n);
3448 if ( trie->maxlen ) {
3449 NEXT_OFF( convert ) = (U16)(tail - convert);
3450 ARG_SET( convert, data_slot );
3451 /* Store the offset to the first unabsorbed branch in
3452 jump[0], which is otherwise unused by the jump logic.
3453 We use this when dumping a trie and during optimisation. */
3455 trie->jump[0] = (U16)(nextbranch - convert);
3457 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3458 * and there is a bitmap
3459 * and the first "jump target" node we found leaves enough room
3460 * then convert the TRIE node into a TRIEC node, with the bitmap
3461 * embedded inline in the opcode - this is hypothetically faster.
3463 if ( !trie->states[trie->startstate].wordnum
3465 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3467 OP( convert ) = TRIEC;
3468 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3469 PerlMemShared_free(trie->bitmap);
3472 OP( convert ) = TRIE;
3474 /* store the type in the flags */
3475 convert->flags = nodetype;
3479 + regarglen[ OP( convert ) ];
3481 /* XXX We really should free up the resource in trie now,
3482 as we won't use them - (which resources?) dmq */
3484 /* needed for dumping*/
3485 DEBUG_r(if (optimize) {
3486 regnode *opt = convert;
3488 while ( ++opt < optimize) {
3489 Set_Node_Offset_Length(opt,0,0);
3492 Try to clean up some of the debris left after the
3495 while( optimize < jumper ) {
3496 mjd_nodelen += Node_Length((optimize));
3497 OP( optimize ) = OPTIMIZED;
3498 Set_Node_Offset_Length(optimize,0,0);
3501 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3503 } /* end node insert */
3505 /* Finish populating the prev field of the wordinfo array. Walk back
3506 * from each accept state until we find another accept state, and if
3507 * so, point the first word's .prev field at the second word. If the
3508 * second already has a .prev field set, stop now. This will be the
3509 * case either if we've already processed that word's accept state,
3510 * or that state had multiple words, and the overspill words were
3511 * already linked up earlier.
3518 for (word=1; word <= trie->wordcount; word++) {
3520 if (trie->wordinfo[word].prev)
3522 state = trie->wordinfo[word].accept;
3524 state = prev_states[state];
3527 prev = trie->states[state].wordnum;
3531 trie->wordinfo[word].prev = prev;
3533 Safefree(prev_states);
3537 /* and now dump out the compressed format */
3538 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3540 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3542 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3543 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3545 SvREFCNT_dec_NN(revcharmap);
3549 : trie->startstate>1
3555 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3557 /* The Trie is constructed and compressed now so we can build a fail array if
3560 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3562 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3566 We find the fail state for each state in the trie, this state is the longest
3567 proper suffix of the current state's 'word' that is also a proper prefix of
3568 another word in our trie. State 1 represents the word '' and is thus the
3569 default fail state. This allows the DFA not to have to restart after its
3570 tried and failed a word at a given point, it simply continues as though it
3571 had been matching the other word in the first place.
3573 'abcdgu'=~/abcdefg|cdgu/
3574 When we get to 'd' we are still matching the first word, we would encounter
3575 'g' which would fail, which would bring us to the state representing 'd' in
3576 the second word where we would try 'g' and succeed, proceeding to match
3579 /* add a fail transition */
3580 const U32 trie_offset = ARG(source);
3581 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3583 const U32 ucharcount = trie->uniquecharcount;
3584 const U32 numstates = trie->statecount;
3585 const U32 ubound = trie->lasttrans + ucharcount;
3589 U32 base = trie->states[ 1 ].trans.base;
3592 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3594 GET_RE_DEBUG_FLAGS_DECL;
3596 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3597 PERL_UNUSED_CONTEXT;
3599 PERL_UNUSED_ARG(depth);
3602 if ( OP(source) == TRIE ) {
3603 struct regnode_1 *op = (struct regnode_1 *)
3604 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3605 StructCopy(source,op,struct regnode_1);
3606 stclass = (regnode *)op;
3608 struct regnode_charclass *op = (struct regnode_charclass *)
3609 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3610 StructCopy(source,op,struct regnode_charclass);
3611 stclass = (regnode *)op;
3613 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3615 ARG_SET( stclass, data_slot );
3616 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3617 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3618 aho->trie=trie_offset;
3619 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3620 Copy( trie->states, aho->states, numstates, reg_trie_state );
3621 Newx( q, numstates, U32);
3622 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3625 /* initialize fail[0..1] to be 1 so that we always have
3626 a valid final fail state */
3627 fail[ 0 ] = fail[ 1 ] = 1;
3629 for ( charid = 0; charid < ucharcount ; charid++ ) {
3630 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3632 q[ q_write ] = newstate;
3633 /* set to point at the root */
3634 fail[ q[ q_write++ ] ]=1;
3637 while ( q_read < q_write) {
3638 const U32 cur = q[ q_read++ % numstates ];
3639 base = trie->states[ cur ].trans.base;
3641 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3642 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3644 U32 fail_state = cur;
3647 fail_state = fail[ fail_state ];
3648 fail_base = aho->states[ fail_state ].trans.base;
3649 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3651 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3652 fail[ ch_state ] = fail_state;
3653 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3655 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3657 q[ q_write++ % numstates] = ch_state;
3661 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3662 when we fail in state 1, this allows us to use the
3663 charclass scan to find a valid start char. This is based on the principle
3664 that theres a good chance the string being searched contains lots of stuff
3665 that cant be a start char.
3667 fail[ 0 ] = fail[ 1 ] = 0;
3668 DEBUG_TRIE_COMPILE_r({
3669 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3670 depth, (UV)numstates
3672 for( q_read=1; q_read<numstates; q_read++ ) {
3673 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3675 Perl_re_printf( aTHX_ "\n");
3678 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3683 /* The below joins as many adjacent EXACTish nodes as possible into a single
3684 * one. The regop may be changed if the node(s) contain certain sequences that
3685 * require special handling. The joining is only done if:
3686 * 1) there is room in the current conglomerated node to entirely contain the
3688 * 2) they are the exact same node type
3690 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3691 * these get optimized out
3693 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3694 * as possible, even if that means splitting an existing node so that its first
3695 * part is moved to the preceeding node. This would maximise the efficiency of
3696 * memEQ during matching. Elsewhere in this file, khw proposes splitting
3697 * EXACTFish nodes into portions that don't change under folding vs those that
3698 * do. Those portions that don't change may be the only things in the pattern that
3699 * could be used to find fixed and floating strings.
3701 * If a node is to match under /i (folded), the number of characters it matches
3702 * can be different than its character length if it contains a multi-character
3703 * fold. *min_subtract is set to the total delta number of characters of the
3706 * And *unfolded_multi_char is set to indicate whether or not the node contains
3707 * an unfolded multi-char fold. This happens when whether the fold is valid or
3708 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3709 * SMALL LETTER SHARP S, as only if the target string being matched against
3710 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3711 * folding rules depend on the locale in force at runtime. (Multi-char folds
3712 * whose components are all above the Latin1 range are not run-time locale
3713 * dependent, and have already been folded by the time this function is
3716 * This is as good a place as any to discuss the design of handling these
3717 * multi-character fold sequences. It's been wrong in Perl for a very long
3718 * time. There are three code points in Unicode whose multi-character folds
3719 * were long ago discovered to mess things up. The previous designs for
3720 * dealing with these involved assigning a special node for them. This
3721 * approach doesn't always work, as evidenced by this example:
3722 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3723 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3724 * would match just the \xDF, it won't be able to handle the case where a
3725 * successful match would have to cross the node's boundary. The new approach
3726 * that hopefully generally solves the problem generates an EXACTFU_SS node
3727 * that is "sss" in this case.
3729 * It turns out that there are problems with all multi-character folds, and not
3730 * just these three. Now the code is general, for all such cases. The
3731 * approach taken is:
3732 * 1) This routine examines each EXACTFish node that could contain multi-
3733 * character folded sequences. Since a single character can fold into
3734 * such a sequence, the minimum match length for this node is less than
3735 * the number of characters in the node. This routine returns in
3736 * *min_subtract how many characters to subtract from the the actual
3737 * length of the string to get a real minimum match length; it is 0 if
3738 * there are no multi-char foldeds. This delta is used by the caller to
3739 * adjust the min length of the match, and the delta between min and max,
3740 * so that the optimizer doesn't reject these possibilities based on size
3742 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3743 * is used for an EXACTFU node that contains at least one "ss" sequence in
3744 * it. For non-UTF-8 patterns and strings, this is the only case where
3745 * there is a possible fold length change. That means that a regular
3746 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3747 * with length changes, and so can be processed faster. regexec.c takes
3748 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3749 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3750 * known until runtime). This saves effort in regex matching. However,
3751 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3752 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3753 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3754 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3755 * possibilities for the non-UTF8 patterns are quite simple, except for
3756 * the sharp s. All the ones that don't involve a UTF-8 target string are
3757 * members of a fold-pair, and arrays are set up for all of them so that
3758 * the other member of the pair can be found quickly. Code elsewhere in
3759 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3760 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3761 * described in the next item.
3762 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3763 * validity of the fold won't be known until runtime, and so must remain
3764 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3765 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3766 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3767 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3768 * The reason this is a problem is that the optimizer part of regexec.c
3769 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3770 * that a character in the pattern corresponds to at most a single
3771 * character in the target string. (And I do mean character, and not byte
3772 * here, unlike other parts of the documentation that have never been
3773 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3774 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3775 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3776 * nodes, violate the assumption, and they are the only instances where it
3777 * is violated. I'm reluctant to try to change the assumption, as the
3778 * code involved is impenetrable to me (khw), so instead the code here
3779 * punts. This routine examines EXACTFL nodes, and (when the pattern
3780 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3781 * boolean indicating whether or not the node contains such a fold. When
3782 * it is true, the caller sets a flag that later causes the optimizer in
3783 * this file to not set values for the floating and fixed string lengths,
3784 * and thus avoids the optimizer code in regexec.c that makes the invalid
3785 * assumption. Thus, there is no optimization based on string lengths for
3786 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3787 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3788 * assumption is wrong only in these cases is that all other non-UTF-8
3789 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3790 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3791 * EXACTF nodes because we don't know at compile time if it actually
3792 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3793 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3794 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3795 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3796 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3797 * string would require the pattern to be forced into UTF-8, the overhead
3798 * of which we want to avoid. Similarly the unfolded multi-char folds in
3799 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3802 * Similarly, the code that generates tries doesn't currently handle
3803 * not-already-folded multi-char folds, and it looks like a pain to change
3804 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3805 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3806 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3807 * using /iaa matching will be doing so almost entirely with ASCII
3808 * strings, so this should rarely be encountered in practice */
3810 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3811 if (PL_regkind[OP(scan)] == EXACT) \
3812 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3815 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3816 UV *min_subtract, bool *unfolded_multi_char,
3817 U32 flags,regnode *val, U32 depth)
3819 /* Merge several consecutive EXACTish nodes into one. */
3820 regnode *n = regnext(scan);
3822 regnode *next = scan + NODE_SZ_STR(scan);
3826 regnode *stop = scan;
3827 GET_RE_DEBUG_FLAGS_DECL;
3829 PERL_UNUSED_ARG(depth);
3832 PERL_ARGS_ASSERT_JOIN_EXACT;
3833 #ifndef EXPERIMENTAL_INPLACESCAN
3834 PERL_UNUSED_ARG(flags);
3835 PERL_UNUSED_ARG(val);
3837 DEBUG_PEEP("join", scan, depth, 0);
3839 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3840 * EXACT ones that are mergeable to the current one. */
3842 && (PL_regkind[OP(n)] == NOTHING
3843 || (stringok && OP(n) == OP(scan)))
3845 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3848 if (OP(n) == TAIL || n > next)
3850 if (PL_regkind[OP(n)] == NOTHING) {
3851 DEBUG_PEEP("skip:", n, depth, 0);
3852 NEXT_OFF(scan) += NEXT_OFF(n);
3853 next = n + NODE_STEP_REGNODE;
3860 else if (stringok) {
3861 const unsigned int oldl = STR_LEN(scan);
3862 regnode * const nnext = regnext(n);
3864 /* XXX I (khw) kind of doubt that this works on platforms (should
3865 * Perl ever run on one) where U8_MAX is above 255 because of lots
3866 * of other assumptions */
3867 /* Don't join if the sum can't fit into a single node */
3868 if (oldl + STR_LEN(n) > U8_MAX)
3871 DEBUG_PEEP("merg", n, depth, 0);
3874 NEXT_OFF(scan) += NEXT_OFF(n);
3875 STR_LEN(scan) += STR_LEN(n);
3876 next = n + NODE_SZ_STR(n);
3877 /* Now we can overwrite *n : */
3878 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3886 #ifdef EXPERIMENTAL_INPLACESCAN
3887 if (flags && !NEXT_OFF(n)) {
3888 DEBUG_PEEP("atch", val, depth, 0);
3889 if (reg_off_by_arg[OP(n)]) {
3890 ARG_SET(n, val - n);
3893 NEXT_OFF(n) = val - n;
3901 *unfolded_multi_char = FALSE;
3903 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3904 * can now analyze for sequences of problematic code points. (Prior to
3905 * this final joining, sequences could have been split over boundaries, and
3906 * hence missed). The sequences only happen in folding, hence for any
3907 * non-EXACT EXACTish node */
3908 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3909 U8* s0 = (U8*) STRING(scan);
3911 U8* s_end = s0 + STR_LEN(scan);
3913 int total_count_delta = 0; /* Total delta number of characters that
3914 multi-char folds expand to */
3916 /* One pass is made over the node's string looking for all the
3917 * possibilities. To avoid some tests in the loop, there are two main
3918 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3923 if (OP(scan) == EXACTFL) {
3926 /* An EXACTFL node would already have been changed to another
3927 * node type unless there is at least one character in it that
3928 * is problematic; likely a character whose fold definition
3929 * won't be known until runtime, and so has yet to be folded.
3930 * For all but the UTF-8 locale, folds are 1-1 in length, but
3931 * to handle the UTF-8 case, we need to create a temporary
3932 * folded copy using UTF-8 locale rules in order to analyze it.
3933 * This is because our macros that look to see if a sequence is
3934 * a multi-char fold assume everything is folded (otherwise the
3935 * tests in those macros would be too complicated and slow).
3936 * Note that here, the non-problematic folds will have already
3937 * been done, so we can just copy such characters. We actually
3938 * don't completely fold the EXACTFL string. We skip the
3939 * unfolded multi-char folds, as that would just create work
3940 * below to figure out the size they already are */
3942 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3945 STRLEN s_len = UTF8SKIP(s);
3946 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3947 Copy(s, d, s_len, U8);
3950 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3951 *unfolded_multi_char = TRUE;
3952 Copy(s, d, s_len, U8);
3955 else if (isASCII(*s)) {
3956 *(d++) = toFOLD(*s);
3960 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
3966 /* Point the remainder of the routine to look at our temporary
3970 } /* End of creating folded copy of EXACTFL string */
3972 /* Examine the string for a multi-character fold sequence. UTF-8
3973 * patterns have all characters pre-folded by the time this code is
3975 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3976 length sequence we are looking for is 2 */
3978 int count = 0; /* How many characters in a multi-char fold */
3979 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3980 if (! len) { /* Not a multi-char fold: get next char */
3985 /* Nodes with 'ss' require special handling, except for
3986 * EXACTFA-ish for which there is no multi-char fold to this */
3987 if (len == 2 && *s == 's' && *(s+1) == 's'
3988 && OP(scan) != EXACTFA
3989 && OP(scan) != EXACTFA_NO_TRIE)
3992 if (OP(scan) != EXACTFL) {
3993 OP(scan) = EXACTFU_SS;
3997 else { /* Here is a generic multi-char fold. */
3998 U8* multi_end = s + len;
4000 /* Count how many characters are in it. In the case of
4001 * /aa, no folds which contain ASCII code points are
4002 * allowed, so check for those, and skip if found. */
4003 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
4004 count = utf8_length(s, multi_end);
4008 while (s < multi_end) {
4011 goto next_iteration;
4021 /* The delta is how long the sequence is minus 1 (1 is how long
4022 * the character that folds to the sequence is) */
4023 total_count_delta += count - 1;
4027 /* We created a temporary folded copy of the string in EXACTFL
4028 * nodes. Therefore we need to be sure it doesn't go below zero,
4029 * as the real string could be shorter */
4030 if (OP(scan) == EXACTFL) {
4031 int total_chars = utf8_length((U8*) STRING(scan),
4032 (U8*) STRING(scan) + STR_LEN(scan));
4033 if (total_count_delta > total_chars) {
4034 total_count_delta = total_chars;
4038 *min_subtract += total_count_delta;
4041 else if (OP(scan) == EXACTFA) {
4043 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
4044 * fold to the ASCII range (and there are no existing ones in the
4045 * upper latin1 range). But, as outlined in the comments preceding
4046 * this function, we need to flag any occurrences of the sharp s.
4047 * This character forbids trie formation (because of added
4049 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4050 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4051 || UNICODE_DOT_DOT_VERSION > 0)
4053 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4054 OP(scan) = EXACTFA_NO_TRIE;
4055 *unfolded_multi_char = TRUE;
4063 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
4064 * folds that are all Latin1. As explained in the comments
4065 * preceding this function, we look also for the sharp s in EXACTF
4066 * and EXACTFL nodes; it can be in the final position. Otherwise
4067 * we can stop looking 1 byte earlier because have to find at least
4068 * two characters for a multi-fold */
4069 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4074 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4075 if (! len) { /* Not a multi-char fold. */
4076 if (*s == LATIN_SMALL_LETTER_SHARP_S
4077 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4079 *unfolded_multi_char = TRUE;
4086 && isALPHA_FOLD_EQ(*s, 's')
4087 && isALPHA_FOLD_EQ(*(s+1), 's'))
4090 /* EXACTF nodes need to know that the minimum length
4091 * changed so that a sharp s in the string can match this
4092 * ss in the pattern, but they remain EXACTF nodes, as they
4093 * won't match this unless the target string is is UTF-8,
4094 * which we don't know until runtime. EXACTFL nodes can't
4095 * transform into EXACTFU nodes */
4096 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4097 OP(scan) = EXACTFU_SS;
4101 *min_subtract += len - 1;
4109 /* Allow dumping but overwriting the collection of skipped
4110 * ops and/or strings with fake optimized ops */
4111 n = scan + NODE_SZ_STR(scan);
4119 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl", scan, depth, 0);});
4123 /* REx optimizer. Converts nodes into quicker variants "in place".
4124 Finds fixed substrings. */
4126 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4127 to the position after last scanned or to NULL. */
4129 #define INIT_AND_WITHP \
4130 assert(!and_withp); \
4131 Newx(and_withp,1, regnode_ssc); \
4132 SAVEFREEPV(and_withp)
4136 S_unwind_scan_frames(pTHX_ const void *p)
4138 scan_frame *f= (scan_frame *)p;
4140 scan_frame *n= f->next_frame;
4148 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4149 SSize_t *minlenp, SSize_t *deltap,
4154 regnode_ssc *and_withp,
4155 U32 flags, U32 depth)
4156 /* scanp: Start here (read-write). */
4157 /* deltap: Write maxlen-minlen here. */
4158 /* last: Stop before this one. */
4159 /* data: string data about the pattern */
4160 /* stopparen: treat close N as END */
4161 /* recursed: which subroutines have we recursed into */
4162 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4164 /* There must be at least this number of characters to match */
4167 regnode *scan = *scanp, *next;
4169 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4170 int is_inf_internal = 0; /* The studied chunk is infinite */
4171 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4172 scan_data_t data_fake;
4173 SV *re_trie_maxbuff = NULL;
4174 regnode *first_non_open = scan;
4175 SSize_t stopmin = SSize_t_MAX;
4176 scan_frame *frame = NULL;
4177 GET_RE_DEBUG_FLAGS_DECL;
4179 PERL_ARGS_ASSERT_STUDY_CHUNK;
4180 RExC_study_started= 1;
4182 Zero(&data_fake, 1, scan_data_t);
4185 while (first_non_open && OP(first_non_open) == OPEN)
4186 first_non_open=regnext(first_non_open);
4192 RExC_study_chunk_recursed_count++;
4194 DEBUG_OPTIMISE_MORE_r(
4196 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4197 depth, (long)stopparen,
4198 (unsigned long)RExC_study_chunk_recursed_count,
4199 (unsigned long)depth, (unsigned long)recursed_depth,
4202 if (recursed_depth) {
4205 for ( j = 0 ; j < recursed_depth ; j++ ) {
4206 for ( i = 0 ; i < (U32)RExC_npar ; i++ ) {
4208 PAREN_TEST(RExC_study_chunk_recursed +
4209 ( j * RExC_study_chunk_recursed_bytes), i )
4212 !PAREN_TEST(RExC_study_chunk_recursed +
4213 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
4216 Perl_re_printf( aTHX_ " %d",(int)i);
4220 if ( j + 1 < recursed_depth ) {
4221 Perl_re_printf( aTHX_ ",");
4225 Perl_re_printf( aTHX_ "\n");
4228 while ( scan && OP(scan) != END && scan < last ){
4229 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4230 node length to get a real minimum (because
4231 the folded version may be shorter) */
4232 bool unfolded_multi_char = FALSE;
4233 /* Peephole optimizer: */
4234 DEBUG_STUDYDATA("Peep", data, depth, is_inf);
4235 DEBUG_PEEP("Peep", scan, depth, flags);
4238 /* The reason we do this here is that we need to deal with things like
4239 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4240 * parsing code, as each (?:..) is handled by a different invocation of
4243 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
4245 /* Follow the next-chain of the current node and optimize
4246 away all the NOTHINGs from it. */
4247 if (OP(scan) != CURLYX) {
4248 const int max = (reg_off_by_arg[OP(scan)]
4250 /* I32 may be smaller than U16 on CRAYs! */
4251 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4252 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
4256 /* Skip NOTHING and LONGJMP. */
4257 while ((n = regnext(n))
4258 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4259 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
4260 && off + noff < max)
4262 if (reg_off_by_arg[OP(scan)])
4265 NEXT_OFF(scan) = off;
4268 /* The principal pseudo-switch. Cannot be a switch, since we
4269 look into several different things. */
4270 if ( OP(scan) == DEFINEP ) {
4272 SSize_t deltanext = 0;
4273 SSize_t fake_last_close = 0;
4274 I32 f = SCF_IN_DEFINE;
4276 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4277 scan = regnext(scan);
4278 assert( OP(scan) == IFTHEN );
4279 DEBUG_PEEP("expect IFTHEN", scan, depth, flags);
4281 data_fake.last_closep= &fake_last_close;
4283 next = regnext(scan);
4284 scan = NEXTOPER(NEXTOPER(scan));
4285 DEBUG_PEEP("scan", scan, depth, flags);
4286 DEBUG_PEEP("next", next, depth, flags);
4288 /* we suppose the run is continuous, last=next...
4289 * NOTE we dont use the return here! */
4290 /* DEFINEP study_chunk() recursion */
4291 (void)study_chunk(pRExC_state, &scan, &minlen,
4292 &deltanext, next, &data_fake, stopparen,
4293 recursed_depth, NULL, f, depth+1);
4298 OP(scan) == BRANCH ||
4299 OP(scan) == BRANCHJ ||
4302 next = regnext(scan);
4305 /* The op(next)==code check below is to see if we
4306 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4307 * IFTHEN is special as it might not appear in pairs.
4308 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4309 * we dont handle it cleanly. */
4310 if (OP(next) == code || code == IFTHEN) {
4311 /* NOTE - There is similar code to this block below for
4312 * handling TRIE nodes on a re-study. If you change stuff here
4313 * check there too. */
4314 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
4316 regnode * const startbranch=scan;
4318 if (flags & SCF_DO_SUBSTR) {
4319 /* Cannot merge strings after this. */
4320 scan_commit(pRExC_state, data, minlenp, is_inf);
4323 if (flags & SCF_DO_STCLASS)
4324 ssc_init_zero(pRExC_state, &accum);
4326 while (OP(scan) == code) {
4327 SSize_t deltanext, minnext, fake;
4329 regnode_ssc this_class;
4331 DEBUG_PEEP("Branch", scan, depth, flags);
4334 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4336 data_fake.whilem_c = data->whilem_c;
4337 data_fake.last_closep = data->last_closep;
4340 data_fake.last_closep = &fake;
4342 data_fake.pos_delta = delta;
4343 next = regnext(scan);
4345 scan = NEXTOPER(scan); /* everything */
4346 if (code != BRANCH) /* everything but BRANCH */
4347 scan = NEXTOPER(scan);
4349 if (flags & SCF_DO_STCLASS) {
4350 ssc_init(pRExC_state, &this_class);
4351 data_fake.start_class = &this_class;
4352 f = SCF_DO_STCLASS_AND;
4354 if (flags & SCF_WHILEM_VISITED_POS)
4355 f |= SCF_WHILEM_VISITED_POS;
4357 /* we suppose the run is continuous, last=next...*/
4358 /* recurse study_chunk() for each BRANCH in an alternation */
4359 minnext = study_chunk(pRExC_state, &scan, minlenp,
4360 &deltanext, next, &data_fake, stopparen,
4361 recursed_depth, NULL, f,depth+1);
4365 if (deltanext == SSize_t_MAX) {
4366 is_inf = is_inf_internal = 1;
4368 } else if (max1 < minnext + deltanext)
4369 max1 = minnext + deltanext;
4371 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4373 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4374 if ( stopmin > minnext)
4375 stopmin = min + min1;
4376 flags &= ~SCF_DO_SUBSTR;
4378 data->flags |= SCF_SEEN_ACCEPT;
4381 if (data_fake.flags & SF_HAS_EVAL)
4382 data->flags |= SF_HAS_EVAL;
4383 data->whilem_c = data_fake.whilem_c;
4385 if (flags & SCF_DO_STCLASS)
4386 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4388 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4390 if (flags & SCF_DO_SUBSTR) {
4391 data->pos_min += min1;
4392 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
4393 data->pos_delta = SSize_t_MAX;
4395 data->pos_delta += max1 - min1;
4396 if (max1 != min1 || is_inf)
4397 data->cur_is_floating = 1;
4400 if (delta == SSize_t_MAX
4401 || SSize_t_MAX - delta - (max1 - min1) < 0)
4402 delta = SSize_t_MAX;
4404 delta += max1 - min1;
4405 if (flags & SCF_DO_STCLASS_OR) {
4406 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4408 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4409 flags &= ~SCF_DO_STCLASS;
4412 else if (flags & SCF_DO_STCLASS_AND) {
4414 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4415 flags &= ~SCF_DO_STCLASS;
4418 /* Switch to OR mode: cache the old value of
4419 * data->start_class */
4421 StructCopy(data->start_class, and_withp, regnode_ssc);
4422 flags &= ~SCF_DO_STCLASS_AND;
4423 StructCopy(&accum, data->start_class, regnode_ssc);
4424 flags |= SCF_DO_STCLASS_OR;
4428 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4429 OP( startbranch ) == BRANCH )
4433 Assuming this was/is a branch we are dealing with: 'scan'
4434 now points at the item that follows the branch sequence,
4435 whatever it is. We now start at the beginning of the
4436 sequence and look for subsequences of
4442 which would be constructed from a pattern like
4445 If we can find such a subsequence we need to turn the first
4446 element into a trie and then add the subsequent branch exact
4447 strings to the trie.
4451 1. patterns where the whole set of branches can be
4454 2. patterns where only a subset can be converted.
4456 In case 1 we can replace the whole set with a single regop
4457 for the trie. In case 2 we need to keep the start and end
4460 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4461 becomes BRANCH TRIE; BRANCH X;
4463 There is an additional case, that being where there is a
4464 common prefix, which gets split out into an EXACT like node
4465 preceding the TRIE node.
4467 If x(1..n)==tail then we can do a simple trie, if not we make
4468 a "jump" trie, such that when we match the appropriate word
4469 we "jump" to the appropriate tail node. Essentially we turn
4470 a nested if into a case structure of sorts.
4475 if (!re_trie_maxbuff) {
4476 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4477 if (!SvIOK(re_trie_maxbuff))
4478 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4480 if ( SvIV(re_trie_maxbuff)>=0 ) {
4482 regnode *first = (regnode *)NULL;
4483 regnode *last = (regnode *)NULL;
4484 regnode *tail = scan;
4488 /* var tail is used because there may be a TAIL
4489 regop in the way. Ie, the exacts will point to the
4490 thing following the TAIL, but the last branch will
4491 point at the TAIL. So we advance tail. If we
4492 have nested (?:) we may have to move through several
4496 while ( OP( tail ) == TAIL ) {
4497 /* this is the TAIL generated by (?:) */
4498 tail = regnext( tail );
4502 DEBUG_TRIE_COMPILE_r({
4503 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4504 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4506 "Looking for TRIE'able sequences. Tail node is ",
4507 (UV)(tail - RExC_emit_start),
4508 SvPV_nolen_const( RExC_mysv )
4514 Step through the branches
4515 cur represents each branch,
4516 noper is the first thing to be matched as part
4518 noper_next is the regnext() of that node.
4520 We normally handle a case like this
4521 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4522 support building with NOJUMPTRIE, which restricts
4523 the trie logic to structures like /FOO|BAR/.
4525 If noper is a trieable nodetype then the branch is
4526 a possible optimization target. If we are building
4527 under NOJUMPTRIE then we require that noper_next is
4528 the same as scan (our current position in the regex
4531 Once we have two or more consecutive such branches
4532 we can create a trie of the EXACT's contents and
4533 stitch it in place into the program.
4535 If the sequence represents all of the branches in
4536 the alternation we replace the entire thing with a
4539 Otherwise when it is a subsequence we need to
4540 stitch it in place and replace only the relevant
4541 branches. This means the first branch has to remain
4542 as it is used by the alternation logic, and its
4543 next pointer, and needs to be repointed at the item
4544 on the branch chain following the last branch we
4545 have optimized away.
4547 This could be either a BRANCH, in which case the
4548 subsequence is internal, or it could be the item
4549 following the branch sequence in which case the
4550 subsequence is at the end (which does not
4551 necessarily mean the first node is the start of the
4554 TRIE_TYPE(X) is a define which maps the optype to a
4558 ----------------+-----------
4562 EXACTFU_SS | EXACTFU
4565 EXACTFLU8 | EXACTFLU8
4569 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4571 : ( EXACT == (X) ) \
4573 : ( EXACTFU == (X) || EXACTFU_SS == (X) ) \
4575 : ( EXACTFA == (X) ) \
4577 : ( EXACTL == (X) ) \
4579 : ( EXACTFLU8 == (X) ) \
4583 /* dont use tail as the end marker for this traverse */
4584 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4585 regnode * const noper = NEXTOPER( cur );
4586 U8 noper_type = OP( noper );
4587 U8 noper_trietype = TRIE_TYPE( noper_type );
4588 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4589 regnode * const noper_next = regnext( noper );
4590 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4591 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4594 DEBUG_TRIE_COMPILE_r({
4595 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4596 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4598 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4600 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4601 Perl_re_printf( aTHX_ " -> %d:%s",
4602 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
4605 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4606 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
4607 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
4609 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
4610 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4611 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4615 /* Is noper a trieable nodetype that can be merged
4616 * with the current trie (if there is one)? */
4620 ( noper_trietype == NOTHING )
4621 || ( trietype == NOTHING )
4622 || ( trietype == noper_trietype )
4625 && noper_next >= tail
4629 /* Handle mergable triable node Either we are
4630 * the first node in a new trieable sequence,
4631 * in which case we do some bookkeeping,
4632 * otherwise we update the end pointer. */
4635 if ( noper_trietype == NOTHING ) {
4636 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4637 regnode * const noper_next = regnext( noper );
4638 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4639 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4642 if ( noper_next_trietype ) {
4643 trietype = noper_next_trietype;
4644 } else if (noper_next_type) {
4645 /* a NOTHING regop is 1 regop wide.
4646 * We need at least two for a trie
4647 * so we can't merge this in */
4651 trietype = noper_trietype;
4654 if ( trietype == NOTHING )
4655 trietype = noper_trietype;
4660 } /* end handle mergable triable node */
4662 /* handle unmergable node -
4663 * noper may either be a triable node which can
4664 * not be tried together with the current trie,
4665 * or a non triable node */
4667 /* If last is set and trietype is not
4668 * NOTHING then we have found at least two
4669 * triable branch sequences in a row of a
4670 * similar trietype so we can turn them
4671 * into a trie. If/when we allow NOTHING to
4672 * start a trie sequence this condition
4673 * will be required, and it isn't expensive
4674 * so we leave it in for now. */
4675 if ( trietype && trietype != NOTHING )
4676 make_trie( pRExC_state,
4677 startbranch, first, cur, tail,
4678 count, trietype, depth+1 );
4679 last = NULL; /* note: we clear/update
4680 first, trietype etc below,
4681 so we dont do it here */
4685 && noper_next >= tail
4688 /* noper is triable, so we can start a new
4692 trietype = noper_trietype;
4694 /* if we already saw a first but the
4695 * current node is not triable then we have
4696 * to reset the first information. */
4701 } /* end handle unmergable node */
4702 } /* loop over branches */
4703 DEBUG_TRIE_COMPILE_r({
4704 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4705 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
4706 depth+1, SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4707 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
4708 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4709 PL_reg_name[trietype]
4713 if ( last && trietype ) {
4714 if ( trietype != NOTHING ) {
4715 /* the last branch of the sequence was part of
4716 * a trie, so we have to construct it here
4717 * outside of the loop */
4718 made= make_trie( pRExC_state, startbranch,
4719 first, scan, tail, count,
4720 trietype, depth+1 );
4721 #ifdef TRIE_STUDY_OPT
4722 if ( ((made == MADE_EXACT_TRIE &&
4723 startbranch == first)
4724 || ( first_non_open == first )) &&
4726 flags |= SCF_TRIE_RESTUDY;
4727 if ( startbranch == first
4730 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4735 /* at this point we know whatever we have is a
4736 * NOTHING sequence/branch AND if 'startbranch'
4737 * is 'first' then we can turn the whole thing
4740 if ( startbranch == first ) {
4742 /* the entire thing is a NOTHING sequence,
4743 * something like this: (?:|) So we can
4744 * turn it into a plain NOTHING op. */
4745 DEBUG_TRIE_COMPILE_r({
4746 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4747 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
4749 SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4752 OP(startbranch)= NOTHING;
4753 NEXT_OFF(startbranch)= tail - startbranch;
4754 for ( opt= startbranch + 1; opt < tail ; opt++ )
4758 } /* end if ( last) */
4759 } /* TRIE_MAXBUF is non zero */
4764 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4765 scan = NEXTOPER(NEXTOPER(scan));
4766 } else /* single branch is optimized. */
4767 scan = NEXTOPER(scan);
4769 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
4771 regnode *start = NULL;
4772 regnode *end = NULL;
4773 U32 my_recursed_depth= recursed_depth;
4775 if (OP(scan) != SUSPEND) { /* GOSUB */
4776 /* Do setup, note this code has side effects beyond
4777 * the rest of this block. Specifically setting
4778 * RExC_recurse[] must happen at least once during
4781 RExC_recurse[ARG2L(scan)] = scan;
4782 start = RExC_open_parens[paren];
4783 end = RExC_close_parens[paren];
4785 /* NOTE we MUST always execute the above code, even
4786 * if we do nothing with a GOSUB */
4788 ( flags & SCF_IN_DEFINE )
4791 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
4793 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
4796 /* no need to do anything here if we are in a define. */
4797 /* or we are after some kind of infinite construct
4798 * so we can skip recursing into this item.
4799 * Since it is infinite we will not change the maxlen
4800 * or delta, and if we miss something that might raise
4801 * the minlen it will merely pessimise a little.
4803 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
4804 * might result in a minlen of 1 and not of 4,
4805 * but this doesn't make us mismatch, just try a bit
4806 * harder than we should.
4808 scan= regnext(scan);
4815 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4817 /* it is quite possible that there are more efficient ways
4818 * to do this. We maintain a bitmap per level of recursion
4819 * of which patterns we have entered so we can detect if a
4820 * pattern creates a possible infinite loop. When we
4821 * recurse down a level we copy the previous levels bitmap
4822 * down. When we are at recursion level 0 we zero the top
4823 * level bitmap. It would be nice to implement a different
4824 * more efficient way of doing this. In particular the top
4825 * level bitmap may be unnecessary.
4827 if (!recursed_depth) {
4828 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4830 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4831 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4832 RExC_study_chunk_recursed_bytes, U8);
4834 /* we havent recursed into this paren yet, so recurse into it */
4835 DEBUG_STUDYDATA("gosub-set", data, depth, is_inf);
4836 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4837 my_recursed_depth= recursed_depth + 1;
4839 DEBUG_STUDYDATA("gosub-inf", data, depth, is_inf);
4840 /* some form of infinite recursion, assume infinite length
4842 if (flags & SCF_DO_SUBSTR) {
4843 scan_commit(pRExC_state, data, minlenp, is_inf);
4844 data->cur_is_floating = 1;
4846 is_inf = is_inf_internal = 1;
4847 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4848 ssc_anything(data->start_class);
4849 flags &= ~SCF_DO_STCLASS;
4851 start= NULL; /* reset start so we dont recurse later on. */
4856 end = regnext(scan);
4859 scan_frame *newframe;
4861 if (!RExC_frame_last) {
4862 Newxz(newframe, 1, scan_frame);
4863 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
4864 RExC_frame_head= newframe;
4866 } else if (!RExC_frame_last->next_frame) {
4867 Newxz(newframe,1,scan_frame);
4868 RExC_frame_last->next_frame= newframe;
4869 newframe->prev_frame= RExC_frame_last;
4872 newframe= RExC_frame_last->next_frame;
4874 RExC_frame_last= newframe;
4876 newframe->next_regnode = regnext(scan);
4877 newframe->last_regnode = last;
4878 newframe->stopparen = stopparen;
4879 newframe->prev_recursed_depth = recursed_depth;
4880 newframe->this_prev_frame= frame;
4882 DEBUG_STUDYDATA("frame-new", data, depth, is_inf);
4883 DEBUG_PEEP("fnew", scan, depth, flags);
4890 recursed_depth= my_recursed_depth;
4895 else if (OP(scan) == EXACT || OP(scan) == EXACTL) {
4896 SSize_t l = STR_LEN(scan);
4900 const U8 * const s = (U8*)STRING(scan);
4901 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4902 l = utf8_length(s, s + l);
4904 uc = *((U8*)STRING(scan));
4907 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4908 /* The code below prefers earlier match for fixed
4909 offset, later match for variable offset. */
4910 if (data->last_end == -1) { /* Update the start info. */
4911 data->last_start_min = data->pos_min;
4912 data->last_start_max = is_inf
4913 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4915 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4917 SvUTF8_on(data->last_found);
4919 SV * const sv = data->last_found;
4920 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4921 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4922 if (mg && mg->mg_len >= 0)
4923 mg->mg_len += utf8_length((U8*)STRING(scan),
4924 (U8*)STRING(scan)+STR_LEN(scan));
4926 data->last_end = data->pos_min + l;
4927 data->pos_min += l; /* As in the first entry. */
4928 data->flags &= ~SF_BEFORE_EOL;
4931 /* ANDing the code point leaves at most it, and not in locale, and
4932 * can't match null string */
4933 if (flags & SCF_DO_STCLASS_AND) {
4934 ssc_cp_and(data->start_class, uc);
4935 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4936 ssc_clear_locale(data->start_class);
4938 else if (flags & SCF_DO_STCLASS_OR) {
4939 ssc_add_cp(data->start_class, uc);
4940 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4942 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4943 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4945 flags &= ~SCF_DO_STCLASS;
4947 else if (PL_regkind[OP(scan)] == EXACT) {
4948 /* But OP != EXACT!, so is EXACTFish */
4949 SSize_t l = STR_LEN(scan);
4950 const U8 * s = (U8*)STRING(scan);
4952 /* Search for fixed substrings supports EXACT only. */
4953 if (flags & SCF_DO_SUBSTR) {
4955 scan_commit(pRExC_state, data, minlenp, is_inf);
4958 l = utf8_length(s, s + l);
4960 if (unfolded_multi_char) {
4961 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4963 min += l - min_subtract;
4965 delta += min_subtract;
4966 if (flags & SCF_DO_SUBSTR) {
4967 data->pos_min += l - min_subtract;
4968 if (data->pos_min < 0) {
4971 data->pos_delta += min_subtract;
4973 data->cur_is_floating = 1; /* float */
4977 if (flags & SCF_DO_STCLASS) {
4978 SV* EXACTF_invlist = _make_exactf_invlist(pRExC_state, scan);
4980 assert(EXACTF_invlist);
4981 if (flags & SCF_DO_STCLASS_AND) {
4982 if (OP(scan) != EXACTFL)
4983 ssc_clear_locale(data->start_class);
4984 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4985 ANYOF_POSIXL_ZERO(data->start_class);
4986 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4988 else { /* SCF_DO_STCLASS_OR */
4989 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4990 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4992 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4993 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4995 flags &= ~SCF_DO_STCLASS;
4996 SvREFCNT_dec(EXACTF_invlist);
4999 else if (REGNODE_VARIES(OP(scan))) {
5000 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
5001 I32 fl = 0, f = flags;
5002 regnode * const oscan = scan;
5003 regnode_ssc this_class;
5004 regnode_ssc *oclass = NULL;
5005 I32 next_is_eval = 0;
5007 switch (PL_regkind[OP(scan)]) {
5008 case WHILEM: /* End of (?:...)* . */
5009 scan = NEXTOPER(scan);
5012 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
5013 next = NEXTOPER(scan);
5014 if (OP(next) == EXACT
5015 || OP(next) == EXACTL
5016 || (flags & SCF_DO_STCLASS))
5019 maxcount = REG_INFTY;
5020 next = regnext(scan);
5021 scan = NEXTOPER(scan);
5025 if (flags & SCF_DO_SUBSTR)
5030 if (flags & SCF_DO_STCLASS) {
5032 maxcount = REG_INFTY;
5033 next = regnext(scan);
5034 scan = NEXTOPER(scan);
5037 if (flags & SCF_DO_SUBSTR) {
5038 scan_commit(pRExC_state, data, minlenp, is_inf);
5039 /* Cannot extend fixed substrings */
5040 data->cur_is_floating = 1; /* float */
5042 is_inf = is_inf_internal = 1;
5043 scan = regnext(scan);
5044 goto optimize_curly_tail;
5046 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
5047 && (scan->flags == stopparen))
5052 mincount = ARG1(scan);
5053 maxcount = ARG2(scan);
5055 next = regnext(scan);
5056 if (OP(scan) == CURLYX) {
5057 I32 lp = (data ? *(data->last_closep) : 0);
5058 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
5060 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
5061 next_is_eval = (OP(scan) == EVAL);
5063 if (flags & SCF_DO_SUBSTR) {
5065 scan_commit(pRExC_state, data, minlenp, is_inf);
5066 /* Cannot extend fixed substrings */
5067 pos_before = data->pos_min;
5071 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
5073 data->flags |= SF_IS_INF;
5075 if (flags & SCF_DO_STCLASS) {
5076 ssc_init(pRExC_state, &this_class);
5077 oclass = data->start_class;
5078 data->start_class = &this_class;
5079 f |= SCF_DO_STCLASS_AND;
5080 f &= ~SCF_DO_STCLASS_OR;
5082 /* Exclude from super-linear cache processing any {n,m}
5083 regops for which the combination of input pos and regex
5084 pos is not enough information to determine if a match
5087 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5088 regex pos at the \s*, the prospects for a match depend not
5089 only on the input position but also on how many (bar\s*)
5090 repeats into the {4,8} we are. */
5091 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5092 f &= ~SCF_WHILEM_VISITED_POS;
5094 /* This will finish on WHILEM, setting scan, or on NULL: */
5095 /* recurse study_chunk() on loop bodies */
5096 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5097 last, data, stopparen, recursed_depth, NULL,
5099 ? (f & ~SCF_DO_SUBSTR)
5103 if (flags & SCF_DO_STCLASS)
5104 data->start_class = oclass;
5105 if (mincount == 0 || minnext == 0) {
5106 if (flags & SCF_DO_STCLASS_OR) {
5107 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5109 else if (flags & SCF_DO_STCLASS_AND) {
5110 /* Switch to OR mode: cache the old value of
5111 * data->start_class */
5113 StructCopy(data->start_class, and_withp, regnode_ssc);
5114 flags &= ~SCF_DO_STCLASS_AND;
5115 StructCopy(&this_class, data->start_class, regnode_ssc);
5116 flags |= SCF_DO_STCLASS_OR;
5117 ANYOF_FLAGS(data->start_class)
5118 |= SSC_MATCHES_EMPTY_STRING;
5120 } else { /* Non-zero len */
5121 if (flags & SCF_DO_STCLASS_OR) {
5122 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5123 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5125 else if (flags & SCF_DO_STCLASS_AND)
5126 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5127 flags &= ~SCF_DO_STCLASS;
5129 if (!scan) /* It was not CURLYX, but CURLY. */
5131 if (((flags & (SCF_TRIE_DOING_RESTUDY|SCF_DO_SUBSTR))==SCF_DO_SUBSTR)
5132 /* ? quantifier ok, except for (?{ ... }) */
5133 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5134 && (minnext == 0) && (deltanext == 0)
5135 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5136 && maxcount <= REG_INFTY/3) /* Complement check for big
5139 /* Fatal warnings may leak the regexp without this: */
5140 SAVEFREESV(RExC_rx_sv);
5141 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5142 "Quantifier unexpected on zero-length expression "
5143 "in regex m/%" UTF8f "/",
5144 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5146 (void)ReREFCNT_inc(RExC_rx_sv);
5149 min += minnext * mincount;
5150 is_inf_internal |= deltanext == SSize_t_MAX
5151 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5152 is_inf |= is_inf_internal;
5154 delta = SSize_t_MAX;
5156 delta += (minnext + deltanext) * maxcount
5157 - minnext * mincount;
5159 /* Try powerful optimization CURLYX => CURLYN. */
5160 if ( OP(oscan) == CURLYX && data
5161 && data->flags & SF_IN_PAR
5162 && !(data->flags & SF_HAS_EVAL)
5163 && !deltanext && minnext == 1 ) {
5164 /* Try to optimize to CURLYN. */
5165 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5166 regnode * const nxt1 = nxt;
5173 if (!REGNODE_SIMPLE(OP(nxt))
5174 && !(PL_regkind[OP(nxt)] == EXACT
5175 && STR_LEN(nxt) == 1))
5181 if (OP(nxt) != CLOSE)
5183 if (RExC_open_parens) {
5184 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5185 RExC_close_parens[ARG(nxt1)]=nxt+2; /*close->while*/
5187 /* Now we know that nxt2 is the only contents: */
5188 oscan->flags = (U8)ARG(nxt);
5190 OP(nxt1) = NOTHING; /* was OPEN. */
5193 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5194 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5195 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5196 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5197 OP(nxt + 1) = OPTIMIZED; /* was count. */
5198 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5203 /* Try optimization CURLYX => CURLYM. */
5204 if ( OP(oscan) == CURLYX && data
5205 && !(data->flags & SF_HAS_PAR)
5206 && !(data->flags & SF_HAS_EVAL)
5207 && !deltanext /* atom is fixed width */
5208 && minnext != 0 /* CURLYM can't handle zero width */
5210 /* Nor characters whose fold at run-time may be
5211 * multi-character */
5212 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5214 /* XXXX How to optimize if data == 0? */
5215 /* Optimize to a simpler form. */
5216 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5220 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5221 && (OP(nxt2) != WHILEM))
5223 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5224 /* Need to optimize away parenths. */
5225 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5226 /* Set the parenth number. */
5227 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5229 oscan->flags = (U8)ARG(nxt);
5230 if (RExC_open_parens) {
5231 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5232 RExC_close_parens[ARG(nxt1)]=nxt2+1; /*close->NOTHING*/
5234 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5235 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5238 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5239 OP(nxt + 1) = OPTIMIZED; /* was count. */
5240 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5241 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5244 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5245 regnode *nnxt = regnext(nxt1);
5247 if (reg_off_by_arg[OP(nxt1)])
5248 ARG_SET(nxt1, nxt2 - nxt1);
5249 else if (nxt2 - nxt1 < U16_MAX)
5250 NEXT_OFF(nxt1) = nxt2 - nxt1;
5252 OP(nxt) = NOTHING; /* Cannot beautify */
5257 /* Optimize again: */
5258 /* recurse study_chunk() on optimised CURLYX => CURLYM */
5259 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5260 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
5265 else if ((OP(oscan) == CURLYX)
5266 && (flags & SCF_WHILEM_VISITED_POS)
5267 /* See the comment on a similar expression above.
5268 However, this time it's not a subexpression
5269 we care about, but the expression itself. */
5270 && (maxcount == REG_INFTY)
5272 /* This stays as CURLYX, we can put the count/of pair. */
5273 /* Find WHILEM (as in regexec.c) */
5274 regnode *nxt = oscan + NEXT_OFF(oscan);
5276 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5278 nxt = PREVOPER(nxt);
5279 if (nxt->flags & 0xf) {
5280 /* we've already set whilem count on this node */
5281 } else if (++data->whilem_c < 16) {
5282 assert(data->whilem_c <= RExC_whilem_seen);
5283 nxt->flags = (U8)(data->whilem_c
5284 | (RExC_whilem_seen << 4)); /* On WHILEM */
5287 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5289 if (flags & SCF_DO_SUBSTR) {
5290 SV *last_str = NULL;
5291 STRLEN last_chrs = 0;
5292 int counted = mincount != 0;
5294 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5296 SSize_t b = pos_before >= data->last_start_min
5297 ? pos_before : data->last_start_min;
5299 const char * const s = SvPV_const(data->last_found, l);
5300 SSize_t old = b - data->last_start_min;
5303 old = utf8_hop((U8*)s, old) - (U8*)s;
5305 /* Get the added string: */
5306 last_str = newSVpvn_utf8(s + old, l, UTF);
5307 last_chrs = UTF ? utf8_length((U8*)(s + old),
5308 (U8*)(s + old + l)) : l;
5309 if (deltanext == 0 && pos_before == b) {
5310 /* What was added is a constant string */
5313 SvGROW(last_str, (mincount * l) + 1);
5314 repeatcpy(SvPVX(last_str) + l,
5315 SvPVX_const(last_str), l,
5317 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5318 /* Add additional parts. */
5319 SvCUR_set(data->last_found,
5320 SvCUR(data->last_found) - l);
5321 sv_catsv(data->last_found, last_str);
5323 SV * sv = data->last_found;
5325 SvUTF8(sv) && SvMAGICAL(sv) ?
5326 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5327 if (mg && mg->mg_len >= 0)
5328 mg->mg_len += last_chrs * (mincount-1);
5330 last_chrs *= mincount;
5331 data->last_end += l * (mincount - 1);
5334 /* start offset must point into the last copy */
5335 data->last_start_min += minnext * (mincount - 1);
5336 data->last_start_max =
5339 : data->last_start_max +
5340 (maxcount - 1) * (minnext + data->pos_delta);
5343 /* It is counted once already... */
5344 data->pos_min += minnext * (mincount - counted);
5346 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5347 " SSize_t_MAX=%" UVuf " minnext=%" UVuf
5348 " maxcount=%" UVuf " mincount=%" UVuf "\n",
5349 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
5351 if (deltanext != SSize_t_MAX)
5352 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5353 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5354 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
5356 if (deltanext == SSize_t_MAX
5357 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
5358 data->pos_delta = SSize_t_MAX;
5360 data->pos_delta += - counted * deltanext +
5361 (minnext + deltanext) * maxcount - minnext * mincount;
5362 if (mincount != maxcount) {
5363 /* Cannot extend fixed substrings found inside
5365 scan_commit(pRExC_state, data, minlenp, is_inf);
5366 if (mincount && last_str) {
5367 SV * const sv = data->last_found;
5368 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5369 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5373 sv_setsv(sv, last_str);
5374 data->last_end = data->pos_min;
5375 data->last_start_min = data->pos_min - last_chrs;
5376 data->last_start_max = is_inf
5378 : data->pos_min + data->pos_delta - last_chrs;
5380 data->cur_is_floating = 1; /* float */
5382 SvREFCNT_dec(last_str);
5384 if (data && (fl & SF_HAS_EVAL))
5385 data->flags |= SF_HAS_EVAL;
5386 optimize_curly_tail:
5387 if (OP(oscan) != CURLYX) {
5388 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
5390 NEXT_OFF(oscan) += NEXT_OFF(next);
5396 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5401 if (flags & SCF_DO_SUBSTR) {
5402 /* Cannot expect anything... */
5403 scan_commit(pRExC_state, data, minlenp, is_inf);
5404 data->cur_is_floating = 1; /* float */
5406 is_inf = is_inf_internal = 1;
5407 if (flags & SCF_DO_STCLASS_OR) {
5408 if (OP(scan) == CLUMP) {
5409 /* Actually is any start char, but very few code points
5410 * aren't start characters */
5411 ssc_match_all_cp(data->start_class);
5414 ssc_anything(data->start_class);
5417 flags &= ~SCF_DO_STCLASS;
5421 else if (OP(scan) == LNBREAK) {
5422 if (flags & SCF_DO_STCLASS) {
5423 if (flags & SCF_DO_STCLASS_AND) {
5424 ssc_intersection(data->start_class,
5425 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5426 ssc_clear_locale(data->start_class);
5427 ANYOF_FLAGS(data->start_class)
5428 &= ~SSC_MATCHES_EMPTY_STRING;
5430 else if (flags & SCF_DO_STCLASS_OR) {
5431 ssc_union(data->start_class,
5432 PL_XPosix_ptrs[_CC_VERTSPACE],
5434 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5436 /* See commit msg for
5437 * 749e076fceedeb708a624933726e7989f2302f6a */
5438 ANYOF_FLAGS(data->start_class)
5439 &= ~SSC_MATCHES_EMPTY_STRING;
5441 flags &= ~SCF_DO_STCLASS;
5444 if (delta != SSize_t_MAX)
5445 delta++; /* Because of the 2 char string cr-lf */
5446 if (flags & SCF_DO_SUBSTR) {
5447 /* Cannot expect anything... */
5448 scan_commit(pRExC_state, data, minlenp, is_inf);
5450 data->pos_delta += 1;
5451 data->cur_is_floating = 1; /* float */
5454 else if (REGNODE_SIMPLE(OP(scan))) {
5456 if (flags & SCF_DO_SUBSTR) {
5457 scan_commit(pRExC_state, data, minlenp, is_inf);
5461 if (flags & SCF_DO_STCLASS) {
5463 SV* my_invlist = NULL;
5466 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5467 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5469 /* Some of the logic below assumes that switching
5470 locale on will only add false positives. */
5475 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5479 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5480 ssc_match_all_cp(data->start_class);
5485 SV* REG_ANY_invlist = _new_invlist(2);
5486 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5488 if (flags & SCF_DO_STCLASS_OR) {
5489 ssc_union(data->start_class,
5491 TRUE /* TRUE => invert, hence all but \n
5495 else if (flags & SCF_DO_STCLASS_AND) {
5496 ssc_intersection(data->start_class,
5498 TRUE /* TRUE => invert */
5500 ssc_clear_locale(data->start_class);
5502 SvREFCNT_dec_NN(REG_ANY_invlist);
5509 if (flags & SCF_DO_STCLASS_AND)
5510 ssc_and(pRExC_state, data->start_class,
5511 (regnode_charclass *) scan);
5513 ssc_or(pRExC_state, data->start_class,
5514 (regnode_charclass *) scan);
5522 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5523 if (flags & SCF_DO_STCLASS_AND) {
5524 bool was_there = cBOOL(
5525 ANYOF_POSIXL_TEST(data->start_class,
5527 ANYOF_POSIXL_ZERO(data->start_class);
5528 if (was_there) { /* Do an AND */
5529 ANYOF_POSIXL_SET(data->start_class, namedclass);
5531 /* No individual code points can now match */
5532 data->start_class->invlist
5533 = sv_2mortal(_new_invlist(0));
5536 int complement = namedclass + ((invert) ? -1 : 1);
5538 assert(flags & SCF_DO_STCLASS_OR);
5540 /* If the complement of this class was already there,
5541 * the result is that they match all code points,
5542 * (\d + \D == everything). Remove the classes from
5543 * future consideration. Locale is not relevant in
5545 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5546 ssc_match_all_cp(data->start_class);
5547 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5548 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5550 else { /* The usual case; just add this class to the
5552 ANYOF_POSIXL_SET(data->start_class, namedclass);
5557 case NPOSIXA: /* For these, we always know the exact set of
5562 if (FLAGS(scan) == _CC_ASCII) {
5563 my_invlist = invlist_clone(PL_XPosix_ptrs[_CC_ASCII]);
5566 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
5567 PL_XPosix_ptrs[_CC_ASCII],
5578 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
5580 /* NPOSIXD matches all upper Latin1 code points unless the
5581 * target string being matched is UTF-8, which is
5582 * unknowable until match time. Since we are going to
5583 * invert, we want to get rid of all of them so that the
5584 * inversion will match all */
5585 if (OP(scan) == NPOSIXD) {
5586 _invlist_subtract(my_invlist, PL_UpperLatin1,
5592 if (flags & SCF_DO_STCLASS_AND) {
5593 ssc_intersection(data->start_class, my_invlist, invert);
5594 ssc_clear_locale(data->start_class);
5597 assert(flags & SCF_DO_STCLASS_OR);
5598 ssc_union(data->start_class, my_invlist, invert);
5600 SvREFCNT_dec(my_invlist);
5602 if (flags & SCF_DO_STCLASS_OR)
5603 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5604 flags &= ~SCF_DO_STCLASS;
5607 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5608 data->flags |= (OP(scan) == MEOL
5611 scan_commit(pRExC_state, data, minlenp, is_inf);
5614 else if ( PL_regkind[OP(scan)] == BRANCHJ
5615 /* Lookbehind, or need to calculate parens/evals/stclass: */
5616 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5617 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5619 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5620 || OP(scan) == UNLESSM )
5622 /* Negative Lookahead/lookbehind
5623 In this case we can't do fixed string optimisation.
5626 SSize_t deltanext, minnext, fake = 0;
5631 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5633 data_fake.whilem_c = data->whilem_c;
5634 data_fake.last_closep = data->last_closep;
5637 data_fake.last_closep = &fake;
5638 data_fake.pos_delta = delta;
5639 if ( flags & SCF_DO_STCLASS && !scan->flags
5640 && OP(scan) == IFMATCH ) { /* Lookahead */
5641 ssc_init(pRExC_state, &intrnl);
5642 data_fake.start_class = &intrnl;
5643 f |= SCF_DO_STCLASS_AND;
5645 if (flags & SCF_WHILEM_VISITED_POS)
5646 f |= SCF_WHILEM_VISITED_POS;
5647 next = regnext(scan);
5648 nscan = NEXTOPER(NEXTOPER(scan));
5650 /* recurse study_chunk() for lookahead body */
5651 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5652 last, &data_fake, stopparen,
5653 recursed_depth, NULL, f, depth+1);
5656 FAIL("Variable length lookbehind not implemented");
5658 else if (minnext > (I32)U8_MAX) {
5659 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5662 scan->flags = (U8)minnext;
5665 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5667 if (data_fake.flags & SF_HAS_EVAL)
5668 data->flags |= SF_HAS_EVAL;
5669 data->whilem_c = data_fake.whilem_c;
5671 if (f & SCF_DO_STCLASS_AND) {
5672 if (flags & SCF_DO_STCLASS_OR) {
5673 /* OR before, AND after: ideally we would recurse with
5674 * data_fake to get the AND applied by study of the
5675 * remainder of the pattern, and then derecurse;
5676 * *** HACK *** for now just treat as "no information".
5677 * See [perl #56690].
5679 ssc_init(pRExC_state, data->start_class);
5681 /* AND before and after: combine and continue. These
5682 * assertions are zero-length, so can match an EMPTY
5684 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5685 ANYOF_FLAGS(data->start_class)
5686 |= SSC_MATCHES_EMPTY_STRING;
5690 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5692 /* Positive Lookahead/lookbehind
5693 In this case we can do fixed string optimisation,
5694 but we must be careful about it. Note in the case of
5695 lookbehind the positions will be offset by the minimum
5696 length of the pattern, something we won't know about
5697 until after the recurse.
5699 SSize_t deltanext, fake = 0;
5703 /* We use SAVEFREEPV so that when the full compile
5704 is finished perl will clean up the allocated
5705 minlens when it's all done. This way we don't
5706 have to worry about freeing them when we know
5707 they wont be used, which would be a pain.
5710 Newx( minnextp, 1, SSize_t );
5711 SAVEFREEPV(minnextp);
5714 StructCopy(data, &data_fake, scan_data_t);
5715 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5718 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5719 data_fake.last_found=newSVsv(data->last_found);
5723 data_fake.last_closep = &fake;
5724 data_fake.flags = 0;
5725 data_fake.substrs[0].flags = 0;
5726 data_fake.substrs[1].flags = 0;
5727 data_fake.pos_delta = delta;
5729 data_fake.flags |= SF_IS_INF;
5730 if ( flags & SCF_DO_STCLASS && !scan->flags
5731 && OP(scan) == IFMATCH ) { /* Lookahead */
5732 ssc_init(pRExC_state, &intrnl);
5733 data_fake.start_class = &intrnl;
5734 f |= SCF_DO_STCLASS_AND;
5736 if (flags & SCF_WHILEM_VISITED_POS)
5737 f |= SCF_WHILEM_VISITED_POS;
5738 next = regnext(scan);
5739 nscan = NEXTOPER(NEXTOPER(scan));
5741 /* positive lookahead study_chunk() recursion */
5742 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5743 &deltanext, last, &data_fake,
5744 stopparen, recursed_depth, NULL,
5748 FAIL("Variable length lookbehind not implemented");
5750 else if (*minnextp > (I32)U8_MAX) {
5751 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5754 scan->flags = (U8)*minnextp;
5759 if (f & SCF_DO_STCLASS_AND) {
5760 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5761 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
5764 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5766 if (data_fake.flags & SF_HAS_EVAL)
5767 data->flags |= SF_HAS_EVAL;
5768 data->whilem_c = data_fake.whilem_c;
5769 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5771 if (RExC_rx->minlen<*minnextp)
5772 RExC_rx->minlen=*minnextp;
5773 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5774 SvREFCNT_dec_NN(data_fake.last_found);
5776 for (i = 0; i < 2; i++) {
5777 if (data_fake.substrs[i].minlenp != minlenp) {
5778 data->substrs[i].min_offset =
5779 data_fake.substrs[i].min_offset;
5780 data->substrs[i].max_offset =
5781 data_fake.substrs[i].max_offset;
5782 data->substrs[i].minlenp =
5783 data_fake.substrs[i].minlenp;
5784 data->substrs[i].lookbehind += scan->flags;
5793 else if (OP(scan) == OPEN) {
5794 if (stopparen != (I32)ARG(scan))
5797 else if (OP(scan) == CLOSE) {
5798 if (stopparen == (I32)ARG(scan)) {
5801 if ((I32)ARG(scan) == is_par) {
5802 next = regnext(scan);
5804 if ( next && (OP(next) != WHILEM) && next < last)
5805 is_par = 0; /* Disable optimization */
5808 *(data->last_closep) = ARG(scan);
5810 else if (OP(scan) == EVAL) {
5812 data->flags |= SF_HAS_EVAL;
5814 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5815 if (flags & SCF_DO_SUBSTR) {
5816 scan_commit(pRExC_state, data, minlenp, is_inf);
5817 flags &= ~SCF_DO_SUBSTR;
5819 if (data && OP(scan)==ACCEPT) {
5820 data->flags |= SCF_SEEN_ACCEPT;
5825 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5827 if (flags & SCF_DO_SUBSTR) {
5828 scan_commit(pRExC_state, data, minlenp, is_inf);
5829 data->cur_is_floating = 1; /* float */
5831 is_inf = is_inf_internal = 1;
5832 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5833 ssc_anything(data->start_class);
5834 flags &= ~SCF_DO_STCLASS;
5836 else if (OP(scan) == GPOS) {
5837 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5838 !(delta || is_inf || (data && data->pos_delta)))
5840 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5841 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5842 if (RExC_rx->gofs < (STRLEN)min)
5843 RExC_rx->gofs = min;
5845 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5849 #ifdef TRIE_STUDY_OPT
5850 #ifdef FULL_TRIE_STUDY
5851 else if (PL_regkind[OP(scan)] == TRIE) {
5852 /* NOTE - There is similar code to this block above for handling
5853 BRANCH nodes on the initial study. If you change stuff here
5855 regnode *trie_node= scan;
5856 regnode *tail= regnext(scan);
5857 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5858 SSize_t max1 = 0, min1 = SSize_t_MAX;
5861 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5862 /* Cannot merge strings after this. */
5863 scan_commit(pRExC_state, data, minlenp, is_inf);
5865 if (flags & SCF_DO_STCLASS)
5866 ssc_init_zero(pRExC_state, &accum);
5872 const regnode *nextbranch= NULL;
5875 for ( word=1 ; word <= trie->wordcount ; word++)
5877 SSize_t deltanext=0, minnext=0, f = 0, fake;
5878 regnode_ssc this_class;
5880 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5882 data_fake.whilem_c = data->whilem_c;
5883 data_fake.last_closep = data->last_closep;
5886 data_fake.last_closep = &fake;
5887 data_fake.pos_delta = delta;
5888 if (flags & SCF_DO_STCLASS) {
5889 ssc_init(pRExC_state, &this_class);
5890 data_fake.start_class = &this_class;
5891 f = SCF_DO_STCLASS_AND;
5893 if (flags & SCF_WHILEM_VISITED_POS)
5894 f |= SCF_WHILEM_VISITED_POS;
5896 if (trie->jump[word]) {
5898 nextbranch = trie_node + trie->jump[0];
5899 scan= trie_node + trie->jump[word];
5900 /* We go from the jump point to the branch that follows
5901 it. Note this means we need the vestigal unused
5902 branches even though they arent otherwise used. */
5903 /* optimise study_chunk() for TRIE */
5904 minnext = study_chunk(pRExC_state, &scan, minlenp,
5905 &deltanext, (regnode *)nextbranch, &data_fake,
5906 stopparen, recursed_depth, NULL, f,depth+1);
5908 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5909 nextbranch= regnext((regnode*)nextbranch);
5911 if (min1 > (SSize_t)(minnext + trie->minlen))
5912 min1 = minnext + trie->minlen;
5913 if (deltanext == SSize_t_MAX) {
5914 is_inf = is_inf_internal = 1;
5916 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5917 max1 = minnext + deltanext + trie->maxlen;
5919 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5921 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5922 if ( stopmin > min + min1)
5923 stopmin = min + min1;
5924 flags &= ~SCF_DO_SUBSTR;
5926 data->flags |= SCF_SEEN_ACCEPT;
5929 if (data_fake.flags & SF_HAS_EVAL)
5930 data->flags |= SF_HAS_EVAL;
5931 data->whilem_c = data_fake.whilem_c;
5933 if (flags & SCF_DO_STCLASS)
5934 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5937 if (flags & SCF_DO_SUBSTR) {
5938 data->pos_min += min1;
5939 data->pos_delta += max1 - min1;
5940 if (max1 != min1 || is_inf)
5941 data->cur_is_floating = 1; /* float */
5944 if (delta != SSize_t_MAX) {
5945 if (SSize_t_MAX - (max1 - min1) >= delta)
5946 delta += max1 - min1;
5948 delta = SSize_t_MAX;
5950 if (flags & SCF_DO_STCLASS_OR) {
5951 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5953 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5954 flags &= ~SCF_DO_STCLASS;
5957 else if (flags & SCF_DO_STCLASS_AND) {
5959 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5960 flags &= ~SCF_DO_STCLASS;
5963 /* Switch to OR mode: cache the old value of
5964 * data->start_class */
5966 StructCopy(data->start_class, and_withp, regnode_ssc);
5967 flags &= ~SCF_DO_STCLASS_AND;
5968 StructCopy(&accum, data->start_class, regnode_ssc);
5969 flags |= SCF_DO_STCLASS_OR;
5976 else if (PL_regkind[OP(scan)] == TRIE) {
5977 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5980 min += trie->minlen;
5981 delta += (trie->maxlen - trie->minlen);
5982 flags &= ~SCF_DO_STCLASS; /* xxx */
5983 if (flags & SCF_DO_SUBSTR) {
5984 /* Cannot expect anything... */
5985 scan_commit(pRExC_state, data, minlenp, is_inf);
5986 data->pos_min += trie->minlen;
5987 data->pos_delta += (trie->maxlen - trie->minlen);
5988 if (trie->maxlen != trie->minlen)
5989 data->cur_is_floating = 1; /* float */
5991 if (trie->jump) /* no more substrings -- for now /grr*/
5992 flags &= ~SCF_DO_SUBSTR;
5994 #endif /* old or new */
5995 #endif /* TRIE_STUDY_OPT */
5997 /* Else: zero-length, ignore. */
5998 scan = regnext(scan);
6003 /* we need to unwind recursion. */
6006 DEBUG_STUDYDATA("frame-end", data, depth, is_inf);
6007 DEBUG_PEEP("fend", scan, depth, flags);
6009 /* restore previous context */
6010 last = frame->last_regnode;
6011 scan = frame->next_regnode;
6012 stopparen = frame->stopparen;
6013 recursed_depth = frame->prev_recursed_depth;
6015 RExC_frame_last = frame->prev_frame;
6016 frame = frame->this_prev_frame;
6017 goto fake_study_recurse;
6021 DEBUG_STUDYDATA("pre-fin", data, depth, is_inf);
6024 *deltap = is_inf_internal ? SSize_t_MAX : delta;
6026 if (flags & SCF_DO_SUBSTR && is_inf)
6027 data->pos_delta = SSize_t_MAX - data->pos_min;
6028 if (is_par > (I32)U8_MAX)
6030 if (is_par && pars==1 && data) {
6031 data->flags |= SF_IN_PAR;
6032 data->flags &= ~SF_HAS_PAR;
6034 else if (pars && data) {
6035 data->flags |= SF_HAS_PAR;
6036 data->flags &= ~SF_IN_PAR;
6038 if (flags & SCF_DO_STCLASS_OR)
6039 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6040 if (flags & SCF_TRIE_RESTUDY)
6041 data->flags |= SCF_TRIE_RESTUDY;
6043 DEBUG_STUDYDATA("post-fin", data, depth, is_inf);
6046 SSize_t final_minlen= min < stopmin ? min : stopmin;
6048 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
6049 if (final_minlen > SSize_t_MAX - delta)
6050 RExC_maxlen = SSize_t_MAX;
6051 else if (RExC_maxlen < final_minlen + delta)
6052 RExC_maxlen = final_minlen + delta;
6054 return final_minlen;
6056 NOT_REACHED; /* NOTREACHED */
6060 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
6062 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
6064 PERL_ARGS_ASSERT_ADD_DATA;
6066 Renewc(RExC_rxi->data,
6067 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6068 char, struct reg_data);
6070 Renew(RExC_rxi->data->what, count + n, U8);
6072 Newx(RExC_rxi->data->what, n, U8);
6073 RExC_rxi->data->count = count + n;
6074 Copy(s, RExC_rxi->data->what + count, n, U8);
6078 /*XXX: todo make this not included in a non debugging perl, but appears to be
6079 * used anyway there, in 'use re' */
6080 #ifndef PERL_IN_XSUB_RE
6082 Perl_reginitcolors(pTHX)
6084 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6086 char *t = savepv(s);
6090 t = strchr(t, '\t');
6096 PL_colors[i] = t = (char *)"";
6101 PL_colors[i++] = (char *)"";
6108 #ifdef TRIE_STUDY_OPT
6109 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6112 (data.flags & SCF_TRIE_RESTUDY) \
6120 #define CHECK_RESTUDY_GOTO_butfirst
6124 * pregcomp - compile a regular expression into internal code
6126 * Decides which engine's compiler to call based on the hint currently in
6130 #ifndef PERL_IN_XSUB_RE
6132 /* return the currently in-scope regex engine (or the default if none) */
6134 regexp_engine const *
6135 Perl_current_re_engine(pTHX)
6137 if (IN_PERL_COMPILETIME) {
6138 HV * const table = GvHV(PL_hintgv);
6141 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6142 return &PL_core_reg_engine;
6143 ptr = hv_fetchs(table, "regcomp", FALSE);
6144 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6145 return &PL_core_reg_engine;
6146 return INT2PTR(regexp_engine*,SvIV(*ptr));
6150 if (!PL_curcop->cop_hints_hash)
6151 return &PL_core_reg_engine;
6152 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6153 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6154 return &PL_core_reg_engine;
6155 return INT2PTR(regexp_engine*,SvIV(ptr));
6161 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6163 regexp_engine const *eng = current_re_engine();
6164 GET_RE_DEBUG_FLAGS_DECL;
6166 PERL_ARGS_ASSERT_PREGCOMP;
6168 /* Dispatch a request to compile a regexp to correct regexp engine. */
6170 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6173 return CALLREGCOMP_ENG(eng, pattern, flags);
6177 /* public(ish) entry point for the perl core's own regex compiling code.
6178 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6179 * pattern rather than a list of OPs, and uses the internal engine rather
6180 * than the current one */
6183 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6185 SV *pat = pattern; /* defeat constness! */
6186 PERL_ARGS_ASSERT_RE_COMPILE;
6187 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6188 #ifdef PERL_IN_XSUB_RE
6191 &PL_core_reg_engine,
6193 NULL, NULL, rx_flags, 0);
6198 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6202 if (--cbs->refcnt > 0)
6204 for (n = 0; n < cbs->count; n++) {
6205 REGEXP *rx = cbs->cb[n].src_regex;
6206 cbs->cb[n].src_regex = NULL;
6214 static struct reg_code_blocks *
6215 S_alloc_code_blocks(pTHX_ int ncode)
6217 struct reg_code_blocks *cbs;
6218 Newx(cbs, 1, struct reg_code_blocks);
6221 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6223 Newx(cbs->cb, ncode, struct reg_code_block);
6230 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6231 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6232 * point to the realloced string and length.
6234 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6238 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6239 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6241 U8 *const src = (U8*)*pat_p;
6246 GET_RE_DEBUG_FLAGS_DECL;
6248 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6249 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6251 Newx(dst, *plen_p * 2 + 1, U8);
6254 while (s < *plen_p) {
6255 append_utf8_from_native_byte(src[s], &d);
6257 if (n < num_code_blocks) {
6258 assert(pRExC_state->code_blocks);
6259 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6260 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6261 assert(*(d - 1) == '(');
6264 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6265 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6266 assert(*(d - 1) == ')');
6275 *pat_p = (char*) dst;
6277 RExC_orig_utf8 = RExC_utf8 = 1;
6282 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6283 * while recording any code block indices, and handling overloading,
6284 * nested qr// objects etc. If pat is null, it will allocate a new
6285 * string, or just return the first arg, if there's only one.
6287 * Returns the malloced/updated pat.
6288 * patternp and pat_count is the array of SVs to be concatted;
6289 * oplist is the optional list of ops that generated the SVs;
6290 * recompile_p is a pointer to a boolean that will be set if
6291 * the regex will need to be recompiled.
6292 * delim, if non-null is an SV that will be inserted between each element
6296 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6297 SV *pat, SV ** const patternp, int pat_count,
6298 OP *oplist, bool *recompile_p, SV *delim)
6302 bool use_delim = FALSE;
6303 bool alloced = FALSE;
6305 /* if we know we have at least two args, create an empty string,
6306 * then concatenate args to that. For no args, return an empty string */
6307 if (!pat && pat_count != 1) {
6313 for (svp = patternp; svp < patternp + pat_count; svp++) {
6316 STRLEN orig_patlen = 0;
6318 SV *msv = use_delim ? delim : *svp;
6319 if (!msv) msv = &PL_sv_undef;
6321 /* if we've got a delimiter, we go round the loop twice for each
6322 * svp slot (except the last), using the delimiter the second
6331 if (SvTYPE(msv) == SVt_PVAV) {
6332 /* we've encountered an interpolated array within
6333 * the pattern, e.g. /...@a..../. Expand the list of elements,
6334 * then recursively append elements.
6335 * The code in this block is based on S_pushav() */
6337 AV *const av = (AV*)msv;
6338 const SSize_t maxarg = AvFILL(av) + 1;
6342 assert(oplist->op_type == OP_PADAV
6343 || oplist->op_type == OP_RV2AV);
6344 oplist = OpSIBLING(oplist);
6347 if (SvRMAGICAL(av)) {
6350 Newx(array, maxarg, SV*);
6352 for (i=0; i < maxarg; i++) {
6353 SV ** const svp = av_fetch(av, i, FALSE);
6354 array[i] = svp ? *svp : &PL_sv_undef;
6358 array = AvARRAY(av);
6360 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6361 array, maxarg, NULL, recompile_p,
6363 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6369 /* we make the assumption here that each op in the list of
6370 * op_siblings maps to one SV pushed onto the stack,
6371 * except for code blocks, with have both an OP_NULL and
6373 * This allows us to match up the list of SVs against the
6374 * list of OPs to find the next code block.
6376 * Note that PUSHMARK PADSV PADSV ..
6378 * PADRANGE PADSV PADSV ..
6379 * so the alignment still works. */
6382 if (oplist->op_type == OP_NULL
6383 && (oplist->op_flags & OPf_SPECIAL))
6385 assert(n < pRExC_state->code_blocks->count);
6386 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6387 pRExC_state->code_blocks->cb[n].block = oplist;
6388 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6391 oplist = OpSIBLING(oplist); /* skip CONST */
6394 oplist = OpSIBLING(oplist);;
6397 /* apply magic and QR overloading to arg */
6400 if (SvROK(msv) && SvAMAGIC(msv)) {
6401 SV *sv = AMG_CALLunary(msv, regexp_amg);
6405 if (SvTYPE(sv) != SVt_REGEXP)
6406 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6411 /* try concatenation overload ... */
6412 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6413 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6416 /* overloading involved: all bets are off over literal
6417 * code. Pretend we haven't seen it */
6419 pRExC_state->code_blocks->count -= n;
6423 /* ... or failing that, try "" overload */
6424 while (SvAMAGIC(msv)
6425 && (sv = AMG_CALLunary(msv, string_amg))
6429 && SvRV(msv) == SvRV(sv))
6434 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6438 /* this is a partially unrolled
6439 * sv_catsv_nomg(pat, msv);
6440 * that allows us to adjust code block indices if
6443 char *dst = SvPV_force_nomg(pat, dlen);
6445 if (SvUTF8(msv) && !SvUTF8(pat)) {
6446 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6447 sv_setpvn(pat, dst, dlen);
6450 sv_catsv_nomg(pat, msv);
6454 /* We have only one SV to process, but we need to verify
6455 * it is properly null terminated or we will fail asserts
6456 * later. In theory we probably shouldn't get such SV's,
6457 * but if we do we should handle it gracefully. */
6458 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) ) {
6459 /* not a string, or a string with a trailing null */
6462 /* a string with no trailing null, we need to copy it
6463 * so it we have a trailing null */
6469 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
6472 /* extract any code blocks within any embedded qr//'s */
6473 if (rx && SvTYPE(rx) == SVt_REGEXP
6474 && RX_ENGINE((REGEXP*)rx)->op_comp)
6477 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6478 if (ri->code_blocks && ri->code_blocks->count) {
6480 /* the presence of an embedded qr// with code means
6481 * we should always recompile: the text of the
6482 * qr// may not have changed, but it may be a
6483 * different closure than last time */
6485 if (pRExC_state->code_blocks) {
6486 int new_count = pRExC_state->code_blocks->count
6487 + ri->code_blocks->count;
6488 Renew(pRExC_state->code_blocks->cb,
6489 new_count, struct reg_code_block);
6490 pRExC_state->code_blocks->count = new_count;
6493 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
6494 ri->code_blocks->count);
6496 for (i=0; i < ri->code_blocks->count; i++) {
6497 struct reg_code_block *src, *dst;
6498 STRLEN offset = orig_patlen
6499 + ReANY((REGEXP *)rx)->pre_prefix;
6500 assert(n < pRExC_state->code_blocks->count);
6501 src = &ri->code_blocks->cb[i];
6502 dst = &pRExC_state->code_blocks->cb[n];
6503 dst->start = src->start + offset;
6504 dst->end = src->end + offset;
6505 dst->block = src->block;
6506 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6515 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6524 /* see if there are any run-time code blocks in the pattern.
6525 * False positives are allowed */
6528 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6529 char *pat, STRLEN plen)
6534 PERL_UNUSED_CONTEXT;
6536 for (s = 0; s < plen; s++) {
6537 if ( pRExC_state->code_blocks
6538 && n < pRExC_state->code_blocks->count
6539 && s == pRExC_state->code_blocks->cb[n].start)
6541 s = pRExC_state->code_blocks->cb[n].end;
6545 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6547 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6549 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6556 /* Handle run-time code blocks. We will already have compiled any direct
6557 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6558 * copy of it, but with any literal code blocks blanked out and
6559 * appropriate chars escaped; then feed it into
6561 * eval "qr'modified_pattern'"
6565 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6569 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6571 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6572 * and merge them with any code blocks of the original regexp.
6574 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6575 * instead, just save the qr and return FALSE; this tells our caller that
6576 * the original pattern needs upgrading to utf8.
6580 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6581 char *pat, STRLEN plen)
6585 GET_RE_DEBUG_FLAGS_DECL;
6587 if (pRExC_state->runtime_code_qr) {
6588 /* this is the second time we've been called; this should
6589 * only happen if the main pattern got upgraded to utf8
6590 * during compilation; re-use the qr we compiled first time
6591 * round (which should be utf8 too)
6593 qr = pRExC_state->runtime_code_qr;
6594 pRExC_state->runtime_code_qr = NULL;
6595 assert(RExC_utf8 && SvUTF8(qr));
6601 int newlen = plen + 7; /* allow for "qr''xx\0" extra chars */
6605 /* determine how many extra chars we need for ' and \ escaping */
6606 for (s = 0; s < plen; s++) {
6607 if (pat[s] == '\'' || pat[s] == '\\')
6611 Newx(newpat, newlen, char);
6613 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6615 for (s = 0; s < plen; s++) {
6616 if ( pRExC_state->code_blocks
6617 && n < pRExC_state->code_blocks->count
6618 && s == pRExC_state->code_blocks->cb[n].start)
6620 /* blank out literal code block */
6621 assert(pat[s] == '(');
6622 while (s <= pRExC_state->code_blocks->cb[n].end) {
6630 if (pat[s] == '\'' || pat[s] == '\\')
6635 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
6637 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
6643 Perl_re_printf( aTHX_
6644 "%sre-parsing pattern for runtime code:%s %s\n",
6645 PL_colors[4],PL_colors[5],newpat);
6648 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
6654 PUSHSTACKi(PERLSI_REQUIRE);
6655 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
6656 * parsing qr''; normally only q'' does this. It also alters
6658 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
6659 SvREFCNT_dec_NN(sv);
6664 SV * const errsv = ERRSV;
6665 if (SvTRUE_NN(errsv))
6666 /* use croak_sv ? */
6667 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
6669 assert(SvROK(qr_ref));
6671 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
6672 /* the leaving below frees the tmp qr_ref.
6673 * Give qr a life of its own */
6681 if (!RExC_utf8 && SvUTF8(qr)) {
6682 /* first time through; the pattern got upgraded; save the
6683 * qr for the next time through */
6684 assert(!pRExC_state->runtime_code_qr);
6685 pRExC_state->runtime_code_qr = qr;
6690 /* extract any code blocks within the returned qr// */
6693 /* merge the main (r1) and run-time (r2) code blocks into one */
6695 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6696 struct reg_code_block *new_block, *dst;
6697 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6701 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
6703 SvREFCNT_dec_NN(qr);
6707 if (!r1->code_blocks)
6708 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
6710 r1c = r1->code_blocks->count;
6711 r2c = r2->code_blocks->count;
6713 Newx(new_block, r1c + r2c, struct reg_code_block);
6717 while (i1 < r1c || i2 < r2c) {
6718 struct reg_code_block *src;
6722 src = &r2->code_blocks->cb[i2++];
6726 src = &r1->code_blocks->cb[i1++];
6727 else if ( r1->code_blocks->cb[i1].start
6728 < r2->code_blocks->cb[i2].start)
6730 src = &r1->code_blocks->cb[i1++];
6731 assert(src->end < r2->code_blocks->cb[i2].start);
6734 assert( r1->code_blocks->cb[i1].start
6735 > r2->code_blocks->cb[i2].start);
6736 src = &r2->code_blocks->cb[i2++];
6738 assert(src->end < r1->code_blocks->cb[i1].start);
6741 assert(pat[src->start] == '(');
6742 assert(pat[src->end] == ')');
6743 dst->start = src->start;
6744 dst->end = src->end;
6745 dst->block = src->block;
6746 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6750 r1->code_blocks->count += r2c;
6751 Safefree(r1->code_blocks->cb);
6752 r1->code_blocks->cb = new_block;
6755 SvREFCNT_dec_NN(qr);
6761 S_setup_longest(pTHX_ RExC_state_t *pRExC_state,
6762 struct reg_substr_datum *rsd,
6763 struct scan_data_substrs *sub,
6764 STRLEN longest_length)
6766 /* This is the common code for setting up the floating and fixed length
6767 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6768 * as to whether succeeded or not */
6772 bool eol = cBOOL(sub->flags & SF_BEFORE_EOL);
6773 bool meol = cBOOL(sub->flags & SF_BEFORE_MEOL);
6775 if (! (longest_length
6776 || (eol /* Can't have SEOL and MULTI */
6777 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6779 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6780 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6785 /* copy the information about the longest from the reg_scan_data
6786 over to the program. */
6787 if (SvUTF8(sub->str)) {
6789 rsd->utf8_substr = sub->str;
6791 rsd->substr = sub->str;
6792 rsd->utf8_substr = NULL;
6794 /* end_shift is how many chars that must be matched that
6795 follow this item. We calculate it ahead of time as once the
6796 lookbehind offset is added in we lose the ability to correctly
6798 ml = sub->minlenp ? *(sub->minlenp) : (SSize_t)longest_length;
6799 rsd->end_shift = ml - sub->min_offset
6801 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
6803 + (SvTAIL(sub->str) != 0)
6807 t = (eol/* Can't have SEOL and MULTI */
6808 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6809 fbm_compile(sub->str, t ? FBMcf_TAIL : 0);
6815 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6816 * regular expression into internal code.
6817 * The pattern may be passed either as:
6818 * a list of SVs (patternp plus pat_count)
6819 * a list of OPs (expr)
6820 * If both are passed, the SV list is used, but the OP list indicates
6821 * which SVs are actually pre-compiled code blocks
6823 * The SVs in the list have magic and qr overloading applied to them (and
6824 * the list may be modified in-place with replacement SVs in the latter
6827 * If the pattern hasn't changed from old_re, then old_re will be
6830 * eng is the current engine. If that engine has an op_comp method, then
6831 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6832 * do the initial concatenation of arguments and pass on to the external
6835 * If is_bare_re is not null, set it to a boolean indicating whether the
6836 * arg list reduced (after overloading) to a single bare regex which has
6837 * been returned (i.e. /$qr/).
6839 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6841 * pm_flags contains the PMf_* flags, typically based on those from the
6842 * pm_flags field of the related PMOP. Currently we're only interested in
6843 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6845 * We can't allocate space until we know how big the compiled form will be,
6846 * but we can't compile it (and thus know how big it is) until we've got a
6847 * place to put the code. So we cheat: we compile it twice, once with code
6848 * generation turned off and size counting turned on, and once "for real".
6849 * This also means that we don't allocate space until we are sure that the
6850 * thing really will compile successfully, and we never have to move the
6851 * code and thus invalidate pointers into it. (Note that it has to be in
6852 * one piece because free() must be able to free it all.) [NB: not true in perl]
6854 * Beware that the optimization-preparation code in here knows about some
6855 * of the structure of the compiled regexp. [I'll say.]
6859 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6860 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6861 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6865 regexp_internal *ri;
6873 SV** new_patternp = patternp;
6875 /* these are all flags - maybe they should be turned
6876 * into a single int with different bit masks */
6877 I32 sawlookahead = 0;
6882 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6884 bool runtime_code = 0;
6886 RExC_state_t RExC_state;
6887 RExC_state_t * const pRExC_state = &RExC_state;
6888 #ifdef TRIE_STUDY_OPT
6890 RExC_state_t copyRExC_state;
6892 GET_RE_DEBUG_FLAGS_DECL;
6894 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6896 DEBUG_r(if (!PL_colorset) reginitcolors());
6898 /* Initialize these here instead of as-needed, as is quick and avoids
6899 * having to test them each time otherwise */
6900 if (! PL_AboveLatin1) {
6902 char * dump_len_string;
6905 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6906 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6907 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6908 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6909 PL_HasMultiCharFold =
6910 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6912 /* This is calculated here, because the Perl program that generates the
6913 * static global ones doesn't currently have access to
6914 * NUM_ANYOF_CODE_POINTS */
6915 PL_InBitmap = _new_invlist(2);
6916 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
6917 NUM_ANYOF_CODE_POINTS - 1);
6919 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
6920 if ( ! dump_len_string
6921 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
6923 PL_dump_re_max_len = 60; /* A reasonable default */
6928 pRExC_state->warn_text = NULL;
6929 pRExC_state->code_blocks = NULL;
6932 *is_bare_re = FALSE;
6934 if (expr && (expr->op_type == OP_LIST ||
6935 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6936 /* allocate code_blocks if needed */
6940 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
6941 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6942 ncode++; /* count of DO blocks */
6945 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
6949 /* compile-time pattern with just OP_CONSTs and DO blocks */
6954 /* find how many CONSTs there are */
6957 if (expr->op_type == OP_CONST)
6960 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6961 if (o->op_type == OP_CONST)
6965 /* fake up an SV array */
6967 assert(!new_patternp);
6968 Newx(new_patternp, n, SV*);
6969 SAVEFREEPV(new_patternp);
6973 if (expr->op_type == OP_CONST)
6974 new_patternp[n] = cSVOPx_sv(expr);
6976 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6977 if (o->op_type == OP_CONST)
6978 new_patternp[n++] = cSVOPo_sv;
6983 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6984 "Assembling pattern from %d elements%s\n", pat_count,
6985 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6987 /* set expr to the first arg op */
6989 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
6990 && expr->op_type != OP_CONST)
6992 expr = cLISTOPx(expr)->op_first;
6993 assert( expr->op_type == OP_PUSHMARK
6994 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6995 || expr->op_type == OP_PADRANGE);
6996 expr = OpSIBLING(expr);
6999 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
7000 expr, &recompile, NULL);
7002 /* handle bare (possibly after overloading) regex: foo =~ $re */
7007 if (SvTYPE(re) == SVt_REGEXP) {
7011 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7012 "Precompiled pattern%s\n",
7013 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7019 exp = SvPV_nomg(pat, plen);
7021 if (!eng->op_comp) {
7022 if ((SvUTF8(pat) && IN_BYTES)
7023 || SvGMAGICAL(pat) || SvAMAGIC(pat))
7025 /* make a temporary copy; either to convert to bytes,
7026 * or to avoid repeating get-magic / overloaded stringify */
7027 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
7028 (IN_BYTES ? 0 : SvUTF8(pat)));
7030 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
7033 /* ignore the utf8ness if the pattern is 0 length */
7034 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
7036 RExC_uni_semantics = 0;
7037 RExC_seen_unfolded_sharp_s = 0;
7038 RExC_contains_locale = 0;
7039 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
7040 RExC_study_started = 0;
7041 pRExC_state->runtime_code_qr = NULL;
7042 RExC_frame_head= NULL;
7043 RExC_frame_last= NULL;
7044 RExC_frame_count= 0;
7047 RExC_mysv1= sv_newmortal();
7048 RExC_mysv2= sv_newmortal();
7051 SV *dsv= sv_newmortal();
7052 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7053 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
7054 PL_colors[4],PL_colors[5],s);
7058 /* we jump here if we have to recompile, e.g., from upgrading the pattern
7061 if ((pm_flags & PMf_USE_RE_EVAL)
7062 /* this second condition covers the non-regex literal case,
7063 * i.e. $foo =~ '(?{})'. */
7064 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
7066 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
7068 /* return old regex if pattern hasn't changed */
7069 /* XXX: note in the below we have to check the flags as well as the
7072 * Things get a touch tricky as we have to compare the utf8 flag
7073 * independently from the compile flags. */
7077 && !!RX_UTF8(old_re) == !!RExC_utf8
7078 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7079 && RX_PRECOMP(old_re)
7080 && RX_PRELEN(old_re) == plen
7081 && memEQ(RX_PRECOMP(old_re), exp, plen)
7082 && !runtime_code /* with runtime code, always recompile */ )
7087 rx_flags = orig_rx_flags;
7089 if ( initial_charset == REGEX_DEPENDS_CHARSET
7090 && (RExC_utf8 ||RExC_uni_semantics))
7093 /* Set to use unicode semantics if the pattern is in utf8 and has the
7094 * 'depends' charset specified, as it means unicode when utf8 */
7095 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7099 RExC_precomp_adj = 0;
7100 RExC_flags = rx_flags;
7101 RExC_pm_flags = pm_flags;
7104 assert(TAINTING_get || !TAINT_get);
7106 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7108 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7109 /* whoops, we have a non-utf8 pattern, whilst run-time code
7110 * got compiled as utf8. Try again with a utf8 pattern */
7111 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7112 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7113 goto redo_first_pass;
7116 assert(!pRExC_state->runtime_code_qr);
7122 RExC_in_lookbehind = 0;
7123 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7126 RExC_recode_x_to_native = 0;
7128 RExC_in_multi_char_class = 0;
7130 /* First pass: determine size, legality. */
7132 RExC_start = RExC_adjusted_start = exp;
7133 RExC_end = exp + plen;
7134 RExC_precomp_end = RExC_end;
7139 RExC_emit = (regnode *) &RExC_emit_dummy;
7140 RExC_whilem_seen = 0;
7141 RExC_open_parens = NULL;
7142 RExC_close_parens = NULL;
7144 RExC_paren_names = NULL;
7146 RExC_paren_name_list = NULL;
7148 RExC_recurse = NULL;
7149 RExC_study_chunk_recursed = NULL;
7150 RExC_study_chunk_recursed_bytes= 0;
7151 RExC_recurse_count = 0;
7152 pRExC_state->code_index = 0;
7154 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7155 * code makes sure the final byte is an uncounted NUL. But should this
7156 * ever not be the case, lots of things could read beyond the end of the
7157 * buffer: loops like
7158 * while(isFOO(*RExC_parse)) RExC_parse++;
7159 * strchr(RExC_parse, "foo");
7160 * etc. So it is worth noting. */
7161 assert(*RExC_end == '\0');
7164 Perl_re_printf( aTHX_ "Starting first pass (sizing)\n");
7166 RExC_lastparse=NULL;
7169 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7170 /* It's possible to write a regexp in ascii that represents Unicode
7171 codepoints outside of the byte range, such as via \x{100}. If we
7172 detect such a sequence we have to convert the entire pattern to utf8
7173 and then recompile, as our sizing calculation will have been based
7174 on 1 byte == 1 character, but we will need to use utf8 to encode
7175 at least some part of the pattern, and therefore must convert the whole
7178 if (flags & RESTART_PASS1) {
7179 if (flags & NEED_UTF8) {
7180 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7181 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7184 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7185 "Need to redo pass 1\n"));
7188 goto redo_first_pass;
7190 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#" UVxf, (UV) flags);
7194 Perl_re_printf( aTHX_
7195 "Required size %" IVdf " nodes\n"
7196 "Starting second pass (creation)\n",
7199 RExC_lastparse=NULL;
7202 /* The first pass could have found things that force Unicode semantics */
7203 if ((RExC_utf8 || RExC_uni_semantics)
7204 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
7206 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7209 /* Small enough for pointer-storage convention?
7210 If extralen==0, this means that we will not need long jumps. */
7211 if (RExC_size >= 0x10000L && RExC_extralen)
7212 RExC_size += RExC_extralen;
7215 if (RExC_whilem_seen > 15)
7216 RExC_whilem_seen = 15;
7218 /* Allocate space and zero-initialize. Note, the two step process
7219 of zeroing when in debug mode, thus anything assigned has to
7220 happen after that */
7221 rx = (REGEXP*) newSV_type(SVt_REGEXP);
7223 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7224 char, regexp_internal);
7225 if ( r == NULL || ri == NULL )
7226 FAIL("Regexp out of space");
7228 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
7229 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7232 /* bulk initialize base fields with 0. */
7233 Zero(ri, sizeof(regexp_internal), char);
7236 /* non-zero initialization begins here */
7239 r->extflags = rx_flags;
7240 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7242 if (pm_flags & PMf_IS_QR) {
7243 ri->code_blocks = pRExC_state->code_blocks;
7244 if (ri->code_blocks)
7245 ri->code_blocks->refcnt++;
7249 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7250 bool has_charset = (get_regex_charset(r->extflags)
7251 != REGEX_DEPENDS_CHARSET);
7253 /* The caret is output if there are any defaults: if not all the STD
7254 * flags are set, or if no character set specifier is needed */
7256 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7258 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7259 == REG_RUN_ON_COMMENT_SEEN);
7260 U8 reganch = (U8)((r->extflags & RXf_PMf_STD_PMMOD)
7261 >> RXf_PMf_STD_PMMOD_SHIFT);
7262 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7265 /* We output all the necessary flags; we never output a minus, as all
7266 * those are defaults, so are
7267 * covered by the caret */
7268 const STRLEN wraplen = plen + has_p + has_runon
7269 + has_default /* If needs a caret */
7270 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7272 /* If needs a character set specifier */
7273 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7274 + (sizeof("(?:)") - 1);
7276 /* make sure PL_bitcount bounds not exceeded */
7277 assert(sizeof(STD_PAT_MODS) <= 8);
7279 p = sv_grow(MUTABLE_SV(rx), wraplen + 1); /* +1 for the ending NUL */
7282 SvFLAGS(rx) |= SVf_UTF8;
7285 /* If a default, cover it using the caret */
7287 *p++= DEFAULT_PAT_MOD;
7291 const char* const name = get_regex_charset_name(r->extflags, &len);
7292 Copy(name, p, len, char);
7296 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7299 while((ch = *fptr++)) {
7307 Copy(RExC_precomp, p, plen, char);
7308 assert ((RX_WRAPPED(rx) - p) < 16);
7309 r->pre_prefix = p - RX_WRAPPED(rx);
7315 SvCUR_set(rx, p - RX_WRAPPED(rx));
7319 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
7321 /* Useful during FAIL. */
7322 #ifdef RE_TRACK_PATTERN_OFFSETS
7323 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
7324 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7325 "%s %" UVuf " bytes for offset annotations.\n",
7326 ri->u.offsets ? "Got" : "Couldn't get",
7327 (UV)((2*RExC_size+1) * sizeof(U32))));
7329 SetProgLen(ri,RExC_size);
7334 /* Second pass: emit code. */
7335 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7336 RExC_pm_flags = pm_flags;
7338 RExC_end = exp + plen;
7340 RExC_emit_start = ri->program;
7341 RExC_emit = ri->program;
7342 RExC_emit_bound = ri->program + RExC_size + 1;
7343 pRExC_state->code_index = 0;
7345 *((char*) RExC_emit++) = (char) REG_MAGIC;
7346 /* setup various meta data about recursion, this all requires
7347 * RExC_npar to be correctly set, and a bit later on we clear it */
7348 if (RExC_seen & REG_RECURSE_SEEN) {
7349 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
7350 "%*s%*s Setting up open/close parens\n",
7351 22, "| |", (int)(0 * 2 + 1), ""));
7353 /* setup RExC_open_parens, which holds the address of each
7354 * OPEN tag, and to make things simpler for the 0 index
7355 * the start of the program - this is used later for offsets */
7356 Newxz(RExC_open_parens, RExC_npar,regnode *);
7357 SAVEFREEPV(RExC_open_parens);
7358 RExC_open_parens[0] = RExC_emit;
7360 /* setup RExC_close_parens, which holds the address of each
7361 * CLOSE tag, and to make things simpler for the 0 index
7362 * the end of the program - this is used later for offsets */
7363 Newxz(RExC_close_parens, RExC_npar,regnode *);
7364 SAVEFREEPV(RExC_close_parens);
7365 /* we dont know where end op starts yet, so we dont
7366 * need to set RExC_close_parens[0] like we do RExC_open_parens[0] above */
7368 /* Note, RExC_npar is 1 + the number of parens in a pattern.
7369 * So its 1 if there are no parens. */
7370 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
7371 ((RExC_npar & 0x07) != 0);
7372 Newx(RExC_study_chunk_recursed,
7373 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7374 SAVEFREEPV(RExC_study_chunk_recursed);
7377 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7379 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#" UVxf, (UV) flags);
7382 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7385 /* XXXX To minimize changes to RE engine we always allocate
7386 3-units-long substrs field. */
7387 Newx(r->substrs, 1, struct reg_substr_data);
7388 if (RExC_recurse_count) {
7389 Newx(RExC_recurse,RExC_recurse_count,regnode *);
7390 SAVEFREEPV(RExC_recurse);
7394 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
7396 RExC_study_chunk_recursed_count= 0;
7398 Zero(r->substrs, 1, struct reg_substr_data);
7399 if (RExC_study_chunk_recursed) {
7400 Zero(RExC_study_chunk_recursed,
7401 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7405 #ifdef TRIE_STUDY_OPT
7407 StructCopy(&zero_scan_data, &data, scan_data_t);
7408 copyRExC_state = RExC_state;
7411 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
7413 RExC_state = copyRExC_state;
7414 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
7415 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
7417 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
7418 StructCopy(&zero_scan_data, &data, scan_data_t);
7421 StructCopy(&zero_scan_data, &data, scan_data_t);
7424 /* Dig out information for optimizations. */
7425 r->extflags = RExC_flags; /* was pm_op */
7426 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
7429 SvUTF8_on(rx); /* Unicode in it? */
7430 ri->regstclass = NULL;
7431 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
7432 r->intflags |= PREGf_NAUGHTY;
7433 scan = ri->program + 1; /* First BRANCH. */
7435 /* testing for BRANCH here tells us whether there is "must appear"
7436 data in the pattern. If there is then we can use it for optimisations */
7437 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
7440 STRLEN longest_length[2];
7441 regnode_ssc ch_class; /* pointed to by data */
7443 SSize_t last_close = 0; /* pointed to by data */
7444 regnode *first= scan;
7445 regnode *first_next= regnext(first);
7449 * Skip introductions and multiplicators >= 1
7450 * so that we can extract the 'meat' of the pattern that must
7451 * match in the large if() sequence following.
7452 * NOTE that EXACT is NOT covered here, as it is normally
7453 * picked up by the optimiser separately.
7455 * This is unfortunate as the optimiser isnt handling lookahead
7456 * properly currently.
7459 while ((OP(first) == OPEN && (sawopen = 1)) ||
7460 /* An OR of *one* alternative - should not happen now. */
7461 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
7462 /* for now we can't handle lookbehind IFMATCH*/
7463 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
7464 (OP(first) == PLUS) ||
7465 (OP(first) == MINMOD) ||
7466 /* An {n,m} with n>0 */
7467 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
7468 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
7471 * the only op that could be a regnode is PLUS, all the rest
7472 * will be regnode_1 or regnode_2.
7474 * (yves doesn't think this is true)
7476 if (OP(first) == PLUS)
7479 if (OP(first) == MINMOD)
7481 first += regarglen[OP(first)];
7483 first = NEXTOPER(first);
7484 first_next= regnext(first);
7487 /* Starting-point info. */
7489 DEBUG_PEEP("first:", first, 0, 0);
7490 /* Ignore EXACT as we deal with it later. */
7491 if (PL_regkind[OP(first)] == EXACT) {
7492 if (OP(first) == EXACT || OP(first) == EXACTL)
7493 NOOP; /* Empty, get anchored substr later. */
7495 ri->regstclass = first;
7498 else if (PL_regkind[OP(first)] == TRIE &&
7499 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
7501 /* this can happen only on restudy */
7502 ri->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
7505 else if (REGNODE_SIMPLE(OP(first)))
7506 ri->regstclass = first;
7507 else if (PL_regkind[OP(first)] == BOUND ||
7508 PL_regkind[OP(first)] == NBOUND)
7509 ri->regstclass = first;
7510 else if (PL_regkind[OP(first)] == BOL) {
7511 r->intflags |= (OP(first) == MBOL
7514 first = NEXTOPER(first);
7517 else if (OP(first) == GPOS) {
7518 r->intflags |= PREGf_ANCH_GPOS;
7519 first = NEXTOPER(first);
7522 else if ((!sawopen || !RExC_sawback) &&
7524 (OP(first) == STAR &&
7525 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
7526 !(r->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
7528 /* turn .* into ^.* with an implied $*=1 */
7530 (OP(NEXTOPER(first)) == REG_ANY)
7533 r->intflags |= (type | PREGf_IMPLICIT);
7534 first = NEXTOPER(first);
7537 if (sawplus && !sawminmod && !sawlookahead
7538 && (!sawopen || !RExC_sawback)
7539 && !pRExC_state->code_blocks) /* May examine pos and $& */
7540 /* x+ must match at the 1st pos of run of x's */
7541 r->intflags |= PREGf_SKIP;
7543 /* Scan is after the zeroth branch, first is atomic matcher. */
7544 #ifdef TRIE_STUDY_OPT
7547 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7548 (IV)(first - scan + 1))
7552 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7553 (IV)(first - scan + 1))
7559 * If there's something expensive in the r.e., find the
7560 * longest literal string that must appear and make it the
7561 * regmust. Resolve ties in favor of later strings, since
7562 * the regstart check works with the beginning of the r.e.
7563 * and avoiding duplication strengthens checking. Not a
7564 * strong reason, but sufficient in the absence of others.
7565 * [Now we resolve ties in favor of the earlier string if
7566 * it happens that c_offset_min has been invalidated, since the
7567 * earlier string may buy us something the later one won't.]
7570 data.substrs[0].str = newSVpvs("");
7571 data.substrs[1].str = newSVpvs("");
7572 data.last_found = newSVpvs("");
7573 data.cur_is_floating = 0; /* initially any found substring is fixed */
7574 ENTER_with_name("study_chunk");
7575 SAVEFREESV(data.substrs[0].str);
7576 SAVEFREESV(data.substrs[1].str);
7577 SAVEFREESV(data.last_found);
7579 if (!ri->regstclass) {
7580 ssc_init(pRExC_state, &ch_class);
7581 data.start_class = &ch_class;
7582 stclass_flag = SCF_DO_STCLASS_AND;
7583 } else /* XXXX Check for BOUND? */
7585 data.last_closep = &last_close;
7589 * MAIN ENTRY FOR study_chunk() FOR m/PATTERN/
7590 * (NO top level branches)
7592 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
7593 scan + RExC_size, /* Up to end */
7595 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
7596 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
7600 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
7603 if ( RExC_npar == 1 && !data.cur_is_floating
7604 && data.last_start_min == 0 && data.last_end > 0
7605 && !RExC_seen_zerolen
7606 && !(RExC_seen & REG_VERBARG_SEEN)
7607 && !(RExC_seen & REG_GPOS_SEEN)
7609 r->extflags |= RXf_CHECK_ALL;
7611 scan_commit(pRExC_state, &data,&minlen,0);
7614 /* XXX this is done in reverse order because that's the way the
7615 * code was before it was parameterised. Don't know whether it
7616 * actually needs doing in reverse order. DAPM */
7617 for (i = 1; i >= 0; i--) {
7618 longest_length[i] = CHR_SVLEN(data.substrs[i].str);
7621 && SvCUR(data.substrs[0].str) /* ok to leave SvCUR */
7622 && data.substrs[0].min_offset
7623 == data.substrs[1].min_offset
7624 && SvCUR(data.substrs[0].str)
7625 == SvCUR(data.substrs[1].str)
7627 && S_setup_longest (aTHX_ pRExC_state,
7628 &(r->substrs->data[i]),
7632 r->substrs->data[i].min_offset =
7633 data.substrs[i].min_offset - data.substrs[i].lookbehind;
7635 r->substrs->data[i].max_offset = data.substrs[i].max_offset;
7636 /* Don't offset infinity */
7637 if (data.substrs[i].max_offset < SSize_t_MAX)
7638 r->substrs->data[i].max_offset -= data.substrs[i].lookbehind;
7639 SvREFCNT_inc_simple_void_NN(data.substrs[i].str);
7642 r->substrs->data[i].substr = NULL;
7643 r->substrs->data[i].utf8_substr = NULL;
7644 longest_length[i] = 0;
7648 LEAVE_with_name("study_chunk");
7651 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
7652 ri->regstclass = NULL;
7654 if ((!(r->substrs->data[0].substr || r->substrs->data[0].utf8_substr)
7655 || r->substrs->data[0].min_offset)
7657 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7658 && is_ssc_worth_it(pRExC_state, data.start_class))
7660 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7662 ssc_finalize(pRExC_state, data.start_class);
7664 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7665 StructCopy(data.start_class,
7666 (regnode_ssc*)RExC_rxi->data->data[n],
7668 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7669 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7670 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
7671 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7672 Perl_re_printf( aTHX_
7673 "synthetic stclass \"%s\".\n",
7674 SvPVX_const(sv));});
7675 data.start_class = NULL;
7678 /* A temporary algorithm prefers floated substr to fixed one of
7679 * same length to dig more info. */
7680 i = (longest_length[0] <= longest_length[1]);
7681 r->substrs->check_ix = i;
7682 r->check_end_shift = r->substrs->data[i].end_shift;
7683 r->check_substr = r->substrs->data[i].substr;
7684 r->check_utf8 = r->substrs->data[i].utf8_substr;
7685 r->check_offset_min = r->substrs->data[i].min_offset;
7686 r->check_offset_max = r->substrs->data[i].max_offset;
7687 if (!i && (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS)))
7688 r->intflags |= PREGf_NOSCAN;
7690 if ((r->check_substr || r->check_utf8) ) {
7691 r->extflags |= RXf_USE_INTUIT;
7692 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
7693 r->extflags |= RXf_INTUIT_TAIL;
7696 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
7697 if ( (STRLEN)minlen < longest_length[1] )
7698 minlen= longest_length[1];
7699 if ( (STRLEN)minlen < longest_length[0] )
7700 minlen= longest_length[0];
7704 /* Several toplevels. Best we can is to set minlen. */
7706 regnode_ssc ch_class;
7707 SSize_t last_close = 0;
7709 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
7711 scan = ri->program + 1;
7712 ssc_init(pRExC_state, &ch_class);
7713 data.start_class = &ch_class;
7714 data.last_closep = &last_close;
7718 * MAIN ENTRY FOR study_chunk() FOR m/P1|P2|.../
7719 * (patterns WITH top level branches)
7721 minlen = study_chunk(pRExC_state,
7722 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7723 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7724 ? SCF_TRIE_DOING_RESTUDY
7728 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7730 r->check_substr = NULL;
7731 r->check_utf8 = NULL;
7732 r->substrs->data[0].substr = NULL;
7733 r->substrs->data[0].utf8_substr = NULL;
7734 r->substrs->data[1].substr = NULL;
7735 r->substrs->data[1].utf8_substr = NULL;
7737 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7738 && is_ssc_worth_it(pRExC_state, data.start_class))
7740 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7742 ssc_finalize(pRExC_state, data.start_class);
7744 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7745 StructCopy(data.start_class,
7746 (regnode_ssc*)RExC_rxi->data->data[n],
7748 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7749 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7750 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7751 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7752 Perl_re_printf( aTHX_
7753 "synthetic stclass \"%s\".\n",
7754 SvPVX_const(sv));});
7755 data.start_class = NULL;
7759 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7760 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7761 r->maxlen = REG_INFTY;
7764 r->maxlen = RExC_maxlen;
7767 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7768 the "real" pattern. */
7770 Perl_re_printf( aTHX_ "minlen: %" IVdf " r->minlen:%" IVdf " maxlen:%" IVdf "\n",
7771 (IV)minlen, (IV)r->minlen, (IV)RExC_maxlen);
7773 r->minlenret = minlen;
7774 if (r->minlen < minlen)
7777 if (RExC_seen & REG_RECURSE_SEEN ) {
7778 r->intflags |= PREGf_RECURSE_SEEN;
7779 Newx(r->recurse_locinput, r->nparens + 1, char *);
7781 if (RExC_seen & REG_GPOS_SEEN)
7782 r->intflags |= PREGf_GPOS_SEEN;
7783 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7784 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7786 if (pRExC_state->code_blocks)
7787 r->extflags |= RXf_EVAL_SEEN;
7788 if (RExC_seen & REG_VERBARG_SEEN)
7790 r->intflags |= PREGf_VERBARG_SEEN;
7791 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7793 if (RExC_seen & REG_CUTGROUP_SEEN)
7794 r->intflags |= PREGf_CUTGROUP_SEEN;
7795 if (pm_flags & PMf_USE_RE_EVAL)
7796 r->intflags |= PREGf_USE_RE_EVAL;
7797 if (RExC_paren_names)
7798 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7800 RXp_PAREN_NAMES(r) = NULL;
7802 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7803 * so it can be used in pp.c */
7804 if (r->intflags & PREGf_ANCH)
7805 r->extflags |= RXf_IS_ANCHORED;
7809 /* this is used to identify "special" patterns that might result
7810 * in Perl NOT calling the regex engine and instead doing the match "itself",
7811 * particularly special cases in split//. By having the regex compiler
7812 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7813 * we avoid weird issues with equivalent patterns resulting in different behavior,
7814 * AND we allow non Perl engines to get the same optimizations by the setting the
7815 * flags appropriately - Yves */
7816 regnode *first = ri->program + 1;
7818 regnode *next = regnext(first);
7821 if (PL_regkind[fop] == NOTHING && nop == END)
7822 r->extflags |= RXf_NULL;
7823 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
7824 /* when fop is SBOL first->flags will be true only when it was
7825 * produced by parsing /\A/, and not when parsing /^/. This is
7826 * very important for the split code as there we want to
7827 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
7828 * See rt #122761 for more details. -- Yves */
7829 r->extflags |= RXf_START_ONLY;
7830 else if (fop == PLUS
7831 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7833 r->extflags |= RXf_WHITE;
7834 else if ( r->extflags & RXf_SPLIT
7835 && (fop == EXACT || fop == EXACTL)
7836 && STR_LEN(first) == 1
7837 && *(STRING(first)) == ' '
7839 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7843 if (RExC_contains_locale) {
7844 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7848 if (RExC_paren_names) {
7849 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7850 ri->data->data[ri->name_list_idx]
7851 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7854 ri->name_list_idx = 0;
7856 while ( RExC_recurse_count > 0 ) {
7857 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
7859 * This data structure is set up in study_chunk() and is used
7860 * to calculate the distance between a GOSUB regopcode and
7861 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
7864 * If for some reason someone writes code that optimises
7865 * away a GOSUB opcode then the assert should be changed to
7866 * an if(scan) to guard the ARG2L_SET() - Yves
7869 assert(scan && OP(scan) == GOSUB);
7870 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - scan );
7873 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7874 /* assume we don't need to swap parens around before we match */
7876 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
7877 (unsigned long)RExC_study_chunk_recursed_count);
7881 Perl_re_printf( aTHX_ "Final program:\n");
7884 #ifdef RE_TRACK_PATTERN_OFFSETS
7885 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7886 const STRLEN len = ri->u.offsets[0];
7888 GET_RE_DEBUG_FLAGS_DECL;
7889 Perl_re_printf( aTHX_
7890 "Offsets: [%" UVuf "]\n\t", (UV)ri->u.offsets[0]);
7891 for (i = 1; i <= len; i++) {
7892 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7893 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7894 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7896 Perl_re_printf( aTHX_ "\n");
7901 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7902 * by setting the regexp SV to readonly-only instead. If the
7903 * pattern's been recompiled, the USEDness should remain. */
7904 if (old_re && SvREADONLY(old_re))
7912 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7915 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7917 PERL_UNUSED_ARG(value);
7919 if (flags & RXapif_FETCH) {
7920 return reg_named_buff_fetch(rx, key, flags);
7921 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7922 Perl_croak_no_modify();
7924 } else if (flags & RXapif_EXISTS) {
7925 return reg_named_buff_exists(rx, key, flags)
7928 } else if (flags & RXapif_REGNAMES) {
7929 return reg_named_buff_all(rx, flags);
7930 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7931 return reg_named_buff_scalar(rx, flags);
7933 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7939 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7942 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7943 PERL_UNUSED_ARG(lastkey);
7945 if (flags & RXapif_FIRSTKEY)
7946 return reg_named_buff_firstkey(rx, flags);
7947 else if (flags & RXapif_NEXTKEY)
7948 return reg_named_buff_nextkey(rx, flags);
7950 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7957 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7961 struct regexp *const rx = ReANY(r);
7963 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7965 if (rx && RXp_PAREN_NAMES(rx)) {
7966 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7969 SV* sv_dat=HeVAL(he_str);
7970 I32 *nums=(I32*)SvPVX(sv_dat);
7971 AV * const retarray = (flags & RXapif_ALL) ? newAV() : NULL;
7972 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7973 if ((I32)(rx->nparens) >= nums[i]
7974 && rx->offs[nums[i]].start != -1
7975 && rx->offs[nums[i]].end != -1)
7978 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7983 ret = newSVsv(&PL_sv_undef);
7986 av_push(retarray, ret);
7989 return newRV_noinc(MUTABLE_SV(retarray));
7996 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7999 struct regexp *const rx = ReANY(r);
8001 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
8003 if (rx && RXp_PAREN_NAMES(rx)) {
8004 if (flags & RXapif_ALL) {
8005 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
8007 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
8009 SvREFCNT_dec_NN(sv);
8021 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
8023 struct regexp *const rx = ReANY(r);
8025 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
8027 if ( rx && RXp_PAREN_NAMES(rx) ) {
8028 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
8030 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
8037 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
8039 struct regexp *const rx = ReANY(r);
8040 GET_RE_DEBUG_FLAGS_DECL;
8042 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
8044 if (rx && RXp_PAREN_NAMES(rx)) {
8045 HV *hv = RXp_PAREN_NAMES(rx);
8047 while ( (temphe = hv_iternext_flags(hv,0)) ) {
8050 SV* sv_dat = HeVAL(temphe);
8051 I32 *nums = (I32*)SvPVX(sv_dat);
8052 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8053 if ((I32)(rx->lastparen) >= nums[i] &&
8054 rx->offs[nums[i]].start != -1 &&
8055 rx->offs[nums[i]].end != -1)
8061 if (parno || flags & RXapif_ALL) {
8062 return newSVhek(HeKEY_hek(temphe));
8070 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8075 struct regexp *const rx = ReANY(r);
8077 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8079 if (rx && RXp_PAREN_NAMES(rx)) {
8080 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8081 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8082 } else if (flags & RXapif_ONE) {
8083 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8084 av = MUTABLE_AV(SvRV(ret));
8085 length = av_tindex(av);
8086 SvREFCNT_dec_NN(ret);
8087 return newSViv(length + 1);
8089 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8094 return &PL_sv_undef;
8098 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8100 struct regexp *const rx = ReANY(r);
8103 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8105 if (rx && RXp_PAREN_NAMES(rx)) {
8106 HV *hv= RXp_PAREN_NAMES(rx);
8108 (void)hv_iterinit(hv);
8109 while ( (temphe = hv_iternext_flags(hv,0)) ) {
8112 SV* sv_dat = HeVAL(temphe);
8113 I32 *nums = (I32*)SvPVX(sv_dat);
8114 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8115 if ((I32)(rx->lastparen) >= nums[i] &&
8116 rx->offs[nums[i]].start != -1 &&
8117 rx->offs[nums[i]].end != -1)
8123 if (parno || flags & RXapif_ALL) {
8124 av_push(av, newSVhek(HeKEY_hek(temphe)));
8129 return newRV_noinc(MUTABLE_SV(av));
8133 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8136 struct regexp *const rx = ReANY(r);
8142 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8144 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8145 || n == RX_BUFF_IDX_CARET_FULLMATCH
8146 || n == RX_BUFF_IDX_CARET_POSTMATCH
8149 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8151 /* on something like
8154 * the KEEPCOPY is set on the PMOP rather than the regex */
8155 if (PL_curpm && r == PM_GETRE(PL_curpm))
8156 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8165 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8166 /* no need to distinguish between them any more */
8167 n = RX_BUFF_IDX_FULLMATCH;
8169 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8170 && rx->offs[0].start != -1)
8172 /* $`, ${^PREMATCH} */
8173 i = rx->offs[0].start;
8177 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8178 && rx->offs[0].end != -1)
8180 /* $', ${^POSTMATCH} */
8181 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8182 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8185 if ( 0 <= n && n <= (I32)rx->nparens &&
8186 (s1 = rx->offs[n].start) != -1 &&
8187 (t1 = rx->offs[n].end) != -1)
8189 /* $&, ${^MATCH}, $1 ... */
8191 s = rx->subbeg + s1 - rx->suboffset;
8196 assert(s >= rx->subbeg);
8197 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8199 #ifdef NO_TAINT_SUPPORT
8200 sv_setpvn(sv, s, i);
8202 const int oldtainted = TAINT_get;
8204 sv_setpvn(sv, s, i);
8205 TAINT_set(oldtainted);
8207 if (RXp_MATCH_UTF8(rx))
8212 if (RXp_MATCH_TAINTED(rx)) {
8213 if (SvTYPE(sv) >= SVt_PVMG) {
8214 MAGIC* const mg = SvMAGIC(sv);
8217 SvMAGIC_set(sv, mg->mg_moremagic);
8219 if ((mgt = SvMAGIC(sv))) {
8220 mg->mg_moremagic = mgt;
8221 SvMAGIC_set(sv, mg);
8238 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8239 SV const * const value)
8241 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8243 PERL_UNUSED_ARG(rx);
8244 PERL_UNUSED_ARG(paren);
8245 PERL_UNUSED_ARG(value);
8248 Perl_croak_no_modify();
8252 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8255 struct regexp *const rx = ReANY(r);
8259 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8261 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8262 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8263 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8266 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8268 /* on something like
8271 * the KEEPCOPY is set on the PMOP rather than the regex */
8272 if (PL_curpm && r == PM_GETRE(PL_curpm))
8273 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8279 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8281 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8282 case RX_BUFF_IDX_PREMATCH: /* $` */
8283 if (rx->offs[0].start != -1) {
8284 i = rx->offs[0].start;
8293 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8294 case RX_BUFF_IDX_POSTMATCH: /* $' */
8295 if (rx->offs[0].end != -1) {
8296 i = rx->sublen - rx->offs[0].end;
8298 s1 = rx->offs[0].end;
8305 default: /* $& / ${^MATCH}, $1, $2, ... */
8306 if (paren <= (I32)rx->nparens &&
8307 (s1 = rx->offs[paren].start) != -1 &&
8308 (t1 = rx->offs[paren].end) != -1)
8314 if (ckWARN(WARN_UNINITIALIZED))
8315 report_uninit((const SV *)sv);
8320 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8321 const char * const s = rx->subbeg - rx->suboffset + s1;
8326 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8333 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8335 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8336 PERL_UNUSED_ARG(rx);
8340 return newSVpvs("Regexp");
8343 /* Scans the name of a named buffer from the pattern.
8344 * If flags is REG_RSN_RETURN_NULL returns null.
8345 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8346 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8347 * to the parsed name as looked up in the RExC_paren_names hash.
8348 * If there is an error throws a vFAIL().. type exception.
8351 #define REG_RSN_RETURN_NULL 0
8352 #define REG_RSN_RETURN_NAME 1
8353 #define REG_RSN_RETURN_DATA 2
8356 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8358 char *name_start = RExC_parse;
8360 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8362 assert (RExC_parse <= RExC_end);
8363 if (RExC_parse == RExC_end) NOOP;
8364 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8365 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8366 * using do...while */
8369 RExC_parse += UTF8SKIP(RExC_parse);
8370 } while ( RExC_parse < RExC_end
8371 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8375 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8377 RExC_parse++; /* so the <- from the vFAIL is after the offending
8379 vFAIL("Group name must start with a non-digit word character");
8383 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8384 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8385 if ( flags == REG_RSN_RETURN_NAME)
8387 else if (flags==REG_RSN_RETURN_DATA) {
8390 if ( ! sv_name ) /* should not happen*/
8391 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8392 if (RExC_paren_names)
8393 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8395 sv_dat = HeVAL(he_str);
8397 vFAIL("Reference to nonexistent named group");
8401 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
8402 (unsigned long) flags);
8404 NOT_REACHED; /* NOTREACHED */
8409 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
8411 if (RExC_lastparse!=RExC_parse) { \
8412 Perl_re_printf( aTHX_ "%s", \
8413 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
8414 RExC_end - RExC_parse, 16, \
8416 PERL_PV_ESCAPE_UNI_DETECT | \
8417 PERL_PV_PRETTY_ELLIPSES | \
8418 PERL_PV_PRETTY_LTGT | \
8419 PERL_PV_ESCAPE_RE | \
8420 PERL_PV_PRETTY_EXACTSIZE \
8424 Perl_re_printf( aTHX_ "%16s",""); \
8427 num = RExC_size + 1; \
8429 num=REG_NODE_NUM(RExC_emit); \
8430 if (RExC_lastnum!=num) \
8431 Perl_re_printf( aTHX_ "|%4d",num); \
8433 Perl_re_printf( aTHX_ "|%4s",""); \
8434 Perl_re_printf( aTHX_ "|%*s%-4s", \
8435 (int)((depth*2)), "", \
8439 RExC_lastparse=RExC_parse; \
8444 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8445 DEBUG_PARSE_MSG((funcname)); \
8446 Perl_re_printf( aTHX_ "%4s","\n"); \
8448 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
8449 DEBUG_PARSE_MSG((funcname)); \
8450 Perl_re_printf( aTHX_ fmt "\n",args); \
8453 /* This section of code defines the inversion list object and its methods. The
8454 * interfaces are highly subject to change, so as much as possible is static to
8455 * this file. An inversion list is here implemented as a malloc'd C UV array
8456 * as an SVt_INVLIST scalar.
8458 * An inversion list for Unicode is an array of code points, sorted by ordinal
8459 * number. Each element gives the code point that begins a range that extends
8460 * up-to but not including the code point given by the next element. The final
8461 * element gives the first code point of a range that extends to the platform's
8462 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
8463 * ...) give ranges whose code points are all in the inversion list. We say
8464 * that those ranges are in the set. The odd-numbered elements give ranges
8465 * whose code points are not in the inversion list, and hence not in the set.
8466 * Thus, element [0] is the first code point in the list. Element [1]
8467 * is the first code point beyond that not in the list; and element [2] is the
8468 * first code point beyond that that is in the list. In other words, the first
8469 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
8470 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
8471 * all code points in that range are not in the inversion list. The third
8472 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
8473 * list, and so forth. Thus every element whose index is divisible by two
8474 * gives the beginning of a range that is in the list, and every element whose
8475 * index is not divisible by two gives the beginning of a range not in the
8476 * list. If the final element's index is divisible by two, the inversion list
8477 * extends to the platform's infinity; otherwise the highest code point in the
8478 * inversion list is the contents of that element minus 1.
8480 * A range that contains just a single code point N will look like
8482 * invlist[i+1] == N+1
8484 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
8485 * impossible to represent, so element [i+1] is omitted. The single element
8487 * invlist[0] == UV_MAX
8488 * contains just UV_MAX, but is interpreted as matching to infinity.
8490 * Taking the complement (inverting) an inversion list is quite simple, if the
8491 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8492 * This implementation reserves an element at the beginning of each inversion
8493 * list to always contain 0; there is an additional flag in the header which
8494 * indicates if the list begins at the 0, or is offset to begin at the next
8495 * element. This means that the inversion list can be inverted without any
8496 * copying; just flip the flag.
8498 * More about inversion lists can be found in "Unicode Demystified"
8499 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8501 * The inversion list data structure is currently implemented as an SV pointing
8502 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8503 * array of UV whose memory management is automatically handled by the existing
8504 * facilities for SV's.
8506 * Some of the methods should always be private to the implementation, and some
8507 * should eventually be made public */
8509 /* The header definitions are in F<invlist_inline.h> */
8511 #ifndef PERL_IN_XSUB_RE
8513 PERL_STATIC_INLINE UV*
8514 S__invlist_array_init(SV* const invlist, const bool will_have_0)
8516 /* Returns a pointer to the first element in the inversion list's array.
8517 * This is called upon initialization of an inversion list. Where the
8518 * array begins depends on whether the list has the code point U+0000 in it
8519 * or not. The other parameter tells it whether the code that follows this
8520 * call is about to put a 0 in the inversion list or not. The first
8521 * element is either the element reserved for 0, if TRUE, or the element
8522 * after it, if FALSE */
8524 bool* offset = get_invlist_offset_addr(invlist);
8525 UV* zero_addr = (UV *) SvPVX(invlist);
8527 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
8530 assert(! _invlist_len(invlist));
8534 /* 1^1 = 0; 1^0 = 1 */
8535 *offset = 1 ^ will_have_0;
8536 return zero_addr + *offset;
8541 PERL_STATIC_INLINE void
8542 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8544 /* Sets the current number of elements stored in the inversion list.
8545 * Updates SvCUR correspondingly */
8546 PERL_UNUSED_CONTEXT;
8547 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8549 assert(SvTYPE(invlist) == SVt_INVLIST);
8554 : TO_INTERNAL_SIZE(len + offset));
8555 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
8558 #ifndef PERL_IN_XSUB_RE
8561 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
8563 /* Replaces the inversion list in 'dest' with the one from 'src'. It
8564 * steals the list from 'src', so 'src' is made to have a NULL list. This
8565 * is similar to what SvSetMagicSV() would do, if it were implemented on
8566 * inversion lists, though this routine avoids a copy */
8568 const UV src_len = _invlist_len(src);
8569 const bool src_offset = *get_invlist_offset_addr(src);
8570 const STRLEN src_byte_len = SvLEN(src);
8571 char * array = SvPVX(src);
8573 const int oldtainted = TAINT_get;
8575 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
8577 assert(SvTYPE(src) == SVt_INVLIST);
8578 assert(SvTYPE(dest) == SVt_INVLIST);
8579 assert(! invlist_is_iterating(src));
8580 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
8582 /* Make sure it ends in the right place with a NUL, as our inversion list
8583 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
8585 array[src_byte_len - 1] = '\0';
8587 TAINT_NOT; /* Otherwise it breaks */
8588 sv_usepvn_flags(dest,
8592 /* This flag is documented to cause a copy to be avoided */
8593 SV_HAS_TRAILING_NUL);
8594 TAINT_set(oldtainted);
8599 /* Finish up copying over the other fields in an inversion list */
8600 *get_invlist_offset_addr(dest) = src_offset;
8601 invlist_set_len(dest, src_len, src_offset);
8602 *get_invlist_previous_index_addr(dest) = 0;
8603 invlist_iterfinish(dest);
8606 PERL_STATIC_INLINE IV*
8607 S_get_invlist_previous_index_addr(SV* invlist)
8609 /* Return the address of the IV that is reserved to hold the cached index
8611 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
8613 assert(SvTYPE(invlist) == SVt_INVLIST);
8615 return &(((XINVLIST*) SvANY(invlist))->prev_index);
8618 PERL_STATIC_INLINE IV
8619 S_invlist_previous_index(SV* const invlist)
8621 /* Returns cached index of previous search */
8623 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
8625 return *get_invlist_previous_index_addr(invlist);
8628 PERL_STATIC_INLINE void
8629 S_invlist_set_previous_index(SV* const invlist, const IV index)
8631 /* Caches <index> for later retrieval */
8633 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
8635 assert(index == 0 || index < (int) _invlist_len(invlist));
8637 *get_invlist_previous_index_addr(invlist) = index;
8640 PERL_STATIC_INLINE void
8641 S_invlist_trim(SV* invlist)
8643 /* Free the not currently-being-used space in an inversion list */
8645 /* But don't free up the space needed for the 0 UV that is always at the
8646 * beginning of the list, nor the trailing NUL */
8647 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
8649 PERL_ARGS_ASSERT_INVLIST_TRIM;
8651 assert(SvTYPE(invlist) == SVt_INVLIST);
8653 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
8656 PERL_STATIC_INLINE void
8657 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
8659 PERL_ARGS_ASSERT_INVLIST_CLEAR;
8661 assert(SvTYPE(invlist) == SVt_INVLIST);
8663 invlist_set_len(invlist, 0, 0);
8664 invlist_trim(invlist);
8667 #endif /* ifndef PERL_IN_XSUB_RE */
8669 PERL_STATIC_INLINE bool
8670 S_invlist_is_iterating(SV* const invlist)
8672 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8674 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8677 #ifndef PERL_IN_XSUB_RE
8679 PERL_STATIC_INLINE UV
8680 S_invlist_max(SV* const invlist)
8682 /* Returns the maximum number of elements storable in the inversion list's
8683 * array, without having to realloc() */
8685 PERL_ARGS_ASSERT_INVLIST_MAX;
8687 assert(SvTYPE(invlist) == SVt_INVLIST);
8689 /* Assumes worst case, in which the 0 element is not counted in the
8690 * inversion list, so subtracts 1 for that */
8691 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
8692 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
8693 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
8696 Perl__new_invlist(pTHX_ IV initial_size)
8699 /* Return a pointer to a newly constructed inversion list, with enough
8700 * space to store 'initial_size' elements. If that number is negative, a
8701 * system default is used instead */
8705 if (initial_size < 0) {
8709 /* Allocate the initial space */
8710 new_list = newSV_type(SVt_INVLIST);
8712 /* First 1 is in case the zero element isn't in the list; second 1 is for
8714 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
8715 invlist_set_len(new_list, 0, 0);
8717 /* Force iterinit() to be used to get iteration to work */
8718 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
8720 *get_invlist_previous_index_addr(new_list) = 0;
8726 Perl__new_invlist_C_array(pTHX_ const UV* const list)
8728 /* Return a pointer to a newly constructed inversion list, initialized to
8729 * point to <list>, which has to be in the exact correct inversion list
8730 * form, including internal fields. Thus this is a dangerous routine that
8731 * should not be used in the wrong hands. The passed in 'list' contains
8732 * several header fields at the beginning that are not part of the
8733 * inversion list body proper */
8735 const STRLEN length = (STRLEN) list[0];
8736 const UV version_id = list[1];
8737 const bool offset = cBOOL(list[2]);
8738 #define HEADER_LENGTH 3
8739 /* If any of the above changes in any way, you must change HEADER_LENGTH
8740 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
8741 * perl -E 'say int(rand 2**31-1)'
8743 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
8744 data structure type, so that one being
8745 passed in can be validated to be an
8746 inversion list of the correct vintage.
8749 SV* invlist = newSV_type(SVt_INVLIST);
8751 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8753 if (version_id != INVLIST_VERSION_ID) {
8754 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8757 /* The generated array passed in includes header elements that aren't part
8758 * of the list proper, so start it just after them */
8759 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8761 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8762 shouldn't touch it */
8764 *(get_invlist_offset_addr(invlist)) = offset;
8766 /* The 'length' passed to us is the physical number of elements in the
8767 * inversion list. But if there is an offset the logical number is one
8769 invlist_set_len(invlist, length - offset, offset);
8771 invlist_set_previous_index(invlist, 0);
8773 /* Initialize the iteration pointer. */
8774 invlist_iterfinish(invlist);
8776 SvREADONLY_on(invlist);
8782 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8784 /* Grow the maximum size of an inversion list */
8786 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8788 assert(SvTYPE(invlist) == SVt_INVLIST);
8790 /* Add one to account for the zero element at the beginning which may not
8791 * be counted by the calling parameters */
8792 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8796 S__append_range_to_invlist(pTHX_ SV* const invlist,
8797 const UV start, const UV end)
8799 /* Subject to change or removal. Append the range from 'start' to 'end' at
8800 * the end of the inversion list. The range must be above any existing
8804 UV max = invlist_max(invlist);
8805 UV len = _invlist_len(invlist);
8808 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8810 if (len == 0) { /* Empty lists must be initialized */
8811 offset = start != 0;
8812 array = _invlist_array_init(invlist, ! offset);
8815 /* Here, the existing list is non-empty. The current max entry in the
8816 * list is generally the first value not in the set, except when the
8817 * set extends to the end of permissible values, in which case it is
8818 * the first entry in that final set, and so this call is an attempt to
8819 * append out-of-order */
8821 UV final_element = len - 1;
8822 array = invlist_array(invlist);
8823 if ( array[final_element] > start
8824 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8826 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",
8827 array[final_element], start,
8828 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8831 /* Here, it is a legal append. If the new range begins 1 above the end
8832 * of the range below it, it is extending the range below it, so the
8833 * new first value not in the set is one greater than the newly
8834 * extended range. */
8835 offset = *get_invlist_offset_addr(invlist);
8836 if (array[final_element] == start) {
8837 if (end != UV_MAX) {
8838 array[final_element] = end + 1;
8841 /* But if the end is the maximum representable on the machine,
8842 * assume that infinity was actually what was meant. Just let
8843 * the range that this would extend to have no end */
8844 invlist_set_len(invlist, len - 1, offset);
8850 /* Here the new range doesn't extend any existing set. Add it */
8852 len += 2; /* Includes an element each for the start and end of range */
8854 /* If wll overflow the existing space, extend, which may cause the array to
8857 invlist_extend(invlist, len);
8859 /* Have to set len here to avoid assert failure in invlist_array() */
8860 invlist_set_len(invlist, len, offset);
8862 array = invlist_array(invlist);
8865 invlist_set_len(invlist, len, offset);
8868 /* The next item on the list starts the range, the one after that is
8869 * one past the new range. */
8870 array[len - 2] = start;
8871 if (end != UV_MAX) {
8872 array[len - 1] = end + 1;
8875 /* But if the end is the maximum representable on the machine, just let
8876 * the range have no end */
8877 invlist_set_len(invlist, len - 1, offset);
8882 Perl__invlist_search(SV* const invlist, const UV cp)
8884 /* Searches the inversion list for the entry that contains the input code
8885 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8886 * return value is the index into the list's array of the range that
8887 * contains <cp>, that is, 'i' such that
8888 * array[i] <= cp < array[i+1]
8893 IV high = _invlist_len(invlist);
8894 const IV highest_element = high - 1;
8897 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8899 /* If list is empty, return failure. */
8904 /* (We can't get the array unless we know the list is non-empty) */
8905 array = invlist_array(invlist);
8907 mid = invlist_previous_index(invlist);
8909 if (mid > highest_element) {
8910 mid = highest_element;
8913 /* <mid> contains the cache of the result of the previous call to this
8914 * function (0 the first time). See if this call is for the same result,
8915 * or if it is for mid-1. This is under the theory that calls to this
8916 * function will often be for related code points that are near each other.
8917 * And benchmarks show that caching gives better results. We also test
8918 * here if the code point is within the bounds of the list. These tests
8919 * replace others that would have had to be made anyway to make sure that
8920 * the array bounds were not exceeded, and these give us extra information
8921 * at the same time */
8922 if (cp >= array[mid]) {
8923 if (cp >= array[highest_element]) {
8924 return highest_element;
8927 /* Here, array[mid] <= cp < array[highest_element]. This means that
8928 * the final element is not the answer, so can exclude it; it also
8929 * means that <mid> is not the final element, so can refer to 'mid + 1'
8931 if (cp < array[mid + 1]) {
8937 else { /* cp < aray[mid] */
8938 if (cp < array[0]) { /* Fail if outside the array */
8942 if (cp >= array[mid - 1]) {
8947 /* Binary search. What we are looking for is <i> such that
8948 * array[i] <= cp < array[i+1]
8949 * The loop below converges on the i+1. Note that there may not be an
8950 * (i+1)th element in the array, and things work nonetheless */
8951 while (low < high) {
8952 mid = (low + high) / 2;
8953 assert(mid <= highest_element);
8954 if (array[mid] <= cp) { /* cp >= array[mid] */
8957 /* We could do this extra test to exit the loop early.
8958 if (cp < array[low]) {
8963 else { /* cp < array[mid] */
8970 invlist_set_previous_index(invlist, high);
8975 Perl__invlist_populate_swatch(SV* const invlist,
8976 const UV start, const UV end, U8* swatch)
8978 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8979 * but is used when the swash has an inversion list. This makes this much
8980 * faster, as it uses a binary search instead of a linear one. This is
8981 * intimately tied to that function, and perhaps should be in utf8.c,
8982 * except it is intimately tied to inversion lists as well. It assumes
8983 * that <swatch> is all 0's on input */
8986 const IV len = _invlist_len(invlist);
8990 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8992 if (len == 0) { /* Empty inversion list */
8996 array = invlist_array(invlist);
8998 /* Find which element it is */
8999 i = _invlist_search(invlist, start);
9001 /* We populate from <start> to <end> */
9002 while (current < end) {
9005 /* The inversion list gives the results for every possible code point
9006 * after the first one in the list. Only those ranges whose index is
9007 * even are ones that the inversion list matches. For the odd ones,
9008 * and if the initial code point is not in the list, we have to skip
9009 * forward to the next element */
9010 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
9012 if (i >= len) { /* Finished if beyond the end of the array */
9016 if (current >= end) { /* Finished if beyond the end of what we
9018 if (LIKELY(end < UV_MAX)) {
9022 /* We get here when the upper bound is the maximum
9023 * representable on the machine, and we are looking for just
9024 * that code point. Have to special case it */
9026 goto join_end_of_list;
9029 assert(current >= start);
9031 /* The current range ends one below the next one, except don't go past
9034 upper = (i < len && array[i] < end) ? array[i] : end;
9036 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
9037 * for each code point in it */
9038 for (; current < upper; current++) {
9039 const STRLEN offset = (STRLEN)(current - start);
9040 swatch[offset >> 3] |= 1 << (offset & 7);
9045 /* Quit if at the end of the list */
9048 /* But first, have to deal with the highest possible code point on
9049 * the platform. The previous code assumes that <end> is one
9050 * beyond where we want to populate, but that is impossible at the
9051 * platform's infinity, so have to handle it specially */
9052 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
9054 const STRLEN offset = (STRLEN)(end - start);
9055 swatch[offset >> 3] |= 1 << (offset & 7);
9060 /* Advance to the next range, which will be for code points not in the
9069 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9070 const bool complement_b, SV** output)
9072 /* Take the union of two inversion lists and point '*output' to it. On
9073 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9074 * even 'a' or 'b'). If to an inversion list, the contents of the original
9075 * list will be replaced by the union. The first list, 'a', may be
9076 * NULL, in which case a copy of the second list is placed in '*output'.
9077 * If 'complement_b' is TRUE, the union is taken of the complement
9078 * (inversion) of 'b' instead of b itself.
9080 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9081 * Richard Gillam, published by Addison-Wesley, and explained at some
9082 * length there. The preface says to incorporate its examples into your
9083 * code at your own risk.
9085 * The algorithm is like a merge sort. */
9087 const UV* array_a; /* a's array */
9089 UV len_a; /* length of a's array */
9092 SV* u; /* the resulting union */
9096 UV i_a = 0; /* current index into a's array */
9100 /* running count, as explained in the algorithm source book; items are
9101 * stopped accumulating and are output when the count changes to/from 0.
9102 * The count is incremented when we start a range that's in an input's set,
9103 * and decremented when we start a range that's not in a set. So this
9104 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9105 * and hence nothing goes into the union; 1, just one of the inputs is in
9106 * its set (and its current range gets added to the union); and 2 when both
9107 * inputs are in their sets. */
9110 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9112 assert(*output == NULL || SvTYPE(*output) == SVt_INVLIST);
9114 len_b = _invlist_len(b);
9117 /* Here, 'b' is empty, hence it's complement is all possible code
9118 * points. So if the union includes the complement of 'b', it includes
9119 * everything, and we need not even look at 'a'. It's easiest to
9120 * create a new inversion list that matches everything. */
9122 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9124 if (*output == NULL) { /* If the output didn't exist, just point it
9126 *output = everything;
9128 else { /* Otherwise, replace its contents with the new list */
9129 invlist_replace_list_destroys_src(*output, everything);
9130 SvREFCNT_dec_NN(everything);
9136 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9137 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9138 * output will be empty */
9140 if (a == NULL || _invlist_len(a) == 0) {
9141 if (*output == NULL) {
9142 *output = _new_invlist(0);
9145 invlist_clear(*output);
9150 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9151 * union. We can just return a copy of 'a' if '*output' doesn't point
9152 * to an existing list */
9153 if (*output == NULL) {
9154 *output = invlist_clone(a);
9158 /* If the output is to overwrite 'a', we have a no-op, as it's
9164 /* Here, '*output' is to be overwritten by 'a' */
9165 u = invlist_clone(a);
9166 invlist_replace_list_destroys_src(*output, u);
9172 /* Here 'b' is not empty. See about 'a' */
9174 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9176 /* Here, 'a' is empty (and b is not). That means the union will come
9177 * entirely from 'b'. If '*output' is NULL, we can directly return a
9178 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9181 SV ** dest = (*output == NULL) ? output : &u;
9182 *dest = invlist_clone(b);
9184 _invlist_invert(*dest);
9188 invlist_replace_list_destroys_src(*output, u);
9195 /* Here both lists exist and are non-empty */
9196 array_a = invlist_array(a);
9197 array_b = invlist_array(b);
9199 /* If are to take the union of 'a' with the complement of b, set it
9200 * up so are looking at b's complement. */
9203 /* To complement, we invert: if the first element is 0, remove it. To
9204 * do this, we just pretend the array starts one later */
9205 if (array_b[0] == 0) {
9211 /* But if the first element is not zero, we pretend the list starts
9212 * at the 0 that is always stored immediately before the array. */
9218 /* Size the union for the worst case: that the sets are completely
9220 u = _new_invlist(len_a + len_b);
9222 /* Will contain U+0000 if either component does */
9223 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9224 || (len_b > 0 && array_b[0] == 0));
9226 /* Go through each input list item by item, stopping when have exhausted
9228 while (i_a < len_a && i_b < len_b) {
9229 UV cp; /* The element to potentially add to the union's array */
9230 bool cp_in_set; /* is it in the the input list's set or not */
9232 /* We need to take one or the other of the two inputs for the union.
9233 * Since we are merging two sorted lists, we take the smaller of the
9234 * next items. In case of a tie, we take first the one that is in its
9235 * set. If we first took the one not in its set, it would decrement
9236 * the count, possibly to 0 which would cause it to be output as ending
9237 * the range, and the next time through we would take the same number,
9238 * and output it again as beginning the next range. By doing it the
9239 * opposite way, there is no possibility that the count will be
9240 * momentarily decremented to 0, and thus the two adjoining ranges will
9241 * be seamlessly merged. (In a tie and both are in the set or both not
9242 * in the set, it doesn't matter which we take first.) */
9243 if ( array_a[i_a] < array_b[i_b]
9244 || ( array_a[i_a] == array_b[i_b]
9245 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9247 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9248 cp = array_a[i_a++];
9251 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9252 cp = array_b[i_b++];
9255 /* Here, have chosen which of the two inputs to look at. Only output
9256 * if the running count changes to/from 0, which marks the
9257 * beginning/end of a range that's in the set */
9260 array_u[i_u++] = cp;
9267 array_u[i_u++] = cp;
9273 /* The loop above increments the index into exactly one of the input lists
9274 * each iteration, and ends when either index gets to its list end. That
9275 * means the other index is lower than its end, and so something is
9276 * remaining in that one. We decrement 'count', as explained below, if
9277 * that list is in its set. (i_a and i_b each currently index the element
9278 * beyond the one we care about.) */
9279 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9280 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9285 /* Above we decremented 'count' if the list that had unexamined elements in
9286 * it was in its set. This has made it so that 'count' being non-zero
9287 * means there isn't anything left to output; and 'count' equal to 0 means
9288 * that what is left to output is precisely that which is left in the
9289 * non-exhausted input list.
9291 * To see why, note first that the exhausted input obviously has nothing
9292 * left to add to the union. If it was in its set at its end, that means
9293 * the set extends from here to the platform's infinity, and hence so does
9294 * the union and the non-exhausted set is irrelevant. The exhausted set
9295 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9296 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9297 * 'count' remains at 1. This is consistent with the decremented 'count'
9298 * != 0 meaning there's nothing left to add to the union.
9300 * But if the exhausted input wasn't in its set, it contributed 0 to
9301 * 'count', and the rest of the union will be whatever the other input is.
9302 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9303 * otherwise it gets decremented to 0. This is consistent with 'count'
9304 * == 0 meaning the remainder of the union is whatever is left in the
9305 * non-exhausted list. */
9310 IV copy_count = len_a - i_a;
9311 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9312 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9314 else { /* The non-exhausted input is b */
9315 copy_count = len_b - i_b;
9316 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9318 len_u = i_u + copy_count;
9321 /* Set the result to the final length, which can change the pointer to
9322 * array_u, so re-find it. (Note that it is unlikely that this will
9323 * change, as we are shrinking the space, not enlarging it) */
9324 if (len_u != _invlist_len(u)) {
9325 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9327 array_u = invlist_array(u);
9330 if (*output == NULL) { /* Simply return the new inversion list */
9334 /* Otherwise, overwrite the inversion list that was in '*output'. We
9335 * could instead free '*output', and then set it to 'u', but experience
9336 * has shown [perl #127392] that if the input is a mortal, we can get a
9337 * huge build-up of these during regex compilation before they get
9339 invlist_replace_list_destroys_src(*output, u);
9347 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9348 const bool complement_b, SV** i)
9350 /* Take the intersection of two inversion lists and point '*i' to it. On
9351 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9352 * even 'a' or 'b'). If to an inversion list, the contents of the original
9353 * list will be replaced by the intersection. The first list, 'a', may be
9354 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9355 * TRUE, the result will be the intersection of 'a' and the complement (or
9356 * inversion) of 'b' instead of 'b' directly.
9358 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9359 * Richard Gillam, published by Addison-Wesley, and explained at some
9360 * length there. The preface says to incorporate its examples into your
9361 * code at your own risk. In fact, it had bugs
9363 * The algorithm is like a merge sort, and is essentially the same as the
9367 const UV* array_a; /* a's array */
9369 UV len_a; /* length of a's array */
9372 SV* r; /* the resulting intersection */
9376 UV i_a = 0; /* current index into a's array */
9380 /* running count of how many of the two inputs are postitioned at ranges
9381 * that are in their sets. As explained in the algorithm source book,
9382 * items are stopped accumulating and are output when the count changes
9383 * to/from 2. The count is incremented when we start a range that's in an
9384 * input's set, and decremented when we start a range that's not in a set.
9385 * Only when it is 2 are we in the intersection. */
9388 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9390 assert(*i == NULL || SvTYPE(*i) == SVt_INVLIST);
9392 /* Special case if either one is empty */
9393 len_a = (a == NULL) ? 0 : _invlist_len(a);
9394 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9395 if (len_a != 0 && complement_b) {
9397 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9398 * must be empty. Here, also we are using 'b's complement, which
9399 * hence must be every possible code point. Thus the intersection
9402 if (*i == a) { /* No-op */
9407 *i = invlist_clone(a);
9411 r = invlist_clone(a);
9412 invlist_replace_list_destroys_src(*i, r);
9417 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9418 * intersection must be empty */
9420 *i = _new_invlist(0);
9428 /* Here both lists exist and are non-empty */
9429 array_a = invlist_array(a);
9430 array_b = invlist_array(b);
9432 /* If are to take the intersection of 'a' with the complement of b, set it
9433 * up so are looking at b's complement. */
9436 /* To complement, we invert: if the first element is 0, remove it. To
9437 * do this, we just pretend the array starts one later */
9438 if (array_b[0] == 0) {
9444 /* But if the first element is not zero, we pretend the list starts
9445 * at the 0 that is always stored immediately before the array. */
9451 /* Size the intersection for the worst case: that the intersection ends up
9452 * fragmenting everything to be completely disjoint */
9453 r= _new_invlist(len_a + len_b);
9455 /* Will contain U+0000 iff both components do */
9456 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9457 && len_b > 0 && array_b[0] == 0);
9459 /* Go through each list item by item, stopping when have exhausted one of
9461 while (i_a < len_a && i_b < len_b) {
9462 UV cp; /* The element to potentially add to the intersection's
9464 bool cp_in_set; /* Is it in the input list's set or not */
9466 /* We need to take one or the other of the two inputs for the
9467 * intersection. Since we are merging two sorted lists, we take the
9468 * smaller of the next items. In case of a tie, we take first the one
9469 * that is not in its set (a difference from the union algorithm). If
9470 * we first took the one in its set, it would increment the count,
9471 * possibly to 2 which would cause it to be output as starting a range
9472 * in the intersection, and the next time through we would take that
9473 * same number, and output it again as ending the set. By doing the
9474 * opposite of this, there is no possibility that the count will be
9475 * momentarily incremented to 2. (In a tie and both are in the set or
9476 * both not in the set, it doesn't matter which we take first.) */
9477 if ( array_a[i_a] < array_b[i_b]
9478 || ( array_a[i_a] == array_b[i_b]
9479 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9481 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9482 cp = array_a[i_a++];
9485 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9489 /* Here, have chosen which of the two inputs to look at. Only output
9490 * if the running count changes to/from 2, which marks the
9491 * beginning/end of a range that's in the intersection */
9495 array_r[i_r++] = cp;
9500 array_r[i_r++] = cp;
9507 /* The loop above increments the index into exactly one of the input lists
9508 * each iteration, and ends when either index gets to its list end. That
9509 * means the other index is lower than its end, and so something is
9510 * remaining in that one. We increment 'count', as explained below, if the
9511 * exhausted list was in its set. (i_a and i_b each currently index the
9512 * element beyond the one we care about.) */
9513 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9514 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9519 /* Above we incremented 'count' if the exhausted list was in its set. This
9520 * has made it so that 'count' being below 2 means there is nothing left to
9521 * output; otheriwse what's left to add to the intersection is precisely
9522 * that which is left in the non-exhausted input list.
9524 * To see why, note first that the exhausted input obviously has nothing
9525 * left to affect the intersection. If it was in its set at its end, that
9526 * means the set extends from here to the platform's infinity, and hence
9527 * anything in the non-exhausted's list will be in the intersection, and
9528 * anything not in it won't be. Hence, the rest of the intersection is
9529 * precisely what's in the non-exhausted list The exhausted set also
9530 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9531 * it means 'count' is now at least 2. This is consistent with the
9532 * incremented 'count' being >= 2 means to add the non-exhausted list to
9535 * But if the exhausted input wasn't in its set, it contributed 0 to
9536 * 'count', and the intersection can't include anything further; the
9537 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9538 * incremented. This is consistent with 'count' being < 2 meaning nothing
9539 * further to add to the intersection. */
9540 if (count < 2) { /* Nothing left to put in the intersection. */
9543 else { /* copy the non-exhausted list, unchanged. */
9544 IV copy_count = len_a - i_a;
9545 if (copy_count > 0) { /* a is the one with stuff left */
9546 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9548 else { /* b is the one with stuff left */
9549 copy_count = len_b - i_b;
9550 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9552 len_r = i_r + copy_count;
9555 /* Set the result to the final length, which can change the pointer to
9556 * array_r, so re-find it. (Note that it is unlikely that this will
9557 * change, as we are shrinking the space, not enlarging it) */
9558 if (len_r != _invlist_len(r)) {
9559 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9561 array_r = invlist_array(r);
9564 if (*i == NULL) { /* Simply return the calculated intersection */
9567 else { /* Otherwise, replace the existing inversion list in '*i'. We could
9568 instead free '*i', and then set it to 'r', but experience has
9569 shown [perl #127392] that if the input is a mortal, we can get a
9570 huge build-up of these during regex compilation before they get
9573 invlist_replace_list_destroys_src(*i, r);
9585 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9587 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9588 * set. A pointer to the inversion list is returned. This may actually be
9589 * a new list, in which case the passed in one has been destroyed. The
9590 * passed-in inversion list can be NULL, in which case a new one is created
9591 * with just the one range in it. The new list is not necessarily
9592 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9593 * result of this function. The gain would not be large, and in many
9594 * cases, this is called multiple times on a single inversion list, so
9595 * anything freed may almost immediately be needed again.
9597 * This used to mostly call the 'union' routine, but that is much more
9598 * heavyweight than really needed for a single range addition */
9600 UV* array; /* The array implementing the inversion list */
9601 UV len; /* How many elements in 'array' */
9602 SSize_t i_s; /* index into the invlist array where 'start'
9604 SSize_t i_e = 0; /* And the index where 'end' should go */
9605 UV cur_highest; /* The highest code point in the inversion list
9606 upon entry to this function */
9608 /* This range becomes the whole inversion list if none already existed */
9609 if (invlist == NULL) {
9610 invlist = _new_invlist(2);
9611 _append_range_to_invlist(invlist, start, end);
9615 /* Likewise, if the inversion list is currently empty */
9616 len = _invlist_len(invlist);
9618 _append_range_to_invlist(invlist, start, end);
9622 /* Starting here, we have to know the internals of the list */
9623 array = invlist_array(invlist);
9625 /* If the new range ends higher than the current highest ... */
9626 cur_highest = invlist_highest(invlist);
9627 if (end > cur_highest) {
9629 /* If the whole range is higher, we can just append it */
9630 if (start > cur_highest) {
9631 _append_range_to_invlist(invlist, start, end);
9635 /* Otherwise, add the portion that is higher ... */
9636 _append_range_to_invlist(invlist, cur_highest + 1, end);
9638 /* ... and continue on below to handle the rest. As a result of the
9639 * above append, we know that the index of the end of the range is the
9640 * final even numbered one of the array. Recall that the final element
9641 * always starts a range that extends to infinity. If that range is in
9642 * the set (meaning the set goes from here to infinity), it will be an
9643 * even index, but if it isn't in the set, it's odd, and the final
9644 * range in the set is one less, which is even. */
9645 if (end == UV_MAX) {
9653 /* We have dealt with appending, now see about prepending. If the new
9654 * range starts lower than the current lowest ... */
9655 if (start < array[0]) {
9657 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
9658 * Let the union code handle it, rather than having to know the
9659 * trickiness in two code places. */
9660 if (UNLIKELY(start == 0)) {
9663 range_invlist = _new_invlist(2);
9664 _append_range_to_invlist(range_invlist, start, end);
9666 _invlist_union(invlist, range_invlist, &invlist);
9668 SvREFCNT_dec_NN(range_invlist);
9673 /* If the whole new range comes before the first entry, and doesn't
9674 * extend it, we have to insert it as an additional range */
9675 if (end < array[0] - 1) {
9677 goto splice_in_new_range;
9680 /* Here the new range adjoins the existing first range, extending it
9684 /* And continue on below to handle the rest. We know that the index of
9685 * the beginning of the range is the first one of the array */
9688 else { /* Not prepending any part of the new range to the existing list.
9689 * Find where in the list it should go. This finds i_s, such that:
9690 * invlist[i_s] <= start < array[i_s+1]
9692 i_s = _invlist_search(invlist, start);
9695 /* At this point, any extending before the beginning of the inversion list
9696 * and/or after the end has been done. This has made it so that, in the
9697 * code below, each endpoint of the new range is either in a range that is
9698 * in the set, or is in a gap between two ranges that are. This means we
9699 * don't have to worry about exceeding the array bounds.
9701 * Find where in the list the new range ends (but we can skip this if we
9702 * have already determined what it is, or if it will be the same as i_s,
9703 * which we already have computed) */
9705 i_e = (start == end)
9707 : _invlist_search(invlist, end);
9710 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
9711 * is a range that goes to infinity there is no element at invlist[i_e+1],
9712 * so only the first relation holds. */
9714 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9716 /* Here, the ranges on either side of the beginning of the new range
9717 * are in the set, and this range starts in the gap between them.
9719 * The new range extends the range above it downwards if the new range
9720 * ends at or above that range's start */
9721 const bool extends_the_range_above = ( end == UV_MAX
9722 || end + 1 >= array[i_s+1]);
9724 /* The new range extends the range below it upwards if it begins just
9725 * after where that range ends */
9726 if (start == array[i_s]) {
9728 /* If the new range fills the entire gap between the other ranges,
9729 * they will get merged together. Other ranges may also get
9730 * merged, depending on how many of them the new range spans. In
9731 * the general case, we do the merge later, just once, after we
9732 * figure out how many to merge. But in the case where the new
9733 * range exactly spans just this one gap (possibly extending into
9734 * the one above), we do the merge here, and an early exit. This
9735 * is done here to avoid having to special case later. */
9736 if (i_e - i_s <= 1) {
9738 /* If i_e - i_s == 1, it means that the new range terminates
9739 * within the range above, and hence 'extends_the_range_above'
9740 * must be true. (If the range above it extends to infinity,
9741 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
9742 * will be 0, so no harm done.) */
9743 if (extends_the_range_above) {
9744 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
9745 invlist_set_len(invlist,
9747 *(get_invlist_offset_addr(invlist)));
9751 /* Here, i_e must == i_s. We keep them in sync, as they apply
9752 * to the same range, and below we are about to decrement i_s
9757 /* Here, the new range is adjacent to the one below. (It may also
9758 * span beyond the range above, but that will get resolved later.)
9759 * Extend the range below to include this one. */
9760 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
9764 else if (extends_the_range_above) {
9766 /* Here the new range only extends the range above it, but not the
9767 * one below. It merges with the one above. Again, we keep i_e
9768 * and i_s in sync if they point to the same range */
9777 /* Here, we've dealt with the new range start extending any adjoining
9780 * If the new range extends to infinity, it is now the final one,
9781 * regardless of what was there before */
9782 if (UNLIKELY(end == UV_MAX)) {
9783 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
9787 /* If i_e started as == i_s, it has also been dealt with,
9788 * and been updated to the new i_s, which will fail the following if */
9789 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
9791 /* Here, the ranges on either side of the end of the new range are in
9792 * the set, and this range ends in the gap between them.
9794 * If this range is adjacent to (hence extends) the range above it, it
9795 * becomes part of that range; likewise if it extends the range below,
9796 * it becomes part of that range */
9797 if (end + 1 == array[i_e+1]) {
9801 else if (start <= array[i_e]) {
9802 array[i_e] = end + 1;
9809 /* If the range fits entirely in an existing range (as possibly already
9810 * extended above), it doesn't add anything new */
9811 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9815 /* Here, no part of the range is in the list. Must add it. It will
9816 * occupy 2 more slots */
9817 splice_in_new_range:
9819 invlist_extend(invlist, len + 2);
9820 array = invlist_array(invlist);
9821 /* Move the rest of the array down two slots. Don't include any
9823 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
9825 /* Do the actual splice */
9826 array[i_e+1] = start;
9827 array[i_e+2] = end + 1;
9828 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
9832 /* Here the new range crossed the boundaries of a pre-existing range. The
9833 * code above has adjusted things so that both ends are in ranges that are
9834 * in the set. This means everything in between must also be in the set.
9835 * Just squash things together */
9836 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
9837 invlist_set_len(invlist,
9839 *(get_invlist_offset_addr(invlist)));
9845 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9846 UV** other_elements_ptr)
9848 /* Create and return an inversion list whose contents are to be populated
9849 * by the caller. The caller gives the number of elements (in 'size') and
9850 * the very first element ('element0'). This function will set
9851 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9854 * Obviously there is some trust involved that the caller will properly
9855 * fill in the other elements of the array.
9857 * (The first element needs to be passed in, as the underlying code does
9858 * things differently depending on whether it is zero or non-zero) */
9860 SV* invlist = _new_invlist(size);
9863 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9865 invlist = add_cp_to_invlist(invlist, element0);
9866 offset = *get_invlist_offset_addr(invlist);
9868 invlist_set_len(invlist, size, offset);
9869 *other_elements_ptr = invlist_array(invlist) + 1;
9875 PERL_STATIC_INLINE SV*
9876 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9877 return _add_range_to_invlist(invlist, cp, cp);
9880 #ifndef PERL_IN_XSUB_RE
9882 Perl__invlist_invert(pTHX_ SV* const invlist)
9884 /* Complement the input inversion list. This adds a 0 if the list didn't
9885 * have a zero; removes it otherwise. As described above, the data
9886 * structure is set up so that this is very efficient */
9888 PERL_ARGS_ASSERT__INVLIST_INVERT;
9890 assert(! invlist_is_iterating(invlist));
9892 /* The inverse of matching nothing is matching everything */
9893 if (_invlist_len(invlist) == 0) {
9894 _append_range_to_invlist(invlist, 0, UV_MAX);
9898 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9903 PERL_STATIC_INLINE SV*
9904 S_invlist_clone(pTHX_ SV* const invlist)
9907 /* Return a new inversion list that is a copy of the input one, which is
9908 * unchanged. The new list will not be mortal even if the old one was. */
9910 /* Need to allocate extra space to accommodate Perl's addition of a
9911 * trailing NUL to SvPV's, since it thinks they are always strings */
9912 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9913 STRLEN physical_length = SvCUR(invlist);
9914 bool offset = *(get_invlist_offset_addr(invlist));
9916 PERL_ARGS_ASSERT_INVLIST_CLONE;
9918 *(get_invlist_offset_addr(new_invlist)) = offset;
9919 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9920 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9925 PERL_STATIC_INLINE STRLEN*
9926 S_get_invlist_iter_addr(SV* invlist)
9928 /* Return the address of the UV that contains the current iteration
9931 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9933 assert(SvTYPE(invlist) == SVt_INVLIST);
9935 return &(((XINVLIST*) SvANY(invlist))->iterator);
9938 PERL_STATIC_INLINE void
9939 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9941 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9943 *get_invlist_iter_addr(invlist) = 0;
9946 PERL_STATIC_INLINE void
9947 S_invlist_iterfinish(SV* invlist)
9949 /* Terminate iterator for invlist. This is to catch development errors.
9950 * Any iteration that is interrupted before completed should call this
9951 * function. Functions that add code points anywhere else but to the end
9952 * of an inversion list assert that they are not in the middle of an
9953 * iteration. If they were, the addition would make the iteration
9954 * problematical: if the iteration hadn't reached the place where things
9955 * were being added, it would be ok */
9957 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
9959 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
9963 S_invlist_iternext(SV* invlist, UV* start, UV* end)
9965 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
9966 * This call sets in <*start> and <*end>, the next range in <invlist>.
9967 * Returns <TRUE> if successful and the next call will return the next
9968 * range; <FALSE> if was already at the end of the list. If the latter,
9969 * <*start> and <*end> are unchanged, and the next call to this function
9970 * will start over at the beginning of the list */
9972 STRLEN* pos = get_invlist_iter_addr(invlist);
9973 UV len = _invlist_len(invlist);
9976 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9979 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9983 array = invlist_array(invlist);
9985 *start = array[(*pos)++];
9991 *end = array[(*pos)++] - 1;
9997 PERL_STATIC_INLINE UV
9998 S_invlist_highest(SV* const invlist)
10000 /* Returns the highest code point that matches an inversion list. This API
10001 * has an ambiguity, as it returns 0 under either the highest is actually
10002 * 0, or if the list is empty. If this distinction matters to you, check
10003 * for emptiness before calling this function */
10005 UV len = _invlist_len(invlist);
10008 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
10014 array = invlist_array(invlist);
10016 /* The last element in the array in the inversion list always starts a
10017 * range that goes to infinity. That range may be for code points that are
10018 * matched in the inversion list, or it may be for ones that aren't
10019 * matched. In the latter case, the highest code point in the set is one
10020 * less than the beginning of this range; otherwise it is the final element
10021 * of this range: infinity */
10022 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
10024 : array[len - 1] - 1;
10028 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
10030 /* Get the contents of an inversion list into a string SV so that they can
10031 * be printed out. If 'traditional_style' is TRUE, it uses the format
10032 * traditionally done for debug tracing; otherwise it uses a format
10033 * suitable for just copying to the output, with blanks between ranges and
10034 * a dash between range components */
10038 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
10039 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
10041 if (traditional_style) {
10042 output = newSVpvs("\n");
10045 output = newSVpvs("");
10048 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
10050 assert(! invlist_is_iterating(invlist));
10052 invlist_iterinit(invlist);
10053 while (invlist_iternext(invlist, &start, &end)) {
10054 if (end == UV_MAX) {
10055 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFINITY%c",
10056 start, intra_range_delimiter,
10057 inter_range_delimiter);
10059 else if (end != start) {
10060 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
10062 intra_range_delimiter,
10063 end, inter_range_delimiter);
10066 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10067 start, inter_range_delimiter);
10071 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10072 SvCUR_set(output, SvCUR(output) - 1);
10078 #ifndef PERL_IN_XSUB_RE
10080 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10081 const char * const indent, SV* const invlist)
10083 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10084 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10085 * the string 'indent'. The output looks like this:
10086 [0] 0x000A .. 0x000D
10088 [4] 0x2028 .. 0x2029
10089 [6] 0x3104 .. INFINITY
10090 * This means that the first range of code points matched by the list are
10091 * 0xA through 0xD; the second range contains only the single code point
10092 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10093 * are used to define each range (except if the final range extends to
10094 * infinity, only a single element is needed). The array index of the
10095 * first element for the corresponding range is given in brackets. */
10100 PERL_ARGS_ASSERT__INVLIST_DUMP;
10102 if (invlist_is_iterating(invlist)) {
10103 Perl_dump_indent(aTHX_ level, file,
10104 "%sCan't dump inversion list because is in middle of iterating\n",
10109 invlist_iterinit(invlist);
10110 while (invlist_iternext(invlist, &start, &end)) {
10111 if (end == UV_MAX) {
10112 Perl_dump_indent(aTHX_ level, file,
10113 "%s[%" UVuf "] 0x%04" UVXf " .. INFINITY\n",
10114 indent, (UV)count, start);
10116 else if (end != start) {
10117 Perl_dump_indent(aTHX_ level, file,
10118 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10119 indent, (UV)count, start, end);
10122 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10123 indent, (UV)count, start);
10130 Perl__load_PL_utf8_foldclosures (pTHX)
10132 assert(! PL_utf8_foldclosures);
10134 /* If the folds haven't been read in, call a fold function
10136 if (! PL_utf8_tofold) {
10137 U8 dummy[UTF8_MAXBYTES_CASE+1];
10138 const U8 hyphen[] = HYPHEN_UTF8;
10140 /* This string is just a short named one above \xff */
10141 toFOLD_utf8_safe(hyphen, hyphen + sizeof(hyphen) - 1, dummy, NULL);
10142 assert(PL_utf8_tofold); /* Verify that worked */
10144 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
10148 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10150 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10152 /* Return a boolean as to if the two passed in inversion lists are
10153 * identical. The final argument, if TRUE, says to take the complement of
10154 * the second inversion list before doing the comparison */
10156 const UV* array_a = invlist_array(a);
10157 const UV* array_b = invlist_array(b);
10158 UV len_a = _invlist_len(a);
10159 UV len_b = _invlist_len(b);
10161 PERL_ARGS_ASSERT__INVLISTEQ;
10163 /* If are to compare 'a' with the complement of b, set it
10164 * up so are looking at b's complement. */
10165 if (complement_b) {
10167 /* The complement of nothing is everything, so <a> would have to have
10168 * just one element, starting at zero (ending at infinity) */
10170 return (len_a == 1 && array_a[0] == 0);
10172 else if (array_b[0] == 0) {
10174 /* Otherwise, to complement, we invert. Here, the first element is
10175 * 0, just remove it. To do this, we just pretend the array starts
10183 /* But if the first element is not zero, we pretend the list starts
10184 * at the 0 that is always stored immediately before the array. */
10190 return len_a == len_b
10191 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10197 * As best we can, determine the characters that can match the start of
10198 * the given EXACTF-ish node.
10200 * Returns the invlist as a new SV*; it is the caller's responsibility to
10201 * call SvREFCNT_dec() when done with it.
10204 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10206 const U8 * s = (U8*)STRING(node);
10207 SSize_t bytelen = STR_LEN(node);
10209 /* Start out big enough for 2 separate code points */
10210 SV* invlist = _new_invlist(4);
10212 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
10217 /* We punt and assume can match anything if the node begins
10218 * with a multi-character fold. Things are complicated. For
10219 * example, /ffi/i could match any of:
10220 * "\N{LATIN SMALL LIGATURE FFI}"
10221 * "\N{LATIN SMALL LIGATURE FF}I"
10222 * "F\N{LATIN SMALL LIGATURE FI}"
10223 * plus several other things; and making sure we have all the
10224 * possibilities is hard. */
10225 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10226 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10229 /* Any Latin1 range character can potentially match any
10230 * other depending on the locale */
10231 if (OP(node) == EXACTFL) {
10232 _invlist_union(invlist, PL_Latin1, &invlist);
10235 /* But otherwise, it matches at least itself. We can
10236 * quickly tell if it has a distinct fold, and if so,
10237 * it matches that as well */
10238 invlist = add_cp_to_invlist(invlist, uc);
10239 if (IS_IN_SOME_FOLD_L1(uc))
10240 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10243 /* Some characters match above-Latin1 ones under /i. This
10244 * is true of EXACTFL ones when the locale is UTF-8 */
10245 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10246 && (! isASCII(uc) || (OP(node) != EXACTFA
10247 && OP(node) != EXACTFA_NO_TRIE)))
10249 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10253 else { /* Pattern is UTF-8 */
10254 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10255 STRLEN foldlen = UTF8SKIP(s);
10256 const U8* e = s + bytelen;
10259 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10261 /* The only code points that aren't folded in a UTF EXACTFish
10262 * node are are the problematic ones in EXACTFL nodes */
10263 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10264 /* We need to check for the possibility that this EXACTFL
10265 * node begins with a multi-char fold. Therefore we fold
10266 * the first few characters of it so that we can make that
10271 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10273 *(d++) = (U8) toFOLD(*s);
10278 toFOLD_utf8_safe(s, e, d, &len);
10284 /* And set up so the code below that looks in this folded
10285 * buffer instead of the node's string */
10287 foldlen = UTF8SKIP(folded);
10291 /* When we reach here 's' points to the fold of the first
10292 * character(s) of the node; and 'e' points to far enough along
10293 * the folded string to be just past any possible multi-char
10294 * fold. 'foldlen' is the length in bytes of the first
10297 * Unlike the non-UTF-8 case, the macro for determining if a
10298 * string is a multi-char fold requires all the characters to
10299 * already be folded. This is because of all the complications
10300 * if not. Note that they are folded anyway, except in EXACTFL
10301 * nodes. Like the non-UTF case above, we punt if the node
10302 * begins with a multi-char fold */
10304 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10305 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10307 else { /* Single char fold */
10309 /* It matches all the things that fold to it, which are
10310 * found in PL_utf8_foldclosures (including itself) */
10311 invlist = add_cp_to_invlist(invlist, uc);
10312 if (! PL_utf8_foldclosures)
10313 _load_PL_utf8_foldclosures();
10314 if ((listp = hv_fetch(PL_utf8_foldclosures,
10315 (char *) s, foldlen, FALSE)))
10317 AV* list = (AV*) *listp;
10319 for (k = 0; k <= av_tindex_skip_len_mg(list); k++) {
10320 SV** c_p = av_fetch(list, k, FALSE);
10326 /* /aa doesn't allow folds between ASCII and non- */
10327 if ((OP(node) == EXACTFA || OP(node) == EXACTFA_NO_TRIE)
10328 && isASCII(c) != isASCII(uc))
10333 invlist = add_cp_to_invlist(invlist, c);
10342 #undef HEADER_LENGTH
10343 #undef TO_INTERNAL_SIZE
10344 #undef FROM_INTERNAL_SIZE
10345 #undef INVLIST_VERSION_ID
10347 /* End of inversion list object */
10350 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10352 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10353 * constructs, and updates RExC_flags with them. On input, RExC_parse
10354 * should point to the first flag; it is updated on output to point to the
10355 * final ')' or ':'. There needs to be at least one flag, or this will
10358 /* for (?g), (?gc), and (?o) warnings; warning
10359 about (?c) will warn about (?g) -- japhy */
10361 #define WASTED_O 0x01
10362 #define WASTED_G 0x02
10363 #define WASTED_C 0x04
10364 #define WASTED_GC (WASTED_G|WASTED_C)
10365 I32 wastedflags = 0x00;
10366 U32 posflags = 0, negflags = 0;
10367 U32 *flagsp = &posflags;
10368 char has_charset_modifier = '\0';
10370 bool has_use_defaults = FALSE;
10371 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10372 int x_mod_count = 0;
10374 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10376 /* '^' as an initial flag sets certain defaults */
10377 if (UCHARAT(RExC_parse) == '^') {
10379 has_use_defaults = TRUE;
10380 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10381 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
10382 ? REGEX_UNICODE_CHARSET
10383 : REGEX_DEPENDS_CHARSET);
10386 cs = get_regex_charset(RExC_flags);
10387 if (cs == REGEX_DEPENDS_CHARSET
10388 && (RExC_utf8 || RExC_uni_semantics))
10390 cs = REGEX_UNICODE_CHARSET;
10393 while (RExC_parse < RExC_end) {
10394 /* && strchr("iogcmsx", *RExC_parse) */
10395 /* (?g), (?gc) and (?o) are useless here
10396 and must be globally applied -- japhy */
10397 switch (*RExC_parse) {
10399 /* Code for the imsxn flags */
10400 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10402 case LOCALE_PAT_MOD:
10403 if (has_charset_modifier) {
10404 goto excess_modifier;
10406 else if (flagsp == &negflags) {
10409 cs = REGEX_LOCALE_CHARSET;
10410 has_charset_modifier = LOCALE_PAT_MOD;
10412 case UNICODE_PAT_MOD:
10413 if (has_charset_modifier) {
10414 goto excess_modifier;
10416 else if (flagsp == &negflags) {
10419 cs = REGEX_UNICODE_CHARSET;
10420 has_charset_modifier = UNICODE_PAT_MOD;
10422 case ASCII_RESTRICT_PAT_MOD:
10423 if (flagsp == &negflags) {
10426 if (has_charset_modifier) {
10427 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10428 goto excess_modifier;
10430 /* Doubled modifier implies more restricted */
10431 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10434 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10436 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10438 case DEPENDS_PAT_MOD:
10439 if (has_use_defaults) {
10440 goto fail_modifiers;
10442 else if (flagsp == &negflags) {
10445 else if (has_charset_modifier) {
10446 goto excess_modifier;
10449 /* The dual charset means unicode semantics if the
10450 * pattern (or target, not known until runtime) are
10451 * utf8, or something in the pattern indicates unicode
10453 cs = (RExC_utf8 || RExC_uni_semantics)
10454 ? REGEX_UNICODE_CHARSET
10455 : REGEX_DEPENDS_CHARSET;
10456 has_charset_modifier = DEPENDS_PAT_MOD;
10460 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10461 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10463 else if (has_charset_modifier == *(RExC_parse - 1)) {
10464 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10465 *(RExC_parse - 1));
10468 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10470 NOT_REACHED; /*NOTREACHED*/
10473 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10474 *(RExC_parse - 1));
10475 NOT_REACHED; /*NOTREACHED*/
10476 case ONCE_PAT_MOD: /* 'o' */
10477 case GLOBAL_PAT_MOD: /* 'g' */
10478 if (PASS2 && ckWARN(WARN_REGEXP)) {
10479 const I32 wflagbit = *RExC_parse == 'o'
10482 if (! (wastedflags & wflagbit) ) {
10483 wastedflags |= wflagbit;
10484 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10487 "Useless (%s%c) - %suse /%c modifier",
10488 flagsp == &negflags ? "?-" : "?",
10490 flagsp == &negflags ? "don't " : "",
10497 case CONTINUE_PAT_MOD: /* 'c' */
10498 if (PASS2 && ckWARN(WARN_REGEXP)) {
10499 if (! (wastedflags & WASTED_C) ) {
10500 wastedflags |= WASTED_GC;
10501 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10504 "Useless (%sc) - %suse /gc modifier",
10505 flagsp == &negflags ? "?-" : "?",
10506 flagsp == &negflags ? "don't " : ""
10511 case KEEPCOPY_PAT_MOD: /* 'p' */
10512 if (flagsp == &negflags) {
10514 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10516 *flagsp |= RXf_PMf_KEEPCOPY;
10520 /* A flag is a default iff it is following a minus, so
10521 * if there is a minus, it means will be trying to
10522 * re-specify a default which is an error */
10523 if (has_use_defaults || flagsp == &negflags) {
10524 goto fail_modifiers;
10526 flagsp = &negflags;
10527 wastedflags = 0; /* reset so (?g-c) warns twice */
10533 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
10534 negflags |= RXf_PMf_EXTENDED_MORE;
10536 RExC_flags |= posflags;
10538 if (negflags & RXf_PMf_EXTENDED) {
10539 negflags |= RXf_PMf_EXTENDED_MORE;
10541 RExC_flags &= ~negflags;
10542 set_regex_charset(&RExC_flags, cs);
10547 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10548 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10549 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
10550 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10551 NOT_REACHED; /*NOTREACHED*/
10554 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10557 vFAIL("Sequence (?... not terminated");
10561 - reg - regular expression, i.e. main body or parenthesized thing
10563 * Caller must absorb opening parenthesis.
10565 * Combining parenthesis handling with the base level of regular expression
10566 * is a trifle forced, but the need to tie the tails of the branches to what
10567 * follows makes it hard to avoid.
10569 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10571 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10573 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10576 PERL_STATIC_INLINE regnode *
10577 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10579 char * parse_start,
10584 char* name_start = RExC_parse;
10586 SV *sv_dat = reg_scan_name(pRExC_state, SIZE_ONLY
10587 ? REG_RSN_RETURN_NULL
10588 : REG_RSN_RETURN_DATA);
10589 GET_RE_DEBUG_FLAGS_DECL;
10591 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10593 if (RExC_parse == name_start || *RExC_parse != ch) {
10594 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10595 vFAIL2("Sequence %.3s... not terminated",parse_start);
10599 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10600 RExC_rxi->data->data[num]=(void*)sv_dat;
10601 SvREFCNT_inc_simple_void(sv_dat);
10604 ret = reganode(pRExC_state,
10607 : (ASCII_FOLD_RESTRICTED)
10609 : (AT_LEAST_UNI_SEMANTICS)
10615 *flagp |= HASWIDTH;
10617 Set_Node_Offset(ret, parse_start+1);
10618 Set_Node_Cur_Length(ret, parse_start);
10620 nextchar(pRExC_state);
10624 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
10625 flags. Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan
10626 needs to be restarted, or'd with NEED_UTF8 if the pattern needs to be
10627 upgraded to UTF-8. Otherwise would only return NULL if regbranch() returns
10628 NULL, which cannot happen. */
10630 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
10631 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
10632 * 2 is like 1, but indicates that nextchar() has been called to advance
10633 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
10634 * this flag alerts us to the need to check for that */
10636 regnode *ret; /* Will be the head of the group. */
10639 regnode *ender = NULL;
10642 U32 oregflags = RExC_flags;
10643 bool have_branch = 0;
10645 I32 freeze_paren = 0;
10646 I32 after_freeze = 0;
10647 I32 num; /* numeric backreferences */
10649 char * parse_start = RExC_parse; /* MJD */
10650 char * const oregcomp_parse = RExC_parse;
10652 GET_RE_DEBUG_FLAGS_DECL;
10654 PERL_ARGS_ASSERT_REG;
10655 DEBUG_PARSE("reg ");
10657 *flagp = 0; /* Tentatively. */
10659 /* Having this true makes it feasible to have a lot fewer tests for the
10660 * parse pointer being in scope. For example, we can write
10661 * while(isFOO(*RExC_parse)) RExC_parse++;
10663 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
10665 assert(*RExC_end == '\0');
10667 /* Make an OPEN node, if parenthesized. */
10670 /* Under /x, space and comments can be gobbled up between the '(' and
10671 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
10672 * intervening space, as the sequence is a token, and a token should be
10674 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
10676 if (RExC_parse >= RExC_end) {
10677 vFAIL("Unmatched (");
10680 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
10681 char *start_verb = RExC_parse + 1;
10683 char *start_arg = NULL;
10684 unsigned char op = 0;
10685 int arg_required = 0;
10686 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
10688 if (has_intervening_patws) {
10689 RExC_parse++; /* past the '*' */
10690 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
10692 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
10693 if ( *RExC_parse == ':' ) {
10694 start_arg = RExC_parse + 1;
10697 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10699 verb_len = RExC_parse - start_verb;
10701 if (RExC_parse >= RExC_end) {
10702 goto unterminated_verb_pattern;
10704 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10705 while ( RExC_parse < RExC_end && *RExC_parse != ')' )
10706 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10707 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10708 unterminated_verb_pattern:
10709 vFAIL("Unterminated verb pattern argument");
10710 if ( RExC_parse == start_arg )
10713 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10714 vFAIL("Unterminated verb pattern");
10717 /* Here, we know that RExC_parse < RExC_end */
10719 switch ( *start_verb ) {
10720 case 'A': /* (*ACCEPT) */
10721 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
10723 internal_argval = RExC_nestroot;
10726 case 'C': /* (*COMMIT) */
10727 if ( memEQs(start_verb,verb_len,"COMMIT") )
10730 case 'F': /* (*FAIL) */
10731 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
10735 case ':': /* (*:NAME) */
10736 case 'M': /* (*MARK:NAME) */
10737 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
10742 case 'P': /* (*PRUNE) */
10743 if ( memEQs(start_verb,verb_len,"PRUNE") )
10746 case 'S': /* (*SKIP) */
10747 if ( memEQs(start_verb,verb_len,"SKIP") )
10750 case 'T': /* (*THEN) */
10751 /* [19:06] <TimToady> :: is then */
10752 if ( memEQs(start_verb,verb_len,"THEN") ) {
10754 RExC_seen |= REG_CUTGROUP_SEEN;
10759 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10761 "Unknown verb pattern '%" UTF8f "'",
10762 UTF8fARG(UTF, verb_len, start_verb));
10764 if ( arg_required && !start_arg ) {
10765 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
10766 verb_len, start_verb);
10768 if (internal_argval == -1) {
10769 ret = reganode(pRExC_state, op, 0);
10771 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
10773 RExC_seen |= REG_VERBARG_SEEN;
10774 if ( ! SIZE_ONLY ) {
10776 SV *sv = newSVpvn( start_arg,
10777 RExC_parse - start_arg);
10778 ARG(ret) = add_data( pRExC_state,
10779 STR_WITH_LEN("S"));
10780 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
10785 if ( internal_argval != -1 )
10786 ARG2L_SET(ret, internal_argval);
10788 nextchar(pRExC_state);
10791 else if (*RExC_parse == '?') { /* (?...) */
10792 bool is_logical = 0;
10793 const char * const seqstart = RExC_parse;
10794 const char * endptr;
10795 if (has_intervening_patws) {
10797 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
10800 RExC_parse++; /* past the '?' */
10801 paren = *RExC_parse; /* might be a trailing NUL, if not
10803 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10804 if (RExC_parse > RExC_end) {
10807 ret = NULL; /* For look-ahead/behind. */
10810 case 'P': /* (?P...) variants for those used to PCRE/Python */
10811 paren = *RExC_parse;
10812 if ( paren == '<') { /* (?P<...>) named capture */
10814 if (RExC_parse >= RExC_end) {
10815 vFAIL("Sequence (?P<... not terminated");
10817 goto named_capture;
10819 else if (paren == '>') { /* (?P>name) named recursion */
10821 if (RExC_parse >= RExC_end) {
10822 vFAIL("Sequence (?P>... not terminated");
10824 goto named_recursion;
10826 else if (paren == '=') { /* (?P=...) named backref */
10828 return handle_named_backref(pRExC_state, flagp,
10831 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10832 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10833 vFAIL3("Sequence (%.*s...) not recognized",
10834 RExC_parse-seqstart, seqstart);
10835 NOT_REACHED; /*NOTREACHED*/
10836 case '<': /* (?<...) */
10837 if (*RExC_parse == '!')
10839 else if (*RExC_parse != '=')
10846 case '\'': /* (?'...') */
10847 name_start = RExC_parse;
10848 svname = reg_scan_name(pRExC_state,
10849 SIZE_ONLY /* reverse test from the others */
10850 ? REG_RSN_RETURN_NAME
10851 : REG_RSN_RETURN_NULL);
10852 if ( RExC_parse == name_start
10853 || RExC_parse >= RExC_end
10854 || *RExC_parse != paren)
10856 vFAIL2("Sequence (?%c... not terminated",
10857 paren=='>' ? '<' : paren);
10862 if (!svname) /* shouldn't happen */
10864 "panic: reg_scan_name returned NULL");
10865 if (!RExC_paren_names) {
10866 RExC_paren_names= newHV();
10867 sv_2mortal(MUTABLE_SV(RExC_paren_names));
10869 RExC_paren_name_list= newAV();
10870 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
10873 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
10875 sv_dat = HeVAL(he_str);
10877 /* croak baby croak */
10879 "panic: paren_name hash element allocation failed");
10880 } else if ( SvPOK(sv_dat) ) {
10881 /* (?|...) can mean we have dupes so scan to check
10882 its already been stored. Maybe a flag indicating
10883 we are inside such a construct would be useful,
10884 but the arrays are likely to be quite small, so
10885 for now we punt -- dmq */
10886 IV count = SvIV(sv_dat);
10887 I32 *pv = (I32*)SvPVX(sv_dat);
10889 for ( i = 0 ; i < count ; i++ ) {
10890 if ( pv[i] == RExC_npar ) {
10896 pv = (I32*)SvGROW(sv_dat,
10897 SvCUR(sv_dat) + sizeof(I32)+1);
10898 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
10899 pv[count] = RExC_npar;
10900 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
10903 (void)SvUPGRADE(sv_dat,SVt_PVNV);
10904 sv_setpvn(sv_dat, (char *)&(RExC_npar),
10907 SvIV_set(sv_dat, 1);
10910 /* Yes this does cause a memory leak in debugging Perls
10912 if (!av_store(RExC_paren_name_list,
10913 RExC_npar, SvREFCNT_inc(svname)))
10914 SvREFCNT_dec_NN(svname);
10917 /*sv_dump(sv_dat);*/
10919 nextchar(pRExC_state);
10921 goto capturing_parens;
10923 RExC_seen |= REG_LOOKBEHIND_SEEN;
10924 RExC_in_lookbehind++;
10926 if (RExC_parse >= RExC_end) {
10927 vFAIL("Sequence (?... not terminated");
10931 case '=': /* (?=...) */
10932 RExC_seen_zerolen++;
10934 case '!': /* (?!...) */
10935 RExC_seen_zerolen++;
10936 /* check if we're really just a "FAIL" assertion */
10937 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
10938 FALSE /* Don't force to /x */ );
10939 if (*RExC_parse == ')') {
10940 ret=reganode(pRExC_state, OPFAIL, 0);
10941 nextchar(pRExC_state);
10945 case '|': /* (?|...) */
10946 /* branch reset, behave like a (?:...) except that
10947 buffers in alternations share the same numbers */
10949 after_freeze = freeze_paren = RExC_npar;
10951 case ':': /* (?:...) */
10952 case '>': /* (?>...) */
10954 case '$': /* (?$...) */
10955 case '@': /* (?@...) */
10956 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
10958 case '0' : /* (?0) */
10959 case 'R' : /* (?R) */
10960 if (RExC_parse == RExC_end || *RExC_parse != ')')
10961 FAIL("Sequence (?R) not terminated");
10963 RExC_seen |= REG_RECURSE_SEEN;
10964 *flagp |= POSTPONED;
10965 goto gen_recurse_regop;
10967 /* named and numeric backreferences */
10968 case '&': /* (?&NAME) */
10969 parse_start = RExC_parse - 1;
10972 SV *sv_dat = reg_scan_name(pRExC_state,
10973 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10974 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10976 if (RExC_parse >= RExC_end || *RExC_parse != ')')
10977 vFAIL("Sequence (?&... not terminated");
10978 goto gen_recurse_regop;
10981 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10983 vFAIL("Illegal pattern");
10985 goto parse_recursion;
10987 case '-': /* (?-1) */
10988 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10989 RExC_parse--; /* rewind to let it be handled later */
10993 case '1': case '2': case '3': case '4': /* (?1) */
10994 case '5': case '6': case '7': case '8': case '9':
10995 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
10998 bool is_neg = FALSE;
11000 parse_start = RExC_parse - 1; /* MJD */
11001 if (*RExC_parse == '-') {
11005 if (grok_atoUV(RExC_parse, &unum, &endptr)
11009 RExC_parse = (char*)endptr;
11013 /* Some limit for num? */
11017 if (*RExC_parse!=')')
11018 vFAIL("Expecting close bracket");
11021 if ( paren == '-' ) {
11023 Diagram of capture buffer numbering.
11024 Top line is the normal capture buffer numbers
11025 Bottom line is the negative indexing as from
11029 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
11033 num = RExC_npar + num;
11036 vFAIL("Reference to nonexistent group");
11038 } else if ( paren == '+' ) {
11039 num = RExC_npar + num - 1;
11041 /* We keep track how many GOSUB items we have produced.
11042 To start off the ARG2L() of the GOSUB holds its "id",
11043 which is used later in conjunction with RExC_recurse
11044 to calculate the offset we need to jump for the GOSUB,
11045 which it will store in the final representation.
11046 We have to defer the actual calculation until much later
11047 as the regop may move.
11050 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
11052 if (num > (I32)RExC_rx->nparens) {
11054 vFAIL("Reference to nonexistent group");
11056 RExC_recurse_count++;
11057 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11058 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
11059 22, "| |", (int)(depth * 2 + 1), "",
11060 (UV)ARG(ret), (IV)ARG2L(ret)));
11062 RExC_seen |= REG_RECURSE_SEEN;
11064 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
11065 Set_Node_Offset(ret, parse_start); /* MJD */
11067 *flagp |= POSTPONED;
11068 assert(*RExC_parse == ')');
11069 nextchar(pRExC_state);
11074 case '?': /* (??...) */
11076 if (*RExC_parse != '{') {
11077 RExC_parse += SKIP_IF_CHAR(RExC_parse);
11078 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11080 "Sequence (%" UTF8f "...) not recognized",
11081 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11082 NOT_REACHED; /*NOTREACHED*/
11084 *flagp |= POSTPONED;
11088 case '{': /* (?{...}) */
11091 struct reg_code_block *cb;
11093 RExC_seen_zerolen++;
11095 if ( !pRExC_state->code_blocks
11096 || pRExC_state->code_index
11097 >= pRExC_state->code_blocks->count
11098 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11099 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11102 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11103 FAIL("panic: Sequence (?{...}): no code block found\n");
11104 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11106 /* this is a pre-compiled code block (?{...}) */
11107 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11108 RExC_parse = RExC_start + cb->end;
11111 if (cb->src_regex) {
11112 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11113 RExC_rxi->data->data[n] =
11114 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11115 RExC_rxi->data->data[n+1] = (void*)o;
11118 n = add_data(pRExC_state,
11119 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11120 RExC_rxi->data->data[n] = (void*)o;
11123 pRExC_state->code_index++;
11124 nextchar(pRExC_state);
11128 ret = reg_node(pRExC_state, LOGICAL);
11130 eval = reg2Lanode(pRExC_state, EVAL,
11133 /* for later propagation into (??{})
11135 RExC_flags & RXf_PMf_COMPILETIME
11140 REGTAIL(pRExC_state, ret, eval);
11141 /* deal with the length of this later - MJD */
11144 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11145 Set_Node_Length(ret, RExC_parse - parse_start + 1);
11146 Set_Node_Offset(ret, parse_start);
11149 case '(': /* (?(?{...})...) and (?(?=...)...) */
11152 const int DEFINE_len = sizeof("DEFINE") - 1;
11153 if (RExC_parse[0] == '?') { /* (?(?...)) */
11154 if ( RExC_parse < RExC_end - 1
11155 && ( RExC_parse[1] == '='
11156 || RExC_parse[1] == '!'
11157 || RExC_parse[1] == '<'
11158 || RExC_parse[1] == '{')
11159 ) { /* Lookahead or eval. */
11163 ret = reg_node(pRExC_state, LOGICAL);
11167 tail = reg(pRExC_state, 1, &flag, depth+1);
11168 if (flag & (RESTART_PASS1|NEED_UTF8)) {
11169 *flagp = flag & (RESTART_PASS1|NEED_UTF8);
11172 REGTAIL(pRExC_state, ret, tail);
11175 /* Fall through to ‘Unknown switch condition’ at the
11176 end of the if/else chain. */
11178 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11179 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11181 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11182 char *name_start= RExC_parse++;
11184 SV *sv_dat=reg_scan_name(pRExC_state,
11185 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11186 if ( RExC_parse == name_start
11187 || RExC_parse >= RExC_end
11188 || *RExC_parse != ch)
11190 vFAIL2("Sequence (?(%c... not terminated",
11191 (ch == '>' ? '<' : ch));
11195 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11196 RExC_rxi->data->data[num]=(void*)sv_dat;
11197 SvREFCNT_inc_simple_void(sv_dat);
11199 ret = reganode(pRExC_state,NGROUPP,num);
11200 goto insert_if_check_paren;
11202 else if (memBEGINs(RExC_parse,
11203 (STRLEN) (RExC_end - RExC_parse),
11206 ret = reganode(pRExC_state,DEFINEP,0);
11207 RExC_parse += DEFINE_len;
11209 goto insert_if_check_paren;
11211 else if (RExC_parse[0] == 'R') {
11213 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11214 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11215 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11218 if (RExC_parse[0] == '0') {
11222 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11224 if (grok_atoUV(RExC_parse, &uv, &endptr)
11227 parno = (I32)uv + 1;
11228 RExC_parse = (char*)endptr;
11230 /* else "Switch condition not recognized" below */
11231 } else if (RExC_parse[0] == '&') {
11234 sv_dat = reg_scan_name(pRExC_state,
11236 ? REG_RSN_RETURN_NULL
11237 : REG_RSN_RETURN_DATA);
11239 /* we should only have a false sv_dat when
11240 * SIZE_ONLY is true, and we always have false
11241 * sv_dat when SIZE_ONLY is true.
11242 * reg_scan_name() will VFAIL() if the name is
11243 * unknown when SIZE_ONLY is false, and otherwise
11244 * will return something, and when SIZE_ONLY is
11245 * true, reg_scan_name() just parses the string,
11246 * and doesnt return anything. (in theory) */
11247 assert(SIZE_ONLY ? !sv_dat : !!sv_dat);
11250 parno = 1 + *((I32 *)SvPVX(sv_dat));
11252 ret = reganode(pRExC_state,INSUBP,parno);
11253 goto insert_if_check_paren;
11255 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11259 if (grok_atoUV(RExC_parse, &uv, &endptr)
11263 RExC_parse = (char*)endptr;
11266 vFAIL("panic: grok_atoUV returned FALSE");
11268 ret = reganode(pRExC_state, GROUPP, parno);
11270 insert_if_check_paren:
11271 if (UCHARAT(RExC_parse) != ')') {
11272 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11273 vFAIL("Switch condition not recognized");
11275 nextchar(pRExC_state);
11277 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
11278 br = regbranch(pRExC_state, &flags, 1,depth+1);
11280 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11281 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11284 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11287 REGTAIL(pRExC_state, br, reganode(pRExC_state,
11289 c = UCHARAT(RExC_parse);
11290 nextchar(pRExC_state);
11291 if (flags&HASWIDTH)
11292 *flagp |= HASWIDTH;
11295 vFAIL("(?(DEFINE)....) does not allow branches");
11297 /* Fake one for optimizer. */
11298 lastbr = reganode(pRExC_state, IFTHEN, 0);
11300 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
11301 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11302 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11305 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11308 REGTAIL(pRExC_state, ret, lastbr);
11309 if (flags&HASWIDTH)
11310 *flagp |= HASWIDTH;
11311 c = UCHARAT(RExC_parse);
11312 nextchar(pRExC_state);
11317 if (RExC_parse >= RExC_end)
11318 vFAIL("Switch (?(condition)... not terminated");
11320 vFAIL("Switch (?(condition)... contains too many branches");
11322 ender = reg_node(pRExC_state, TAIL);
11323 REGTAIL(pRExC_state, br, ender);
11325 REGTAIL(pRExC_state, lastbr, ender);
11326 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11329 REGTAIL(pRExC_state, ret, ender);
11330 RExC_size++; /* XXX WHY do we need this?!!
11331 For large programs it seems to be required
11332 but I can't figure out why. -- dmq*/
11335 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11336 vFAIL("Unknown switch condition (?(...))");
11338 case '[': /* (?[ ... ]) */
11339 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
11341 case 0: /* A NUL */
11342 RExC_parse--; /* for vFAIL to print correctly */
11343 vFAIL("Sequence (? incomplete");
11345 default: /* e.g., (?i) */
11346 RExC_parse = (char *) seqstart + 1;
11348 parse_lparen_question_flags(pRExC_state);
11349 if (UCHARAT(RExC_parse) != ':') {
11350 if (RExC_parse < RExC_end)
11351 nextchar(pRExC_state);
11356 nextchar(pRExC_state);
11361 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
11366 ret = reganode(pRExC_state, OPEN, parno);
11368 if (!RExC_nestroot)
11369 RExC_nestroot = parno;
11370 if (RExC_open_parens && !RExC_open_parens[parno])
11372 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11373 "%*s%*s Setting open paren #%" IVdf " to %d\n",
11374 22, "| |", (int)(depth * 2 + 1), "",
11375 (IV)parno, REG_NODE_NUM(ret)));
11376 RExC_open_parens[parno]= ret;
11379 Set_Node_Length(ret, 1); /* MJD */
11380 Set_Node_Offset(ret, RExC_parse); /* MJD */
11383 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
11392 /* Pick up the branches, linking them together. */
11393 parse_start = RExC_parse; /* MJD */
11394 br = regbranch(pRExC_state, &flags, 1,depth+1);
11396 /* branch_len = (paren != 0); */
11399 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11400 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11403 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11405 if (*RExC_parse == '|') {
11406 if (!SIZE_ONLY && RExC_extralen) {
11407 reginsert(pRExC_state, BRANCHJ, br, depth+1);
11410 reginsert(pRExC_state, BRANCH, br, depth+1);
11411 Set_Node_Length(br, paren != 0);
11412 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
11416 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
11418 else if (paren == ':') {
11419 *flagp |= flags&SIMPLE;
11421 if (is_open) { /* Starts with OPEN. */
11422 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
11424 else if (paren != '?') /* Not Conditional */
11426 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11428 while (*RExC_parse == '|') {
11429 if (!SIZE_ONLY && RExC_extralen) {
11430 ender = reganode(pRExC_state, LONGJMP,0);
11432 /* Append to the previous. */
11433 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11436 RExC_extralen += 2; /* Account for LONGJMP. */
11437 nextchar(pRExC_state);
11438 if (freeze_paren) {
11439 if (RExC_npar > after_freeze)
11440 after_freeze = RExC_npar;
11441 RExC_npar = freeze_paren;
11443 br = regbranch(pRExC_state, &flags, 0, depth+1);
11446 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11447 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11450 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11452 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
11454 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11457 if (have_branch || paren != ':') {
11458 /* Make a closing node, and hook it on the end. */
11461 ender = reg_node(pRExC_state, TAIL);
11464 ender = reganode(pRExC_state, CLOSE, parno);
11465 if ( RExC_close_parens ) {
11466 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11467 "%*s%*s Setting close paren #%" IVdf " to %d\n",
11468 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
11469 RExC_close_parens[parno]= ender;
11470 if (RExC_nestroot == parno)
11473 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
11474 Set_Node_Length(ender,1); /* MJD */
11480 *flagp &= ~HASWIDTH;
11483 ender = reg_node(pRExC_state, SUCCEED);
11486 ender = reg_node(pRExC_state, END);
11488 assert(!RExC_end_op); /* there can only be one! */
11489 RExC_end_op = ender;
11490 if (RExC_close_parens) {
11491 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11492 "%*s%*s Setting close paren #0 (END) to %d\n",
11493 22, "| |", (int)(depth * 2 + 1), "", REG_NODE_NUM(ender)));
11495 RExC_close_parens[0]= ender;
11500 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11501 DEBUG_PARSE_MSG("lsbr");
11502 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
11503 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11504 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11505 SvPV_nolen_const(RExC_mysv1),
11506 (IV)REG_NODE_NUM(lastbr),
11507 SvPV_nolen_const(RExC_mysv2),
11508 (IV)REG_NODE_NUM(ender),
11509 (IV)(ender - lastbr)
11512 REGTAIL(pRExC_state, lastbr, ender);
11514 if (have_branch && !SIZE_ONLY) {
11515 char is_nothing= 1;
11517 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
11519 /* Hook the tails of the branches to the closing node. */
11520 for (br = ret; br; br = regnext(br)) {
11521 const U8 op = PL_regkind[OP(br)];
11522 if (op == BRANCH) {
11523 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
11524 if ( OP(NEXTOPER(br)) != NOTHING
11525 || regnext(NEXTOPER(br)) != ender)
11528 else if (op == BRANCHJ) {
11529 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
11530 /* for now we always disable this optimisation * /
11531 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
11532 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
11538 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
11539 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11540 DEBUG_PARSE_MSG("NADA");
11541 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
11542 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11543 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11544 SvPV_nolen_const(RExC_mysv1),
11545 (IV)REG_NODE_NUM(ret),
11546 SvPV_nolen_const(RExC_mysv2),
11547 (IV)REG_NODE_NUM(ender),
11552 if (OP(ender) == TAIL) {
11557 for ( opt= br + 1; opt < ender ; opt++ )
11558 OP(opt)= OPTIMIZED;
11559 NEXT_OFF(br)= ender - br;
11567 static const char parens[] = "=!<,>";
11569 if (paren && (p = strchr(parens, paren))) {
11570 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
11571 int flag = (p - parens) > 1;
11574 node = SUSPEND, flag = 0;
11575 reginsert(pRExC_state, node,ret, depth+1);
11576 Set_Node_Cur_Length(ret, parse_start);
11577 Set_Node_Offset(ret, parse_start + 1);
11579 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
11583 /* Check for proper termination. */
11585 /* restore original flags, but keep (?p) and, if we've changed from /d
11586 * rules to /u, keep the /u */
11587 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
11588 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
11589 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
11591 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
11592 RExC_parse = oregcomp_parse;
11593 vFAIL("Unmatched (");
11595 nextchar(pRExC_state);
11597 else if (!paren && RExC_parse < RExC_end) {
11598 if (*RExC_parse == ')') {
11600 vFAIL("Unmatched )");
11603 FAIL("Junk on end of regexp"); /* "Can't happen". */
11604 NOT_REACHED; /* NOTREACHED */
11607 if (RExC_in_lookbehind) {
11608 RExC_in_lookbehind--;
11610 if (after_freeze > RExC_npar)
11611 RExC_npar = after_freeze;
11616 - regbranch - one alternative of an | operator
11618 * Implements the concatenation operator.
11620 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11621 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11624 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
11627 regnode *chain = NULL;
11629 I32 flags = 0, c = 0;
11630 GET_RE_DEBUG_FLAGS_DECL;
11632 PERL_ARGS_ASSERT_REGBRANCH;
11634 DEBUG_PARSE("brnc");
11639 if (!SIZE_ONLY && RExC_extralen)
11640 ret = reganode(pRExC_state, BRANCHJ,0);
11642 ret = reg_node(pRExC_state, BRANCH);
11643 Set_Node_Length(ret, 1);
11647 if (!first && SIZE_ONLY)
11648 RExC_extralen += 1; /* BRANCHJ */
11650 *flagp = WORST; /* Tentatively. */
11652 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11653 FALSE /* Don't force to /x */ );
11654 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
11655 flags &= ~TRYAGAIN;
11656 latest = regpiece(pRExC_state, &flags,depth+1);
11657 if (latest == NULL) {
11658 if (flags & TRYAGAIN)
11660 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11661 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11664 FAIL2("panic: regpiece returned NULL, flags=%#" UVxf, (UV) flags);
11666 else if (ret == NULL)
11668 *flagp |= flags&(HASWIDTH|POSTPONED);
11669 if (chain == NULL) /* First piece. */
11670 *flagp |= flags&SPSTART;
11672 /* FIXME adding one for every branch after the first is probably
11673 * excessive now we have TRIE support. (hv) */
11675 REGTAIL(pRExC_state, chain, latest);
11680 if (chain == NULL) { /* Loop ran zero times. */
11681 chain = reg_node(pRExC_state, NOTHING);
11686 *flagp |= flags&SIMPLE;
11693 - regpiece - something followed by possible quantifier * + ? {n,m}
11695 * Note that the branching code sequences used for ? and the general cases
11696 * of * and + are somewhat optimized: they use the same NOTHING node as
11697 * both the endmarker for their branch list and the body of the last branch.
11698 * It might seem that this node could be dispensed with entirely, but the
11699 * endmarker role is not redundant.
11701 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
11703 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11704 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11707 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11713 const char * const origparse = RExC_parse;
11715 I32 max = REG_INFTY;
11716 #ifdef RE_TRACK_PATTERN_OFFSETS
11719 const char *maxpos = NULL;
11722 /* Save the original in case we change the emitted regop to a FAIL. */
11723 regnode * const orig_emit = RExC_emit;
11725 GET_RE_DEBUG_FLAGS_DECL;
11727 PERL_ARGS_ASSERT_REGPIECE;
11729 DEBUG_PARSE("piec");
11731 ret = regatom(pRExC_state, &flags,depth+1);
11733 if (flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8))
11734 *flagp |= flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8);
11736 FAIL2("panic: regatom returned NULL, flags=%#" UVxf, (UV) flags);
11742 if (op == '{' && regcurly(RExC_parse)) {
11744 #ifdef RE_TRACK_PATTERN_OFFSETS
11745 parse_start = RExC_parse; /* MJD */
11747 next = RExC_parse + 1;
11748 while (isDIGIT(*next) || *next == ',') {
11749 if (*next == ',') {
11757 if (*next == '}') { /* got one */
11758 const char* endptr;
11762 if (isDIGIT(*RExC_parse)) {
11763 if (!grok_atoUV(RExC_parse, &uv, &endptr))
11764 vFAIL("Invalid quantifier in {,}");
11765 if (uv >= REG_INFTY)
11766 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11771 if (*maxpos == ',')
11774 maxpos = RExC_parse;
11775 if (isDIGIT(*maxpos)) {
11776 if (!grok_atoUV(maxpos, &uv, &endptr))
11777 vFAIL("Invalid quantifier in {,}");
11778 if (uv >= REG_INFTY)
11779 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11782 max = REG_INFTY; /* meaning "infinity" */
11785 nextchar(pRExC_state);
11786 if (max < min) { /* If can't match, warn and optimize to fail
11788 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
11790 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
11791 NEXT_OFF(orig_emit)= regarglen[OPFAIL] + NODE_STEP_REGNODE;
11795 else if (min == max && *RExC_parse == '?')
11798 ckWARN2reg(RExC_parse + 1,
11799 "Useless use of greediness modifier '%c'",
11805 if ((flags&SIMPLE)) {
11806 if (min == 0 && max == REG_INFTY) {
11807 reginsert(pRExC_state, STAR, ret, depth+1);
11809 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11812 if (min == 1 && max == REG_INFTY) {
11813 reginsert(pRExC_state, PLUS, ret, depth+1);
11815 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11818 MARK_NAUGHTY_EXP(2, 2);
11819 reginsert(pRExC_state, CURLY, ret, depth+1);
11820 Set_Node_Offset(ret, parse_start+1); /* MJD */
11821 Set_Node_Cur_Length(ret, parse_start);
11824 regnode * const w = reg_node(pRExC_state, WHILEM);
11827 REGTAIL(pRExC_state, ret, w);
11828 if (!SIZE_ONLY && RExC_extralen) {
11829 reginsert(pRExC_state, LONGJMP,ret, depth+1);
11830 reginsert(pRExC_state, NOTHING,ret, depth+1);
11831 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
11833 reginsert(pRExC_state, CURLYX,ret, depth+1);
11835 Set_Node_Offset(ret, parse_start+1);
11836 Set_Node_Length(ret,
11837 op == '{' ? (RExC_parse - parse_start) : 1);
11839 if (!SIZE_ONLY && RExC_extralen)
11840 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
11841 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
11843 RExC_whilem_seen++, RExC_extralen += 3;
11844 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
11851 *flagp |= HASWIDTH;
11853 ARG1_SET(ret, (U16)min);
11854 ARG2_SET(ret, (U16)max);
11856 if (max == REG_INFTY)
11857 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11863 if (!ISMULT1(op)) {
11868 #if 0 /* Now runtime fix should be reliable. */
11870 /* if this is reinstated, don't forget to put this back into perldiag:
11872 =item Regexp *+ operand could be empty at {#} in regex m/%s/
11874 (F) The part of the regexp subject to either the * or + quantifier
11875 could match an empty string. The {#} shows in the regular
11876 expression about where the problem was discovered.
11880 if (!(flags&HASWIDTH) && op != '?')
11881 vFAIL("Regexp *+ operand could be empty");
11884 #ifdef RE_TRACK_PATTERN_OFFSETS
11885 parse_start = RExC_parse;
11887 nextchar(pRExC_state);
11889 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
11895 else if (op == '+') {
11899 else if (op == '?') {
11904 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
11905 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
11906 ckWARN2reg(RExC_parse,
11907 "%" UTF8f " matches null string many times",
11908 UTF8fARG(UTF, (RExC_parse >= origparse
11909 ? RExC_parse - origparse
11912 (void)ReREFCNT_inc(RExC_rx_sv);
11915 if (*RExC_parse == '?') {
11916 nextchar(pRExC_state);
11917 reginsert(pRExC_state, MINMOD, ret, depth+1);
11918 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
11920 else if (*RExC_parse == '+') {
11922 nextchar(pRExC_state);
11923 ender = reg_node(pRExC_state, SUCCEED);
11924 REGTAIL(pRExC_state, ret, ender);
11925 reginsert(pRExC_state, SUSPEND, ret, depth+1);
11926 ender = reg_node(pRExC_state, TAIL);
11927 REGTAIL(pRExC_state, ret, ender);
11930 if (ISMULT2(RExC_parse)) {
11932 vFAIL("Nested quantifiers");
11939 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
11948 /* This routine teases apart the various meanings of \N and returns
11949 * accordingly. The input parameters constrain which meaning(s) is/are valid
11950 * in the current context.
11952 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
11954 * If <code_point_p> is not NULL, the context is expecting the result to be a
11955 * single code point. If this \N instance turns out to a single code point,
11956 * the function returns TRUE and sets *code_point_p to that code point.
11958 * If <node_p> is not NULL, the context is expecting the result to be one of
11959 * the things representable by a regnode. If this \N instance turns out to be
11960 * one such, the function generates the regnode, returns TRUE and sets *node_p
11961 * to point to that regnode.
11963 * If this instance of \N isn't legal in any context, this function will
11964 * generate a fatal error and not return.
11966 * On input, RExC_parse should point to the first char following the \N at the
11967 * time of the call. On successful return, RExC_parse will have been updated
11968 * to point to just after the sequence identified by this routine. Also
11969 * *flagp has been updated as needed.
11971 * When there is some problem with the current context and this \N instance,
11972 * the function returns FALSE, without advancing RExC_parse, nor setting
11973 * *node_p, nor *code_point_p, nor *flagp.
11975 * If <cp_count> is not NULL, the caller wants to know the length (in code
11976 * points) that this \N sequence matches. This is set even if the function
11977 * returns FALSE, as detailed below.
11979 * There are 5 possibilities here, as detailed in the next 5 paragraphs.
11981 * Probably the most common case is for the \N to specify a single code point.
11982 * *cp_count will be set to 1, and *code_point_p will be set to that code
11985 * Another possibility is for the input to be an empty \N{}, which for
11986 * backwards compatibility we accept. *cp_count will be set to 0. *node_p
11987 * will be set to a generated NOTHING node.
11989 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
11990 * set to 0. *node_p will be set to a generated REG_ANY node.
11992 * The fourth possibility is that \N resolves to a sequence of more than one
11993 * code points. *cp_count will be set to the number of code points in the
11994 * sequence. *node_p * will be set to a generated node returned by this
11995 * function calling S_reg().
11997 * The final possibility is that it is premature to be calling this function;
11998 * that pass1 needs to be restarted. This can happen when this changes from
11999 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
12000 * latter occurs only when the fourth possibility would otherwise be in
12001 * effect, and is because one of those code points requires the pattern to be
12002 * recompiled as UTF-8. The function returns FALSE, and sets the
12003 * RESTART_PASS1 and NEED_UTF8 flags in *flagp, as appropriate. When this
12004 * happens, the caller needs to desist from continuing parsing, and return
12005 * this information to its caller. This is not set for when there is only one
12006 * code point, as this can be called as part of an ANYOF node, and they can
12007 * store above-Latin1 code points without the pattern having to be in UTF-8.
12009 * For non-single-quoted regexes, the tokenizer has resolved character and
12010 * sequence names inside \N{...} into their Unicode values, normalizing the
12011 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
12012 * hex-represented code points in the sequence. This is done there because
12013 * the names can vary based on what charnames pragma is in scope at the time,
12014 * so we need a way to take a snapshot of what they resolve to at the time of
12015 * the original parse. [perl #56444].
12017 * That parsing is skipped for single-quoted regexes, so we may here get
12018 * '\N{NAME}'. This is a fatal error. These names have to be resolved by the
12019 * parser. But if the single-quoted regex is something like '\N{U+41}', that
12020 * is legal and handled here. The code point is Unicode, and has to be
12021 * translated into the native character set for non-ASCII platforms.
12024 char * endbrace; /* points to '}' following the name */
12025 char *endchar; /* Points to '.' or '}' ending cur char in the input
12027 char* p = RExC_parse; /* Temporary */
12029 GET_RE_DEBUG_FLAGS_DECL;
12031 PERL_ARGS_ASSERT_GROK_BSLASH_N;
12033 GET_RE_DEBUG_FLAGS;
12035 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
12036 assert(! (node_p && cp_count)); /* At most 1 should be set */
12038 if (cp_count) { /* Initialize return for the most common case */
12042 /* The [^\n] meaning of \N ignores spaces and comments under the /x
12043 * modifier. The other meanings do not, so use a temporary until we find
12044 * out which we are being called with */
12045 skip_to_be_ignored_text(pRExC_state, &p,
12046 FALSE /* Don't force to /x */ );
12048 /* Disambiguate between \N meaning a named character versus \N meaning
12049 * [^\n]. The latter is assumed when the {...} following the \N is a legal
12050 * quantifier, or there is no '{' at all */
12051 if (*p != '{' || regcurly(p)) {
12061 *node_p = reg_node(pRExC_state, REG_ANY);
12062 *flagp |= HASWIDTH|SIMPLE;
12064 Set_Node_Length(*node_p, 1); /* MJD */
12068 /* Here, we have decided it should be a named character or sequence */
12070 /* The test above made sure that the next real character is a '{', but
12071 * under the /x modifier, it could be separated by space (or a comment and
12072 * \n) and this is not allowed (for consistency with \x{...} and the
12073 * tokenizer handling of \N{NAME}). */
12074 if (*RExC_parse != '{') {
12075 vFAIL("Missing braces on \\N{}");
12078 RExC_parse++; /* Skip past the '{' */
12080 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
12081 if (! endbrace) { /* no trailing brace */
12082 vFAIL2("Missing right brace on \\%c{}", 'N');
12084 else if (!( endbrace == RExC_parse /* nothing between the {} */
12085 || memBEGINs(RExC_parse, /* U+ (bad hex is checked below
12086 for a better error msg) */
12087 (STRLEN) (RExC_end - RExC_parse),
12090 RExC_parse = endbrace; /* position msg's '<--HERE' */
12091 vFAIL("\\N{NAME} must be resolved by the lexer");
12094 REQUIRE_UNI_RULES(flagp, FALSE); /* Unicode named chars imply Unicode
12097 if (endbrace == RExC_parse) { /* empty: \N{} */
12099 RExC_parse++; /* Position after the "}" */
12100 vFAIL("Zero length \\N{}");
12105 nextchar(pRExC_state);
12110 *node_p = reg_node(pRExC_state,NOTHING);
12114 RExC_parse += 2; /* Skip past the 'U+' */
12116 /* Because toke.c has generated a special construct for us guaranteed not
12117 * to have NULs, we can use a str function */
12118 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12120 /* Code points are separated by dots. If none, there is only one code
12121 * point, and is terminated by the brace */
12123 if (endchar >= endbrace) {
12124 STRLEN length_of_hex;
12125 I32 grok_hex_flags;
12127 /* Here, exactly one code point. If that isn't what is wanted, fail */
12128 if (! code_point_p) {
12133 /* Convert code point from hex */
12134 length_of_hex = (STRLEN)(endchar - RExC_parse);
12135 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
12136 | PERL_SCAN_DISALLOW_PREFIX
12138 /* No errors in the first pass (See [perl
12139 * #122671].) We let the code below find the
12140 * errors when there are multiple chars. */
12142 ? PERL_SCAN_SILENT_ILLDIGIT
12145 /* This routine is the one place where both single- and double-quotish
12146 * \N{U+xxxx} are evaluated. The value is a Unicode code point which
12147 * must be converted to native. */
12148 *code_point_p = UNI_TO_NATIVE(grok_hex(RExC_parse,
12153 /* The tokenizer should have guaranteed validity, but it's possible to
12154 * bypass it by using single quoting, so check. Don't do the check
12155 * here when there are multiple chars; we do it below anyway. */
12156 if (length_of_hex == 0
12157 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
12159 RExC_parse += length_of_hex; /* Includes all the valid */
12160 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
12161 ? UTF8SKIP(RExC_parse)
12163 /* Guard against malformed utf8 */
12164 if (RExC_parse >= endchar) {
12165 RExC_parse = endchar;
12167 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12170 RExC_parse = endbrace + 1;
12173 else { /* Is a multiple character sequence */
12174 SV * substitute_parse;
12176 char *orig_end = RExC_end;
12177 char *save_start = RExC_start;
12180 /* Count the code points, if desired, in the sequence */
12183 while (RExC_parse < endbrace) {
12184 /* Point to the beginning of the next character in the sequence. */
12185 RExC_parse = endchar + 1;
12186 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12191 /* Fail if caller doesn't want to handle a multi-code-point sequence.
12192 * But don't backup up the pointer if the caller wants to know how many
12193 * code points there are (they can then handle things) */
12201 /* What is done here is to convert this to a sub-pattern of the form
12202 * \x{char1}\x{char2}... and then call reg recursively to parse it
12203 * (enclosing in "(?: ... )" ). That way, it retains its atomicness,
12204 * while not having to worry about special handling that some code
12205 * points may have. */
12207 substitute_parse = newSVpvs("?:");
12209 while (RExC_parse < endbrace) {
12211 /* Convert to notation the rest of the code understands */
12212 sv_catpv(substitute_parse, "\\x{");
12213 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
12214 sv_catpv(substitute_parse, "}");
12216 /* Point to the beginning of the next character in the sequence. */
12217 RExC_parse = endchar + 1;
12218 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12221 sv_catpv(substitute_parse, ")");
12223 len = SvCUR(substitute_parse);
12225 /* Don't allow empty number */
12226 if (len < (STRLEN) 8) {
12227 RExC_parse = endbrace;
12228 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12231 RExC_parse = RExC_start = RExC_adjusted_start
12232 = SvPV_nolen(substitute_parse);
12233 RExC_end = RExC_parse + len;
12235 /* The values are Unicode, and therefore not subject to recoding, but
12236 * have to be converted to native on a non-Unicode (meaning non-ASCII)
12239 RExC_recode_x_to_native = 1;
12242 *node_p = reg(pRExC_state, 1, &flags, depth+1);
12244 /* Restore the saved values */
12245 RExC_start = RExC_adjusted_start = save_start;
12246 RExC_parse = endbrace;
12247 RExC_end = orig_end;
12249 RExC_recode_x_to_native = 0;
12251 SvREFCNT_dec_NN(substitute_parse);
12254 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12255 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12258 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#" UVxf,
12261 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12263 nextchar(pRExC_state);
12270 PERL_STATIC_INLINE U8
12271 S_compute_EXACTish(RExC_state_t *pRExC_state)
12275 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
12283 op = get_regex_charset(RExC_flags);
12284 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
12285 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
12286 been, so there is no hole */
12289 return op + EXACTF;
12292 PERL_STATIC_INLINE void
12293 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
12294 regnode *node, I32* flagp, STRLEN len, UV code_point,
12297 /* This knows the details about sizing an EXACTish node, setting flags for
12298 * it (by setting <*flagp>, and potentially populating it with a single
12301 * If <len> (the length in bytes) is non-zero, this function assumes that
12302 * the node has already been populated, and just does the sizing. In this
12303 * case <code_point> should be the final code point that has already been
12304 * placed into the node. This value will be ignored except that under some
12305 * circumstances <*flagp> is set based on it.
12307 * If <len> is zero, the function assumes that the node is to contain only
12308 * the single character given by <code_point> and calculates what <len>
12309 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
12310 * additionally will populate the node's STRING with <code_point> or its
12313 * In both cases <*flagp> is appropriately set
12315 * It knows that under FOLD, the Latin Sharp S and UTF characters above
12316 * 255, must be folded (the former only when the rules indicate it can
12319 * When it does the populating, it looks at the flag 'downgradable'. If
12320 * true with a node that folds, it checks if the single code point
12321 * participates in a fold, and if not downgrades the node to an EXACT.
12322 * This helps the optimizer */
12324 bool len_passed_in = cBOOL(len != 0);
12325 U8 character[UTF8_MAXBYTES_CASE+1];
12327 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
12329 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
12330 * sizing difference, and is extra work that is thrown away */
12331 if (downgradable && ! PASS2) {
12332 downgradable = FALSE;
12335 if (! len_passed_in) {
12337 if (UVCHR_IS_INVARIANT(code_point)) {
12338 if (LOC || ! FOLD) { /* /l defers folding until runtime */
12339 *character = (U8) code_point;
12341 else { /* Here is /i and not /l. (toFOLD() is defined on just
12342 ASCII, which isn't the same thing as INVARIANT on
12343 EBCDIC, but it works there, as the extra invariants
12344 fold to themselves) */
12345 *character = toFOLD((U8) code_point);
12347 /* We can downgrade to an EXACT node if this character
12348 * isn't a folding one. Note that this assumes that
12349 * nothing above Latin1 folds to some other invariant than
12350 * one of these alphabetics; otherwise we would also have
12352 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12353 * || ASCII_FOLD_RESTRICTED))
12355 if (downgradable && PL_fold[code_point] == code_point) {
12361 else if (FOLD && (! LOC
12362 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
12363 { /* Folding, and ok to do so now */
12364 UV folded = _to_uni_fold_flags(
12368 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12369 ? FOLD_FLAGS_NOMIX_ASCII
12372 && folded == code_point /* This quickly rules out many
12373 cases, avoiding the
12374 _invlist_contains_cp() overhead
12376 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
12383 else if (code_point <= MAX_UTF8_TWO_BYTE) {
12385 /* Not folding this cp, and can output it directly */
12386 *character = UTF8_TWO_BYTE_HI(code_point);
12387 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
12391 uvchr_to_utf8( character, code_point);
12392 len = UTF8SKIP(character);
12394 } /* Else pattern isn't UTF8. */
12396 *character = (U8) code_point;
12398 } /* Else is folded non-UTF8 */
12399 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12400 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12401 || UNICODE_DOT_DOT_VERSION > 0)
12402 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
12406 /* We don't fold any non-UTF8 except possibly the Sharp s (see
12407 * comments at join_exact()); */
12408 *character = (U8) code_point;
12411 /* Can turn into an EXACT node if we know the fold at compile time,
12412 * and it folds to itself and doesn't particpate in other folds */
12415 && PL_fold_latin1[code_point] == code_point
12416 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12417 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
12421 } /* else is Sharp s. May need to fold it */
12422 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
12424 *(character + 1) = 's';
12428 *character = LATIN_SMALL_LETTER_SHARP_S;
12434 RExC_size += STR_SZ(len);
12437 RExC_emit += STR_SZ(len);
12438 STR_LEN(node) = len;
12439 if (! len_passed_in) {
12440 Copy((char *) character, STRING(node), len, char);
12444 *flagp |= HASWIDTH;
12446 /* A single character node is SIMPLE, except for the special-cased SHARP S
12448 if ((len == 1 || (UTF && len == UVCHR_SKIP(code_point)))
12449 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12450 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12451 || UNICODE_DOT_DOT_VERSION > 0)
12452 && ( code_point != LATIN_SMALL_LETTER_SHARP_S
12453 || ! FOLD || ! DEPENDS_SEMANTICS)
12459 /* The OP may not be well defined in PASS1 */
12460 if (PASS2 && OP(node) == EXACTFL) {
12461 RExC_contains_locale = 1;
12466 S_new_regcurly(const char *s, const char *e)
12468 /* This is a temporary function designed to match the most lenient form of
12469 * a {m,n} quantifier we ever envision, with either number omitted, and
12470 * spaces anywhere between/before/after them.
12472 * If this function fails, then the string it matches is very unlikely to
12473 * ever be considered a valid quantifier, so we can allow the '{' that
12474 * begins it to be considered as a literal */
12476 bool has_min = FALSE;
12477 bool has_max = FALSE;
12479 PERL_ARGS_ASSERT_NEW_REGCURLY;
12481 if (s >= e || *s++ != '{')
12484 while (s < e && isSPACE(*s)) {
12487 while (s < e && isDIGIT(*s)) {
12491 while (s < e && isSPACE(*s)) {
12497 while (s < e && isSPACE(*s)) {
12500 while (s < e && isDIGIT(*s)) {
12504 while (s < e && isSPACE(*s)) {
12509 return s < e && *s == '}' && (has_min || has_max);
12512 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
12513 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
12516 S_backref_value(char *p)
12518 const char* endptr;
12520 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
12527 - regatom - the lowest level
12529 Try to identify anything special at the start of the current parse position.
12530 If there is, then handle it as required. This may involve generating a
12531 single regop, such as for an assertion; or it may involve recursing, such as
12532 to handle a () structure.
12534 If the string doesn't start with something special then we gobble up
12535 as much literal text as we can. If we encounter a quantifier, we have to
12536 back off the final literal character, as that quantifier applies to just it
12537 and not to the whole string of literals.
12539 Once we have been able to handle whatever type of thing started the
12540 sequence, we return.
12542 Note: we have to be careful with escapes, as they can be both literal
12543 and special, and in the case of \10 and friends, context determines which.
12545 A summary of the code structure is:
12547 switch (first_byte) {
12548 cases for each special:
12549 handle this special;
12552 switch (2nd byte) {
12553 cases for each unambiguous special:
12554 handle this special;
12556 cases for each ambigous special/literal:
12558 if (special) handle here
12560 default: // unambiguously literal:
12563 default: // is a literal char
12566 create EXACTish node for literal;
12567 while (more input and node isn't full) {
12568 switch (input_byte) {
12569 cases for each special;
12570 make sure parse pointer is set so that the next call to
12571 regatom will see this special first
12572 goto loopdone; // EXACTish node terminated by prev. char
12574 append char to EXACTISH node;
12576 get next input byte;
12580 return the generated node;
12582 Specifically there are two separate switches for handling
12583 escape sequences, with the one for handling literal escapes requiring
12584 a dummy entry for all of the special escapes that are actually handled
12587 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
12589 Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
12590 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
12591 Otherwise does not return NULL.
12595 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12597 regnode *ret = NULL;
12604 GET_RE_DEBUG_FLAGS_DECL;
12606 *flagp = WORST; /* Tentatively. */
12608 DEBUG_PARSE("atom");
12610 PERL_ARGS_ASSERT_REGATOM;
12613 parse_start = RExC_parse;
12614 assert(RExC_parse < RExC_end);
12615 switch ((U8)*RExC_parse) {
12617 RExC_seen_zerolen++;
12618 nextchar(pRExC_state);
12619 if (RExC_flags & RXf_PMf_MULTILINE)
12620 ret = reg_node(pRExC_state, MBOL);
12622 ret = reg_node(pRExC_state, SBOL);
12623 Set_Node_Length(ret, 1); /* MJD */
12626 nextchar(pRExC_state);
12628 RExC_seen_zerolen++;
12629 if (RExC_flags & RXf_PMf_MULTILINE)
12630 ret = reg_node(pRExC_state, MEOL);
12632 ret = reg_node(pRExC_state, SEOL);
12633 Set_Node_Length(ret, 1); /* MJD */
12636 nextchar(pRExC_state);
12637 if (RExC_flags & RXf_PMf_SINGLELINE)
12638 ret = reg_node(pRExC_state, SANY);
12640 ret = reg_node(pRExC_state, REG_ANY);
12641 *flagp |= HASWIDTH|SIMPLE;
12643 Set_Node_Length(ret, 1); /* MJD */
12647 char * const oregcomp_parse = ++RExC_parse;
12648 ret = regclass(pRExC_state, flagp,depth+1,
12649 FALSE, /* means parse the whole char class */
12650 TRUE, /* allow multi-char folds */
12651 FALSE, /* don't silence non-portable warnings. */
12652 (bool) RExC_strict,
12653 TRUE, /* Allow an optimized regnode result */
12657 if (*flagp & (RESTART_PASS1|NEED_UTF8))
12659 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12662 if (*RExC_parse != ']') {
12663 RExC_parse = oregcomp_parse;
12664 vFAIL("Unmatched [");
12666 nextchar(pRExC_state);
12667 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
12671 nextchar(pRExC_state);
12672 ret = reg(pRExC_state, 2, &flags,depth+1);
12674 if (flags & TRYAGAIN) {
12675 if (RExC_parse >= RExC_end) {
12676 /* Make parent create an empty node if needed. */
12677 *flagp |= TRYAGAIN;
12682 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12683 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12686 FAIL2("panic: reg returned NULL to regatom, flags=%#" UVxf,
12689 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12693 if (flags & TRYAGAIN) {
12694 *flagp |= TRYAGAIN;
12697 vFAIL("Internal urp");
12698 /* Supposed to be caught earlier. */
12704 vFAIL("Quantifier follows nothing");
12709 This switch handles escape sequences that resolve to some kind
12710 of special regop and not to literal text. Escape sequnces that
12711 resolve to literal text are handled below in the switch marked
12714 Every entry in this switch *must* have a corresponding entry
12715 in the literal escape switch. However, the opposite is not
12716 required, as the default for this switch is to jump to the
12717 literal text handling code.
12720 switch ((U8)*RExC_parse) {
12721 /* Special Escapes */
12723 RExC_seen_zerolen++;
12724 ret = reg_node(pRExC_state, SBOL);
12725 /* SBOL is shared with /^/ so we set the flags so we can tell
12726 * /\A/ from /^/ in split. We check ret because first pass we
12727 * have no regop struct to set the flags on. */
12731 goto finish_meta_pat;
12733 ret = reg_node(pRExC_state, GPOS);
12734 RExC_seen |= REG_GPOS_SEEN;
12736 goto finish_meta_pat;
12738 RExC_seen_zerolen++;
12739 ret = reg_node(pRExC_state, KEEPS);
12741 /* XXX:dmq : disabling in-place substitution seems to
12742 * be necessary here to avoid cases of memory corruption, as
12743 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
12745 RExC_seen |= REG_LOOKBEHIND_SEEN;
12746 goto finish_meta_pat;
12748 ret = reg_node(pRExC_state, SEOL);
12750 RExC_seen_zerolen++; /* Do not optimize RE away */
12751 goto finish_meta_pat;
12753 ret = reg_node(pRExC_state, EOS);
12755 RExC_seen_zerolen++; /* Do not optimize RE away */
12756 goto finish_meta_pat;
12758 vFAIL("\\C no longer supported");
12760 ret = reg_node(pRExC_state, CLUMP);
12761 *flagp |= HASWIDTH;
12762 goto finish_meta_pat;
12768 arg = ANYOF_WORDCHAR;
12776 regex_charset charset = get_regex_charset(RExC_flags);
12778 RExC_seen_zerolen++;
12779 RExC_seen |= REG_LOOKBEHIND_SEEN;
12780 op = BOUND + charset;
12782 if (op == BOUNDL) {
12783 RExC_contains_locale = 1;
12786 ret = reg_node(pRExC_state, op);
12788 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
12789 FLAGS(ret) = TRADITIONAL_BOUND;
12790 if (PASS2 && op > BOUNDA) { /* /aa is same as /a */
12796 char name = *RExC_parse;
12797 char * endbrace = NULL;
12799 if (RExC_parse < RExC_end) {
12800 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
12804 vFAIL2("Missing right brace on \\%c{}", name);
12806 /* XXX Need to decide whether to take spaces or not. Should be
12807 * consistent with \p{}, but that currently is SPACE, which
12808 * means vertical too, which seems wrong
12809 * while (isBLANK(*RExC_parse)) {
12812 if (endbrace == RExC_parse) {
12813 RExC_parse++; /* After the '}' */
12814 vFAIL2("Empty \\%c{}", name);
12816 length = endbrace - RExC_parse;
12817 /*while (isBLANK(*(RExC_parse + length - 1))) {
12820 switch (*RExC_parse) {
12823 && (memNEs(RExC_parse + 1, length - 1, "cb")))
12825 goto bad_bound_type;
12827 FLAGS(ret) = GCB_BOUND;
12830 if (length != 2 || *(RExC_parse + 1) != 'b') {
12831 goto bad_bound_type;
12833 FLAGS(ret) = LB_BOUND;
12836 if (length != 2 || *(RExC_parse + 1) != 'b') {
12837 goto bad_bound_type;
12839 FLAGS(ret) = SB_BOUND;
12842 if (length != 2 || *(RExC_parse + 1) != 'b') {
12843 goto bad_bound_type;
12845 FLAGS(ret) = WB_BOUND;
12849 RExC_parse = endbrace;
12851 "'%" UTF8f "' is an unknown bound type",
12852 UTF8fARG(UTF, length, endbrace - length));
12853 NOT_REACHED; /*NOTREACHED*/
12855 RExC_parse = endbrace;
12856 REQUIRE_UNI_RULES(flagp, NULL);
12858 if (PASS2 && op >= BOUNDA) { /* /aa is same as /a */
12862 /* Don't have to worry about UTF-8, in this message because
12863 * to get here the contents of the \b must be ASCII */
12864 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
12865 "Using /u for '%.*s' instead of /%s",
12867 endbrace - length + 1,
12868 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
12869 ? ASCII_RESTRICT_PAT_MODS
12870 : ASCII_MORE_RESTRICT_PAT_MODS);
12874 if (PASS2 && invert) {
12875 OP(ret) += NBOUND - BOUND;
12877 goto finish_meta_pat;
12885 if (! DEPENDS_SEMANTICS) {
12889 /* \d doesn't have any matches in the upper Latin1 range, hence /d
12890 * is equivalent to /u. Changing to /u saves some branches at
12893 goto join_posix_op_known;
12896 ret = reg_node(pRExC_state, LNBREAK);
12897 *flagp |= HASWIDTH|SIMPLE;
12898 goto finish_meta_pat;
12906 goto join_posix_op_known;
12912 arg = ANYOF_VERTWS;
12914 goto join_posix_op_known;
12924 op = POSIXD + get_regex_charset(RExC_flags);
12925 if (op > POSIXA) { /* /aa is same as /a */
12928 else if (op == POSIXL) {
12929 RExC_contains_locale = 1;
12932 join_posix_op_known:
12935 op += NPOSIXD - POSIXD;
12938 ret = reg_node(pRExC_state, op);
12940 FLAGS(ret) = namedclass_to_classnum(arg);
12943 *flagp |= HASWIDTH|SIMPLE;
12947 if ( UCHARAT(RExC_parse + 1) == '{'
12948 && UNLIKELY(! new_regcurly(RExC_parse + 1, RExC_end)))
12951 vFAIL("Unescaped left brace in regex is illegal here");
12953 nextchar(pRExC_state);
12954 Set_Node_Length(ret, 2); /* MJD */
12960 ret = regclass(pRExC_state, flagp,depth+1,
12961 TRUE, /* means just parse this element */
12962 FALSE, /* don't allow multi-char folds */
12963 FALSE, /* don't silence non-portable warnings. It
12964 would be a bug if these returned
12966 (bool) RExC_strict,
12967 TRUE, /* Allow an optimized regnode result */
12970 if (*flagp & RESTART_PASS1)
12972 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
12973 * multi-char folds are allowed. */
12975 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12980 Set_Node_Offset(ret, parse_start);
12981 Set_Node_Cur_Length(ret, parse_start - 2);
12982 nextchar(pRExC_state);
12985 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
12986 * \N{...} evaluates to a sequence of more than one code points).
12987 * The function call below returns a regnode, which is our result.
12988 * The parameters cause it to fail if the \N{} evaluates to a
12989 * single code point; we handle those like any other literal. The
12990 * reason that the multicharacter case is handled here and not as
12991 * part of the EXACtish code is because of quantifiers. In
12992 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
12993 * this way makes that Just Happen. dmq.
12994 * join_exact() will join this up with adjacent EXACTish nodes
12995 * later on, if appropriate. */
12997 if (grok_bslash_N(pRExC_state,
12998 &ret, /* Want a regnode returned */
12999 NULL, /* Fail if evaluates to a single code
13001 NULL, /* Don't need a count of how many code
13010 if (*flagp & RESTART_PASS1)
13013 /* Here, evaluates to a single code point. Go get that */
13014 RExC_parse = parse_start;
13017 case 'k': /* Handle \k<NAME> and \k'NAME' */
13021 if ( RExC_parse >= RExC_end - 1
13022 || (( ch = RExC_parse[1]) != '<'
13027 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
13028 vFAIL2("Sequence %.2s... not terminated",parse_start);
13031 ret = handle_named_backref(pRExC_state,
13043 case '1': case '2': case '3': case '4':
13044 case '5': case '6': case '7': case '8': case '9':
13049 if (*RExC_parse == 'g') {
13053 if (*RExC_parse == '{') {
13057 if (*RExC_parse == '-') {
13061 if (hasbrace && !isDIGIT(*RExC_parse)) {
13062 if (isrel) RExC_parse--;
13064 goto parse_named_seq;
13067 if (RExC_parse >= RExC_end) {
13068 goto unterminated_g;
13070 num = S_backref_value(RExC_parse);
13072 vFAIL("Reference to invalid group 0");
13073 else if (num == I32_MAX) {
13074 if (isDIGIT(*RExC_parse))
13075 vFAIL("Reference to nonexistent group");
13078 vFAIL("Unterminated \\g... pattern");
13082 num = RExC_npar - num;
13084 vFAIL("Reference to nonexistent or unclosed group");
13088 num = S_backref_value(RExC_parse);
13089 /* bare \NNN might be backref or octal - if it is larger
13090 * than or equal RExC_npar then it is assumed to be an
13091 * octal escape. Note RExC_npar is +1 from the actual
13092 * number of parens. */
13093 /* Note we do NOT check if num == I32_MAX here, as that is
13094 * handled by the RExC_npar check */
13097 /* any numeric escape < 10 is always a backref */
13099 /* any numeric escape < RExC_npar is a backref */
13100 && num >= RExC_npar
13101 /* cannot be an octal escape if it starts with 8 */
13102 && *RExC_parse != '8'
13103 /* cannot be an octal escape it it starts with 9 */
13104 && *RExC_parse != '9'
13107 /* Probably not a backref, instead likely to be an
13108 * octal character escape, e.g. \35 or \777.
13109 * The above logic should make it obvious why using
13110 * octal escapes in patterns is problematic. - Yves */
13111 RExC_parse = parse_start;
13116 /* At this point RExC_parse points at a numeric escape like
13117 * \12 or \88 or something similar, which we should NOT treat
13118 * as an octal escape. It may or may not be a valid backref
13119 * escape. For instance \88888888 is unlikely to be a valid
13121 while (isDIGIT(*RExC_parse))
13124 if (*RExC_parse != '}')
13125 vFAIL("Unterminated \\g{...} pattern");
13129 if (num > (I32)RExC_rx->nparens)
13130 vFAIL("Reference to nonexistent group");
13133 ret = reganode(pRExC_state,
13136 : (ASCII_FOLD_RESTRICTED)
13138 : (AT_LEAST_UNI_SEMANTICS)
13144 *flagp |= HASWIDTH;
13146 /* override incorrect value set in reganode MJD */
13147 Set_Node_Offset(ret, parse_start);
13148 Set_Node_Cur_Length(ret, parse_start-1);
13149 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13150 FALSE /* Don't force to /x */ );
13154 if (RExC_parse >= RExC_end)
13155 FAIL("Trailing \\");
13158 /* Do not generate "unrecognized" warnings here, we fall
13159 back into the quick-grab loop below */
13160 RExC_parse = parse_start;
13162 } /* end of switch on a \foo sequence */
13167 /* '#' comments should have been spaced over before this function was
13169 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13171 if (RExC_flags & RXf_PMf_EXTENDED) {
13172 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13173 if (RExC_parse < RExC_end)
13183 /* Here, we have determined that the next thing is probably a
13184 * literal character. RExC_parse points to the first byte of its
13185 * definition. (It still may be an escape sequence that evaluates
13186 * to a single character) */
13192 #define MAX_NODE_STRING_SIZE 127
13193 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
13195 U8 upper_parse = MAX_NODE_STRING_SIZE;
13196 U8 node_type = compute_EXACTish(pRExC_state);
13197 bool next_is_quantifier;
13198 char * oldp = NULL;
13200 /* We can convert EXACTF nodes to EXACTFU if they contain only
13201 * characters that match identically regardless of the target
13202 * string's UTF8ness. The reason to do this is that EXACTF is not
13203 * trie-able, EXACTFU is.
13205 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13206 * contain only above-Latin1 characters (hence must be in UTF8),
13207 * which don't participate in folds with Latin1-range characters,
13208 * as the latter's folds aren't known until runtime. (We don't
13209 * need to figure this out until pass 2) */
13210 bool maybe_exactfu = PASS2
13211 && (node_type == EXACTF || node_type == EXACTFL);
13213 /* If a folding node contains only code points that don't
13214 * participate in folds, it can be changed into an EXACT node,
13215 * which allows the optimizer more things to look for */
13218 ret = reg_node(pRExC_state, node_type);
13220 /* In pass1, folded, we use a temporary buffer instead of the
13221 * actual node, as the node doesn't exist yet */
13222 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
13228 /* We look for the EXACTFish to EXACT node optimizaton only if
13229 * folding. (And we don't need to figure this out until pass 2).
13230 * XXX It might actually make sense to split the node into portions
13231 * that are exact and ones that aren't, so that we could later use
13232 * the exact ones to find the longest fixed and floating strings.
13233 * One would want to join them back into a larger node. One could
13234 * use a pseudo regnode like 'EXACT_ORIG_FOLD' */
13235 maybe_exact = FOLD && PASS2;
13237 /* XXX The node can hold up to 255 bytes, yet this only goes to
13238 * 127. I (khw) do not know why. Keeping it somewhat less than
13239 * 255 allows us to not have to worry about overflow due to
13240 * converting to utf8 and fold expansion, but that value is
13241 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
13242 * split up by this limit into a single one using the real max of
13243 * 255. Even at 127, this breaks under rare circumstances. If
13244 * folding, we do not want to split a node at a character that is a
13245 * non-final in a multi-char fold, as an input string could just
13246 * happen to want to match across the node boundary. The join
13247 * would solve that problem if the join actually happens. But a
13248 * series of more than two nodes in a row each of 127 would cause
13249 * the first join to succeed to get to 254, but then there wouldn't
13250 * be room for the next one, which could at be one of those split
13251 * multi-char folds. I don't know of any fool-proof solution. One
13252 * could back off to end with only a code point that isn't such a
13253 * non-final, but it is possible for there not to be any in the
13256 assert( ! UTF /* Is at the beginning of a character */
13257 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13258 || UTF8_IS_START(UCHARAT(RExC_parse)));
13260 /* Here, we have a literal character. Find the maximal string of
13261 * them in the input that we can fit into a single EXACTish node.
13262 * We quit at the first non-literal or when the node gets full */
13263 for (p = RExC_parse;
13264 len < upper_parse && p < RExC_end;
13269 /* White space has already been ignored */
13270 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13271 || ! is_PATWS_safe((p), RExC_end, UTF));
13283 /* Literal Escapes Switch
13285 This switch is meant to handle escape sequences that
13286 resolve to a literal character.
13288 Every escape sequence that represents something
13289 else, like an assertion or a char class, is handled
13290 in the switch marked 'Special Escapes' above in this
13291 routine, but also has an entry here as anything that
13292 isn't explicitly mentioned here will be treated as
13293 an unescaped equivalent literal.
13296 switch ((U8)*++p) {
13297 /* These are all the special escapes. */
13298 case 'A': /* Start assertion */
13299 case 'b': case 'B': /* Word-boundary assertion*/
13300 case 'C': /* Single char !DANGEROUS! */
13301 case 'd': case 'D': /* digit class */
13302 case 'g': case 'G': /* generic-backref, pos assertion */
13303 case 'h': case 'H': /* HORIZWS */
13304 case 'k': case 'K': /* named backref, keep marker */
13305 case 'p': case 'P': /* Unicode property */
13306 case 'R': /* LNBREAK */
13307 case 's': case 'S': /* space class */
13308 case 'v': case 'V': /* VERTWS */
13309 case 'w': case 'W': /* word class */
13310 case 'X': /* eXtended Unicode "combining
13311 character sequence" */
13312 case 'z': case 'Z': /* End of line/string assertion */
13316 /* Anything after here is an escape that resolves to a
13317 literal. (Except digits, which may or may not)
13323 case 'N': /* Handle a single-code point named character. */
13324 RExC_parse = p + 1;
13325 if (! grok_bslash_N(pRExC_state,
13326 NULL, /* Fail if evaluates to
13327 anything other than a
13328 single code point */
13329 &ender, /* The returned single code
13331 NULL, /* Don't need a count of
13332 how many code points */
13337 if (*flagp & NEED_UTF8)
13338 FAIL("panic: grok_bslash_N set NEED_UTF8");
13339 if (*flagp & RESTART_PASS1)
13342 /* Here, it wasn't a single code point. Go close
13343 * up this EXACTish node. The switch() prior to
13344 * this switch handles the other cases */
13345 RExC_parse = p = oldp;
13349 RExC_parse = parse_start;
13350 if (ender > 0xff) {
13351 REQUIRE_UTF8(flagp);
13367 ender = ESC_NATIVE;
13377 const char* error_msg;
13379 bool valid = grok_bslash_o(&p,
13383 PASS2, /* out warnings */
13384 (bool) RExC_strict,
13385 TRUE, /* Output warnings
13390 RExC_parse = p; /* going to die anyway; point
13391 to exact spot of failure */
13395 if (ender > 0xff) {
13396 REQUIRE_UTF8(flagp);
13402 UV result = UV_MAX; /* initialize to erroneous
13404 const char* error_msg;
13406 bool valid = grok_bslash_x(&p,
13410 PASS2, /* out warnings */
13411 (bool) RExC_strict,
13412 TRUE, /* Silence warnings
13417 RExC_parse = p; /* going to die anyway; point
13418 to exact spot of failure */
13423 if (ender < 0x100) {
13425 if (RExC_recode_x_to_native) {
13426 ender = LATIN1_TO_NATIVE(ender);
13431 REQUIRE_UTF8(flagp);
13437 ender = grok_bslash_c(*p++, PASS2);
13439 case '8': case '9': /* must be a backreference */
13441 /* we have an escape like \8 which cannot be an octal escape
13442 * so we exit the loop, and let the outer loop handle this
13443 * escape which may or may not be a legitimate backref. */
13445 case '1': case '2': case '3':case '4':
13446 case '5': case '6': case '7':
13447 /* When we parse backslash escapes there is ambiguity
13448 * between backreferences and octal escapes. Any escape
13449 * from \1 - \9 is a backreference, any multi-digit
13450 * escape which does not start with 0 and which when
13451 * evaluated as decimal could refer to an already
13452 * parsed capture buffer is a back reference. Anything
13455 * Note this implies that \118 could be interpreted as
13456 * 118 OR as "\11" . "8" depending on whether there
13457 * were 118 capture buffers defined already in the
13460 /* NOTE, RExC_npar is 1 more than the actual number of
13461 * parens we have seen so far, hence the < RExC_npar below. */
13463 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
13464 { /* Not to be treated as an octal constant, go
13472 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13474 ender = grok_oct(p, &numlen, &flags, NULL);
13475 if (ender > 0xff) {
13476 REQUIRE_UTF8(flagp);
13479 if (PASS2 /* like \08, \178 */
13481 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
13483 reg_warn_non_literal_string(
13485 form_short_octal_warning(p, numlen));
13491 FAIL("Trailing \\");
13494 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
13495 /* Include any left brace following the alpha to emphasize
13496 * that it could be part of an escape at some point
13498 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
13499 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
13501 goto normal_default;
13502 } /* End of switch on '\' */
13505 /* Currently we allow an lbrace at the start of a construct
13506 * without raising a warning. This is because we think we
13507 * will never want such a brace to be meant to be other
13508 * than taken literally. */
13509 if (len || (p > RExC_start && isALPHA_A(*(p - 1)))) {
13511 /* But, we raise a fatal warning otherwise, as the
13512 * deprecation cycle has come and gone. Except that it
13513 * turns out that some heavily-relied on upstream
13514 * software, notably GNU Autoconf, have failed to fix
13515 * their uses. For these, don't make it fatal unless
13516 * we anticipate using the '{' for something else.
13517 * This happens after any alpha, and for a looser {m,n}
13518 * quantifier specification */
13520 || ( p > parse_start + 1
13521 && isALPHA_A(*(p - 1))
13522 && *(p - 2) == '\\')
13523 || new_regcurly(p, RExC_end))
13525 RExC_parse = p + 1;
13526 vFAIL("Unescaped left brace in regex is "
13530 ckWARNregdep(p + 1,
13531 "Unescaped left brace in regex is "
13532 "deprecated here (and will be fatal "
13533 "in Perl 5.30), passed through");
13536 goto normal_default;
13539 if (PASS2 && p > RExC_parse && RExC_strict) {
13540 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
13543 default: /* A literal character */
13545 if (! UTF8_IS_INVARIANT(*p) && UTF) {
13547 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
13548 &numlen, UTF8_ALLOW_DEFAULT);
13554 } /* End of switch on the literal */
13556 /* Here, have looked at the literal character and <ender>
13557 * contains its ordinal, <p> points to the character after it.
13558 * We need to check if the next non-ignored thing is a
13559 * quantifier. Move <p> to after anything that should be
13560 * ignored, which, as a side effect, positions <p> for the next
13561 * loop iteration */
13562 skip_to_be_ignored_text(pRExC_state, &p,
13563 FALSE /* Don't force to /x */ );
13565 /* If the next thing is a quantifier, it applies to this
13566 * character only, which means that this character has to be in
13567 * its own node and can't just be appended to the string in an
13568 * existing node, so if there are already other characters in
13569 * the node, close the node with just them, and set up to do
13570 * this character again next time through, when it will be the
13571 * only thing in its new node */
13573 next_is_quantifier = LIKELY(p < RExC_end)
13574 && UNLIKELY(ISMULT2(p));
13576 if (next_is_quantifier && LIKELY(len)) {
13581 /* Ready to add 'ender' to the node */
13583 if (! FOLD) { /* The simple case, just append the literal */
13585 /* In the sizing pass, we need only the size of the
13586 * character we are appending, hence we can delay getting
13587 * its representation until PASS2. */
13589 if (UTF && ! UVCHR_IS_INVARIANT(ender)) {
13590 const STRLEN unilen = UVCHR_SKIP(ender);
13593 /* We have to subtract 1 just below (and again in
13594 * the corresponding PASS2 code) because the loop
13595 * increments <len> each time, as all but this path
13596 * (and one other) through it add a single byte to
13597 * the EXACTish node. But these paths would change
13598 * len to be the correct final value, so cancel out
13599 * the increment that follows */
13605 } else { /* PASS2 */
13607 if (UTF && ! UVCHR_IS_INVARIANT(ender)) {
13608 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
13609 len += (char *) new_s - s - 1;
13610 s = (char *) new_s;
13613 *(s++) = (char) ender;
13617 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
13619 /* Here are folding under /l, and the code point is
13620 * problematic. First, we know we can't simplify things */
13621 maybe_exact = FALSE;
13622 maybe_exactfu = FALSE;
13624 /* A problematic code point in this context means that its
13625 * fold isn't known until runtime, so we can't fold it now.
13626 * (The non-problematic code points are the above-Latin1
13627 * ones that fold to also all above-Latin1. Their folds
13628 * don't vary no matter what the locale is.) But here we
13629 * have characters whose fold depends on the locale.
13630 * Unlike the non-folding case above, we have to keep track
13631 * of these in the sizing pass, so that we can make sure we
13632 * don't split too-long nodes in the middle of a potential
13633 * multi-char fold. And unlike the regular fold case
13634 * handled in the else clauses below, we don't actually
13635 * fold and don't have special cases to consider. What we
13636 * do for both passes is the PASS2 code for non-folding */
13637 goto not_fold_common;
13639 else /* A regular FOLD code point */
13641 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13642 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13643 || UNICODE_DOT_DOT_VERSION > 0)
13644 /* See comments for join_exact() as to why we fold
13645 * this non-UTF at compile time */
13646 || ( node_type == EXACTFU
13647 && ender == LATIN_SMALL_LETTER_SHARP_S)
13650 /* Here, are folding and are not UTF-8 encoded; therefore
13651 * the character must be in the range 0-255, and is not /l
13652 * (Not /l because we already handled these under /l in
13653 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
13654 if (IS_IN_SOME_FOLD_L1(ender)) {
13655 maybe_exact = FALSE;
13657 /* See if the character's fold differs between /d and
13658 * /u. This includes the multi-char fold SHARP S to
13660 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
13661 RExC_seen_unfolded_sharp_s = 1;
13662 maybe_exactfu = FALSE;
13664 else if (maybe_exactfu
13665 && (PL_fold[ender] != PL_fold_latin1[ender]
13666 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13667 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13668 || UNICODE_DOT_DOT_VERSION > 0)
13670 && isALPHA_FOLD_EQ(ender, 's')
13671 && isALPHA_FOLD_EQ(*(s-1), 's'))
13674 maybe_exactfu = FALSE;
13678 /* Even when folding, we store just the input character, as
13679 * we have an array that finds its fold quickly */
13680 *(s++) = (char) ender;
13682 else { /* FOLD, and UTF (or sharp s) */
13683 /* Unlike the non-fold case, we do actually have to
13684 * calculate the results here in pass 1. This is for two
13685 * reasons, the folded length may be longer than the
13686 * unfolded, and we have to calculate how many EXACTish
13687 * nodes it will take; and we may run out of room in a node
13688 * in the middle of a potential multi-char fold, and have
13689 * to back off accordingly. */
13692 if (isASCII_uni(ender)) {
13693 folded = toFOLD(ender);
13694 *(s)++ = (U8) folded;
13699 folded = _to_uni_fold_flags(
13703 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
13704 ? FOLD_FLAGS_NOMIX_ASCII
13708 /* The loop increments <len> each time, as all but this
13709 * path (and one other) through it add a single byte to
13710 * the EXACTish node. But this one has changed len to
13711 * be the correct final value, so subtract one to
13712 * cancel out the increment that follows */
13713 len += foldlen - 1;
13715 /* If this node only contains non-folding code points so
13716 * far, see if this new one is also non-folding */
13718 if (folded != ender) {
13719 maybe_exact = FALSE;
13722 /* Here the fold is the original; we have to check
13723 * further to see if anything folds to it */
13724 if (_invlist_contains_cp(PL_utf8_foldable,
13727 maybe_exact = FALSE;
13734 if (next_is_quantifier) {
13736 /* Here, the next input is a quantifier, and to get here,
13737 * the current character is the only one in the node.
13738 * Also, here <len> doesn't include the final byte for this
13744 } /* End of loop through literal characters */
13746 /* Here we have either exhausted the input or ran out of room in
13747 * the node. (If we encountered a character that can't be in the
13748 * node, transfer is made directly to <loopdone>, and so we
13749 * wouldn't have fallen off the end of the loop.) In the latter
13750 * case, we artificially have to split the node into two, because
13751 * we just don't have enough space to hold everything. This
13752 * creates a problem if the final character participates in a
13753 * multi-character fold in the non-final position, as a match that
13754 * should have occurred won't, due to the way nodes are matched,
13755 * and our artificial boundary. So back off until we find a non-
13756 * problematic character -- one that isn't at the beginning or
13757 * middle of such a fold. (Either it doesn't participate in any
13758 * folds, or appears only in the final position of all the folds it
13759 * does participate in.) A better solution with far fewer false
13760 * positives, and that would fill the nodes more completely, would
13761 * be to actually have available all the multi-character folds to
13762 * test against, and to back-off only far enough to be sure that
13763 * this node isn't ending with a partial one. <upper_parse> is set
13764 * further below (if we need to reparse the node) to include just
13765 * up through that final non-problematic character that this code
13766 * identifies, so when it is set to less than the full node, we can
13767 * skip the rest of this */
13768 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
13770 const STRLEN full_len = len;
13772 assert(len >= MAX_NODE_STRING_SIZE);
13774 /* Here, <s> points to the final byte of the final character.
13775 * Look backwards through the string until find a non-
13776 * problematic character */
13780 /* This has no multi-char folds to non-UTF characters */
13781 if (ASCII_FOLD_RESTRICTED) {
13785 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
13789 if (! PL_NonL1NonFinalFold) {
13790 PL_NonL1NonFinalFold = _new_invlist_C_array(
13791 NonL1_Perl_Non_Final_Folds_invlist);
13794 /* Point to the first byte of the final character */
13795 s = (char *) utf8_hop((U8 *) s, -1);
13797 while (s >= s0) { /* Search backwards until find
13798 non-problematic char */
13799 if (UTF8_IS_INVARIANT(*s)) {
13801 /* There are no ascii characters that participate
13802 * in multi-char folds under /aa. In EBCDIC, the
13803 * non-ascii invariants are all control characters,
13804 * so don't ever participate in any folds. */
13805 if (ASCII_FOLD_RESTRICTED
13806 || ! IS_NON_FINAL_FOLD(*s))
13811 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
13812 if (! IS_NON_FINAL_FOLD(EIGHT_BIT_UTF8_TO_NATIVE(
13818 else if (! _invlist_contains_cp(
13819 PL_NonL1NonFinalFold,
13820 valid_utf8_to_uvchr((U8 *) s, NULL)))
13825 /* Here, the current character is problematic in that
13826 * it does occur in the non-final position of some
13827 * fold, so try the character before it, but have to
13828 * special case the very first byte in the string, so
13829 * we don't read outside the string */
13830 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
13831 } /* End of loop backwards through the string */
13833 /* If there were only problematic characters in the string,
13834 * <s> will point to before s0, in which case the length
13835 * should be 0, otherwise include the length of the
13836 * non-problematic character just found */
13837 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
13840 /* Here, have found the final character, if any, that is
13841 * non-problematic as far as ending the node without splitting
13842 * it across a potential multi-char fold. <len> contains the
13843 * number of bytes in the node up-to and including that
13844 * character, or is 0 if there is no such character, meaning
13845 * the whole node contains only problematic characters. In
13846 * this case, give up and just take the node as-is. We can't
13851 /* If the node ends in an 's' we make sure it stays EXACTF,
13852 * as if it turns into an EXACTFU, it could later get
13853 * joined with another 's' that would then wrongly match
13855 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
13857 maybe_exactfu = FALSE;
13861 /* Here, the node does contain some characters that aren't
13862 * problematic. If one such is the final character in the
13863 * node, we are done */
13864 if (len == full_len) {
13867 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
13869 /* If the final character is problematic, but the
13870 * penultimate is not, back-off that last character to
13871 * later start a new node with it */
13876 /* Here, the final non-problematic character is earlier
13877 * in the input than the penultimate character. What we do
13878 * is reparse from the beginning, going up only as far as
13879 * this final ok one, thus guaranteeing that the node ends
13880 * in an acceptable character. The reason we reparse is
13881 * that we know how far in the character is, but we don't
13882 * know how to correlate its position with the input parse.
13883 * An alternate implementation would be to build that
13884 * correlation as we go along during the original parse,
13885 * but that would entail extra work for every node, whereas
13886 * this code gets executed only when the string is too
13887 * large for the node, and the final two characters are
13888 * problematic, an infrequent occurrence. Yet another
13889 * possible strategy would be to save the tail of the
13890 * string, and the next time regatom is called, initialize
13891 * with that. The problem with this is that unless you
13892 * back off one more character, you won't be guaranteed
13893 * regatom will get called again, unless regbranch,
13894 * regpiece ... are also changed. If you do back off that
13895 * extra character, so that there is input guaranteed to
13896 * force calling regatom, you can't handle the case where
13897 * just the first character in the node is acceptable. I
13898 * (khw) decided to try this method which doesn't have that
13899 * pitfall; if performance issues are found, we can do a
13900 * combination of the current approach plus that one */
13906 } /* End of verifying node ends with an appropriate char */
13908 loopdone: /* Jumped to when encounters something that shouldn't be
13911 /* I (khw) don't know if you can get here with zero length, but the
13912 * old code handled this situation by creating a zero-length EXACT
13913 * node. Might as well be NOTHING instead */
13919 /* If 'maybe_exact' is still set here, means there are no
13920 * code points in the node that participate in folds;
13921 * similarly for 'maybe_exactfu' and code points that match
13922 * differently depending on UTF8ness of the target string
13923 * (for /u), or depending on locale for /l */
13929 else if (maybe_exactfu) {
13935 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
13936 FALSE /* Don't look to see if could
13937 be turned into an EXACT
13938 node, as we have already
13943 RExC_parse = p - 1;
13944 Set_Node_Cur_Length(ret, parse_start);
13947 /* len is STRLEN which is unsigned, need to copy to signed */
13950 vFAIL("Internal disaster");
13953 } /* End of label 'defchar:' */
13955 } /* End of giant switch on input character */
13957 /* Position parse to next real character */
13958 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13959 FALSE /* Don't force to /x */ );
13960 if (PASS2 && *RExC_parse == '{' && OP(ret) != SBOL && ! regcurly(RExC_parse)) {
13961 ckWARNregdep(RExC_parse + 1, "Unescaped left brace in regex is deprecated here (and will be fatal in Perl 5.30), passed through");
13969 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
13971 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
13972 * sets up the bitmap and any flags, removing those code points from the
13973 * inversion list, setting it to NULL should it become completely empty */
13975 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
13976 assert(PL_regkind[OP(node)] == ANYOF);
13978 ANYOF_BITMAP_ZERO(node);
13979 if (*invlist_ptr) {
13981 /* This gets set if we actually need to modify things */
13982 bool change_invlist = FALSE;
13986 /* Start looking through *invlist_ptr */
13987 invlist_iterinit(*invlist_ptr);
13988 while (invlist_iternext(*invlist_ptr, &start, &end)) {
13992 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
13993 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
13996 /* Quit if are above what we should change */
13997 if (start >= NUM_ANYOF_CODE_POINTS) {
14001 change_invlist = TRUE;
14003 /* Set all the bits in the range, up to the max that we are doing */
14004 high = (end < NUM_ANYOF_CODE_POINTS - 1)
14006 : NUM_ANYOF_CODE_POINTS - 1;
14007 for (i = start; i <= (int) high; i++) {
14008 if (! ANYOF_BITMAP_TEST(node, i)) {
14009 ANYOF_BITMAP_SET(node, i);
14013 invlist_iterfinish(*invlist_ptr);
14015 /* Done with loop; remove any code points that are in the bitmap from
14016 * *invlist_ptr; similarly for code points above the bitmap if we have
14017 * a flag to match all of them anyways */
14018 if (change_invlist) {
14019 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
14021 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
14022 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
14025 /* If have completely emptied it, remove it completely */
14026 if (_invlist_len(*invlist_ptr) == 0) {
14027 SvREFCNT_dec_NN(*invlist_ptr);
14028 *invlist_ptr = NULL;
14033 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
14034 Character classes ([:foo:]) can also be negated ([:^foo:]).
14035 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
14036 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
14037 but trigger failures because they are currently unimplemented. */
14039 #define POSIXCC_DONE(c) ((c) == ':')
14040 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
14041 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
14042 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
14044 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
14045 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
14046 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
14048 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
14050 /* 'posix_warnings' and 'warn_text' are names of variables in the following
14052 #define ADD_POSIX_WARNING(p, text) STMT_START { \
14053 if (posix_warnings) { \
14054 if (! RExC_warn_text ) RExC_warn_text = (AV *) sv_2mortal((SV *) newAV()); \
14055 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
14059 REPORT_LOCATION_ARGS(p))); \
14062 #define CLEAR_POSIX_WARNINGS() \
14064 if (posix_warnings && RExC_warn_text) \
14065 av_clear(RExC_warn_text); \
14068 #define CLEAR_POSIX_WARNINGS_AND_RETURN(ret) \
14070 CLEAR_POSIX_WARNINGS(); \
14075 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
14077 const char * const s, /* Where the putative posix class begins.
14078 Normally, this is one past the '['. This
14079 parameter exists so it can be somewhere
14080 besides RExC_parse. */
14081 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
14083 AV ** posix_warnings, /* Where to place any generated warnings, or
14085 const bool check_only /* Don't die if error */
14088 /* This parses what the caller thinks may be one of the three POSIX
14090 * 1) a character class, like [:blank:]
14091 * 2) a collating symbol, like [. .]
14092 * 3) an equivalence class, like [= =]
14093 * In the latter two cases, it croaks if it finds a syntactically legal
14094 * one, as these are not handled by Perl.
14096 * The main purpose is to look for a POSIX character class. It returns:
14097 * a) the class number
14098 * if it is a completely syntactically and semantically legal class.
14099 * 'updated_parse_ptr', if not NULL, is set to point to just after the
14100 * closing ']' of the class
14101 * b) OOB_NAMEDCLASS
14102 * if it appears that one of the three POSIX constructs was meant, but
14103 * its specification was somehow defective. 'updated_parse_ptr', if
14104 * not NULL, is set to point to the character just after the end
14105 * character of the class. See below for handling of warnings.
14106 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
14107 * if it doesn't appear that a POSIX construct was intended.
14108 * 'updated_parse_ptr' is not changed. No warnings nor errors are
14111 * In b) there may be errors or warnings generated. If 'check_only' is
14112 * TRUE, then any errors are discarded. Warnings are returned to the
14113 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
14114 * instead it is NULL, warnings are suppressed. This is done in all
14115 * passes. The reason for this is that the rest of the parsing is heavily
14116 * dependent on whether this routine found a valid posix class or not. If
14117 * it did, the closing ']' is absorbed as part of the class. If no class,
14118 * or an invalid one is found, any ']' will be considered the terminator of
14119 * the outer bracketed character class, leading to very different results.
14120 * In particular, a '(?[ ])' construct will likely have a syntax error if
14121 * the class is parsed other than intended, and this will happen in pass1,
14122 * before the warnings would normally be output. This mechanism allows the
14123 * caller to output those warnings in pass1 just before dieing, giving a
14124 * much better clue as to what is wrong.
14126 * The reason for this function, and its complexity is that a bracketed
14127 * character class can contain just about anything. But it's easy to
14128 * mistype the very specific posix class syntax but yielding a valid
14129 * regular bracketed class, so it silently gets compiled into something
14130 * quite unintended.
14132 * The solution adopted here maintains backward compatibility except that
14133 * it adds a warning if it looks like a posix class was intended but
14134 * improperly specified. The warning is not raised unless what is input
14135 * very closely resembles one of the 14 legal posix classes. To do this,
14136 * it uses fuzzy parsing. It calculates how many single-character edits it
14137 * would take to transform what was input into a legal posix class. Only
14138 * if that number is quite small does it think that the intention was a
14139 * posix class. Obviously these are heuristics, and there will be cases
14140 * where it errs on one side or another, and they can be tweaked as
14141 * experience informs.
14143 * The syntax for a legal posix class is:
14145 * qr/(?xa: \[ : \^? [[:lower:]]{4,6} : \] )/
14147 * What this routine considers syntactically to be an intended posix class
14148 * is this (the comments indicate some restrictions that the pattern
14151 * qr/(?x: \[? # The left bracket, possibly
14153 * \h* # possibly followed by blanks
14154 * (?: \^ \h* )? # possibly a misplaced caret
14155 * [:;]? # The opening class character,
14156 * # possibly omitted. A typo
14157 * # semi-colon can also be used.
14159 * \^? # possibly a correctly placed
14160 * # caret, but not if there was also
14161 * # a misplaced one
14163 * .{3,15} # The class name. If there are
14164 * # deviations from the legal syntax,
14165 * # its edit distance must be close
14166 * # to a real class name in order
14167 * # for it to be considered to be
14168 * # an intended posix class.
14170 * [[:punct:]]? # The closing class character,
14171 * # possibly omitted. If not a colon
14172 * # nor semi colon, the class name
14173 * # must be even closer to a valid
14176 * \]? # The right bracket, possibly
14180 * In the above, \h must be ASCII-only.
14182 * These are heuristics, and can be tweaked as field experience dictates.
14183 * There will be cases when someone didn't intend to specify a posix class
14184 * that this warns as being so. The goal is to minimize these, while
14185 * maximizing the catching of things intended to be a posix class that
14186 * aren't parsed as such.
14190 const char * const e = RExC_end;
14191 unsigned complement = 0; /* If to complement the class */
14192 bool found_problem = FALSE; /* Assume OK until proven otherwise */
14193 bool has_opening_bracket = FALSE;
14194 bool has_opening_colon = FALSE;
14195 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
14197 const char * possible_end = NULL; /* used for a 2nd parse pass */
14198 const char* name_start; /* ptr to class name first char */
14200 /* If the number of single-character typos the input name is away from a
14201 * legal name is no more than this number, it is considered to have meant
14202 * the legal name */
14203 int max_distance = 2;
14205 /* to store the name. The size determines the maximum length before we
14206 * decide that no posix class was intended. Should be at least
14207 * sizeof("alphanumeric") */
14209 STATIC_ASSERT_DECL(C_ARRAY_LENGTH(input_text) >= sizeof "alphanumeric");
14211 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
14213 CLEAR_POSIX_WARNINGS();
14216 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14219 if (*(p - 1) != '[') {
14220 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
14221 found_problem = TRUE;
14224 has_opening_bracket = TRUE;
14227 /* They could be confused and think you can put spaces between the
14230 found_problem = TRUE;
14234 } while (p < e && isBLANK(*p));
14236 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14239 /* For [. .] and [= =]. These are quite different internally from [: :],
14240 * so they are handled separately. */
14241 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
14242 and 1 for at least one char in it
14245 const char open_char = *p;
14246 const char * temp_ptr = p + 1;
14248 /* These two constructs are not handled by perl, and if we find a
14249 * syntactically valid one, we croak. khw, who wrote this code, finds
14250 * this explanation of them very unclear:
14251 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
14252 * And searching the rest of the internet wasn't very helpful either.
14253 * It looks like just about any byte can be in these constructs,
14254 * depending on the locale. But unless the pattern is being compiled
14255 * under /l, which is very rare, Perl runs under the C or POSIX locale.
14256 * In that case, it looks like [= =] isn't allowed at all, and that
14257 * [. .] could be any single code point, but for longer strings the
14258 * constituent characters would have to be the ASCII alphabetics plus
14259 * the minus-hyphen. Any sensible locale definition would limit itself
14260 * to these. And any portable one definitely should. Trying to parse
14261 * the general case is a nightmare (see [perl #127604]). So, this code
14262 * looks only for interiors of these constructs that match:
14264 * Using \w relaxes the apparent rules a little, without adding much
14265 * danger of mistaking something else for one of these constructs.
14267 * [. .] in some implementations described on the internet is usable to
14268 * escape a character that otherwise is special in bracketed character
14269 * classes. For example [.].] means a literal right bracket instead of
14270 * the ending of the class
14272 * [= =] can legitimately contain a [. .] construct, but we don't
14273 * handle this case, as that [. .] construct will later get parsed
14274 * itself and croak then. And [= =] is checked for even when not under
14275 * /l, as Perl has long done so.
14277 * The code below relies on there being a trailing NUL, so it doesn't
14278 * have to keep checking if the parse ptr < e.
14280 if (temp_ptr[1] == open_char) {
14283 else while ( temp_ptr < e
14284 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
14289 if (*temp_ptr == open_char) {
14291 if (*temp_ptr == ']') {
14293 if (! found_problem && ! check_only) {
14294 RExC_parse = (char *) temp_ptr;
14295 vFAIL3("POSIX syntax [%c %c] is reserved for future "
14296 "extensions", open_char, open_char);
14299 /* Here, the syntax wasn't completely valid, or else the call
14300 * is to check-only */
14301 if (updated_parse_ptr) {
14302 *updated_parse_ptr = (char *) temp_ptr;
14305 CLEAR_POSIX_WARNINGS_AND_RETURN(OOB_NAMEDCLASS);
14309 /* If we find something that started out to look like one of these
14310 * constructs, but isn't, we continue below so that it can be checked
14311 * for being a class name with a typo of '.' or '=' instead of a colon.
14315 /* Here, we think there is a possibility that a [: :] class was meant, and
14316 * we have the first real character. It could be they think the '^' comes
14319 found_problem = TRUE;
14320 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
14325 found_problem = TRUE;
14329 } while (p < e && isBLANK(*p));
14331 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14335 /* But the first character should be a colon, which they could have easily
14336 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
14337 * distinguish from a colon, so treat that as a colon). */
14340 has_opening_colon = TRUE;
14342 else if (*p == ';') {
14343 found_problem = TRUE;
14345 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14346 has_opening_colon = TRUE;
14349 found_problem = TRUE;
14350 ADD_POSIX_WARNING(p, "there must be a starting ':'");
14352 /* Consider an initial punctuation (not one of the recognized ones) to
14353 * be a left terminator */
14354 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
14359 /* They may think that you can put spaces between the components */
14361 found_problem = TRUE;
14365 } while (p < e && isBLANK(*p));
14367 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14372 /* We consider something like [^:^alnum:]] to not have been intended to
14373 * be a posix class, but XXX maybe we should */
14375 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14382 /* Again, they may think that you can put spaces between the components */
14384 found_problem = TRUE;
14388 } while (p < e && isBLANK(*p));
14390 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14395 /* XXX This ']' may be a typo, and something else was meant. But
14396 * treating it as such creates enough complications, that that
14397 * possibility isn't currently considered here. So we assume that the
14398 * ']' is what is intended, and if we've already found an initial '[',
14399 * this leaves this construct looking like [:] or [:^], which almost
14400 * certainly weren't intended to be posix classes */
14401 if (has_opening_bracket) {
14402 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14405 /* But this function can be called when we parse the colon for
14406 * something like qr/[alpha:]]/, so we back up to look for the
14411 found_problem = TRUE;
14412 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14414 else if (*p != ':') {
14416 /* XXX We are currently very restrictive here, so this code doesn't
14417 * consider the possibility that, say, /[alpha.]]/ was intended to
14418 * be a posix class. */
14419 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14422 /* Here we have something like 'foo:]'. There was no initial colon,
14423 * and we back up over 'foo. XXX Unlike the going forward case, we
14424 * don't handle typos of non-word chars in the middle */
14425 has_opening_colon = FALSE;
14428 while (p > RExC_start && isWORDCHAR(*p)) {
14433 /* Here, we have positioned ourselves to where we think the first
14434 * character in the potential class is */
14437 /* Now the interior really starts. There are certain key characters that
14438 * can end the interior, or these could just be typos. To catch both
14439 * cases, we may have to do two passes. In the first pass, we keep on
14440 * going unless we come to a sequence that matches
14441 * qr/ [[:punct:]] [[:blank:]]* \] /xa
14442 * This means it takes a sequence to end the pass, so two typos in a row if
14443 * that wasn't what was intended. If the class is perfectly formed, just
14444 * this one pass is needed. We also stop if there are too many characters
14445 * being accumulated, but this number is deliberately set higher than any
14446 * real class. It is set high enough so that someone who thinks that
14447 * 'alphanumeric' is a correct name would get warned that it wasn't.
14448 * While doing the pass, we keep track of where the key characters were in
14449 * it. If we don't find an end to the class, and one of the key characters
14450 * was found, we redo the pass, but stop when we get to that character.
14451 * Thus the key character was considered a typo in the first pass, but a
14452 * terminator in the second. If two key characters are found, we stop at
14453 * the second one in the first pass. Again this can miss two typos, but
14454 * catches a single one
14456 * In the first pass, 'possible_end' starts as NULL, and then gets set to
14457 * point to the first key character. For the second pass, it starts as -1.
14463 bool has_blank = FALSE;
14464 bool has_upper = FALSE;
14465 bool has_terminating_colon = FALSE;
14466 bool has_terminating_bracket = FALSE;
14467 bool has_semi_colon = FALSE;
14468 unsigned int name_len = 0;
14469 int punct_count = 0;
14473 /* Squeeze out blanks when looking up the class name below */
14474 if (isBLANK(*p) ) {
14476 found_problem = TRUE;
14481 /* The name will end with a punctuation */
14483 const char * peek = p + 1;
14485 /* Treat any non-']' punctuation followed by a ']' (possibly
14486 * with intervening blanks) as trying to terminate the class.
14487 * ']]' is very likely to mean a class was intended (but
14488 * missing the colon), but the warning message that gets
14489 * generated shows the error position better if we exit the
14490 * loop at the bottom (eventually), so skip it here. */
14492 if (peek < e && isBLANK(*peek)) {
14494 found_problem = TRUE;
14497 } while (peek < e && isBLANK(*peek));
14500 if (peek < e && *peek == ']') {
14501 has_terminating_bracket = TRUE;
14503 has_terminating_colon = TRUE;
14505 else if (*p == ';') {
14506 has_semi_colon = TRUE;
14507 has_terminating_colon = TRUE;
14510 found_problem = TRUE;
14517 /* Here we have punctuation we thought didn't end the class.
14518 * Keep track of the position of the key characters that are
14519 * more likely to have been class-enders */
14520 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
14522 /* Allow just one such possible class-ender not actually
14523 * ending the class. */
14524 if (possible_end) {
14530 /* If we have too many punctuation characters, no use in
14532 if (++punct_count > max_distance) {
14536 /* Treat the punctuation as a typo. */
14537 input_text[name_len++] = *p;
14540 else if (isUPPER(*p)) { /* Use lowercase for lookup */
14541 input_text[name_len++] = toLOWER(*p);
14543 found_problem = TRUE;
14545 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
14546 input_text[name_len++] = *p;
14550 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
14554 /* The declaration of 'input_text' is how long we allow a potential
14555 * class name to be, before saying they didn't mean a class name at
14557 if (name_len >= C_ARRAY_LENGTH(input_text)) {
14562 /* We get to here when the possible class name hasn't been properly
14563 * terminated before:
14564 * 1) we ran off the end of the pattern; or
14565 * 2) found two characters, each of which might have been intended to
14566 * be the name's terminator
14567 * 3) found so many punctuation characters in the purported name,
14568 * that the edit distance to a valid one is exceeded
14569 * 4) we decided it was more characters than anyone could have
14570 * intended to be one. */
14572 found_problem = TRUE;
14574 /* In the final two cases, we know that looking up what we've
14575 * accumulated won't lead to a match, even a fuzzy one. */
14576 if ( name_len >= C_ARRAY_LENGTH(input_text)
14577 || punct_count > max_distance)
14579 /* If there was an intermediate key character that could have been
14580 * an intended end, redo the parse, but stop there */
14581 if (possible_end && possible_end != (char *) -1) {
14582 possible_end = (char *) -1; /* Special signal value to say
14583 we've done a first pass */
14588 /* Otherwise, it can't have meant to have been a class */
14589 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14592 /* If we ran off the end, and the final character was a punctuation
14593 * one, back up one, to look at that final one just below. Later, we
14594 * will restore the parse pointer if appropriate */
14595 if (name_len && p == e && isPUNCT(*(p-1))) {
14600 if (p < e && isPUNCT(*p)) {
14602 has_terminating_bracket = TRUE;
14604 /* If this is a 2nd ']', and the first one is just below this
14605 * one, consider that to be the real terminator. This gives a
14606 * uniform and better positioning for the warning message */
14608 && possible_end != (char *) -1
14609 && *possible_end == ']'
14610 && name_len && input_text[name_len - 1] == ']')
14615 /* And this is actually equivalent to having done the 2nd
14616 * pass now, so set it to not try again */
14617 possible_end = (char *) -1;
14622 has_terminating_colon = TRUE;
14624 else if (*p == ';') {
14625 has_semi_colon = TRUE;
14626 has_terminating_colon = TRUE;
14634 /* Here, we have a class name to look up. We can short circuit the
14635 * stuff below for short names that can't possibly be meant to be a
14636 * class name. (We can do this on the first pass, as any second pass
14637 * will yield an even shorter name) */
14638 if (name_len < 3) {
14639 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14642 /* Find which class it is. Initially switch on the length of the name.
14644 switch (name_len) {
14646 if (memEQs(name_start, 4, "word")) {
14647 /* this is not POSIX, this is the Perl \w */
14648 class_number = ANYOF_WORDCHAR;
14652 /* Names all of length 5: alnum alpha ascii blank cntrl digit
14653 * graph lower print punct space upper
14654 * Offset 4 gives the best switch position. */
14655 switch (name_start[4]) {
14657 if (memBEGINs(name_start, 5, "alph")) /* alpha */
14658 class_number = ANYOF_ALPHA;
14661 if (memBEGINs(name_start, 5, "spac")) /* space */
14662 class_number = ANYOF_SPACE;
14665 if (memBEGINs(name_start, 5, "grap")) /* graph */
14666 class_number = ANYOF_GRAPH;
14669 if (memBEGINs(name_start, 5, "asci")) /* ascii */
14670 class_number = ANYOF_ASCII;
14673 if (memBEGINs(name_start, 5, "blan")) /* blank */
14674 class_number = ANYOF_BLANK;
14677 if (memBEGINs(name_start, 5, "cntr")) /* cntrl */
14678 class_number = ANYOF_CNTRL;
14681 if (memBEGINs(name_start, 5, "alnu")) /* alnum */
14682 class_number = ANYOF_ALPHANUMERIC;
14685 if (memBEGINs(name_start, 5, "lowe")) /* lower */
14686 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
14687 else if (memBEGINs(name_start, 5, "uppe")) /* upper */
14688 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
14691 if (memBEGINs(name_start, 5, "digi")) /* digit */
14692 class_number = ANYOF_DIGIT;
14693 else if (memBEGINs(name_start, 5, "prin")) /* print */
14694 class_number = ANYOF_PRINT;
14695 else if (memBEGINs(name_start, 5, "punc")) /* punct */
14696 class_number = ANYOF_PUNCT;
14701 if (memEQs(name_start, 6, "xdigit"))
14702 class_number = ANYOF_XDIGIT;
14706 /* If the name exactly matches a posix class name the class number will
14707 * here be set to it, and the input almost certainly was meant to be a
14708 * posix class, so we can skip further checking. If instead the syntax
14709 * is exactly correct, but the name isn't one of the legal ones, we
14710 * will return that as an error below. But if neither of these apply,
14711 * it could be that no posix class was intended at all, or that one
14712 * was, but there was a typo. We tease these apart by doing fuzzy
14713 * matching on the name */
14714 if (class_number == OOB_NAMEDCLASS && found_problem) {
14715 const UV posix_names[][6] = {
14716 { 'a', 'l', 'n', 'u', 'm' },
14717 { 'a', 'l', 'p', 'h', 'a' },
14718 { 'a', 's', 'c', 'i', 'i' },
14719 { 'b', 'l', 'a', 'n', 'k' },
14720 { 'c', 'n', 't', 'r', 'l' },
14721 { 'd', 'i', 'g', 'i', 't' },
14722 { 'g', 'r', 'a', 'p', 'h' },
14723 { 'l', 'o', 'w', 'e', 'r' },
14724 { 'p', 'r', 'i', 'n', 't' },
14725 { 'p', 'u', 'n', 'c', 't' },
14726 { 's', 'p', 'a', 'c', 'e' },
14727 { 'u', 'p', 'p', 'e', 'r' },
14728 { 'w', 'o', 'r', 'd' },
14729 { 'x', 'd', 'i', 'g', 'i', 't' }
14731 /* The names of the above all have added NULs to make them the same
14732 * size, so we need to also have the real lengths */
14733 const UV posix_name_lengths[] = {
14734 sizeof("alnum") - 1,
14735 sizeof("alpha") - 1,
14736 sizeof("ascii") - 1,
14737 sizeof("blank") - 1,
14738 sizeof("cntrl") - 1,
14739 sizeof("digit") - 1,
14740 sizeof("graph") - 1,
14741 sizeof("lower") - 1,
14742 sizeof("print") - 1,
14743 sizeof("punct") - 1,
14744 sizeof("space") - 1,
14745 sizeof("upper") - 1,
14746 sizeof("word") - 1,
14747 sizeof("xdigit")- 1
14750 int temp_max = max_distance; /* Use a temporary, so if we
14751 reparse, we haven't changed the
14754 /* Use a smaller max edit distance if we are missing one of the
14756 if ( has_opening_bracket + has_opening_colon < 2
14757 || has_terminating_bracket + has_terminating_colon < 2)
14762 /* See if the input name is close to a legal one */
14763 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
14765 /* Short circuit call if the lengths are too far apart to be
14767 if (abs( (int) (name_len - posix_name_lengths[i]))
14773 if (edit_distance(input_text,
14776 posix_name_lengths[i],
14780 { /* If it is close, it probably was intended to be a class */
14781 goto probably_meant_to_be;
14785 /* Here the input name is not close enough to a valid class name
14786 * for us to consider it to be intended to be a posix class. If
14787 * we haven't already done so, and the parse found a character that
14788 * could have been terminators for the name, but which we absorbed
14789 * as typos during the first pass, repeat the parse, signalling it
14790 * to stop at that character */
14791 if (possible_end && possible_end != (char *) -1) {
14792 possible_end = (char *) -1;
14797 /* Here neither pass found a close-enough class name */
14798 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14801 probably_meant_to_be:
14803 /* Here we think that a posix specification was intended. Update any
14805 if (updated_parse_ptr) {
14806 *updated_parse_ptr = (char *) p;
14809 /* If a posix class name was intended but incorrectly specified, we
14810 * output or return the warnings */
14811 if (found_problem) {
14813 /* We set flags for these issues in the parse loop above instead of
14814 * adding them to the list of warnings, because we can parse it
14815 * twice, and we only want one warning instance */
14817 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
14820 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14822 if (has_semi_colon) {
14823 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14825 else if (! has_terminating_colon) {
14826 ADD_POSIX_WARNING(p, "there is no terminating ':'");
14828 if (! has_terminating_bracket) {
14829 ADD_POSIX_WARNING(p, "there is no terminating ']'");
14832 if (posix_warnings && RExC_warn_text && av_top_index(RExC_warn_text) > -1) {
14833 *posix_warnings = RExC_warn_text;
14836 else if (class_number != OOB_NAMEDCLASS) {
14837 /* If it is a known class, return the class. The class number
14838 * #defines are structured so each complement is +1 to the normal
14840 CLEAR_POSIX_WARNINGS_AND_RETURN(class_number + complement);
14842 else if (! check_only) {
14844 /* Here, it is an unrecognized class. This is an error (unless the
14845 * call is to check only, which we've already handled above) */
14846 const char * const complement_string = (complement)
14849 RExC_parse = (char *) p;
14850 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
14852 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
14856 return OOB_NAMEDCLASS;
14858 #undef ADD_POSIX_WARNING
14860 STATIC unsigned int
14861 S_regex_set_precedence(const U8 my_operator) {
14863 /* Returns the precedence in the (?[...]) construct of the input operator,
14864 * specified by its character representation. The precedence follows
14865 * general Perl rules, but it extends this so that ')' and ']' have (low)
14866 * precedence even though they aren't really operators */
14868 switch (my_operator) {
14884 NOT_REACHED; /* NOTREACHED */
14885 return 0; /* Silence compiler warning */
14889 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
14890 I32 *flagp, U32 depth,
14891 char * const oregcomp_parse)
14893 /* Handle the (?[...]) construct to do set operations */
14895 U8 curchar; /* Current character being parsed */
14896 UV start, end; /* End points of code point ranges */
14897 SV* final = NULL; /* The end result inversion list */
14898 SV* result_string; /* 'final' stringified */
14899 AV* stack; /* stack of operators and operands not yet
14901 AV* fence_stack = NULL; /* A stack containing the positions in
14902 'stack' of where the undealt-with left
14903 parens would be if they were actually
14905 /* The 'volatile' is a workaround for an optimiser bug
14906 * in Solaris Studio 12.3. See RT #127455 */
14907 volatile IV fence = 0; /* Position of where most recent undealt-
14908 with left paren in stack is; -1 if none.
14910 STRLEN len; /* Temporary */
14911 regnode* node; /* Temporary, and final regnode returned by
14913 const bool save_fold = FOLD; /* Temporary */
14914 char *save_end, *save_parse; /* Temporaries */
14915 const bool in_locale = LOC; /* we turn off /l during processing */
14916 AV* posix_warnings = NULL;
14918 GET_RE_DEBUG_FLAGS_DECL;
14920 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
14923 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
14926 REQUIRE_UNI_RULES(flagp, NULL); /* The use of this operator implies /u.
14927 This is required so that the compile
14928 time values are valid in all runtime
14931 /* This will return only an ANYOF regnode, or (unlikely) something smaller
14932 * (such as EXACT). Thus we can skip most everything if just sizing. We
14933 * call regclass to handle '[]' so as to not have to reinvent its parsing
14934 * rules here (throwing away the size it computes each time). And, we exit
14935 * upon an unescaped ']' that isn't one ending a regclass. To do both
14936 * these things, we need to realize that something preceded by a backslash
14937 * is escaped, so we have to keep track of backslashes */
14939 UV depth = 0; /* how many nested (?[...]) constructs */
14941 while (RExC_parse < RExC_end) {
14942 SV* current = NULL;
14944 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14945 TRUE /* Force /x */ );
14947 switch (*RExC_parse) {
14949 if (RExC_parse[1] == '[') depth++, RExC_parse++;
14954 /* Skip past this, so the next character gets skipped, after
14957 if (*RExC_parse == 'c') {
14958 /* Skip the \cX notation for control characters */
14959 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14965 /* See if this is a [:posix:] class. */
14966 bool is_posix_class = (OOB_NAMEDCLASS
14967 < handle_possible_posix(pRExC_state,
14971 TRUE /* checking only */));
14972 /* If it is a posix class, leave the parse pointer at the
14973 * '[' to fool regclass() into thinking it is part of a
14974 * '[[:posix:]]'. */
14975 if (! is_posix_class) {
14979 /* regclass() can only return RESTART_PASS1 and NEED_UTF8
14980 * if multi-char folds are allowed. */
14981 if (!regclass(pRExC_state, flagp,depth+1,
14982 is_posix_class, /* parse the whole char
14983 class only if not a
14985 FALSE, /* don't allow multi-char folds */
14986 TRUE, /* silence non-portable warnings. */
14988 FALSE, /* Require return to be an ANYOF */
14992 FAIL2("panic: regclass returned NULL to handle_sets, "
14993 "flags=%#" UVxf, (UV) *flagp);
14995 /* function call leaves parse pointing to the ']', except
14996 * if we faked it */
14997 if (is_posix_class) {
15001 SvREFCNT_dec(current); /* In case it returned something */
15006 if (depth--) break;
15008 if (*RExC_parse == ')') {
15009 node = reganode(pRExC_state, ANYOF, 0);
15010 RExC_size += ANYOF_SKIP;
15011 nextchar(pRExC_state);
15012 Set_Node_Length(node,
15013 RExC_parse - oregcomp_parse + 1); /* MJD */
15015 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
15023 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
15027 /* We output the messages even if warnings are off, because we'll fail
15028 * the very next thing, and these give a likely diagnosis for that */
15029 if (posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
15030 output_or_return_posix_warnings(pRExC_state, posix_warnings, NULL);
15033 FAIL("Syntax error in (?[...])");
15036 /* Pass 2 only after this. */
15037 Perl_ck_warner_d(aTHX_
15038 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
15039 "The regex_sets feature is experimental" REPORT_LOCATION,
15040 REPORT_LOCATION_ARGS(RExC_parse));
15042 /* Everything in this construct is a metacharacter. Operands begin with
15043 * either a '\' (for an escape sequence), or a '[' for a bracketed
15044 * character class. Any other character should be an operator, or
15045 * parenthesis for grouping. Both types of operands are handled by calling
15046 * regclass() to parse them. It is called with a parameter to indicate to
15047 * return the computed inversion list. The parsing here is implemented via
15048 * a stack. Each entry on the stack is a single character representing one
15049 * of the operators; or else a pointer to an operand inversion list. */
15051 #define IS_OPERATOR(a) SvIOK(a)
15052 #define IS_OPERAND(a) (! IS_OPERATOR(a))
15054 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
15055 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
15056 * with pronouncing it called it Reverse Polish instead, but now that YOU
15057 * know how to pronounce it you can use the correct term, thus giving due
15058 * credit to the person who invented it, and impressing your geek friends.
15059 * Wikipedia says that the pronounciation of "Ł" has been changing so that
15060 * it is now more like an English initial W (as in wonk) than an L.)
15062 * This means that, for example, 'a | b & c' is stored on the stack as
15070 * where the numbers in brackets give the stack [array] element number.
15071 * In this implementation, parentheses are not stored on the stack.
15072 * Instead a '(' creates a "fence" so that the part of the stack below the
15073 * fence is invisible except to the corresponding ')' (this allows us to
15074 * replace testing for parens, by using instead subtraction of the fence
15075 * position). As new operands are processed they are pushed onto the stack
15076 * (except as noted in the next paragraph). New operators of higher
15077 * precedence than the current final one are inserted on the stack before
15078 * the lhs operand (so that when the rhs is pushed next, everything will be
15079 * in the correct positions shown above. When an operator of equal or
15080 * lower precedence is encountered in parsing, all the stacked operations
15081 * of equal or higher precedence are evaluated, leaving the result as the
15082 * top entry on the stack. This makes higher precedence operations
15083 * evaluate before lower precedence ones, and causes operations of equal
15084 * precedence to left associate.
15086 * The only unary operator '!' is immediately pushed onto the stack when
15087 * encountered. When an operand is encountered, if the top of the stack is
15088 * a '!", the complement is immediately performed, and the '!' popped. The
15089 * resulting value is treated as a new operand, and the logic in the
15090 * previous paragraph is executed. Thus in the expression
15092 * the stack looks like
15098 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
15105 * A ')' is treated as an operator with lower precedence than all the
15106 * aforementioned ones, which causes all operations on the stack above the
15107 * corresponding '(' to be evaluated down to a single resultant operand.
15108 * Then the fence for the '(' is removed, and the operand goes through the
15109 * algorithm above, without the fence.
15111 * A separate stack is kept of the fence positions, so that the position of
15112 * the latest so-far unbalanced '(' is at the top of it.
15114 * The ']' ending the construct is treated as the lowest operator of all,
15115 * so that everything gets evaluated down to a single operand, which is the
15118 sv_2mortal((SV *)(stack = newAV()));
15119 sv_2mortal((SV *)(fence_stack = newAV()));
15121 while (RExC_parse < RExC_end) {
15122 I32 top_index; /* Index of top-most element in 'stack' */
15123 SV** top_ptr; /* Pointer to top 'stack' element */
15124 SV* current = NULL; /* To contain the current inversion list
15126 SV* only_to_avoid_leaks;
15128 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
15129 TRUE /* Force /x */ );
15130 if (RExC_parse >= RExC_end) {
15131 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
15134 curchar = UCHARAT(RExC_parse);
15138 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15139 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
15140 DEBUG_U(dump_regex_sets_structures(pRExC_state,
15141 stack, fence, fence_stack));
15144 top_index = av_tindex_skip_len_mg(stack);
15147 SV** stacked_ptr; /* Ptr to something already on 'stack' */
15148 char stacked_operator; /* The topmost operator on the 'stack'. */
15149 SV* lhs; /* Operand to the left of the operator */
15150 SV* rhs; /* Operand to the right of the operator */
15151 SV* fence_ptr; /* Pointer to top element of the fence
15156 if ( RExC_parse < RExC_end - 1
15157 && (UCHARAT(RExC_parse + 1) == '?'))
15159 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
15160 * This happens when we have some thing like
15162 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
15164 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
15166 * Here we would be handling the interpolated
15167 * '$thai_or_lao'. We handle this by a recursive call to
15168 * ourselves which returns the inversion list the
15169 * interpolated expression evaluates to. We use the flags
15170 * from the interpolated pattern. */
15171 U32 save_flags = RExC_flags;
15172 const char * save_parse;
15174 RExC_parse += 2; /* Skip past the '(?' */
15175 save_parse = RExC_parse;
15177 /* Parse any flags for the '(?' */
15178 parse_lparen_question_flags(pRExC_state);
15180 if (RExC_parse == save_parse /* Makes sure there was at
15181 least one flag (or else
15182 this embedding wasn't
15184 || RExC_parse >= RExC_end - 4
15185 || UCHARAT(RExC_parse) != ':'
15186 || UCHARAT(++RExC_parse) != '('
15187 || UCHARAT(++RExC_parse) != '?'
15188 || UCHARAT(++RExC_parse) != '[')
15191 /* In combination with the above, this moves the
15192 * pointer to the point just after the first erroneous
15193 * character (or if there are no flags, to where they
15194 * should have been) */
15195 if (RExC_parse >= RExC_end - 4) {
15196 RExC_parse = RExC_end;
15198 else if (RExC_parse != save_parse) {
15199 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15201 vFAIL("Expecting '(?flags:(?[...'");
15204 /* Recurse, with the meat of the embedded expression */
15206 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
15207 depth+1, oregcomp_parse);
15209 /* Here, 'current' contains the embedded expression's
15210 * inversion list, and RExC_parse points to the trailing
15211 * ']'; the next character should be the ')' */
15213 assert(UCHARAT(RExC_parse) == ')');
15215 /* Then the ')' matching the original '(' handled by this
15216 * case: statement */
15218 assert(UCHARAT(RExC_parse) == ')');
15221 RExC_flags = save_flags;
15222 goto handle_operand;
15225 /* A regular '('. Look behind for illegal syntax */
15226 if (top_index - fence >= 0) {
15227 /* If the top entry on the stack is an operator, it had
15228 * better be a '!', otherwise the entry below the top
15229 * operand should be an operator */
15230 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
15231 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
15232 || ( IS_OPERAND(*top_ptr)
15233 && ( top_index - fence < 1
15234 || ! (stacked_ptr = av_fetch(stack,
15237 || ! IS_OPERATOR(*stacked_ptr))))
15240 vFAIL("Unexpected '(' with no preceding operator");
15244 /* Stack the position of this undealt-with left paren */
15245 av_push(fence_stack, newSViv(fence));
15246 fence = top_index + 1;
15250 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15251 * multi-char folds are allowed. */
15252 if (!regclass(pRExC_state, flagp,depth+1,
15253 TRUE, /* means parse just the next thing */
15254 FALSE, /* don't allow multi-char folds */
15255 FALSE, /* don't silence non-portable warnings. */
15257 FALSE, /* Require return to be an ANYOF */
15261 FAIL2("panic: regclass returned NULL to handle_sets, "
15262 "flags=%#" UVxf, (UV) *flagp);
15265 /* regclass() will return with parsing just the \ sequence,
15266 * leaving the parse pointer at the next thing to parse */
15268 goto handle_operand;
15270 case '[': /* Is a bracketed character class */
15272 /* See if this is a [:posix:] class. */
15273 bool is_posix_class = (OOB_NAMEDCLASS
15274 < handle_possible_posix(pRExC_state,
15278 TRUE /* checking only */));
15279 /* If it is a posix class, leave the parse pointer at the '['
15280 * to fool regclass() into thinking it is part of a
15281 * '[[:posix:]]'. */
15282 if (! is_posix_class) {
15286 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15287 * multi-char folds are allowed. */
15288 if (!regclass(pRExC_state, flagp,depth+1,
15289 is_posix_class, /* parse the whole char
15290 class only if not a
15292 FALSE, /* don't allow multi-char folds */
15293 TRUE, /* silence non-portable warnings. */
15295 FALSE, /* Require return to be an ANYOF */
15300 FAIL2("panic: regclass returned NULL to handle_sets, "
15301 "flags=%#" UVxf, (UV) *flagp);
15304 /* function call leaves parse pointing to the ']', except if we
15306 if (is_posix_class) {
15310 goto handle_operand;
15314 if (top_index >= 1) {
15315 goto join_operators;
15318 /* Only a single operand on the stack: are done */
15322 if (av_tindex_skip_len_mg(fence_stack) < 0) {
15324 vFAIL("Unexpected ')'");
15327 /* If nothing after the fence, is missing an operand */
15328 if (top_index - fence < 0) {
15332 /* If at least two things on the stack, treat this as an
15334 if (top_index - fence >= 1) {
15335 goto join_operators;
15338 /* Here only a single thing on the fenced stack, and there is a
15339 * fence. Get rid of it */
15340 fence_ptr = av_pop(fence_stack);
15342 fence = SvIV(fence_ptr) - 1;
15343 SvREFCNT_dec_NN(fence_ptr);
15350 /* Having gotten rid of the fence, we pop the operand at the
15351 * stack top and process it as a newly encountered operand */
15352 current = av_pop(stack);
15353 if (IS_OPERAND(current)) {
15354 goto handle_operand;
15366 /* These binary operators should have a left operand already
15368 if ( top_index - fence < 0
15369 || top_index - fence == 1
15370 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
15371 || ! IS_OPERAND(*top_ptr))
15373 goto unexpected_binary;
15376 /* If only the one operand is on the part of the stack visible
15377 * to us, we just place this operator in the proper position */
15378 if (top_index - fence < 2) {
15380 /* Place the operator before the operand */
15382 SV* lhs = av_pop(stack);
15383 av_push(stack, newSVuv(curchar));
15384 av_push(stack, lhs);
15388 /* But if there is something else on the stack, we need to
15389 * process it before this new operator if and only if the
15390 * stacked operation has equal or higher precedence than the
15395 /* The operator on the stack is supposed to be below both its
15397 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
15398 || IS_OPERAND(*stacked_ptr))
15400 /* But if not, it's legal and indicates we are completely
15401 * done if and only if we're currently processing a ']',
15402 * which should be the final thing in the expression */
15403 if (curchar == ']') {
15409 vFAIL2("Unexpected binary operator '%c' with no "
15410 "preceding operand", curchar);
15412 stacked_operator = (char) SvUV(*stacked_ptr);
15414 if (regex_set_precedence(curchar)
15415 > regex_set_precedence(stacked_operator))
15417 /* Here, the new operator has higher precedence than the
15418 * stacked one. This means we need to add the new one to
15419 * the stack to await its rhs operand (and maybe more
15420 * stuff). We put it before the lhs operand, leaving
15421 * untouched the stacked operator and everything below it
15423 lhs = av_pop(stack);
15424 assert(IS_OPERAND(lhs));
15426 av_push(stack, newSVuv(curchar));
15427 av_push(stack, lhs);
15431 /* Here, the new operator has equal or lower precedence than
15432 * what's already there. This means the operation already
15433 * there should be performed now, before the new one. */
15435 rhs = av_pop(stack);
15436 if (! IS_OPERAND(rhs)) {
15438 /* This can happen when a ! is not followed by an operand,
15439 * like in /(?[\t &!])/ */
15443 lhs = av_pop(stack);
15445 if (! IS_OPERAND(lhs)) {
15447 /* This can happen when there is an empty (), like in
15448 * /(?[[0]+()+])/ */
15452 switch (stacked_operator) {
15454 _invlist_intersection(lhs, rhs, &rhs);
15459 _invlist_union(lhs, rhs, &rhs);
15463 _invlist_subtract(lhs, rhs, &rhs);
15466 case '^': /* The union minus the intersection */
15471 _invlist_union(lhs, rhs, &u);
15472 _invlist_intersection(lhs, rhs, &i);
15473 _invlist_subtract(u, i, &rhs);
15474 SvREFCNT_dec_NN(i);
15475 SvREFCNT_dec_NN(u);
15481 /* Here, the higher precedence operation has been done, and the
15482 * result is in 'rhs'. We overwrite the stacked operator with
15483 * the result. Then we redo this code to either push the new
15484 * operator onto the stack or perform any higher precedence
15485 * stacked operation */
15486 only_to_avoid_leaks = av_pop(stack);
15487 SvREFCNT_dec(only_to_avoid_leaks);
15488 av_push(stack, rhs);
15491 case '!': /* Highest priority, right associative */
15493 /* If what's already at the top of the stack is another '!",
15494 * they just cancel each other out */
15495 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
15496 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
15498 only_to_avoid_leaks = av_pop(stack);
15499 SvREFCNT_dec(only_to_avoid_leaks);
15501 else { /* Otherwise, since it's right associative, just push
15503 av_push(stack, newSVuv(curchar));
15508 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15509 vFAIL("Unexpected character");
15513 /* Here 'current' is the operand. If something is already on the
15514 * stack, we have to check if it is a !. But first, the code above
15515 * may have altered the stack in the time since we earlier set
15518 top_index = av_tindex_skip_len_mg(stack);
15519 if (top_index - fence >= 0) {
15520 /* If the top entry on the stack is an operator, it had better
15521 * be a '!', otherwise the entry below the top operand should
15522 * be an operator */
15523 top_ptr = av_fetch(stack, top_index, FALSE);
15525 if (IS_OPERATOR(*top_ptr)) {
15527 /* The only permissible operator at the top of the stack is
15528 * '!', which is applied immediately to this operand. */
15529 curchar = (char) SvUV(*top_ptr);
15530 if (curchar != '!') {
15531 SvREFCNT_dec(current);
15532 vFAIL2("Unexpected binary operator '%c' with no "
15533 "preceding operand", curchar);
15536 _invlist_invert(current);
15538 only_to_avoid_leaks = av_pop(stack);
15539 SvREFCNT_dec(only_to_avoid_leaks);
15541 /* And we redo with the inverted operand. This allows
15542 * handling multiple ! in a row */
15543 goto handle_operand;
15545 /* Single operand is ok only for the non-binary ')'
15547 else if ((top_index - fence == 0 && curchar != ')')
15548 || (top_index - fence > 0
15549 && (! (stacked_ptr = av_fetch(stack,
15552 || IS_OPERAND(*stacked_ptr))))
15554 SvREFCNT_dec(current);
15555 vFAIL("Operand with no preceding operator");
15559 /* Here there was nothing on the stack or the top element was
15560 * another operand. Just add this new one */
15561 av_push(stack, current);
15563 } /* End of switch on next parse token */
15565 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15566 } /* End of loop parsing through the construct */
15569 if (av_tindex_skip_len_mg(fence_stack) >= 0) {
15570 vFAIL("Unmatched (");
15573 if (av_tindex_skip_len_mg(stack) < 0 /* Was empty */
15574 || ((final = av_pop(stack)) == NULL)
15575 || ! IS_OPERAND(final)
15576 || SvTYPE(final) != SVt_INVLIST
15577 || av_tindex_skip_len_mg(stack) >= 0) /* More left on stack */
15580 SvREFCNT_dec(final);
15581 vFAIL("Incomplete expression within '(?[ ])'");
15584 /* Here, 'final' is the resultant inversion list from evaluating the
15585 * expression. Return it if so requested */
15586 if (return_invlist) {
15587 *return_invlist = final;
15591 /* Otherwise generate a resultant node, based on 'final'. regclass() is
15592 * expecting a string of ranges and individual code points */
15593 invlist_iterinit(final);
15594 result_string = newSVpvs("");
15595 while (invlist_iternext(final, &start, &end)) {
15596 if (start == end) {
15597 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
15600 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%" UVXf "}",
15605 /* About to generate an ANYOF (or similar) node from the inversion list we
15606 * have calculated */
15607 save_parse = RExC_parse;
15608 RExC_parse = SvPV(result_string, len);
15609 save_end = RExC_end;
15610 RExC_end = RExC_parse + len;
15612 /* We turn off folding around the call, as the class we have constructed
15613 * already has all folding taken into consideration, and we don't want
15614 * regclass() to add to that */
15615 RExC_flags &= ~RXf_PMf_FOLD;
15616 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if multi-char
15617 * folds are allowed. */
15618 node = regclass(pRExC_state, flagp,depth+1,
15619 FALSE, /* means parse the whole char class */
15620 FALSE, /* don't allow multi-char folds */
15621 TRUE, /* silence non-portable warnings. The above may very
15622 well have generated non-portable code points, but
15623 they're valid on this machine */
15624 FALSE, /* similarly, no need for strict */
15625 FALSE, /* Require return to be an ANYOF */
15630 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#" UVxf,
15633 /* Fix up the node type if we are in locale. (We have pretended we are
15634 * under /u for the purposes of regclass(), as this construct will only
15635 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
15636 * as to cause any warnings about bad locales to be output in regexec.c),
15637 * and add the flag that indicates to check if not in a UTF-8 locale. The
15638 * reason we above forbid optimization into something other than an ANYOF
15639 * node is simply to minimize the number of code changes in regexec.c.
15640 * Otherwise we would have to create new EXACTish node types and deal with
15641 * them. This decision could be revisited should this construct become
15644 * (One might think we could look at the resulting ANYOF node and suppress
15645 * the flag if everything is above 255, as those would be UTF-8 only,
15646 * but this isn't true, as the components that led to that result could
15647 * have been locale-affected, and just happen to cancel each other out
15648 * under UTF-8 locales.) */
15650 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
15652 assert(OP(node) == ANYOF);
15656 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
15660 RExC_flags |= RXf_PMf_FOLD;
15663 RExC_parse = save_parse + 1;
15664 RExC_end = save_end;
15665 SvREFCNT_dec_NN(final);
15666 SvREFCNT_dec_NN(result_string);
15668 nextchar(pRExC_state);
15669 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
15673 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15676 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
15677 AV * stack, const IV fence, AV * fence_stack)
15678 { /* Dumps the stacks in handle_regex_sets() */
15680 const SSize_t stack_top = av_tindex_skip_len_mg(stack);
15681 const SSize_t fence_stack_top = av_tindex_skip_len_mg(fence_stack);
15684 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
15686 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
15688 if (stack_top < 0) {
15689 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
15692 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
15693 for (i = stack_top; i >= 0; i--) {
15694 SV ** element_ptr = av_fetch(stack, i, FALSE);
15695 if (! element_ptr) {
15698 if (IS_OPERATOR(*element_ptr)) {
15699 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
15700 (int) i, (int) SvIV(*element_ptr));
15703 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
15704 sv_dump(*element_ptr);
15709 if (fence_stack_top < 0) {
15710 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
15713 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
15714 for (i = fence_stack_top; i >= 0; i--) {
15715 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
15716 if (! element_ptr) {
15719 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
15720 (int) i, (int) SvIV(*element_ptr));
15731 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
15733 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
15734 * innocent-looking character class, like /[ks]/i won't have to go out to
15735 * disk to find the possible matches.
15737 * This should be called only for a Latin1-range code points, cp, which is
15738 * known to be involved in a simple fold with other code points above
15739 * Latin1. It would give false results if /aa has been specified.
15740 * Multi-char folds are outside the scope of this, and must be handled
15743 * XXX It would be better to generate these via regen, in case a new
15744 * version of the Unicode standard adds new mappings, though that is not
15745 * really likely, and may be caught by the default: case of the switch
15748 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
15750 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
15756 add_cp_to_invlist(*invlist, KELVIN_SIGN);
15760 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
15763 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
15764 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
15766 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
15767 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
15768 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
15770 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
15771 *invlist = add_cp_to_invlist(*invlist,
15772 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
15775 #ifdef LATIN_CAPITAL_LETTER_SHARP_S /* not defined in early Unicode releases */
15777 case LATIN_SMALL_LETTER_SHARP_S:
15778 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
15783 #if UNICODE_MAJOR_VERSION < 3 \
15784 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0)
15786 /* In 3.0 and earlier, U+0130 folded simply to 'i'; and in 3.0.1 so did
15791 add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
15792 # if UNICODE_DOT_DOT_VERSION == 1
15793 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_DOTLESS_I);
15799 /* Use deprecated warning to increase the chances of this being
15802 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
15809 S_output_or_return_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings, AV** return_posix_warnings)
15811 /* If the final parameter is NULL, output the elements of the array given
15812 * by '*posix_warnings' as REGEXP warnings. Otherwise, the elements are
15813 * pushed onto it, (creating if necessary) */
15816 const bool first_is_fatal = ! return_posix_warnings
15817 && ckDEAD(packWARN(WARN_REGEXP));
15819 PERL_ARGS_ASSERT_OUTPUT_OR_RETURN_POSIX_WARNINGS;
15821 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
15822 if (return_posix_warnings) {
15823 if (! *return_posix_warnings) { /* mortalize to not leak if
15824 warnings are fatal */
15825 *return_posix_warnings = (AV *) sv_2mortal((SV *) newAV());
15827 av_push(*return_posix_warnings, msg);
15830 if (first_is_fatal) { /* Avoid leaking this */
15831 av_undef(posix_warnings); /* This isn't necessary if the
15832 array is mortal, but is a
15834 (void) sv_2mortal(msg);
15836 SAVEFREESV(RExC_rx_sv);
15839 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
15840 SvREFCNT_dec_NN(msg);
15846 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
15848 /* This adds the string scalar <multi_string> to the array
15849 * <multi_char_matches>. <multi_string> is known to have exactly
15850 * <cp_count> code points in it. This is used when constructing a
15851 * bracketed character class and we find something that needs to match more
15852 * than a single character.
15854 * <multi_char_matches> is actually an array of arrays. Each top-level
15855 * element is an array that contains all the strings known so far that are
15856 * the same length. And that length (in number of code points) is the same
15857 * as the index of the top-level array. Hence, the [2] element is an
15858 * array, each element thereof is a string containing TWO code points;
15859 * while element [3] is for strings of THREE characters, and so on. Since
15860 * this is for multi-char strings there can never be a [0] nor [1] element.
15862 * When we rewrite the character class below, we will do so such that the
15863 * longest strings are written first, so that it prefers the longest
15864 * matching strings first. This is done even if it turns out that any
15865 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
15866 * Christiansen has agreed that this is ok. This makes the test for the
15867 * ligature 'ffi' come before the test for 'ff', for example */
15870 AV** this_array_ptr;
15872 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
15874 if (! multi_char_matches) {
15875 multi_char_matches = newAV();
15878 if (av_exists(multi_char_matches, cp_count)) {
15879 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
15880 this_array = *this_array_ptr;
15883 this_array = newAV();
15884 av_store(multi_char_matches, cp_count,
15887 av_push(this_array, multi_string);
15889 return multi_char_matches;
15892 /* The names of properties whose definitions are not known at compile time are
15893 * stored in this SV, after a constant heading. So if the length has been
15894 * changed since initialization, then there is a run-time definition. */
15895 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
15896 (SvCUR(listsv) != initial_listsv_len)
15898 /* There is a restricted set of white space characters that are legal when
15899 * ignoring white space in a bracketed character class. This generates the
15900 * code to skip them.
15902 * There is a line below that uses the same white space criteria but is outside
15903 * this macro. Both here and there must use the same definition */
15904 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
15907 while (isBLANK_A(UCHARAT(p))) \
15915 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
15916 const bool stop_at_1, /* Just parse the next thing, don't
15917 look for a full character class */
15918 bool allow_multi_folds,
15919 const bool silence_non_portable, /* Don't output warnings
15923 bool optimizable, /* ? Allow a non-ANYOF return
15925 SV** ret_invlist, /* Return an inversion list, not a node */
15926 AV** return_posix_warnings
15929 /* parse a bracketed class specification. Most of these will produce an
15930 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
15931 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
15932 * under /i with multi-character folds: it will be rewritten following the
15933 * paradigm of this example, where the <multi-fold>s are characters which
15934 * fold to multiple character sequences:
15935 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
15936 * gets effectively rewritten as:
15937 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
15938 * reg() gets called (recursively) on the rewritten version, and this
15939 * function will return what it constructs. (Actually the <multi-fold>s
15940 * aren't physically removed from the [abcdefghi], it's just that they are
15941 * ignored in the recursion by means of a flag:
15942 * <RExC_in_multi_char_class>.)
15944 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
15945 * characters, with the corresponding bit set if that character is in the
15946 * list. For characters above this, a range list or swash is used. There
15947 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
15948 * determinable at compile time
15950 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs
15951 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded
15952 * to UTF-8. This can only happen if ret_invlist is non-NULL.
15955 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
15957 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
15960 int namedclass = OOB_NAMEDCLASS;
15961 char *rangebegin = NULL;
15962 bool need_class = 0;
15964 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
15965 than just initialized. */
15966 SV* properties = NULL; /* Code points that match \p{} \P{} */
15967 SV* posixes = NULL; /* Code points that match classes like [:word:],
15968 extended beyond the Latin1 range. These have to
15969 be kept separate from other code points for much
15970 of this function because their handling is
15971 different under /i, and for most classes under
15973 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
15974 separate for a while from the non-complemented
15975 versions because of complications with /d
15977 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
15978 treated more simply than the general case,
15979 leading to less compilation and execution
15981 UV element_count = 0; /* Number of distinct elements in the class.
15982 Optimizations may be possible if this is tiny */
15983 AV * multi_char_matches = NULL; /* Code points that fold to more than one
15984 character; used under /i */
15986 char * stop_ptr = RExC_end; /* where to stop parsing */
15988 /* ignore unescaped whitespace? */
15989 const bool skip_white = cBOOL( ret_invlist
15990 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
15992 /* Unicode properties are stored in a swash; this holds the current one
15993 * being parsed. If this swash is the only above-latin1 component of the
15994 * character class, an optimization is to pass it directly on to the
15995 * execution engine. Otherwise, it is set to NULL to indicate that there
15996 * are other things in the class that have to be dealt with at execution
15998 SV* swash = NULL; /* Code points that match \p{} \P{} */
16000 /* Set if a component of this character class is user-defined; just passed
16001 * on to the engine */
16002 bool has_user_defined_property = FALSE;
16004 /* inversion list of code points this node matches only when the target
16005 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
16007 SV* has_upper_latin1_only_utf8_matches = NULL;
16009 /* Inversion list of code points this node matches regardless of things
16010 * like locale, folding, utf8ness of the target string */
16011 SV* cp_list = NULL;
16013 /* Like cp_list, but code points on this list need to be checked for things
16014 * that fold to/from them under /i */
16015 SV* cp_foldable_list = NULL;
16017 /* Like cp_list, but code points on this list are valid only when the
16018 * runtime locale is UTF-8 */
16019 SV* only_utf8_locale_list = NULL;
16021 /* In a range, if one of the endpoints is non-character-set portable,
16022 * meaning that it hard-codes a code point that may mean a different
16023 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
16024 * mnemonic '\t' which each mean the same character no matter which
16025 * character set the platform is on. */
16026 unsigned int non_portable_endpoint = 0;
16028 /* Is the range unicode? which means on a platform that isn't 1-1 native
16029 * to Unicode (i.e. non-ASCII), each code point in it should be considered
16030 * to be a Unicode value. */
16031 bool unicode_range = FALSE;
16032 bool invert = FALSE; /* Is this class to be complemented */
16034 bool warn_super = ALWAYS_WARN_SUPER;
16036 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
16037 case we need to change the emitted regop to an EXACT. */
16038 const char * orig_parse = RExC_parse;
16039 const SSize_t orig_size = RExC_size;
16040 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
16042 /* This variable is used to mark where the end in the input is of something
16043 * that looks like a POSIX construct but isn't. During the parse, when
16044 * something looks like it could be such a construct is encountered, it is
16045 * checked for being one, but not if we've already checked this area of the
16046 * input. Only after this position is reached do we check again */
16047 char *not_posix_region_end = RExC_parse - 1;
16049 AV* posix_warnings = NULL;
16050 const bool do_posix_warnings = return_posix_warnings
16051 || (PASS2 && ckWARN(WARN_REGEXP));
16053 GET_RE_DEBUG_FLAGS_DECL;
16055 PERL_ARGS_ASSERT_REGCLASS;
16057 PERL_UNUSED_ARG(depth);
16060 DEBUG_PARSE("clas");
16062 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
16063 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
16064 && UNICODE_DOT_DOT_VERSION == 0)
16065 allow_multi_folds = FALSE;
16068 /* Assume we are going to generate an ANYOF node. */
16069 ret = reganode(pRExC_state,
16076 RExC_size += ANYOF_SKIP;
16077 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
16080 ANYOF_FLAGS(ret) = 0;
16082 RExC_emit += ANYOF_SKIP;
16083 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
16084 initial_listsv_len = SvCUR(listsv);
16085 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
16088 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16090 assert(RExC_parse <= RExC_end);
16092 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
16095 allow_multi_folds = FALSE;
16097 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16100 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
16101 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
16102 int maybe_class = handle_possible_posix(pRExC_state,
16104 ¬_posix_region_end,
16106 TRUE /* checking only */);
16107 if (PASS2 && maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
16108 SAVEFREESV(RExC_rx_sv);
16109 ckWARN4reg(not_posix_region_end,
16110 "POSIX syntax [%c %c] belongs inside character classes%s",
16111 *RExC_parse, *RExC_parse,
16112 (maybe_class == OOB_NAMEDCLASS)
16113 ? ((POSIXCC_NOTYET(*RExC_parse))
16114 ? " (but this one isn't implemented)"
16115 : " (but this one isn't fully valid)")
16118 (void)ReREFCNT_inc(RExC_rx_sv);
16122 /* If the caller wants us to just parse a single element, accomplish this
16123 * by faking the loop ending condition */
16124 if (stop_at_1 && RExC_end > RExC_parse) {
16125 stop_ptr = RExC_parse + 1;
16128 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
16129 if (UCHARAT(RExC_parse) == ']')
16130 goto charclassloop;
16134 if ( posix_warnings
16135 && av_tindex_skip_len_mg(posix_warnings) >= 0
16136 && RExC_parse > not_posix_region_end)
16138 /* Warnings about posix class issues are considered tentative until
16139 * we are far enough along in the parse that we can no longer
16140 * change our mind, at which point we either output them or add
16141 * them, if it has so specified, to what gets returned to the
16142 * caller. This is done each time through the loop so that a later
16143 * class won't zap them before they have been dealt with. */
16144 output_or_return_posix_warnings(pRExC_state, posix_warnings,
16145 return_posix_warnings);
16148 if (RExC_parse >= stop_ptr) {
16152 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16154 if (UCHARAT(RExC_parse) == ']') {
16160 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
16161 save_value = value;
16162 save_prevvalue = prevvalue;
16165 rangebegin = RExC_parse;
16167 non_portable_endpoint = 0;
16169 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
16170 value = utf8n_to_uvchr((U8*)RExC_parse,
16171 RExC_end - RExC_parse,
16172 &numlen, UTF8_ALLOW_DEFAULT);
16173 RExC_parse += numlen;
16176 value = UCHARAT(RExC_parse++);
16178 if (value == '[') {
16179 char * posix_class_end;
16180 namedclass = handle_possible_posix(pRExC_state,
16183 do_posix_warnings ? &posix_warnings : NULL,
16184 FALSE /* die if error */);
16185 if (namedclass > OOB_NAMEDCLASS) {
16187 /* If there was an earlier attempt to parse this particular
16188 * posix class, and it failed, it was a false alarm, as this
16189 * successful one proves */
16190 if ( posix_warnings
16191 && av_tindex_skip_len_mg(posix_warnings) >= 0
16192 && not_posix_region_end >= RExC_parse
16193 && not_posix_region_end <= posix_class_end)
16195 av_undef(posix_warnings);
16198 RExC_parse = posix_class_end;
16200 else if (namedclass == OOB_NAMEDCLASS) {
16201 not_posix_region_end = posix_class_end;
16204 namedclass = OOB_NAMEDCLASS;
16207 else if ( RExC_parse - 1 > not_posix_region_end
16208 && MAYBE_POSIXCC(value))
16210 (void) handle_possible_posix(
16212 RExC_parse - 1, /* -1 because parse has already been
16214 ¬_posix_region_end,
16215 do_posix_warnings ? &posix_warnings : NULL,
16216 TRUE /* checking only */);
16218 else if (value == '\\') {
16219 /* Is a backslash; get the code point of the char after it */
16221 if (RExC_parse >= RExC_end) {
16222 vFAIL("Unmatched [");
16225 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
16226 value = utf8n_to_uvchr((U8*)RExC_parse,
16227 RExC_end - RExC_parse,
16228 &numlen, UTF8_ALLOW_DEFAULT);
16229 RExC_parse += numlen;
16232 value = UCHARAT(RExC_parse++);
16234 /* Some compilers cannot handle switching on 64-bit integer
16235 * values, therefore value cannot be an UV. Yes, this will
16236 * be a problem later if we want switch on Unicode.
16237 * A similar issue a little bit later when switching on
16238 * namedclass. --jhi */
16240 /* If the \ is escaping white space when white space is being
16241 * skipped, it means that that white space is wanted literally, and
16242 * is already in 'value'. Otherwise, need to translate the escape
16243 * into what it signifies. */
16244 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
16246 case 'w': namedclass = ANYOF_WORDCHAR; break;
16247 case 'W': namedclass = ANYOF_NWORDCHAR; break;
16248 case 's': namedclass = ANYOF_SPACE; break;
16249 case 'S': namedclass = ANYOF_NSPACE; break;
16250 case 'd': namedclass = ANYOF_DIGIT; break;
16251 case 'D': namedclass = ANYOF_NDIGIT; break;
16252 case 'v': namedclass = ANYOF_VERTWS; break;
16253 case 'V': namedclass = ANYOF_NVERTWS; break;
16254 case 'h': namedclass = ANYOF_HORIZWS; break;
16255 case 'H': namedclass = ANYOF_NHORIZWS; break;
16256 case 'N': /* Handle \N{NAME} in class */
16258 const char * const backslash_N_beg = RExC_parse - 2;
16261 if (! grok_bslash_N(pRExC_state,
16262 NULL, /* No regnode */
16263 &value, /* Yes single value */
16264 &cp_count, /* Multiple code pt count */
16270 if (*flagp & NEED_UTF8)
16271 FAIL("panic: grok_bslash_N set NEED_UTF8");
16272 if (*flagp & RESTART_PASS1)
16275 if (cp_count < 0) {
16276 vFAIL("\\N in a character class must be a named character: \\N{...}");
16278 else if (cp_count == 0) {
16280 ckWARNreg(RExC_parse,
16281 "Ignoring zero length \\N{} in character class");
16284 else { /* cp_count > 1 */
16285 if (! RExC_in_multi_char_class) {
16286 if (invert || range || *RExC_parse == '-') {
16289 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
16292 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
16294 break; /* <value> contains the first code
16295 point. Drop out of the switch to
16299 SV * multi_char_N = newSVpvn(backslash_N_beg,
16300 RExC_parse - backslash_N_beg);
16302 = add_multi_match(multi_char_matches,
16307 } /* End of cp_count != 1 */
16309 /* This element should not be processed further in this
16312 value = save_value;
16313 prevvalue = save_prevvalue;
16314 continue; /* Back to top of loop to get next char */
16317 /* Here, is a single code point, and <value> contains it */
16318 unicode_range = TRUE; /* \N{} are Unicode */
16326 /* We will handle any undefined properties ourselves */
16327 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
16328 /* And we actually would prefer to get
16329 * the straight inversion list of the
16330 * swash, since we will be accessing it
16331 * anyway, to save a little time */
16332 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
16334 if (RExC_parse >= RExC_end)
16335 vFAIL2("Empty \\%c", (U8)value);
16336 if (*RExC_parse == '{') {
16337 const U8 c = (U8)value;
16338 e = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
16341 vFAIL2("Missing right brace on \\%c{}", c);
16345 while (isSPACE(*RExC_parse)) {
16349 if (UCHARAT(RExC_parse) == '^') {
16351 /* toggle. (The rhs xor gets the single bit that
16352 * differs between P and p; the other xor inverts just
16354 value ^= 'P' ^ 'p';
16357 while (isSPACE(*RExC_parse)) {
16362 if (e == RExC_parse)
16363 vFAIL2("Empty \\%c{}", c);
16365 n = e - RExC_parse;
16366 while (isSPACE(*(RExC_parse + n - 1)))
16368 } /* The \p isn't immediately followed by a '{' */
16369 else if (! isALPHA(*RExC_parse)) {
16370 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16371 vFAIL2("Character following \\%c must be '{' or a "
16372 "single-character Unicode property name",
16382 char* base_name; /* name after any packages are stripped */
16383 char* lookup_name = NULL;
16384 const char * const colon_colon = "::";
16386 /* Try to get the definition of the property into
16387 * <invlist>. If /i is in effect, the effective property
16388 * will have its name be <__NAME_i>. The design is
16389 * discussed in commit
16390 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
16391 name = savepv(Perl_form(aTHX_ "%.*s", (int)n, RExC_parse));
16394 lookup_name = savepv(Perl_form(aTHX_ "__%s_i", name));
16396 /* The function call just below that uses this can fail
16397 * to return, leaking memory if we don't do this */
16398 SAVEFREEPV(lookup_name);
16401 /* Look up the property name, and get its swash and
16402 * inversion list, if the property is found */
16403 SvREFCNT_dec(swash); /* Free any left-overs */
16404 swash = _core_swash_init("utf8",
16411 NULL, /* No inversion list */
16414 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
16415 HV* curpkg = (IN_PERL_COMPILETIME)
16417 : CopSTASH(PL_curcop);
16421 if (swash) { /* Got a swash but no inversion list.
16422 Something is likely wrong that will
16423 be sorted-out later */
16424 SvREFCNT_dec_NN(swash);
16428 /* Here didn't find it. It could be a an error (like a
16429 * typo) in specifying a Unicode property, or it could
16430 * be a user-defined property that will be available at
16431 * run-time. The names of these must begin with 'In'
16432 * or 'Is' (after any packages are stripped off). So
16433 * if not one of those, or if we accept only
16434 * compile-time properties, is an error; otherwise add
16435 * it to the list for run-time look up. */
16436 if ((base_name = rninstr(name, name + n,
16437 colon_colon, colon_colon + 2)))
16438 { /* Has ::. We know this must be a user-defined
16441 final_n -= base_name - name;
16450 || base_name[0] != 'I'
16451 || (base_name[1] != 's' && base_name[1] != 'n')
16454 const char * const msg
16456 ? "Illegal user-defined property name"
16457 : "Can't find Unicode property definition";
16458 RExC_parse = e + 1;
16460 /* diag_listed_as: Can't find Unicode property definition "%s" */
16461 vFAIL3utf8f("%s \"%" UTF8f "\"",
16462 msg, UTF8fARG(UTF, n, name));
16465 /* If the property name doesn't already have a package
16466 * name, add the current one to it so that it can be
16467 * referred to outside it. [perl #121777] */
16468 if (! has_pkg && curpkg) {
16469 char* pkgname = HvNAME(curpkg);
16470 if (memNEs(pkgname, HvNAMELEN(curpkg), "main")) {
16471 char* full_name = Perl_form(aTHX_
16475 n = strlen(full_name);
16476 name = savepvn(full_name, n);
16480 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s%" UTF8f "%s\n",
16481 (value == 'p' ? '+' : '!'),
16482 (FOLD) ? "__" : "",
16483 UTF8fARG(UTF, n, name),
16484 (FOLD) ? "_i" : "");
16485 has_user_defined_property = TRUE;
16486 optimizable = FALSE; /* Will have to leave this an
16489 /* We don't know yet what this matches, so have to flag
16491 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
16495 /* Here, did get the swash and its inversion list. If
16496 * the swash is from a user-defined property, then this
16497 * whole character class should be regarded as such */
16498 if (swash_init_flags
16499 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
16501 has_user_defined_property = TRUE;
16504 /* We warn on matching an above-Unicode code point
16505 * if the match would return true, except don't
16506 * warn for \p{All}, which has exactly one element
16508 (_invlist_contains_cp(invlist, 0x110000)
16509 && (! (_invlist_len(invlist) == 1
16510 && *invlist_array(invlist) == 0)))
16516 /* Invert if asking for the complement */
16517 if (value == 'P') {
16518 _invlist_union_complement_2nd(properties,
16522 /* The swash can't be used as-is, because we've
16523 * inverted things; delay removing it to here after
16524 * have copied its invlist above */
16525 SvREFCNT_dec_NN(swash);
16529 _invlist_union(properties, invlist, &properties);
16533 RExC_parse = e + 1;
16534 namedclass = ANYOF_UNIPROP; /* no official name, but it's
16537 /* \p means they want Unicode semantics */
16538 REQUIRE_UNI_RULES(flagp, NULL);
16541 case 'n': value = '\n'; break;
16542 case 'r': value = '\r'; break;
16543 case 't': value = '\t'; break;
16544 case 'f': value = '\f'; break;
16545 case 'b': value = '\b'; break;
16546 case 'e': value = ESC_NATIVE; break;
16547 case 'a': value = '\a'; break;
16549 RExC_parse--; /* function expects to be pointed at the 'o' */
16551 const char* error_msg;
16552 bool valid = grok_bslash_o(&RExC_parse,
16556 PASS2, /* warnings only in
16559 silence_non_portable,
16565 non_portable_endpoint++;
16568 RExC_parse--; /* function expects to be pointed at the 'x' */
16570 const char* error_msg;
16571 bool valid = grok_bslash_x(&RExC_parse,
16575 PASS2, /* Output warnings */
16577 silence_non_portable,
16583 non_portable_endpoint++;
16586 value = grok_bslash_c(*RExC_parse++, PASS2);
16587 non_portable_endpoint++;
16589 case '0': case '1': case '2': case '3': case '4':
16590 case '5': case '6': case '7':
16592 /* Take 1-3 octal digits */
16593 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
16594 numlen = (strict) ? 4 : 3;
16595 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
16596 RExC_parse += numlen;
16599 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16600 vFAIL("Need exactly 3 octal digits");
16602 else if (! SIZE_ONLY /* like \08, \178 */
16604 && RExC_parse < RExC_end
16605 && isDIGIT(*RExC_parse)
16606 && ckWARN(WARN_REGEXP))
16608 SAVEFREESV(RExC_rx_sv);
16609 reg_warn_non_literal_string(
16611 form_short_octal_warning(RExC_parse, numlen));
16612 (void)ReREFCNT_inc(RExC_rx_sv);
16615 non_portable_endpoint++;
16619 /* Allow \_ to not give an error */
16620 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
16622 vFAIL2("Unrecognized escape \\%c in character class",
16626 SAVEFREESV(RExC_rx_sv);
16627 ckWARN2reg(RExC_parse,
16628 "Unrecognized escape \\%c in character class passed through",
16630 (void)ReREFCNT_inc(RExC_rx_sv);
16634 } /* End of switch on char following backslash */
16635 } /* end of handling backslash escape sequences */
16637 /* Here, we have the current token in 'value' */
16639 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
16642 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
16643 * literal, as is the character that began the false range, i.e.
16644 * the 'a' in the examples */
16647 const int w = (RExC_parse >= rangebegin)
16648 ? RExC_parse - rangebegin
16652 "False [] range \"%" UTF8f "\"",
16653 UTF8fARG(UTF, w, rangebegin));
16656 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
16657 ckWARN2reg(RExC_parse,
16658 "False [] range \"%" UTF8f "\"",
16659 UTF8fARG(UTF, w, rangebegin));
16660 (void)ReREFCNT_inc(RExC_rx_sv);
16661 cp_list = add_cp_to_invlist(cp_list, '-');
16662 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
16667 range = 0; /* this was not a true range */
16668 element_count += 2; /* So counts for three values */
16671 classnum = namedclass_to_classnum(namedclass);
16673 if (LOC && namedclass < ANYOF_POSIXL_MAX
16674 #ifndef HAS_ISASCII
16675 && classnum != _CC_ASCII
16678 /* What the Posix classes (like \w, [:space:]) match in locale
16679 * isn't knowable under locale until actual match time. Room
16680 * must be reserved (one time per outer bracketed class) to
16681 * store such classes. The space will contain a bit for each
16682 * named class that is to be matched against. This isn't
16683 * needed for \p{} and pseudo-classes, as they are not affected
16684 * by locale, and hence are dealt with separately */
16685 if (! need_class) {
16688 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16691 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16693 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
16694 ANYOF_POSIXL_ZERO(ret);
16696 /* We can't change this into some other type of node
16697 * (unless this is the only element, in which case there
16698 * are nodes that mean exactly this) as has runtime
16700 optimizable = FALSE;
16703 /* Coverity thinks it is possible for this to be negative; both
16704 * jhi and khw think it's not, but be safer */
16705 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16706 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
16708 /* See if it already matches the complement of this POSIX
16710 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16711 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
16715 posixl_matches_all = TRUE;
16716 break; /* No need to continue. Since it matches both
16717 e.g., \w and \W, it matches everything, and the
16718 bracketed class can be optimized into qr/./s */
16721 /* Add this class to those that should be checked at runtime */
16722 ANYOF_POSIXL_SET(ret, namedclass);
16724 /* The above-Latin1 characters are not subject to locale rules.
16725 * Just add them, in the second pass, to the
16726 * unconditionally-matched list */
16728 SV* scratch_list = NULL;
16730 /* Get the list of the above-Latin1 code points this
16732 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
16733 PL_XPosix_ptrs[classnum],
16735 /* Odd numbers are complements, like
16736 * NDIGIT, NASCII, ... */
16737 namedclass % 2 != 0,
16739 /* Checking if 'cp_list' is NULL first saves an extra
16740 * clone. Its reference count will be decremented at the
16741 * next union, etc, or if this is the only instance, at the
16742 * end of the routine */
16744 cp_list = scratch_list;
16747 _invlist_union(cp_list, scratch_list, &cp_list);
16748 SvREFCNT_dec_NN(scratch_list);
16750 continue; /* Go get next character */
16753 else if (! SIZE_ONLY) {
16755 /* Here, not in pass1 (in that pass we skip calculating the
16756 * contents of this class), and is not /l, or is a POSIX class
16757 * for which /l doesn't matter (or is a Unicode property, which
16758 * is skipped here). */
16759 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
16760 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
16762 /* Here, should be \h, \H, \v, or \V. None of /d, /i
16763 * nor /l make a difference in what these match,
16764 * therefore we just add what they match to cp_list. */
16765 if (classnum != _CC_VERTSPACE) {
16766 assert( namedclass == ANYOF_HORIZWS
16767 || namedclass == ANYOF_NHORIZWS);
16769 /* It turns out that \h is just a synonym for
16771 classnum = _CC_BLANK;
16774 _invlist_union_maybe_complement_2nd(
16776 PL_XPosix_ptrs[classnum],
16777 namedclass % 2 != 0, /* Complement if odd
16778 (NHORIZWS, NVERTWS)
16783 else if ( UNI_SEMANTICS
16784 || classnum == _CC_ASCII
16785 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
16786 || classnum == _CC_XDIGIT)))
16788 /* We usually have to worry about /d and /a affecting what
16789 * POSIX classes match, with special code needed for /d
16790 * because we won't know until runtime what all matches.
16791 * But there is no extra work needed under /u, and
16792 * [:ascii:] is unaffected by /a and /d; and :digit: and
16793 * :xdigit: don't have runtime differences under /d. So we
16794 * can special case these, and avoid some extra work below,
16795 * and at runtime. */
16796 _invlist_union_maybe_complement_2nd(
16798 PL_XPosix_ptrs[classnum],
16799 namedclass % 2 != 0,
16802 else { /* Garden variety class. If is NUPPER, NALPHA, ...
16803 complement and use nposixes */
16804 SV** posixes_ptr = namedclass % 2 == 0
16807 _invlist_union_maybe_complement_2nd(
16809 PL_XPosix_ptrs[classnum],
16810 namedclass % 2 != 0,
16814 } /* end of namedclass \blah */
16816 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16818 /* If 'range' is set, 'value' is the ending of a range--check its
16819 * validity. (If value isn't a single code point in the case of a
16820 * range, we should have figured that out above in the code that
16821 * catches false ranges). Later, we will handle each individual code
16822 * point in the range. If 'range' isn't set, this could be the
16823 * beginning of a range, so check for that by looking ahead to see if
16824 * the next real character to be processed is the range indicator--the
16829 /* For unicode ranges, we have to test that the Unicode as opposed
16830 * to the native values are not decreasing. (Above 255, there is
16831 * no difference between native and Unicode) */
16832 if (unicode_range && prevvalue < 255 && value < 255) {
16833 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
16834 goto backwards_range;
16839 if (prevvalue > value) /* b-a */ {
16844 w = RExC_parse - rangebegin;
16846 "Invalid [] range \"%" UTF8f "\"",
16847 UTF8fARG(UTF, w, rangebegin));
16848 NOT_REACHED; /* NOTREACHED */
16852 prevvalue = value; /* save the beginning of the potential range */
16853 if (! stop_at_1 /* Can't be a range if parsing just one thing */
16854 && *RExC_parse == '-')
16856 char* next_char_ptr = RExC_parse + 1;
16858 /* Get the next real char after the '-' */
16859 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
16861 /* If the '-' is at the end of the class (just before the ']',
16862 * it is a literal minus; otherwise it is a range */
16863 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
16864 RExC_parse = next_char_ptr;
16866 /* a bad range like \w-, [:word:]- ? */
16867 if (namedclass > OOB_NAMEDCLASS) {
16868 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
16869 const int w = RExC_parse >= rangebegin
16870 ? RExC_parse - rangebegin
16873 vFAIL4("False [] range \"%*.*s\"",
16878 "False [] range \"%*.*s\"",
16883 cp_list = add_cp_to_invlist(cp_list, '-');
16887 range = 1; /* yeah, it's a range! */
16888 continue; /* but do it the next time */
16893 if (namedclass > OOB_NAMEDCLASS) {
16897 /* Here, we have a single value this time through the loop, and
16898 * <prevvalue> is the beginning of the range, if any; or <value> if
16901 /* non-Latin1 code point implies unicode semantics. Must be set in
16902 * pass1 so is there for the whole of pass 2 */
16904 REQUIRE_UNI_RULES(flagp, NULL);
16907 /* Ready to process either the single value, or the completed range.
16908 * For single-valued non-inverted ranges, we consider the possibility
16909 * of multi-char folds. (We made a conscious decision to not do this
16910 * for the other cases because it can often lead to non-intuitive
16911 * results. For example, you have the peculiar case that:
16912 * "s s" =~ /^[^\xDF]+$/i => Y
16913 * "ss" =~ /^[^\xDF]+$/i => N
16915 * See [perl #89750] */
16916 if (FOLD && allow_multi_folds && value == prevvalue) {
16917 if (value == LATIN_SMALL_LETTER_SHARP_S
16918 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
16921 /* Here <value> is indeed a multi-char fold. Get what it is */
16923 U8 foldbuf[UTF8_MAXBYTES_CASE];
16926 UV folded = _to_uni_fold_flags(
16930 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
16931 ? FOLD_FLAGS_NOMIX_ASCII
16935 /* Here, <folded> should be the first character of the
16936 * multi-char fold of <value>, with <foldbuf> containing the
16937 * whole thing. But, if this fold is not allowed (because of
16938 * the flags), <fold> will be the same as <value>, and should
16939 * be processed like any other character, so skip the special
16941 if (folded != value) {
16943 /* Skip if we are recursed, currently parsing the class
16944 * again. Otherwise add this character to the list of
16945 * multi-char folds. */
16946 if (! RExC_in_multi_char_class) {
16947 STRLEN cp_count = utf8_length(foldbuf,
16948 foldbuf + foldlen);
16949 SV* multi_fold = sv_2mortal(newSVpvs(""));
16951 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
16954 = add_multi_match(multi_char_matches,
16960 /* This element should not be processed further in this
16963 value = save_value;
16964 prevvalue = save_prevvalue;
16970 if (strict && PASS2 && ckWARN(WARN_REGEXP)) {
16973 /* If the range starts above 255, everything is portable and
16974 * likely to be so for any forseeable character set, so don't
16976 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
16977 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
16979 else if (prevvalue != value) {
16981 /* Under strict, ranges that stop and/or end in an ASCII
16982 * printable should have each end point be a portable value
16983 * for it (preferably like 'A', but we don't warn if it is
16984 * a (portable) Unicode name or code point), and the range
16985 * must be be all digits or all letters of the same case.
16986 * Otherwise, the range is non-portable and unclear as to
16987 * what it contains */
16988 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
16989 && ( non_portable_endpoint
16990 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
16991 || (isLOWER_A(prevvalue) && isLOWER_A(value))
16992 || (isUPPER_A(prevvalue) && isUPPER_A(value))
16994 vWARN(RExC_parse, "Ranges of ASCII printables should"
16995 " be some subset of \"0-9\","
16996 " \"A-Z\", or \"a-z\"");
16998 else if (prevvalue >= 0x660) { /* ARABIC_INDIC_DIGIT_ZERO */
16999 SSize_t index_start;
17000 SSize_t index_final;
17002 /* But the nature of Unicode and languages mean we
17003 * can't do the same checks for above-ASCII ranges,
17004 * except in the case of digit ones. These should
17005 * contain only digits from the same group of 10. The
17006 * ASCII case is handled just above. 0x660 is the
17007 * first digit character beyond ASCII. Hence here, the
17008 * range could be a range of digits. First some
17009 * unlikely special cases. Grandfather in that a range
17010 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
17011 * if its starting value is one of the 10 digits prior
17012 * to it. This is because it is an alternate way of
17013 * writing 19D1, and some people may expect it to be in
17014 * that group. But it is bad, because it won't give
17015 * the expected results. In Unicode 5.2 it was
17016 * considered to be in that group (of 11, hence), but
17017 * this was fixed in the next version */
17019 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
17020 goto warn_bad_digit_range;
17022 else if (UNLIKELY( prevvalue >= 0x1D7CE
17023 && value <= 0x1D7FF))
17025 /* This is the only other case currently in Unicode
17026 * where the algorithm below fails. The code
17027 * points just above are the end points of a single
17028 * range containing only decimal digits. It is 5
17029 * different series of 0-9. All other ranges of
17030 * digits currently in Unicode are just a single
17031 * series. (And mktables will notify us if a later
17032 * Unicode version breaks this.)
17034 * If the range being checked is at most 9 long,
17035 * and the digit values represented are in
17036 * numerical order, they are from the same series.
17038 if ( value - prevvalue > 9
17039 || ((( value - 0x1D7CE) % 10)
17040 <= (prevvalue - 0x1D7CE) % 10))
17042 goto warn_bad_digit_range;
17047 /* For all other ranges of digits in Unicode, the
17048 * algorithm is just to check if both end points
17049 * are in the same series, which is the same range.
17051 index_start = _invlist_search(
17052 PL_XPosix_ptrs[_CC_DIGIT],
17055 /* Warn if the range starts and ends with a digit,
17056 * and they are not in the same group of 10. */
17057 if ( index_start >= 0
17058 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
17060 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
17061 value)) != index_start
17062 && index_final >= 0
17063 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
17065 warn_bad_digit_range:
17066 vWARN(RExC_parse, "Ranges of digits should be"
17067 " from the same group of"
17074 if ((! range || prevvalue == value) && non_portable_endpoint) {
17075 if (isPRINT_A(value)) {
17078 if (isBACKSLASHED_PUNCT(value)) {
17079 literal[d++] = '\\';
17081 literal[d++] = (char) value;
17082 literal[d++] = '\0';
17085 "\"%.*s\" is more clearly written simply as \"%s\"",
17086 (int) (RExC_parse - rangebegin),
17091 else if isMNEMONIC_CNTRL(value) {
17093 "\"%.*s\" is more clearly written simply as \"%s\"",
17094 (int) (RExC_parse - rangebegin),
17096 cntrl_to_mnemonic((U8) value)
17102 /* Deal with this element of the class */
17106 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17109 /* On non-ASCII platforms, for ranges that span all of 0..255, and
17110 * ones that don't require special handling, we can just add the
17111 * range like we do for ASCII platforms */
17112 if ((UNLIKELY(prevvalue == 0) && value >= 255)
17113 || ! (prevvalue < 256
17115 || (! non_portable_endpoint
17116 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
17117 || (isUPPER_A(prevvalue)
17118 && isUPPER_A(value)))))))
17120 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17124 /* Here, requires special handling. This can be because it is
17125 * a range whose code points are considered to be Unicode, and
17126 * so must be individually translated into native, or because
17127 * its a subrange of 'A-Z' or 'a-z' which each aren't
17128 * contiguous in EBCDIC, but we have defined them to include
17129 * only the "expected" upper or lower case ASCII alphabetics.
17130 * Subranges above 255 are the same in native and Unicode, so
17131 * can be added as a range */
17132 U8 start = NATIVE_TO_LATIN1(prevvalue);
17134 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
17135 for (j = start; j <= end; j++) {
17136 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
17139 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17146 range = 0; /* this range (if it was one) is done now */
17147 } /* End of loop through all the text within the brackets */
17150 if ( posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
17151 output_or_return_posix_warnings(pRExC_state, posix_warnings,
17152 return_posix_warnings);
17155 /* If anything in the class expands to more than one character, we have to
17156 * deal with them by building up a substitute parse string, and recursively
17157 * calling reg() on it, instead of proceeding */
17158 if (multi_char_matches) {
17159 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
17162 char *save_end = RExC_end;
17163 char *save_parse = RExC_parse;
17164 char *save_start = RExC_start;
17165 STRLEN prefix_end = 0; /* We copy the character class after a
17166 prefix supplied here. This is the size
17167 + 1 of that prefix */
17168 bool first_time = TRUE; /* First multi-char occurrence doesn't get
17173 assert(RExC_precomp_adj == 0); /* Only one level of recursion allowed */
17175 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
17176 because too confusing */
17178 sv_catpv(substitute_parse, "(?:");
17182 /* Look at the longest folds first */
17183 for (cp_count = av_tindex_skip_len_mg(multi_char_matches);
17188 if (av_exists(multi_char_matches, cp_count)) {
17189 AV** this_array_ptr;
17192 this_array_ptr = (AV**) av_fetch(multi_char_matches,
17194 while ((this_sequence = av_pop(*this_array_ptr)) !=
17197 if (! first_time) {
17198 sv_catpv(substitute_parse, "|");
17200 first_time = FALSE;
17202 sv_catpv(substitute_parse, SvPVX(this_sequence));
17207 /* If the character class contains anything else besides these
17208 * multi-character folds, have to include it in recursive parsing */
17209 if (element_count) {
17210 sv_catpv(substitute_parse, "|[");
17211 prefix_end = SvCUR(substitute_parse);
17212 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
17214 /* Put in a closing ']' only if not going off the end, as otherwise
17215 * we are adding something that really isn't there */
17216 if (RExC_parse < RExC_end) {
17217 sv_catpv(substitute_parse, "]");
17221 sv_catpv(substitute_parse, ")");
17224 /* This is a way to get the parse to skip forward a whole named
17225 * sequence instead of matching the 2nd character when it fails the
17227 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
17231 /* Set up the data structure so that any errors will be properly
17232 * reported. See the comments at the definition of
17233 * REPORT_LOCATION_ARGS for details */
17234 RExC_precomp_adj = orig_parse - RExC_precomp;
17235 RExC_start = RExC_parse = SvPV(substitute_parse, len);
17236 RExC_adjusted_start = RExC_start + prefix_end;
17237 RExC_end = RExC_parse + len;
17238 RExC_in_multi_char_class = 1;
17239 RExC_emit = (regnode *)orig_emit;
17241 ret = reg(pRExC_state, 1, ®_flags, depth+1);
17243 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PASS1|NEED_UTF8);
17245 /* And restore so can parse the rest of the pattern */
17246 RExC_parse = save_parse;
17247 RExC_start = RExC_adjusted_start = save_start;
17248 RExC_precomp_adj = 0;
17249 RExC_end = save_end;
17250 RExC_in_multi_char_class = 0;
17251 SvREFCNT_dec_NN(multi_char_matches);
17255 /* Here, we've gone through the entire class and dealt with multi-char
17256 * folds. We are now in a position that we can do some checks to see if we
17257 * can optimize this ANYOF node into a simpler one, even in Pass 1.
17258 * Currently we only do two checks:
17259 * 1) is in the unlikely event that the user has specified both, eg. \w and
17260 * \W under /l, then the class matches everything. (This optimization
17261 * is done only to make the optimizer code run later work.)
17262 * 2) if the character class contains only a single element (including a
17263 * single range), we see if there is an equivalent node for it.
17264 * Other checks are possible */
17266 && ! ret_invlist /* Can't optimize if returning the constructed
17268 && (UNLIKELY(posixl_matches_all) || element_count == 1))
17273 if (UNLIKELY(posixl_matches_all)) {
17276 else if (namedclass > OOB_NAMEDCLASS) { /* this is a single named
17277 class, like \w or [:digit:]
17280 /* All named classes are mapped into POSIXish nodes, with its FLAG
17281 * argument giving which class it is */
17282 switch ((I32)namedclass) {
17283 case ANYOF_UNIPROP:
17286 /* These don't depend on the charset modifiers. They always
17287 * match under /u rules */
17288 case ANYOF_NHORIZWS:
17289 case ANYOF_HORIZWS:
17290 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
17293 case ANYOF_NVERTWS:
17298 /* The actual POSIXish node for all the rest depends on the
17299 * charset modifier. The ones in the first set depend only on
17300 * ASCII or, if available on this platform, also locale */
17304 op = (LOC) ? POSIXL : POSIXA;
17310 /* The following don't have any matches in the upper Latin1
17311 * range, hence /d is equivalent to /u for them. Making it /u
17312 * saves some branches at runtime */
17316 case ANYOF_NXDIGIT:
17317 if (! DEPENDS_SEMANTICS) {
17318 goto treat_as_default;
17324 /* The following change to CASED under /i */
17330 namedclass = ANYOF_CASED + (namedclass % 2);
17334 /* The rest have more possibilities depending on the charset.
17335 * We take advantage of the enum ordering of the charset
17336 * modifiers to get the exact node type, */
17339 op = POSIXD + get_regex_charset(RExC_flags);
17340 if (op > POSIXA) { /* /aa is same as /a */
17345 /* The odd numbered ones are the complements of the
17346 * next-lower even number one */
17347 if (namedclass % 2 == 1) {
17351 arg = namedclass_to_classnum(namedclass);
17355 else if (value == prevvalue) {
17357 /* Here, the class consists of just a single code point */
17360 if (! LOC && value == '\n') {
17361 op = REG_ANY; /* Optimize [^\n] */
17362 *flagp |= HASWIDTH|SIMPLE;
17366 else if (value < 256 || UTF) {
17368 /* Optimize a single value into an EXACTish node, but not if it
17369 * would require converting the pattern to UTF-8. */
17370 op = compute_EXACTish(pRExC_state);
17372 } /* Otherwise is a range */
17373 else if (! LOC) { /* locale could vary these */
17374 if (prevvalue == '0') {
17375 if (value == '9') {
17380 else if (! FOLD || ASCII_FOLD_RESTRICTED) {
17381 /* We can optimize A-Z or a-z, but not if they could match
17382 * something like the KELVIN SIGN under /i. */
17383 if (prevvalue == 'A') {
17386 && ! non_portable_endpoint
17389 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
17393 else if (prevvalue == 'a') {
17396 && ! non_portable_endpoint
17399 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
17406 /* Here, we have changed <op> away from its initial value iff we found
17407 * an optimization */
17410 /* Throw away this ANYOF regnode, and emit the calculated one,
17411 * which should correspond to the beginning, not current, state of
17413 const char * cur_parse = RExC_parse;
17414 RExC_parse = (char *)orig_parse;
17418 /* To get locale nodes to not use the full ANYOF size would
17419 * require moving the code above that writes the portions
17420 * of it that aren't in other nodes to after this point.
17421 * e.g. ANYOF_POSIXL_SET */
17422 RExC_size = orig_size;
17426 RExC_emit = (regnode *)orig_emit;
17427 if (PL_regkind[op] == POSIXD) {
17428 if (op == POSIXL) {
17429 RExC_contains_locale = 1;
17432 op += NPOSIXD - POSIXD;
17437 ret = reg_node(pRExC_state, op);
17439 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17443 *flagp |= HASWIDTH|SIMPLE;
17445 else if (PL_regkind[op] == EXACT) {
17446 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17447 TRUE /* downgradable to EXACT */
17451 RExC_parse = (char *) cur_parse;
17453 SvREFCNT_dec(posixes);
17454 SvREFCNT_dec(nposixes);
17455 SvREFCNT_dec(simple_posixes);
17456 SvREFCNT_dec(cp_list);
17457 SvREFCNT_dec(cp_foldable_list);
17464 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
17466 /* If folding, we calculate all characters that could fold to or from the
17467 * ones already on the list */
17468 if (cp_foldable_list) {
17470 UV start, end; /* End points of code point ranges */
17472 SV* fold_intersection = NULL;
17475 /* Our calculated list will be for Unicode rules. For locale
17476 * matching, we have to keep a separate list that is consulted at
17477 * runtime only when the locale indicates Unicode rules. For
17478 * non-locale, we just use the general list */
17480 use_list = &only_utf8_locale_list;
17483 use_list = &cp_list;
17486 /* Only the characters in this class that participate in folds need
17487 * be checked. Get the intersection of this class and all the
17488 * possible characters that are foldable. This can quickly narrow
17489 * down a large class */
17490 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
17491 &fold_intersection);
17493 /* The folds for all the Latin1 characters are hard-coded into this
17494 * program, but we have to go out to disk to get the others. */
17495 if (invlist_highest(cp_foldable_list) >= 256) {
17497 /* This is a hash that for a particular fold gives all
17498 * characters that are involved in it */
17499 if (! PL_utf8_foldclosures) {
17500 _load_PL_utf8_foldclosures();
17504 /* Now look at the foldable characters in this class individually */
17505 invlist_iterinit(fold_intersection);
17506 while (invlist_iternext(fold_intersection, &start, &end)) {
17509 /* Look at every character in the range */
17510 for (j = start; j <= end; j++) {
17511 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17517 if (IS_IN_SOME_FOLD_L1(j)) {
17519 /* ASCII is always matched; non-ASCII is matched
17520 * only under Unicode rules (which could happen
17521 * under /l if the locale is a UTF-8 one */
17522 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
17523 *use_list = add_cp_to_invlist(*use_list,
17524 PL_fold_latin1[j]);
17527 has_upper_latin1_only_utf8_matches
17528 = add_cp_to_invlist(
17529 has_upper_latin1_only_utf8_matches,
17530 PL_fold_latin1[j]);
17534 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
17535 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
17537 add_above_Latin1_folds(pRExC_state,
17544 /* Here is an above Latin1 character. We don't have the
17545 * rules hard-coded for it. First, get its fold. This is
17546 * the simple fold, as the multi-character folds have been
17547 * handled earlier and separated out */
17548 _to_uni_fold_flags(j, foldbuf, &foldlen,
17549 (ASCII_FOLD_RESTRICTED)
17550 ? FOLD_FLAGS_NOMIX_ASCII
17553 /* Single character fold of above Latin1. Add everything in
17554 * its fold closure to the list that this node should match.
17555 * The fold closures data structure is a hash with the keys
17556 * being the UTF-8 of every character that is folded to, like
17557 * 'k', and the values each an array of all code points that
17558 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
17559 * Multi-character folds are not included */
17560 if ((listp = hv_fetch(PL_utf8_foldclosures,
17561 (char *) foldbuf, foldlen, FALSE)))
17563 AV* list = (AV*) *listp;
17565 for (k = 0; k <= av_tindex_skip_len_mg(list); k++) {
17566 SV** c_p = av_fetch(list, k, FALSE);
17572 /* /aa doesn't allow folds between ASCII and non- */
17573 if ((ASCII_FOLD_RESTRICTED
17574 && (isASCII(c) != isASCII(j))))
17579 /* Folds under /l which cross the 255/256 boundary
17580 * are added to a separate list. (These are valid
17581 * only when the locale is UTF-8.) */
17582 if (c < 256 && LOC) {
17583 *use_list = add_cp_to_invlist(*use_list, c);
17587 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
17589 cp_list = add_cp_to_invlist(cp_list, c);
17592 /* Similarly folds involving non-ascii Latin1
17593 * characters under /d are added to their list */
17594 has_upper_latin1_only_utf8_matches
17595 = add_cp_to_invlist(
17596 has_upper_latin1_only_utf8_matches,
17603 SvREFCNT_dec_NN(fold_intersection);
17606 /* Now that we have finished adding all the folds, there is no reason
17607 * to keep the foldable list separate */
17608 _invlist_union(cp_list, cp_foldable_list, &cp_list);
17609 SvREFCNT_dec_NN(cp_foldable_list);
17612 /* And combine the result (if any) with any inversion lists from posix
17613 * classes. The lists are kept separate up to now because we don't want to
17614 * fold the classes (folding of those is automatically handled by the swash
17615 * fetching code) */
17616 if (simple_posixes) { /* These are the classes known to be unaffected by
17619 _invlist_union(cp_list, simple_posixes, &cp_list);
17620 SvREFCNT_dec_NN(simple_posixes);
17623 cp_list = simple_posixes;
17626 if (posixes || nposixes) {
17628 /* We have to adjust /a and /aa */
17629 if (AT_LEAST_ASCII_RESTRICTED) {
17631 /* Under /a and /aa, nothing above ASCII matches these */
17633 _invlist_intersection(posixes,
17634 PL_XPosix_ptrs[_CC_ASCII],
17638 /* Under /a and /aa, everything above ASCII matches these
17641 _invlist_union_complement_2nd(nposixes,
17642 PL_XPosix_ptrs[_CC_ASCII],
17647 if (! DEPENDS_SEMANTICS) {
17649 /* For everything but /d, we can just add the current 'posixes' and
17650 * 'nposixes' to the main list */
17653 _invlist_union(cp_list, posixes, &cp_list);
17654 SvREFCNT_dec_NN(posixes);
17662 _invlist_union(cp_list, nposixes, &cp_list);
17663 SvREFCNT_dec_NN(nposixes);
17666 cp_list = nposixes;
17671 /* Under /d, things like \w match upper Latin1 characters only if
17672 * the target string is in UTF-8. But things like \W match all the
17673 * upper Latin1 characters if the target string is not in UTF-8.
17675 * Handle the case where there something like \W separately */
17677 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1);
17679 /* A complemented posix class matches all upper Latin1
17680 * characters if not in UTF-8. And it matches just certain
17681 * ones when in UTF-8. That means those certain ones are
17682 * matched regardless, so can just be added to the
17683 * unconditional list */
17685 _invlist_union(cp_list, nposixes, &cp_list);
17686 SvREFCNT_dec_NN(nposixes);
17690 cp_list = nposixes;
17693 /* Likewise for 'posixes' */
17694 _invlist_union(posixes, cp_list, &cp_list);
17696 /* Likewise for anything else in the range that matched only
17698 if (has_upper_latin1_only_utf8_matches) {
17699 _invlist_union(cp_list,
17700 has_upper_latin1_only_utf8_matches,
17702 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17703 has_upper_latin1_only_utf8_matches = NULL;
17706 /* If we don't match all the upper Latin1 characters regardless
17707 * of UTF-8ness, we have to set a flag to match the rest when
17709 _invlist_subtract(only_non_utf8_list, cp_list,
17710 &only_non_utf8_list);
17711 if (_invlist_len(only_non_utf8_list) != 0) {
17712 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17716 /* Here there were no complemented posix classes. That means
17717 * the upper Latin1 characters in 'posixes' match only when the
17718 * target string is in UTF-8. So we have to add them to the
17719 * list of those types of code points, while adding the
17720 * remainder to the unconditional list.
17722 * First calculate what they are */
17723 SV* nonascii_but_latin1_properties = NULL;
17724 _invlist_intersection(posixes, PL_UpperLatin1,
17725 &nonascii_but_latin1_properties);
17727 /* And add them to the final list of such characters. */
17728 _invlist_union(has_upper_latin1_only_utf8_matches,
17729 nonascii_but_latin1_properties,
17730 &has_upper_latin1_only_utf8_matches);
17732 /* Remove them from what now becomes the unconditional list */
17733 _invlist_subtract(posixes, nonascii_but_latin1_properties,
17736 /* And add those unconditional ones to the final list */
17738 _invlist_union(cp_list, posixes, &cp_list);
17739 SvREFCNT_dec_NN(posixes);
17746 SvREFCNT_dec(nonascii_but_latin1_properties);
17748 /* Get rid of any characters that we now know are matched
17749 * unconditionally from the conditional list, which may make
17750 * that list empty */
17751 _invlist_subtract(has_upper_latin1_only_utf8_matches,
17753 &has_upper_latin1_only_utf8_matches);
17754 if (_invlist_len(has_upper_latin1_only_utf8_matches) == 0) {
17755 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17756 has_upper_latin1_only_utf8_matches = NULL;
17762 /* And combine the result (if any) with any inversion list from properties.
17763 * The lists are kept separate up to now so that we can distinguish the two
17764 * in regards to matching above-Unicode. A run-time warning is generated
17765 * if a Unicode property is matched against a non-Unicode code point. But,
17766 * we allow user-defined properties to match anything, without any warning,
17767 * and we also suppress the warning if there is a portion of the character
17768 * class that isn't a Unicode property, and which matches above Unicode, \W
17769 * or [\x{110000}] for example.
17770 * (Note that in this case, unlike the Posix one above, there is no
17771 * <has_upper_latin1_only_utf8_matches>, because having a Unicode property
17772 * forces Unicode semantics */
17776 /* If it matters to the final outcome, see if a non-property
17777 * component of the class matches above Unicode. If so, the
17778 * warning gets suppressed. This is true even if just a single
17779 * such code point is specified, as, though not strictly correct if
17780 * another such code point is matched against, the fact that they
17781 * are using above-Unicode code points indicates they should know
17782 * the issues involved */
17784 warn_super = ! (invert
17785 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
17788 _invlist_union(properties, cp_list, &cp_list);
17789 SvREFCNT_dec_NN(properties);
17792 cp_list = properties;
17797 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17799 /* Because an ANYOF node is the only one that warns, this node
17800 * can't be optimized into something else */
17801 optimizable = FALSE;
17805 /* Here, we have calculated what code points should be in the character
17808 * Now we can see about various optimizations. Fold calculation (which we
17809 * did above) needs to take place before inversion. Otherwise /[^k]/i
17810 * would invert to include K, which under /i would match k, which it
17811 * shouldn't. Therefore we can't invert folded locale now, as it won't be
17812 * folded until runtime */
17814 /* If we didn't do folding, it's because some information isn't available
17815 * until runtime; set the run-time fold flag for these. (We don't have to
17816 * worry about properties folding, as that is taken care of by the swash
17817 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
17818 * locales, or the class matches at least one 0-255 range code point */
17821 /* Some things on the list might be unconditionally included because of
17822 * other components. Remove them, and clean up the list if it goes to
17824 if (only_utf8_locale_list && cp_list) {
17825 _invlist_subtract(only_utf8_locale_list, cp_list,
17826 &only_utf8_locale_list);
17828 if (_invlist_len(only_utf8_locale_list) == 0) {
17829 SvREFCNT_dec_NN(only_utf8_locale_list);
17830 only_utf8_locale_list = NULL;
17833 if (only_utf8_locale_list) {
17836 |ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
17838 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
17840 invlist_iterinit(cp_list);
17841 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
17842 ANYOF_FLAGS(ret) |= ANYOFL_FOLD;
17844 invlist_iterfinish(cp_list);
17847 else if ( DEPENDS_SEMANTICS
17848 && ( has_upper_latin1_only_utf8_matches
17849 || (ANYOF_FLAGS(ret) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
17852 optimizable = FALSE;
17856 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
17857 * at compile time. Besides not inverting folded locale now, we can't
17858 * invert if there are things such as \w, which aren't known until runtime
17862 && OP(ret) != ANYOFD
17863 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
17864 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17866 _invlist_invert(cp_list);
17868 /* Any swash can't be used as-is, because we've inverted things */
17870 SvREFCNT_dec_NN(swash);
17874 /* Clear the invert flag since have just done it here */
17881 *ret_invlist = cp_list;
17882 SvREFCNT_dec(swash);
17884 /* Discard the generated node */
17886 RExC_size = orig_size;
17889 RExC_emit = orig_emit;
17894 /* Some character classes are equivalent to other nodes. Such nodes take
17895 * up less room and generally fewer operations to execute than ANYOF nodes.
17896 * Above, we checked for and optimized into some such equivalents for
17897 * certain common classes that are easy to test. Getting to this point in
17898 * the code means that the class didn't get optimized there. Since this
17899 * code is only executed in Pass 2, it is too late to save space--it has
17900 * been allocated in Pass 1, and currently isn't given back. But turning
17901 * things into an EXACTish node can allow the optimizer to join it to any
17902 * adjacent such nodes. And if the class is equivalent to things like /./,
17903 * expensive run-time swashes can be avoided. Now that we have more
17904 * complete information, we can find things necessarily missed by the
17905 * earlier code. Another possible "optimization" that isn't done is that
17906 * something like [Ee] could be changed into an EXACTFU. khw tried this
17907 * and found that the ANYOF is faster, including for code points not in the
17908 * bitmap. This still might make sense to do, provided it got joined with
17909 * an adjacent node(s) to create a longer EXACTFU one. This could be
17910 * accomplished by creating a pseudo ANYOF_EXACTFU node type that the join
17911 * routine would know is joinable. If that didn't happen, the node type
17912 * could then be made a straight ANYOF */
17914 if (optimizable && cp_list && ! invert) {
17916 U8 op = END; /* The optimzation node-type */
17917 int posix_class = -1; /* Illegal value */
17918 const char * cur_parse= RExC_parse;
17920 invlist_iterinit(cp_list);
17921 if (! invlist_iternext(cp_list, &start, &end)) {
17923 /* Here, the list is empty. This happens, for example, when a
17924 * Unicode property that doesn't match anything is the only element
17925 * in the character class (perluniprops.pod notes such properties).
17928 *flagp |= HASWIDTH|SIMPLE;
17930 else if (start == end) { /* The range is a single code point */
17931 if (! invlist_iternext(cp_list, &start, &end)
17933 /* Don't do this optimization if it would require changing
17934 * the pattern to UTF-8 */
17935 && (start < 256 || UTF))
17937 /* Here, the list contains a single code point. Can optimize
17938 * into an EXACTish node */
17949 /* A locale node under folding with one code point can be
17950 * an EXACTFL, as its fold won't be calculated until
17956 /* Here, we are generally folding, but there is only one
17957 * code point to match. If we have to, we use an EXACT
17958 * node, but it would be better for joining with adjacent
17959 * nodes in the optimization pass if we used the same
17960 * EXACTFish node that any such are likely to be. We can
17961 * do this iff the code point doesn't participate in any
17962 * folds. For example, an EXACTF of a colon is the same as
17963 * an EXACT one, since nothing folds to or from a colon. */
17965 if (IS_IN_SOME_FOLD_L1(value)) {
17970 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
17975 /* If we haven't found the node type, above, it means we
17976 * can use the prevailing one */
17978 op = compute_EXACTish(pRExC_state);
17982 } /* End of first range contains just a single code point */
17983 else if (start == 0) {
17984 if (end == UV_MAX) {
17986 *flagp |= HASWIDTH|SIMPLE;
17989 else if (end == '\n' - 1
17990 && invlist_iternext(cp_list, &start, &end)
17991 && start == '\n' + 1 && end == UV_MAX)
17994 *flagp |= HASWIDTH|SIMPLE;
17998 invlist_iterfinish(cp_list);
18001 const UV cp_list_len = _invlist_len(cp_list);
18002 const UV* cp_list_array = invlist_array(cp_list);
18004 /* Here, didn't find an optimization. See if this matches any of
18005 * the POSIX classes. These run slightly faster for above-Unicode
18006 * code points, so don't bother with POSIXA ones nor the 2 that
18007 * have no above-Unicode matches. We can avoid these checks unless
18008 * the ANYOF matches at least as high as the lowest POSIX one
18009 * (which was manually found to be \v. The actual code point may
18010 * increase in later Unicode releases, if a higher code point is
18011 * assigned to be \v, but this code will never break. It would
18012 * just mean we could execute the checks for posix optimizations
18013 * unnecessarily) */
18015 if (cp_list_array[cp_list_len-1] > 0x2029) {
18016 for (posix_class = 0;
18017 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
18021 if (posix_class == _CC_ASCII || posix_class == _CC_CNTRL) {
18024 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
18026 /* Check if matches normal or inverted */
18027 if (_invlistEQ(cp_list,
18028 PL_XPosix_ptrs[posix_class],
18031 op = (try_inverted)
18034 *flagp |= HASWIDTH|SIMPLE;
18044 RExC_parse = (char *)orig_parse;
18045 RExC_emit = (regnode *)orig_emit;
18047 if (regarglen[op]) {
18048 ret = reganode(pRExC_state, op, 0);
18050 ret = reg_node(pRExC_state, op);
18053 RExC_parse = (char *)cur_parse;
18055 if (PL_regkind[op] == EXACT) {
18056 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
18057 TRUE /* downgradable to EXACT */
18060 else if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
18061 FLAGS(ret) = posix_class;
18064 SvREFCNT_dec_NN(cp_list);
18069 /* Here, <cp_list> contains all the code points we can determine at
18070 * compile time that match under all conditions. Go through it, and
18071 * for things that belong in the bitmap, put them there, and delete from
18072 * <cp_list>. While we are at it, see if everything above 255 is in the
18073 * list, and if so, set a flag to speed up execution */
18075 populate_ANYOF_from_invlist(ret, &cp_list);
18078 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
18081 /* Here, the bitmap has been populated with all the Latin1 code points that
18082 * always match. Can now add to the overall list those that match only
18083 * when the target string is UTF-8 (<has_upper_latin1_only_utf8_matches>).
18085 if (has_upper_latin1_only_utf8_matches) {
18087 _invlist_union(cp_list,
18088 has_upper_latin1_only_utf8_matches,
18090 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
18093 cp_list = has_upper_latin1_only_utf8_matches;
18095 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
18098 /* If there is a swash and more than one element, we can't use the swash in
18099 * the optimization below. */
18100 if (swash && element_count > 1) {
18101 SvREFCNT_dec_NN(swash);
18105 /* Note that the optimization of using 'swash' if it is the only thing in
18106 * the class doesn't have us change swash at all, so it can include things
18107 * that are also in the bitmap; otherwise we have purposely deleted that
18108 * duplicate information */
18109 set_ANYOF_arg(pRExC_state, ret, cp_list,
18110 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
18112 only_utf8_locale_list,
18113 swash, has_user_defined_property);
18115 *flagp |= HASWIDTH|SIMPLE;
18117 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
18118 RExC_contains_locale = 1;
18124 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
18127 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
18128 regnode* const node,
18130 SV* const runtime_defns,
18131 SV* const only_utf8_locale_list,
18133 const bool has_user_defined_property)
18135 /* Sets the arg field of an ANYOF-type node 'node', using information about
18136 * the node passed-in. If there is nothing outside the node's bitmap, the
18137 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
18138 * the count returned by add_data(), having allocated and stored an array,
18139 * av, that that count references, as follows:
18140 * av[0] stores the character class description in its textual form.
18141 * This is used later (regexec.c:Perl_regclass_swash()) to
18142 * initialize the appropriate swash, and is also useful for dumping
18143 * the regnode. This is set to &PL_sv_undef if the textual
18144 * description is not needed at run-time (as happens if the other
18145 * elements completely define the class)
18146 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
18147 * computed from av[0]. But if no further computation need be done,
18148 * the swash is stored here now (and av[0] is &PL_sv_undef).
18149 * av[2] stores the inversion list of code points that match only if the
18150 * current locale is UTF-8
18151 * av[3] stores the cp_list inversion list for use in addition or instead
18152 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
18153 * (Otherwise everything needed is already in av[0] and av[1])
18154 * av[4] is set if any component of the class is from a user-defined
18155 * property; used only if av[3] exists */
18159 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
18161 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
18162 assert(! (ANYOF_FLAGS(node)
18163 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
18164 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
18167 AV * const av = newAV();
18170 av_store(av, 0, (runtime_defns)
18171 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
18174 av_store(av, 1, swash);
18175 SvREFCNT_dec_NN(cp_list);
18178 av_store(av, 1, &PL_sv_undef);
18180 av_store(av, 3, cp_list);
18181 av_store(av, 4, newSVuv(has_user_defined_property));
18185 if (only_utf8_locale_list) {
18186 av_store(av, 2, only_utf8_locale_list);
18189 av_store(av, 2, &PL_sv_undef);
18192 rv = newRV_noinc(MUTABLE_SV(av));
18193 n = add_data(pRExC_state, STR_WITH_LEN("s"));
18194 RExC_rxi->data->data[n] = (void*)rv;
18199 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
18201 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
18202 const regnode* node,
18205 SV** only_utf8_locale_ptr,
18206 SV** output_invlist)
18209 /* For internal core use only.
18210 * Returns the swash for the input 'node' in the regex 'prog'.
18211 * If <doinit> is 'true', will attempt to create the swash if not already
18213 * If <listsvp> is non-null, will return the printable contents of the
18214 * swash. This can be used to get debugging information even before the
18215 * swash exists, by calling this function with 'doinit' set to false, in
18216 * which case the components that will be used to eventually create the
18217 * swash are returned (in a printable form).
18218 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
18219 * store an inversion list of code points that should match only if the
18220 * execution-time locale is a UTF-8 one.
18221 * If <output_invlist> is not NULL, it is where this routine is to store an
18222 * inversion list of the code points that would be instead returned in
18223 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
18224 * when this parameter is used, is just the non-code point data that
18225 * will go into creating the swash. This currently should be just
18226 * user-defined properties whose definitions were not known at compile
18227 * time. Using this parameter allows for easier manipulation of the
18228 * swash's data by the caller. It is illegal to call this function with
18229 * this parameter set, but not <listsvp>
18231 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
18232 * that, in spite of this function's name, the swash it returns may include
18233 * the bitmap data as well */
18236 SV *si = NULL; /* Input swash initialization string */
18237 SV* invlist = NULL;
18239 RXi_GET_DECL(prog,progi);
18240 const struct reg_data * const data = prog ? progi->data : NULL;
18242 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
18243 assert(! output_invlist || listsvp);
18245 if (data && data->count) {
18246 const U32 n = ARG(node);
18248 if (data->what[n] == 's') {
18249 SV * const rv = MUTABLE_SV(data->data[n]);
18250 AV * const av = MUTABLE_AV(SvRV(rv));
18251 SV **const ary = AvARRAY(av);
18252 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
18254 si = *ary; /* ary[0] = the string to initialize the swash with */
18256 if (av_tindex_skip_len_mg(av) >= 2) {
18257 if (only_utf8_locale_ptr
18259 && ary[2] != &PL_sv_undef)
18261 *only_utf8_locale_ptr = ary[2];
18264 assert(only_utf8_locale_ptr);
18265 *only_utf8_locale_ptr = NULL;
18268 /* Elements 3 and 4 are either both present or both absent. [3]
18269 * is any inversion list generated at compile time; [4]
18270 * indicates if that inversion list has any user-defined
18271 * properties in it. */
18272 if (av_tindex_skip_len_mg(av) >= 3) {
18274 if (SvUV(ary[4])) {
18275 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
18283 /* Element [1] is reserved for the set-up swash. If already there,
18284 * return it; if not, create it and store it there */
18285 if (ary[1] && SvROK(ary[1])) {
18288 else if (doinit && ((si && si != &PL_sv_undef)
18289 || (invlist && invlist != &PL_sv_undef))) {
18291 sw = _core_swash_init("utf8", /* the utf8 package */
18295 0, /* not from tr/// */
18297 &swash_init_flags);
18298 (void)av_store(av, 1, sw);
18303 /* If requested, return a printable version of what this swash matches */
18305 SV* matches_string = NULL;
18307 /* The swash should be used, if possible, to get the data, as it
18308 * contains the resolved data. But this function can be called at
18309 * compile-time, before everything gets resolved, in which case we
18310 * return the currently best available information, which is the string
18311 * that will eventually be used to do that resolving, 'si' */
18312 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
18313 && (si && si != &PL_sv_undef))
18315 /* Here, we only have 'si' (and possibly some passed-in data in
18316 * 'invlist', which is handled below) If the caller only wants
18317 * 'si', use that. */
18318 if (! output_invlist) {
18319 matches_string = newSVsv(si);
18322 /* But if the caller wants an inversion list of the node, we
18323 * need to parse 'si' and place as much as possible in the
18324 * desired output inversion list, making 'matches_string' only
18325 * contain the currently unresolvable things */
18326 const char *si_string = SvPVX(si);
18327 STRLEN remaining = SvCUR(si);
18331 /* Ignore everything before the first new-line */
18332 while (*si_string != '\n' && remaining > 0) {
18336 assert(remaining > 0);
18341 while (remaining > 0) {
18343 /* The data consists of just strings defining user-defined
18344 * property names, but in prior incarnations, and perhaps
18345 * somehow from pluggable regex engines, it could still
18346 * hold hex code point definitions. Each component of a
18347 * range would be separated by a tab, and each range by a
18348 * new-line. If these are found, instead add them to the
18349 * inversion list */
18350 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
18351 |PERL_SCAN_SILENT_NON_PORTABLE;
18352 STRLEN len = remaining;
18353 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
18355 /* If the hex decode routine found something, it should go
18356 * up to the next \n */
18357 if ( *(si_string + len) == '\n') {
18358 if (count) { /* 2nd code point on line */
18359 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
18362 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
18365 goto prepare_for_next_iteration;
18368 /* If the hex decode was instead for the lower range limit,
18369 * save it, and go parse the upper range limit */
18370 if (*(si_string + len) == '\t') {
18371 assert(count == 0);
18375 prepare_for_next_iteration:
18376 si_string += len + 1;
18377 remaining -= len + 1;
18381 /* Here, didn't find a legal hex number. Just add it from
18382 * here to the next \n */
18385 while (*(si_string + len) != '\n' && remaining > 0) {
18389 if (*(si_string + len) == '\n') {
18393 if (matches_string) {
18394 sv_catpvn(matches_string, si_string, len - 1);
18397 matches_string = newSVpvn(si_string, len - 1);
18400 sv_catpvs(matches_string, " ");
18401 } /* end of loop through the text */
18403 assert(matches_string);
18404 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
18405 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
18407 } /* end of has an 'si' but no swash */
18410 /* If we have a swash in place, its equivalent inversion list was above
18411 * placed into 'invlist'. If not, this variable may contain a stored
18412 * inversion list which is information beyond what is in 'si' */
18415 /* Again, if the caller doesn't want the output inversion list, put
18416 * everything in 'matches-string' */
18417 if (! output_invlist) {
18418 if ( ! matches_string) {
18419 matches_string = newSVpvs("\n");
18421 sv_catsv(matches_string, invlist_contents(invlist,
18422 TRUE /* traditional style */
18425 else if (! *output_invlist) {
18426 *output_invlist = invlist_clone(invlist);
18429 _invlist_union(*output_invlist, invlist, output_invlist);
18433 *listsvp = matches_string;
18438 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
18440 /* reg_skipcomment()
18442 Absorbs an /x style # comment from the input stream,
18443 returning a pointer to the first character beyond the comment, or if the
18444 comment terminates the pattern without anything following it, this returns
18445 one past the final character of the pattern (in other words, RExC_end) and
18446 sets the REG_RUN_ON_COMMENT_SEEN flag.
18448 Note it's the callers responsibility to ensure that we are
18449 actually in /x mode
18453 PERL_STATIC_INLINE char*
18454 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
18456 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
18460 while (p < RExC_end) {
18461 if (*(++p) == '\n') {
18466 /* we ran off the end of the pattern without ending the comment, so we have
18467 * to add an \n when wrapping */
18468 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
18473 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
18475 const bool force_to_xmod
18478 /* If the text at the current parse position '*p' is a '(?#...)' comment,
18479 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
18480 * is /x whitespace, advance '*p' so that on exit it points to the first
18481 * byte past all such white space and comments */
18483 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
18485 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
18487 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
18490 if (RExC_end - (*p) >= 3
18492 && *(*p + 1) == '?'
18493 && *(*p + 2) == '#')
18495 while (*(*p) != ')') {
18496 if ((*p) == RExC_end)
18497 FAIL("Sequence (?#... not terminated");
18505 const char * save_p = *p;
18506 while ((*p) < RExC_end) {
18508 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
18511 else if (*(*p) == '#') {
18512 (*p) = reg_skipcomment(pRExC_state, (*p));
18518 if (*p != save_p) {
18531 Advances the parse position by one byte, unless that byte is the beginning
18532 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
18533 those two cases, the parse position is advanced beyond all such comments and
18536 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
18540 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
18542 PERL_ARGS_ASSERT_NEXTCHAR;
18544 if (RExC_parse < RExC_end) {
18546 || UTF8_IS_INVARIANT(*RExC_parse)
18547 || UTF8_IS_START(*RExC_parse));
18549 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
18551 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
18552 FALSE /* Don't force /x */ );
18557 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
18559 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
18560 * space. In pass1, it aligns and increments RExC_size; in pass2,
18563 regnode * const ret = RExC_emit;
18564 GET_RE_DEBUG_FLAGS_DECL;
18566 PERL_ARGS_ASSERT_REGNODE_GUTS;
18568 assert(extra_size >= regarglen[op]);
18571 SIZE_ALIGN(RExC_size);
18572 RExC_size += 1 + extra_size;
18575 if (RExC_emit >= RExC_emit_bound)
18576 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
18577 op, (void*)RExC_emit, (void*)RExC_emit_bound);
18579 NODE_ALIGN_FILL(ret);
18580 #ifndef RE_TRACK_PATTERN_OFFSETS
18581 PERL_UNUSED_ARG(name);
18583 if (RExC_offsets) { /* MJD */
18585 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
18588 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
18589 ? "Overwriting end of array!\n" : "OK",
18590 (UV)(RExC_emit - RExC_emit_start),
18591 (UV)(RExC_parse - RExC_start),
18592 (UV)RExC_offsets[0]));
18593 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
18600 - reg_node - emit a node
18602 STATIC regnode * /* Location. */
18603 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
18605 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
18607 PERL_ARGS_ASSERT_REG_NODE;
18609 assert(regarglen[op] == 0);
18612 regnode *ptr = ret;
18613 FILL_ADVANCE_NODE(ptr, op);
18620 - reganode - emit a node with an argument
18622 STATIC regnode * /* Location. */
18623 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
18625 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
18627 PERL_ARGS_ASSERT_REGANODE;
18629 assert(regarglen[op] == 1);
18632 regnode *ptr = ret;
18633 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
18640 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
18642 /* emit a node with U32 and I32 arguments */
18644 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
18646 PERL_ARGS_ASSERT_REG2LANODE;
18648 assert(regarglen[op] == 2);
18651 regnode *ptr = ret;
18652 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
18659 - reginsert - insert an operator in front of already-emitted operand
18661 * Means relocating the operand.
18663 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
18664 * set up NEXT_OFF() of the inserted node if needed. Something like this:
18666 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
18668 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
18670 * ALSO NOTE - operand->flags will be set to 0 as well.
18673 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *operand, U32 depth)
18678 const int offset = regarglen[(U8)op];
18679 const int size = NODE_STEP_REGNODE + offset;
18680 GET_RE_DEBUG_FLAGS_DECL;
18682 PERL_ARGS_ASSERT_REGINSERT;
18683 PERL_UNUSED_CONTEXT;
18684 PERL_UNUSED_ARG(depth);
18685 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
18686 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
18691 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
18692 studying. If this is wrong then we need to adjust RExC_recurse
18693 below like we do with RExC_open_parens/RExC_close_parens. */
18697 if (RExC_open_parens) {
18699 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
18700 /* remember that RExC_npar is rex->nparens + 1,
18701 * iow it is 1 more than the number of parens seen in
18702 * the pattern so far. */
18703 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
18704 /* note, RExC_open_parens[0] is the start of the
18705 * regex, it can't move. RExC_close_parens[0] is the end
18706 * of the regex, it *can* move. */
18707 if ( paren && RExC_open_parens[paren] >= operand ) {
18708 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
18709 RExC_open_parens[paren] += size;
18711 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
18713 if ( RExC_close_parens[paren] >= operand ) {
18714 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
18715 RExC_close_parens[paren] += size;
18717 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
18722 RExC_end_op += size;
18724 while (src > operand) {
18725 StructCopy(--src, --dst, regnode);
18726 #ifdef RE_TRACK_PATTERN_OFFSETS
18727 if (RExC_offsets) { /* MJD 20010112 */
18729 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
18733 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
18734 ? "Overwriting end of array!\n" : "OK",
18735 (UV)(src - RExC_emit_start),
18736 (UV)(dst - RExC_emit_start),
18737 (UV)RExC_offsets[0]));
18738 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
18739 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
18744 place = operand; /* Op node, where operand used to be. */
18745 #ifdef RE_TRACK_PATTERN_OFFSETS
18746 if (RExC_offsets) { /* MJD */
18748 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
18752 (UV)(place - RExC_emit_start) > RExC_offsets[0]
18753 ? "Overwriting end of array!\n" : "OK",
18754 (UV)(place - RExC_emit_start),
18755 (UV)(RExC_parse - RExC_start),
18756 (UV)RExC_offsets[0]));
18757 Set_Node_Offset(place, RExC_parse);
18758 Set_Node_Length(place, 1);
18761 src = NEXTOPER(place);
18763 FILL_ADVANCE_NODE(place, op);
18764 Zero(src, offset, regnode);
18768 - regtail - set the next-pointer at the end of a node chain of p to val.
18769 - SEE ALSO: regtail_study
18772 S_regtail(pTHX_ RExC_state_t * pRExC_state,
18773 const regnode * const p,
18774 const regnode * const val,
18778 GET_RE_DEBUG_FLAGS_DECL;
18780 PERL_ARGS_ASSERT_REGTAIL;
18782 PERL_UNUSED_ARG(depth);
18788 /* Find last node. */
18789 scan = (regnode *) p;
18791 regnode * const temp = regnext(scan);
18793 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
18794 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18795 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
18796 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
18797 (temp == NULL ? "->" : ""),
18798 (temp == NULL ? PL_reg_name[OP(val)] : "")
18806 if (reg_off_by_arg[OP(scan)]) {
18807 ARG_SET(scan, val - scan);
18810 NEXT_OFF(scan) = val - scan;
18816 - regtail_study - set the next-pointer at the end of a node chain of p to val.
18817 - Look for optimizable sequences at the same time.
18818 - currently only looks for EXACT chains.
18820 This is experimental code. The idea is to use this routine to perform
18821 in place optimizations on branches and groups as they are constructed,
18822 with the long term intention of removing optimization from study_chunk so
18823 that it is purely analytical.
18825 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
18826 to control which is which.
18829 /* TODO: All four parms should be const */
18832 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
18833 const regnode *val,U32 depth)
18837 #ifdef EXPERIMENTAL_INPLACESCAN
18840 GET_RE_DEBUG_FLAGS_DECL;
18842 PERL_ARGS_ASSERT_REGTAIL_STUDY;
18848 /* Find last node. */
18852 regnode * const temp = regnext(scan);
18853 #ifdef EXPERIMENTAL_INPLACESCAN
18854 if (PL_regkind[OP(scan)] == EXACT) {
18855 bool unfolded_multi_char; /* Unexamined in this routine */
18856 if (join_exact(pRExC_state, scan, &min,
18857 &unfolded_multi_char, 1, val, depth+1))
18862 switch (OP(scan)) {
18866 case EXACTFA_NO_TRIE:
18872 if( exact == PSEUDO )
18874 else if ( exact != OP(scan) )
18883 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
18884 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18885 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
18886 SvPV_nolen_const(RExC_mysv),
18887 REG_NODE_NUM(scan),
18888 PL_reg_name[exact]);
18895 DEBUG_PARSE_MSG("");
18896 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
18897 Perl_re_printf( aTHX_
18898 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
18899 SvPV_nolen_const(RExC_mysv),
18900 (IV)REG_NODE_NUM(val),
18904 if (reg_off_by_arg[OP(scan)]) {
18905 ARG_SET(scan, val - scan);
18908 NEXT_OFF(scan) = val - scan;
18916 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
18921 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
18926 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18928 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
18929 if (flags & (1<<bit)) {
18930 if (!set++ && lead)
18931 Perl_re_printf( aTHX_ "%s",lead);
18932 Perl_re_printf( aTHX_ "%s ",PL_reg_intflags_name[bit]);
18937 Perl_re_printf( aTHX_ "\n");
18939 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18944 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
18950 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18952 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
18953 if (flags & (1<<bit)) {
18954 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
18957 if (!set++ && lead)
18958 Perl_re_printf( aTHX_ "%s",lead);
18959 Perl_re_printf( aTHX_ "%s ",PL_reg_extflags_name[bit]);
18962 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
18963 if (!set++ && lead) {
18964 Perl_re_printf( aTHX_ "%s",lead);
18967 case REGEX_UNICODE_CHARSET:
18968 Perl_re_printf( aTHX_ "UNICODE");
18970 case REGEX_LOCALE_CHARSET:
18971 Perl_re_printf( aTHX_ "LOCALE");
18973 case REGEX_ASCII_RESTRICTED_CHARSET:
18974 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
18976 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
18977 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
18980 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
18986 Perl_re_printf( aTHX_ "\n");
18988 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18994 Perl_regdump(pTHX_ const regexp *r)
18998 SV * const sv = sv_newmortal();
18999 SV *dsv= sv_newmortal();
19000 RXi_GET_DECL(r,ri);
19001 GET_RE_DEBUG_FLAGS_DECL;
19003 PERL_ARGS_ASSERT_REGDUMP;
19005 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
19007 /* Header fields of interest. */
19008 for (i = 0; i < 2; i++) {
19009 if (r->substrs->data[i].substr) {
19010 RE_PV_QUOTED_DECL(s, 0, dsv,
19011 SvPVX_const(r->substrs->data[i].substr),
19012 RE_SV_DUMPLEN(r->substrs->data[i].substr),
19013 PL_dump_re_max_len);
19014 Perl_re_printf( aTHX_
19015 "%s %s%s at %" IVdf "..%" UVuf " ",
19016 i ? "floating" : "anchored",
19018 RE_SV_TAIL(r->substrs->data[i].substr),
19019 (IV)r->substrs->data[i].min_offset,
19020 (UV)r->substrs->data[i].max_offset);
19022 else if (r->substrs->data[i].utf8_substr) {
19023 RE_PV_QUOTED_DECL(s, 1, dsv,
19024 SvPVX_const(r->substrs->data[i].utf8_substr),
19025 RE_SV_DUMPLEN(r->substrs->data[i].utf8_substr),
19027 Perl_re_printf( aTHX_
19028 "%s utf8 %s%s at %" IVdf "..%" UVuf " ",
19029 i ? "floating" : "anchored",
19031 RE_SV_TAIL(r->substrs->data[i].utf8_substr),
19032 (IV)r->substrs->data[i].min_offset,
19033 (UV)r->substrs->data[i].max_offset);
19037 if (r->check_substr || r->check_utf8)
19038 Perl_re_printf( aTHX_
19040 ( r->check_substr == r->substrs->data[1].substr
19041 && r->check_utf8 == r->substrs->data[1].utf8_substr
19042 ? "(checking floating" : "(checking anchored"));
19043 if (r->intflags & PREGf_NOSCAN)
19044 Perl_re_printf( aTHX_ " noscan");
19045 if (r->extflags & RXf_CHECK_ALL)
19046 Perl_re_printf( aTHX_ " isall");
19047 if (r->check_substr || r->check_utf8)
19048 Perl_re_printf( aTHX_ ") ");
19050 if (ri->regstclass) {
19051 regprop(r, sv, ri->regstclass, NULL, NULL);
19052 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
19054 if (r->intflags & PREGf_ANCH) {
19055 Perl_re_printf( aTHX_ "anchored");
19056 if (r->intflags & PREGf_ANCH_MBOL)
19057 Perl_re_printf( aTHX_ "(MBOL)");
19058 if (r->intflags & PREGf_ANCH_SBOL)
19059 Perl_re_printf( aTHX_ "(SBOL)");
19060 if (r->intflags & PREGf_ANCH_GPOS)
19061 Perl_re_printf( aTHX_ "(GPOS)");
19062 Perl_re_printf( aTHX_ " ");
19064 if (r->intflags & PREGf_GPOS_SEEN)
19065 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
19066 if (r->intflags & PREGf_SKIP)
19067 Perl_re_printf( aTHX_ "plus ");
19068 if (r->intflags & PREGf_IMPLICIT)
19069 Perl_re_printf( aTHX_ "implicit ");
19070 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
19071 if (r->extflags & RXf_EVAL_SEEN)
19072 Perl_re_printf( aTHX_ "with eval ");
19073 Perl_re_printf( aTHX_ "\n");
19075 regdump_extflags("r->extflags: ",r->extflags);
19076 regdump_intflags("r->intflags: ",r->intflags);
19079 PERL_ARGS_ASSERT_REGDUMP;
19080 PERL_UNUSED_CONTEXT;
19081 PERL_UNUSED_ARG(r);
19082 #endif /* DEBUGGING */
19085 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
19088 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
19089 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
19090 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
19091 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
19092 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
19093 || _CC_VERTSPACE != 15
19094 # error Need to adjust order of anyofs[]
19096 static const char * const anyofs[] = {
19133 - regprop - printable representation of opcode, with run time support
19137 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
19141 RXi_GET_DECL(prog,progi);
19142 GET_RE_DEBUG_FLAGS_DECL;
19144 PERL_ARGS_ASSERT_REGPROP;
19148 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
19149 /* It would be nice to FAIL() here, but this may be called from
19150 regexec.c, and it would be hard to supply pRExC_state. */
19151 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
19152 (int)OP(o), (int)REGNODE_MAX);
19153 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
19155 k = PL_regkind[OP(o)];
19158 sv_catpvs(sv, " ");
19159 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
19160 * is a crude hack but it may be the best for now since
19161 * we have no flag "this EXACTish node was UTF-8"
19163 pv_pretty(sv, STRING(o), STR_LEN(o), PL_dump_re_max_len,
19164 PL_colors[0], PL_colors[1],
19165 PERL_PV_ESCAPE_UNI_DETECT |
19166 PERL_PV_ESCAPE_NONASCII |
19167 PERL_PV_PRETTY_ELLIPSES |
19168 PERL_PV_PRETTY_LTGT |
19169 PERL_PV_PRETTY_NOCLEAR
19171 } else if (k == TRIE) {
19172 /* print the details of the trie in dumpuntil instead, as
19173 * progi->data isn't available here */
19174 const char op = OP(o);
19175 const U32 n = ARG(o);
19176 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
19177 (reg_ac_data *)progi->data->data[n] :
19179 const reg_trie_data * const trie
19180 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
19182 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
19183 DEBUG_TRIE_COMPILE_r({
19185 sv_catpvs(sv, "(JUMP)");
19186 Perl_sv_catpvf(aTHX_ sv,
19187 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
19188 (UV)trie->startstate,
19189 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
19190 (UV)trie->wordcount,
19193 (UV)TRIE_CHARCOUNT(trie),
19194 (UV)trie->uniquecharcount
19197 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
19198 sv_catpvs(sv, "[");
19199 (void) put_charclass_bitmap_innards(sv,
19200 ((IS_ANYOF_TRIE(op))
19202 : TRIE_BITMAP(trie)),
19208 sv_catpvs(sv, "]");
19210 } else if (k == CURLY) {
19211 U32 lo = ARG1(o), hi = ARG2(o);
19212 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
19213 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
19214 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
19215 if (hi == REG_INFTY)
19216 sv_catpvs(sv, "INFTY");
19218 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
19219 sv_catpvs(sv, "}");
19221 else if (k == WHILEM && o->flags) /* Ordinal/of */
19222 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
19223 else if (k == REF || k == OPEN || k == CLOSE
19224 || k == GROUPP || OP(o)==ACCEPT)
19226 AV *name_list= NULL;
19227 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
19228 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
19229 if ( RXp_PAREN_NAMES(prog) ) {
19230 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19231 } else if ( pRExC_state ) {
19232 name_list= RExC_paren_name_list;
19235 if ( k != REF || (OP(o) < NREF)) {
19236 SV **name= av_fetch(name_list, parno, 0 );
19238 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19241 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
19242 I32 *nums=(I32*)SvPVX(sv_dat);
19243 SV **name= av_fetch(name_list, nums[0], 0 );
19246 for ( n=0; n<SvIVX(sv_dat); n++ ) {
19247 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
19248 (n ? "," : ""), (IV)nums[n]);
19250 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19254 if ( k == REF && reginfo) {
19255 U32 n = ARG(o); /* which paren pair */
19256 I32 ln = prog->offs[n].start;
19257 if (prog->lastparen < n || ln == -1)
19258 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
19259 else if (ln == prog->offs[n].end)
19260 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
19262 const char *s = reginfo->strbeg + ln;
19263 Perl_sv_catpvf(aTHX_ sv, ": ");
19264 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
19265 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
19268 } else if (k == GOSUB) {
19269 AV *name_list= NULL;
19270 if ( RXp_PAREN_NAMES(prog) ) {
19271 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19272 } else if ( pRExC_state ) {
19273 name_list= RExC_paren_name_list;
19276 /* Paren and offset */
19277 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
19278 (int)((o + (int)ARG2L(o)) - progi->program) );
19280 SV **name= av_fetch(name_list, ARG(o), 0 );
19282 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19285 else if (k == LOGICAL)
19286 /* 2: embedded, otherwise 1 */
19287 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
19288 else if (k == ANYOF) {
19289 const U8 flags = ANYOF_FLAGS(o);
19290 bool do_sep = FALSE; /* Do we need to separate various components of
19292 /* Set if there is still an unresolved user-defined property */
19293 SV *unresolved = NULL;
19295 /* Things that are ignored except when the runtime locale is UTF-8 */
19296 SV *only_utf8_locale_invlist = NULL;
19298 /* Code points that don't fit in the bitmap */
19299 SV *nonbitmap_invlist = NULL;
19301 /* And things that aren't in the bitmap, but are small enough to be */
19302 SV* bitmap_range_not_in_bitmap = NULL;
19304 const bool inverted = flags & ANYOF_INVERT;
19306 if (OP(o) == ANYOFL) {
19307 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
19308 sv_catpvs(sv, "{utf8-locale-reqd}");
19310 if (flags & ANYOFL_FOLD) {
19311 sv_catpvs(sv, "{i}");
19315 /* If there is stuff outside the bitmap, get it */
19316 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
19317 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
19319 &only_utf8_locale_invlist,
19320 &nonbitmap_invlist);
19321 /* The non-bitmap data may contain stuff that could fit in the
19322 * bitmap. This could come from a user-defined property being
19323 * finally resolved when this call was done; or much more likely
19324 * because there are matches that require UTF-8 to be valid, and so
19325 * aren't in the bitmap. This is teased apart later */
19326 _invlist_intersection(nonbitmap_invlist,
19328 &bitmap_range_not_in_bitmap);
19329 /* Leave just the things that don't fit into the bitmap */
19330 _invlist_subtract(nonbitmap_invlist,
19332 &nonbitmap_invlist);
19335 /* Obey this flag to add all above-the-bitmap code points */
19336 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
19337 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
19338 NUM_ANYOF_CODE_POINTS,
19342 /* Ready to start outputting. First, the initial left bracket */
19343 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
19345 /* Then all the things that could fit in the bitmap */
19346 do_sep = put_charclass_bitmap_innards(sv,
19348 bitmap_range_not_in_bitmap,
19349 only_utf8_locale_invlist,
19352 /* Can't try inverting for a
19353 * better display if there are
19354 * things that haven't been
19356 unresolved != NULL);
19357 SvREFCNT_dec(bitmap_range_not_in_bitmap);
19359 /* If there are user-defined properties which haven't been defined yet,
19360 * output them. If the result is not to be inverted, it is clearest to
19361 * output them in a separate [] from the bitmap range stuff. If the
19362 * result is to be complemented, we have to show everything in one [],
19363 * as the inversion applies to the whole thing. Use {braces} to
19364 * separate them from anything in the bitmap and anything above the
19368 if (! do_sep) { /* If didn't output anything in the bitmap */
19369 sv_catpvs(sv, "^");
19371 sv_catpvs(sv, "{");
19374 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19376 sv_catsv(sv, unresolved);
19378 sv_catpvs(sv, "}");
19380 do_sep = ! inverted;
19383 /* And, finally, add the above-the-bitmap stuff */
19384 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
19387 /* See if truncation size is overridden */
19388 const STRLEN dump_len = (PL_dump_re_max_len > 256)
19389 ? PL_dump_re_max_len
19392 /* This is output in a separate [] */
19394 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19397 /* And, for easy of understanding, it is shown in the
19398 * uncomplemented form if possible. The one exception being if
19399 * there are unresolved items, where the inversion has to be
19400 * delayed until runtime */
19401 if (inverted && ! unresolved) {
19402 _invlist_invert(nonbitmap_invlist);
19403 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
19406 contents = invlist_contents(nonbitmap_invlist,
19407 FALSE /* output suitable for catsv */
19410 /* If the output is shorter than the permissible maximum, just do it. */
19411 if (SvCUR(contents) <= dump_len) {
19412 sv_catsv(sv, contents);
19415 const char * contents_string = SvPVX(contents);
19416 STRLEN i = dump_len;
19418 /* Otherwise, start at the permissible max and work back to the
19419 * first break possibility */
19420 while (i > 0 && contents_string[i] != ' ') {
19423 if (i == 0) { /* Fail-safe. Use the max if we couldn't
19424 find a legal break */
19428 sv_catpvn(sv, contents_string, i);
19429 sv_catpvs(sv, "...");
19432 SvREFCNT_dec_NN(contents);
19433 SvREFCNT_dec_NN(nonbitmap_invlist);
19436 /* And finally the matching, closing ']' */
19437 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
19439 SvREFCNT_dec(unresolved);
19441 else if (k == POSIXD || k == NPOSIXD) {
19442 U8 index = FLAGS(o) * 2;
19443 if (index < C_ARRAY_LENGTH(anyofs)) {
19444 if (*anyofs[index] != '[') {
19447 sv_catpv(sv, anyofs[index]);
19448 if (*anyofs[index] != '[') {
19453 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
19456 else if (k == BOUND || k == NBOUND) {
19457 /* Must be synced with order of 'bound_type' in regcomp.h */
19458 const char * const bounds[] = {
19459 "", /* Traditional */
19465 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
19466 sv_catpv(sv, bounds[FLAGS(o)]);
19468 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
19469 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
19470 else if (OP(o) == SBOL)
19471 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
19473 /* add on the verb argument if there is one */
19474 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
19476 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
19477 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
19479 sv_catpvs(sv, ":NULL");
19482 PERL_UNUSED_CONTEXT;
19483 PERL_UNUSED_ARG(sv);
19484 PERL_UNUSED_ARG(o);
19485 PERL_UNUSED_ARG(prog);
19486 PERL_UNUSED_ARG(reginfo);
19487 PERL_UNUSED_ARG(pRExC_state);
19488 #endif /* DEBUGGING */
19494 Perl_re_intuit_string(pTHX_ REGEXP * const r)
19495 { /* Assume that RE_INTUIT is set */
19496 struct regexp *const prog = ReANY(r);
19497 GET_RE_DEBUG_FLAGS_DECL;
19499 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
19500 PERL_UNUSED_CONTEXT;
19504 const char * const s = SvPV_nolen_const(RX_UTF8(r)
19505 ? prog->check_utf8 : prog->check_substr);
19507 if (!PL_colorset) reginitcolors();
19508 Perl_re_printf( aTHX_
19509 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
19511 RX_UTF8(r) ? "utf8 " : "",
19512 PL_colors[5],PL_colors[0],
19515 (strlen(s) > PL_dump_re_max_len ? "..." : ""));
19518 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
19519 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
19525 handles refcounting and freeing the perl core regexp structure. When
19526 it is necessary to actually free the structure the first thing it
19527 does is call the 'free' method of the regexp_engine associated to
19528 the regexp, allowing the handling of the void *pprivate; member
19529 first. (This routine is not overridable by extensions, which is why
19530 the extensions free is called first.)
19532 See regdupe and regdupe_internal if you change anything here.
19534 #ifndef PERL_IN_XSUB_RE
19536 Perl_pregfree(pTHX_ REGEXP *r)
19542 Perl_pregfree2(pTHX_ REGEXP *rx)
19544 struct regexp *const r = ReANY(rx);
19545 GET_RE_DEBUG_FLAGS_DECL;
19547 PERL_ARGS_ASSERT_PREGFREE2;
19549 if (r->mother_re) {
19550 ReREFCNT_dec(r->mother_re);
19552 CALLREGFREE_PVT(rx); /* free the private data */
19553 SvREFCNT_dec(RXp_PAREN_NAMES(r));
19557 for (i = 0; i < 2; i++) {
19558 SvREFCNT_dec(r->substrs->data[i].substr);
19559 SvREFCNT_dec(r->substrs->data[i].utf8_substr);
19561 Safefree(r->substrs);
19563 RX_MATCH_COPY_FREE(rx);
19564 #ifdef PERL_ANY_COW
19565 SvREFCNT_dec(r->saved_copy);
19568 SvREFCNT_dec(r->qr_anoncv);
19569 if (r->recurse_locinput)
19570 Safefree(r->recurse_locinput);
19576 Copy ssv to dsv, both of which should of type SVt_REGEXP or SVt_PVLV,
19577 except that dsv will be created if NULL.
19579 This function is used in two main ways. First to implement
19580 $r = qr/....; $s = $$r;
19582 Secondly, it is used as a hacky workaround to the structural issue of
19584 being stored in the regexp structure which is in turn stored in
19585 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
19586 could be PL_curpm in multiple contexts, and could require multiple
19587 result sets being associated with the pattern simultaneously, such
19588 as when doing a recursive match with (??{$qr})
19590 The solution is to make a lightweight copy of the regexp structure
19591 when a qr// is returned from the code executed by (??{$qr}) this
19592 lightweight copy doesn't actually own any of its data except for
19593 the starp/end and the actual regexp structure itself.
19599 Perl_reg_temp_copy(pTHX_ REGEXP *dsv, REGEXP *ssv)
19601 struct regexp *drx;
19602 struct regexp *const srx = ReANY(ssv);
19603 const bool islv = dsv && SvTYPE(dsv) == SVt_PVLV;
19605 PERL_ARGS_ASSERT_REG_TEMP_COPY;
19608 dsv = (REGEXP*) newSV_type(SVt_REGEXP);
19610 SvOK_off((SV *)dsv);
19612 /* For PVLVs, the head (sv_any) points to an XPVLV, while
19613 * the LV's xpvlenu_rx will point to a regexp body, which
19614 * we allocate here */
19615 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
19616 assert(!SvPVX(dsv));
19617 ((XPV*)SvANY(dsv))->xpv_len_u.xpvlenu_rx = temp->sv_any;
19618 temp->sv_any = NULL;
19619 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
19620 SvREFCNT_dec_NN(temp);
19621 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
19622 ing below will not set it. */
19623 SvCUR_set(dsv, SvCUR(ssv));
19626 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
19627 sv_force_normal(sv) is called. */
19631 SvFLAGS(dsv) |= SvFLAGS(ssv) & (SVf_POK|SVp_POK|SVf_UTF8);
19632 SvPV_set(dsv, RX_WRAPPED(ssv));
19633 /* We share the same string buffer as the original regexp, on which we
19634 hold a reference count, incremented when mother_re is set below.
19635 The string pointer is copied here, being part of the regexp struct.
19637 memcpy(&(drx->xpv_cur), &(srx->xpv_cur),
19638 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
19642 const I32 npar = srx->nparens+1;
19643 Newx(drx->offs, npar, regexp_paren_pair);
19644 Copy(srx->offs, drx->offs, npar, regexp_paren_pair);
19646 if (srx->substrs) {
19648 Newx(drx->substrs, 1, struct reg_substr_data);
19649 StructCopy(srx->substrs, drx->substrs, struct reg_substr_data);
19651 for (i = 0; i < 2; i++) {
19652 SvREFCNT_inc_void(drx->substrs->data[i].substr);
19653 SvREFCNT_inc_void(drx->substrs->data[i].utf8_substr);
19656 /* check_substr and check_utf8, if non-NULL, point to either their
19657 anchored or float namesakes, and don't hold a second reference. */
19659 RX_MATCH_COPIED_off(dsv);
19660 #ifdef PERL_ANY_COW
19661 drx->saved_copy = NULL;
19663 drx->mother_re = ReREFCNT_inc(srx->mother_re ? srx->mother_re : ssv);
19664 SvREFCNT_inc_void(drx->qr_anoncv);
19665 if (srx->recurse_locinput)
19666 Newx(drx->recurse_locinput,srx->nparens + 1,char *);
19673 /* regfree_internal()
19675 Free the private data in a regexp. This is overloadable by
19676 extensions. Perl takes care of the regexp structure in pregfree(),
19677 this covers the *pprivate pointer which technically perl doesn't
19678 know about, however of course we have to handle the
19679 regexp_internal structure when no extension is in use.
19681 Note this is called before freeing anything in the regexp
19686 Perl_regfree_internal(pTHX_ REGEXP * const rx)
19688 struct regexp *const r = ReANY(rx);
19689 RXi_GET_DECL(r,ri);
19690 GET_RE_DEBUG_FLAGS_DECL;
19692 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
19698 SV *dsv= sv_newmortal();
19699 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
19700 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), PL_dump_re_max_len);
19701 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
19702 PL_colors[4],PL_colors[5],s);
19705 #ifdef RE_TRACK_PATTERN_OFFSETS
19707 Safefree(ri->u.offsets); /* 20010421 MJD */
19709 if (ri->code_blocks)
19710 S_free_codeblocks(aTHX_ ri->code_blocks);
19713 int n = ri->data->count;
19716 /* If you add a ->what type here, update the comment in regcomp.h */
19717 switch (ri->data->what[n]) {
19723 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
19726 Safefree(ri->data->data[n]);
19732 { /* Aho Corasick add-on structure for a trie node.
19733 Used in stclass optimization only */
19735 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
19736 #ifdef USE_ITHREADS
19740 refcount = --aho->refcount;
19743 PerlMemShared_free(aho->states);
19744 PerlMemShared_free(aho->fail);
19745 /* do this last!!!! */
19746 PerlMemShared_free(ri->data->data[n]);
19747 /* we should only ever get called once, so
19748 * assert as much, and also guard the free
19749 * which /might/ happen twice. At the least
19750 * it will make code anlyzers happy and it
19751 * doesn't cost much. - Yves */
19752 assert(ri->regstclass);
19753 if (ri->regstclass) {
19754 PerlMemShared_free(ri->regstclass);
19755 ri->regstclass = 0;
19762 /* trie structure. */
19764 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
19765 #ifdef USE_ITHREADS
19769 refcount = --trie->refcount;
19772 PerlMemShared_free(trie->charmap);
19773 PerlMemShared_free(trie->states);
19774 PerlMemShared_free(trie->trans);
19776 PerlMemShared_free(trie->bitmap);
19778 PerlMemShared_free(trie->jump);
19779 PerlMemShared_free(trie->wordinfo);
19780 /* do this last!!!! */
19781 PerlMemShared_free(ri->data->data[n]);
19786 Perl_croak(aTHX_ "panic: regfree data code '%c'",
19787 ri->data->what[n]);
19790 Safefree(ri->data->what);
19791 Safefree(ri->data);
19797 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
19798 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
19799 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
19802 re_dup_guts - duplicate a regexp.
19804 This routine is expected to clone a given regexp structure. It is only
19805 compiled under USE_ITHREADS.
19807 After all of the core data stored in struct regexp is duplicated
19808 the regexp_engine.dupe method is used to copy any private data
19809 stored in the *pprivate pointer. This allows extensions to handle
19810 any duplication it needs to do.
19812 See pregfree() and regfree_internal() if you change anything here.
19814 #if defined(USE_ITHREADS)
19815 #ifndef PERL_IN_XSUB_RE
19817 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
19821 const struct regexp *r = ReANY(sstr);
19822 struct regexp *ret = ReANY(dstr);
19824 PERL_ARGS_ASSERT_RE_DUP_GUTS;
19826 npar = r->nparens+1;
19827 Newx(ret->offs, npar, regexp_paren_pair);
19828 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19830 if (ret->substrs) {
19831 /* Do it this way to avoid reading from *r after the StructCopy().
19832 That way, if any of the sv_dup_inc()s dislodge *r from the L1
19833 cache, it doesn't matter. */
19835 const bool anchored = r->check_substr
19836 ? r->check_substr == r->substrs->data[0].substr
19837 : r->check_utf8 == r->substrs->data[0].utf8_substr;
19838 Newx(ret->substrs, 1, struct reg_substr_data);
19839 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19841 for (i = 0; i < 2; i++) {
19842 ret->substrs->data[i].substr =
19843 sv_dup_inc(ret->substrs->data[i].substr, param);
19844 ret->substrs->data[i].utf8_substr =
19845 sv_dup_inc(ret->substrs->data[i].utf8_substr, param);
19848 /* check_substr and check_utf8, if non-NULL, point to either their
19849 anchored or float namesakes, and don't hold a second reference. */
19851 if (ret->check_substr) {
19853 assert(r->check_utf8 == r->substrs->data[0].utf8_substr);
19855 ret->check_substr = ret->substrs->data[0].substr;
19856 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
19858 assert(r->check_substr == r->substrs->data[1].substr);
19859 assert(r->check_utf8 == r->substrs->data[1].utf8_substr);
19861 ret->check_substr = ret->substrs->data[1].substr;
19862 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
19864 } else if (ret->check_utf8) {
19866 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
19868 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
19873 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
19874 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
19875 if (r->recurse_locinput)
19876 Newx(ret->recurse_locinput,r->nparens + 1,char *);
19879 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
19881 if (RX_MATCH_COPIED(dstr))
19882 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
19884 ret->subbeg = NULL;
19885 #ifdef PERL_ANY_COW
19886 ret->saved_copy = NULL;
19889 /* Whether mother_re be set or no, we need to copy the string. We
19890 cannot refrain from copying it when the storage points directly to
19891 our mother regexp, because that's
19892 1: a buffer in a different thread
19893 2: something we no longer hold a reference on
19894 so we need to copy it locally. */
19895 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED_const(sstr), SvCUR(sstr)+1);
19896 ret->mother_re = NULL;
19898 #endif /* PERL_IN_XSUB_RE */
19903 This is the internal complement to regdupe() which is used to copy
19904 the structure pointed to by the *pprivate pointer in the regexp.
19905 This is the core version of the extension overridable cloning hook.
19906 The regexp structure being duplicated will be copied by perl prior
19907 to this and will be provided as the regexp *r argument, however
19908 with the /old/ structures pprivate pointer value. Thus this routine
19909 may override any copying normally done by perl.
19911 It returns a pointer to the new regexp_internal structure.
19915 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
19918 struct regexp *const r = ReANY(rx);
19919 regexp_internal *reti;
19921 RXi_GET_DECL(r,ri);
19923 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
19927 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
19928 char, regexp_internal);
19929 Copy(ri->program, reti->program, len+1, regnode);
19932 if (ri->code_blocks) {
19934 Newx(reti->code_blocks, 1, struct reg_code_blocks);
19935 Newx(reti->code_blocks->cb, ri->code_blocks->count,
19936 struct reg_code_block);
19937 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
19938 ri->code_blocks->count, struct reg_code_block);
19939 for (n = 0; n < ri->code_blocks->count; n++)
19940 reti->code_blocks->cb[n].src_regex = (REGEXP*)
19941 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
19942 reti->code_blocks->count = ri->code_blocks->count;
19943 reti->code_blocks->refcnt = 1;
19946 reti->code_blocks = NULL;
19948 reti->regstclass = NULL;
19951 struct reg_data *d;
19952 const int count = ri->data->count;
19955 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
19956 char, struct reg_data);
19957 Newx(d->what, count, U8);
19960 for (i = 0; i < count; i++) {
19961 d->what[i] = ri->data->what[i];
19962 switch (d->what[i]) {
19963 /* see also regcomp.h and regfree_internal() */
19964 case 'a': /* actually an AV, but the dup function is identical.
19965 values seem to be "plain sv's" generally. */
19966 case 'r': /* a compiled regex (but still just another SV) */
19967 case 's': /* an RV (currently only used for an RV to an AV by the ANYOF code)
19968 this use case should go away, the code could have used
19969 'a' instead - see S_set_ANYOF_arg() for array contents. */
19970 case 'S': /* actually an SV, but the dup function is identical. */
19971 case 'u': /* actually an HV, but the dup function is identical.
19972 values are "plain sv's" */
19973 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
19976 /* Synthetic Start Class - "Fake" charclass we generate to optimize
19977 * patterns which could start with several different things. Pre-TRIE
19978 * this was more important than it is now, however this still helps
19979 * in some places, for instance /x?a+/ might produce a SSC equivalent
19980 * to [xa]. This is used by Perl_re_intuit_start() and S_find_byclass()
19983 /* This is cheating. */
19984 Newx(d->data[i], 1, regnode_ssc);
19985 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
19986 reti->regstclass = (regnode*)d->data[i];
19989 /* AHO-CORASICK fail table */
19990 /* Trie stclasses are readonly and can thus be shared
19991 * without duplication. We free the stclass in pregfree
19992 * when the corresponding reg_ac_data struct is freed.
19994 reti->regstclass= ri->regstclass;
19997 /* TRIE transition table */
19999 ((reg_trie_data*)ri->data->data[i])->refcount++;
20002 case 'l': /* (?{...}) or (??{ ... }) code (cb->block) */
20003 case 'L': /* same when RExC_pm_flags & PMf_HAS_CV and code
20004 is not from another regexp */
20005 d->data[i] = ri->data->data[i];
20008 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
20009 ri->data->what[i]);
20018 reti->name_list_idx = ri->name_list_idx;
20020 #ifdef RE_TRACK_PATTERN_OFFSETS
20021 if (ri->u.offsets) {
20022 Newx(reti->u.offsets, 2*len+1, U32);
20023 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
20026 SetProgLen(reti,len);
20029 return (void*)reti;
20032 #endif /* USE_ITHREADS */
20034 #ifndef PERL_IN_XSUB_RE
20037 - regnext - dig the "next" pointer out of a node
20040 Perl_regnext(pTHX_ regnode *p)
20047 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
20048 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
20049 (int)OP(p), (int)REGNODE_MAX);
20052 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
20061 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
20064 STRLEN l1 = strlen(pat1);
20065 STRLEN l2 = strlen(pat2);
20068 const char *message;
20070 PERL_ARGS_ASSERT_RE_CROAK2;
20076 Copy(pat1, buf, l1 , char);
20077 Copy(pat2, buf + l1, l2 , char);
20078 buf[l1 + l2] = '\n';
20079 buf[l1 + l2 + 1] = '\0';
20080 va_start(args, pat2);
20081 msv = vmess(buf, &args);
20083 message = SvPV_const(msv,l1);
20086 Copy(message, buf, l1 , char);
20087 /* l1-1 to avoid \n */
20088 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, l1-1, buf));
20091 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
20093 #ifndef PERL_IN_XSUB_RE
20095 Perl_save_re_context(pTHX)
20100 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
20103 const REGEXP * const rx = PM_GETRE(PL_curpm);
20105 nparens = RX_NPARENS(rx);
20108 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
20109 * that PL_curpm will be null, but that utf8.pm and the modules it
20110 * loads will only use $1..$3.
20111 * The t/porting/re_context.t test file checks this assumption.
20116 for (i = 1; i <= nparens; i++) {
20117 char digits[TYPE_CHARS(long)];
20118 const STRLEN len = my_snprintf(digits, sizeof(digits),
20120 GV *const *const gvp
20121 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
20124 GV * const gv = *gvp;
20125 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
20135 S_put_code_point(pTHX_ SV *sv, UV c)
20137 PERL_ARGS_ASSERT_PUT_CODE_POINT;
20140 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
20142 else if (isPRINT(c)) {
20143 const char string = (char) c;
20145 /* We use {phrase} as metanotation in the class, so also escape literal
20147 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
20148 sv_catpvs(sv, "\\");
20149 sv_catpvn(sv, &string, 1);
20151 else if (isMNEMONIC_CNTRL(c)) {
20152 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
20155 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
20159 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
20162 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
20164 /* Appends to 'sv' a displayable version of the range of code points from
20165 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
20166 * that have them, when they occur at the beginning or end of the range.
20167 * It uses hex to output the remaining code points, unless 'allow_literals'
20168 * is true, in which case the printable ASCII ones are output as-is (though
20169 * some of these will be escaped by put_code_point()).
20171 * NOTE: This is designed only for printing ranges of code points that fit
20172 * inside an ANYOF bitmap. Higher code points are simply suppressed
20175 const unsigned int min_range_count = 3;
20177 assert(start <= end);
20179 PERL_ARGS_ASSERT_PUT_RANGE;
20181 while (start <= end) {
20183 const char * format;
20185 if (end - start < min_range_count) {
20187 /* Output chars individually when they occur in short ranges */
20188 for (; start <= end; start++) {
20189 put_code_point(sv, start);
20194 /* If permitted by the input options, and there is a possibility that
20195 * this range contains a printable literal, look to see if there is
20197 if (allow_literals && start <= MAX_PRINT_A) {
20199 /* If the character at the beginning of the range isn't an ASCII
20200 * printable, effectively split the range into two parts:
20201 * 1) the portion before the first such printable,
20203 * and output them separately. */
20204 if (! isPRINT_A(start)) {
20205 UV temp_end = start + 1;
20207 /* There is no point looking beyond the final possible
20208 * printable, in MAX_PRINT_A */
20209 UV max = MIN(end, MAX_PRINT_A);
20211 while (temp_end <= max && ! isPRINT_A(temp_end)) {
20215 /* Here, temp_end points to one beyond the first printable if
20216 * found, or to one beyond 'max' if not. If none found, make
20217 * sure that we use the entire range */
20218 if (temp_end > MAX_PRINT_A) {
20219 temp_end = end + 1;
20222 /* Output the first part of the split range: the part that
20223 * doesn't have printables, with the parameter set to not look
20224 * for literals (otherwise we would infinitely recurse) */
20225 put_range(sv, start, temp_end - 1, FALSE);
20227 /* The 2nd part of the range (if any) starts here. */
20230 /* We do a continue, instead of dropping down, because even if
20231 * the 2nd part is non-empty, it could be so short that we want
20232 * to output it as individual characters, as tested for at the
20233 * top of this loop. */
20237 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
20238 * output a sub-range of just the digits or letters, then process
20239 * the remaining portion as usual. */
20240 if (isALPHANUMERIC_A(start)) {
20241 UV mask = (isDIGIT_A(start))
20246 UV temp_end = start + 1;
20248 /* Find the end of the sub-range that includes just the
20249 * characters in the same class as the first character in it */
20250 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
20255 /* For short ranges, don't duplicate the code above to output
20256 * them; just call recursively */
20257 if (temp_end - start < min_range_count) {
20258 put_range(sv, start, temp_end, FALSE);
20260 else { /* Output as a range */
20261 put_code_point(sv, start);
20262 sv_catpvs(sv, "-");
20263 put_code_point(sv, temp_end);
20265 start = temp_end + 1;
20269 /* We output any other printables as individual characters */
20270 if (isPUNCT_A(start) || isSPACE_A(start)) {
20271 while (start <= end && (isPUNCT_A(start)
20272 || isSPACE_A(start)))
20274 put_code_point(sv, start);
20279 } /* End of looking for literals */
20281 /* Here is not to output as a literal. Some control characters have
20282 * mnemonic names. Split off any of those at the beginning and end of
20283 * the range to print mnemonically. It isn't possible for many of
20284 * these to be in a row, so this won't overwhelm with output */
20286 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
20288 while (isMNEMONIC_CNTRL(start) && start <= end) {
20289 put_code_point(sv, start);
20293 /* If this didn't take care of the whole range ... */
20294 if (start <= end) {
20296 /* Look backwards from the end to find the final non-mnemonic
20299 while (isMNEMONIC_CNTRL(temp_end)) {
20303 /* And separately output the interior range that doesn't start
20304 * or end with mnemonics */
20305 put_range(sv, start, temp_end, FALSE);
20307 /* Then output the mnemonic trailing controls */
20308 start = temp_end + 1;
20309 while (start <= end) {
20310 put_code_point(sv, start);
20317 /* As a final resort, output the range or subrange as hex. */
20319 this_end = (end < NUM_ANYOF_CODE_POINTS)
20321 : NUM_ANYOF_CODE_POINTS - 1;
20322 #if NUM_ANYOF_CODE_POINTS > 256
20323 format = (this_end < 256)
20324 ? "\\x%02" UVXf "-\\x%02" UVXf
20325 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
20327 format = "\\x%02" UVXf "-\\x%02" UVXf;
20329 GCC_DIAG_IGNORE(-Wformat-nonliteral);
20330 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
20337 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
20339 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
20343 bool allow_literals = TRUE;
20345 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
20347 /* Generally, it is more readable if printable characters are output as
20348 * literals, but if a range (nearly) spans all of them, it's best to output
20349 * it as a single range. This code will use a single range if all but 2
20350 * ASCII printables are in it */
20351 invlist_iterinit(invlist);
20352 while (invlist_iternext(invlist, &start, &end)) {
20354 /* If the range starts beyond the final printable, it doesn't have any
20356 if (start > MAX_PRINT_A) {
20360 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
20361 * all but two, the range must start and end no later than 2 from
20363 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
20364 if (end > MAX_PRINT_A) {
20370 if (end - start >= MAX_PRINT_A - ' ' - 2) {
20371 allow_literals = FALSE;
20376 invlist_iterfinish(invlist);
20378 /* Here we have figured things out. Output each range */
20379 invlist_iterinit(invlist);
20380 while (invlist_iternext(invlist, &start, &end)) {
20381 if (start >= NUM_ANYOF_CODE_POINTS) {
20384 put_range(sv, start, end, allow_literals);
20386 invlist_iterfinish(invlist);
20392 S_put_charclass_bitmap_innards_common(pTHX_
20393 SV* invlist, /* The bitmap */
20394 SV* posixes, /* Under /l, things like [:word:], \S */
20395 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
20396 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
20397 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
20398 const bool invert /* Is the result to be inverted? */
20401 /* Create and return an SV containing a displayable version of the bitmap
20402 * and associated information determined by the input parameters. If the
20403 * output would have been only the inversion indicator '^', NULL is instead
20408 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
20411 output = newSVpvs("^");
20414 output = newSVpvs("");
20417 /* First, the code points in the bitmap that are unconditionally there */
20418 put_charclass_bitmap_innards_invlist(output, invlist);
20420 /* Traditionally, these have been placed after the main code points */
20422 sv_catsv(output, posixes);
20425 if (only_utf8 && _invlist_len(only_utf8)) {
20426 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
20427 put_charclass_bitmap_innards_invlist(output, only_utf8);
20430 if (not_utf8 && _invlist_len(not_utf8)) {
20431 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
20432 put_charclass_bitmap_innards_invlist(output, not_utf8);
20435 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
20436 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
20437 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
20439 /* This is the only list in this routine that can legally contain code
20440 * points outside the bitmap range. The call just above to
20441 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
20442 * output them here. There's about a half-dozen possible, and none in
20443 * contiguous ranges longer than 2 */
20444 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20446 SV* above_bitmap = NULL;
20448 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
20450 invlist_iterinit(above_bitmap);
20451 while (invlist_iternext(above_bitmap, &start, &end)) {
20454 for (i = start; i <= end; i++) {
20455 put_code_point(output, i);
20458 invlist_iterfinish(above_bitmap);
20459 SvREFCNT_dec_NN(above_bitmap);
20463 if (invert && SvCUR(output) == 1) {
20471 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
20473 SV *nonbitmap_invlist,
20474 SV *only_utf8_locale_invlist,
20475 const regnode * const node,
20476 const bool force_as_is_display)
20478 /* Appends to 'sv' a displayable version of the innards of the bracketed
20479 * character class defined by the other arguments:
20480 * 'bitmap' points to the bitmap.
20481 * 'nonbitmap_invlist' is an inversion list of the code points that are in
20482 * the bitmap range, but for some reason aren't in the bitmap; NULL if
20483 * none. The reasons for this could be that they require some
20484 * condition such as the target string being or not being in UTF-8
20485 * (under /d), or because they came from a user-defined property that
20486 * was not resolved at the time of the regex compilation (under /u)
20487 * 'only_utf8_locale_invlist' is an inversion list of the code points that
20488 * are valid only if the runtime locale is a UTF-8 one; NULL if none
20489 * 'node' is the regex pattern node. It is needed only when the above two
20490 * parameters are not null, and is passed so that this routine can
20491 * tease apart the various reasons for them.
20492 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
20493 * to invert things to see if that leads to a cleaner display. If
20494 * FALSE, this routine is free to use its judgment about doing this.
20496 * It returns TRUE if there was actually something output. (It may be that
20497 * the bitmap, etc is empty.)
20499 * When called for outputting the bitmap of a non-ANYOF node, just pass the
20500 * bitmap, with the succeeding parameters set to NULL, and the final one to
20504 /* In general, it tries to display the 'cleanest' representation of the
20505 * innards, choosing whether to display them inverted or not, regardless of
20506 * whether the class itself is to be inverted. However, there are some
20507 * cases where it can't try inverting, as what actually matches isn't known
20508 * until runtime, and hence the inversion isn't either. */
20509 bool inverting_allowed = ! force_as_is_display;
20512 STRLEN orig_sv_cur = SvCUR(sv);
20514 SV* invlist; /* Inversion list we accumulate of code points that
20515 are unconditionally matched */
20516 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
20518 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
20520 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
20521 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
20524 SV* as_is_display; /* The output string when we take the inputs
20526 SV* inverted_display; /* The output string when we invert the inputs */
20528 U8 flags = (node) ? ANYOF_FLAGS(node) : 0;
20530 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
20532 /* We are biased in favor of displaying things without them being inverted,
20533 * as that is generally easier to understand */
20534 const int bias = 5;
20536 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
20538 /* Start off with whatever code points are passed in. (We clone, so we
20539 * don't change the caller's list) */
20540 if (nonbitmap_invlist) {
20541 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
20542 invlist = invlist_clone(nonbitmap_invlist);
20544 else { /* Worst case size is every other code point is matched */
20545 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
20549 if (OP(node) == ANYOFD) {
20551 /* This flag indicates that the code points below 0x100 in the
20552 * nonbitmap list are precisely the ones that match only when the
20553 * target is UTF-8 (they should all be non-ASCII). */
20554 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
20556 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
20557 _invlist_subtract(invlist, only_utf8, &invlist);
20560 /* And this flag for matching all non-ASCII 0xFF and below */
20561 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
20563 not_utf8 = invlist_clone(PL_UpperLatin1);
20566 else if (OP(node) == ANYOFL) {
20568 /* If either of these flags are set, what matches isn't
20569 * determinable except during execution, so don't know enough here
20571 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
20572 inverting_allowed = FALSE;
20575 /* What the posix classes match also varies at runtime, so these
20576 * will be output symbolically. */
20577 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
20580 posixes = newSVpvs("");
20581 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
20582 if (ANYOF_POSIXL_TEST(node,i)) {
20583 sv_catpv(posixes, anyofs[i]);
20590 /* Accumulate the bit map into the unconditional match list */
20591 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
20592 if (BITMAP_TEST(bitmap, i)) {
20594 for (; i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i); i++) {
20597 invlist = _add_range_to_invlist(invlist, start, i-1);
20601 /* Make sure that the conditional match lists don't have anything in them
20602 * that match unconditionally; otherwise the output is quite confusing.
20603 * This could happen if the code that populates these misses some
20606 _invlist_subtract(only_utf8, invlist, &only_utf8);
20609 _invlist_subtract(not_utf8, invlist, ¬_utf8);
20612 if (only_utf8_locale_invlist) {
20614 /* Since this list is passed in, we have to make a copy before
20616 only_utf8_locale = invlist_clone(only_utf8_locale_invlist);
20618 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
20620 /* And, it can get really weird for us to try outputting an inverted
20621 * form of this list when it has things above the bitmap, so don't even
20623 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20624 inverting_allowed = FALSE;
20628 /* Calculate what the output would be if we take the input as-is */
20629 as_is_display = put_charclass_bitmap_innards_common(invlist,
20636 /* If have to take the output as-is, just do that */
20637 if (! inverting_allowed) {
20638 if (as_is_display) {
20639 sv_catsv(sv, as_is_display);
20640 SvREFCNT_dec_NN(as_is_display);
20643 else { /* But otherwise, create the output again on the inverted input, and
20644 use whichever version is shorter */
20646 int inverted_bias, as_is_bias;
20648 /* We will apply our bias to whichever of the the results doesn't have
20658 inverted_bias = bias;
20661 /* Now invert each of the lists that contribute to the output,
20662 * excluding from the result things outside the possible range */
20664 /* For the unconditional inversion list, we have to add in all the
20665 * conditional code points, so that when inverted, they will be gone
20667 _invlist_union(only_utf8, invlist, &invlist);
20668 _invlist_union(not_utf8, invlist, &invlist);
20669 _invlist_union(only_utf8_locale, invlist, &invlist);
20670 _invlist_invert(invlist);
20671 _invlist_intersection(invlist, PL_InBitmap, &invlist);
20674 _invlist_invert(only_utf8);
20675 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
20677 else if (not_utf8) {
20679 /* If a code point matches iff the target string is not in UTF-8,
20680 * then complementing the result has it not match iff not in UTF-8,
20681 * which is the same thing as matching iff it is UTF-8. */
20682 only_utf8 = not_utf8;
20686 if (only_utf8_locale) {
20687 _invlist_invert(only_utf8_locale);
20688 _invlist_intersection(only_utf8_locale,
20690 &only_utf8_locale);
20693 inverted_display = put_charclass_bitmap_innards_common(
20698 only_utf8_locale, invert);
20700 /* Use the shortest representation, taking into account our bias
20701 * against showing it inverted */
20702 if ( inverted_display
20703 && ( ! as_is_display
20704 || ( SvCUR(inverted_display) + inverted_bias
20705 < SvCUR(as_is_display) + as_is_bias)))
20707 sv_catsv(sv, inverted_display);
20709 else if (as_is_display) {
20710 sv_catsv(sv, as_is_display);
20713 SvREFCNT_dec(as_is_display);
20714 SvREFCNT_dec(inverted_display);
20717 SvREFCNT_dec_NN(invlist);
20718 SvREFCNT_dec(only_utf8);
20719 SvREFCNT_dec(not_utf8);
20720 SvREFCNT_dec(posixes);
20721 SvREFCNT_dec(only_utf8_locale);
20723 return SvCUR(sv) > orig_sv_cur;
20726 #define CLEAR_OPTSTART \
20727 if (optstart) STMT_START { \
20728 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
20729 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
20733 #define DUMPUNTIL(b,e) \
20735 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
20737 STATIC const regnode *
20738 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
20739 const regnode *last, const regnode *plast,
20740 SV* sv, I32 indent, U32 depth)
20742 U8 op = PSEUDO; /* Arbitrary non-END op. */
20743 const regnode *next;
20744 const regnode *optstart= NULL;
20746 RXi_GET_DECL(r,ri);
20747 GET_RE_DEBUG_FLAGS_DECL;
20749 PERL_ARGS_ASSERT_DUMPUNTIL;
20751 #ifdef DEBUG_DUMPUNTIL
20752 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n",indent,node-start,
20753 last ? last-start : 0,plast ? plast-start : 0);
20756 if (plast && plast < last)
20759 while (PL_regkind[op] != END && (!last || node < last)) {
20761 /* While that wasn't END last time... */
20764 if (op == CLOSE || op == WHILEM)
20766 next = regnext((regnode *)node);
20769 if (OP(node) == OPTIMIZED) {
20770 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
20777 regprop(r, sv, node, NULL, NULL);
20778 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
20779 (int)(2*indent + 1), "", SvPVX_const(sv));
20781 if (OP(node) != OPTIMIZED) {
20782 if (next == NULL) /* Next ptr. */
20783 Perl_re_printf( aTHX_ " (0)");
20784 else if (PL_regkind[(U8)op] == BRANCH
20785 && PL_regkind[OP(next)] != BRANCH )
20786 Perl_re_printf( aTHX_ " (FAIL)");
20788 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
20789 Perl_re_printf( aTHX_ "\n");
20793 if (PL_regkind[(U8)op] == BRANCHJ) {
20796 const regnode *nnode = (OP(next) == LONGJMP
20797 ? regnext((regnode *)next)
20799 if (last && nnode > last)
20801 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
20804 else if (PL_regkind[(U8)op] == BRANCH) {
20806 DUMPUNTIL(NEXTOPER(node), next);
20808 else if ( PL_regkind[(U8)op] == TRIE ) {
20809 const regnode *this_trie = node;
20810 const char op = OP(node);
20811 const U32 n = ARG(node);
20812 const reg_ac_data * const ac = op>=AHOCORASICK ?
20813 (reg_ac_data *)ri->data->data[n] :
20815 const reg_trie_data * const trie =
20816 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
20818 AV *const trie_words
20819 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
20821 const regnode *nextbranch= NULL;
20824 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
20825 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
20827 Perl_re_indentf( aTHX_ "%s ",
20830 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
20831 SvCUR(*elem_ptr), PL_dump_re_max_len,
20832 PL_colors[0], PL_colors[1],
20834 ? PERL_PV_ESCAPE_UNI
20836 | PERL_PV_PRETTY_ELLIPSES
20837 | PERL_PV_PRETTY_LTGT
20842 U16 dist= trie->jump[word_idx+1];
20843 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
20844 (UV)((dist ? this_trie + dist : next) - start));
20847 nextbranch= this_trie + trie->jump[0];
20848 DUMPUNTIL(this_trie + dist, nextbranch);
20850 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
20851 nextbranch= regnext((regnode *)nextbranch);
20853 Perl_re_printf( aTHX_ "\n");
20856 if (last && next > last)
20861 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
20862 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
20863 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
20865 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
20867 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
20869 else if ( op == PLUS || op == STAR) {
20870 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
20872 else if (PL_regkind[(U8)op] == ANYOF) {
20873 /* arglen 1 + class block */
20874 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
20875 ? ANYOF_POSIXL_SKIP
20877 node = NEXTOPER(node);
20879 else if (PL_regkind[(U8)op] == EXACT) {
20880 /* Literal string, where present. */
20881 node += NODE_SZ_STR(node) - 1;
20882 node = NEXTOPER(node);
20885 node = NEXTOPER(node);
20886 node += regarglen[(U8)op];
20888 if (op == CURLYX || op == OPEN)
20892 #ifdef DEBUG_DUMPUNTIL
20893 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
20898 #endif /* DEBUGGING */
20901 * ex: set ts=8 sts=4 sw=4 et: