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 ANYOF_POSIXL_ZERO(&temp);
1705 for (i = 0; i < ANYOF_MAX; i++) {
1707 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1708 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1710 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1711 ANYOF_POSIXL_SET(&temp, i + add);
1713 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1715 ANYOF_POSIXL_AND(&temp, ssc);
1717 } /* else ssc already has no posixes */
1718 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1719 in its initial state */
1720 else if (! is_ANYOF_SYNTHETIC(and_with)
1721 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1723 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1724 * copy it over 'ssc' */
1725 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1726 if (is_ANYOF_SYNTHETIC(and_with)) {
1727 StructCopy(and_with, ssc, regnode_ssc);
1730 ssc->invlist = anded_cp_list;
1731 ANYOF_POSIXL_ZERO(ssc);
1732 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1733 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1737 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1738 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1740 /* One or the other of P1, P2 is non-empty. */
1741 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1742 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1744 ssc_union(ssc, anded_cp_list, FALSE);
1746 else { /* P1 = P2 = empty */
1747 ssc_intersection(ssc, anded_cp_list, FALSE);
1753 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1754 const regnode_charclass *or_with)
1756 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1757 * another SSC or a regular ANYOF class. Can create false positives if
1758 * 'or_with' is to be inverted. */
1763 PERL_ARGS_ASSERT_SSC_OR;
1765 assert(is_ANYOF_SYNTHETIC(ssc));
1767 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1768 * the code point inversion list and just the relevant flags */
1769 if (is_ANYOF_SYNTHETIC(or_with)) {
1770 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1771 ored_flags = ANYOF_FLAGS(or_with);
1774 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1775 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1776 if (OP(or_with) != ANYOFD) {
1778 |= ANYOF_FLAGS(or_with)
1779 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1780 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1781 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1783 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1788 ANYOF_FLAGS(ssc) |= ored_flags;
1790 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1791 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1792 * 'or_with' may be inverted. When not inverted, we have the simple
1793 * situation of computing:
1794 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1795 * If P1|P2 yields a situation with both a class and its complement are
1796 * set, like having both \w and \W, this matches all code points, and we
1797 * can delete these from the P component of the ssc going forward. XXX We
1798 * might be able to delete all the P components, but I (khw) am not certain
1799 * about this, and it is better to be safe.
1802 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1803 * <= (C1 | P1) | ~C2
1804 * <= (C1 | ~C2) | P1
1805 * (which results in actually simpler code than the non-inverted case)
1808 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1809 && ! is_ANYOF_SYNTHETIC(or_with))
1811 /* We ignore P2, leaving P1 going forward */
1812 } /* else Not inverted */
1813 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1814 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1815 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1817 for (i = 0; i < ANYOF_MAX; i += 2) {
1818 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1820 ssc_match_all_cp(ssc);
1821 ANYOF_POSIXL_CLEAR(ssc, i);
1822 ANYOF_POSIXL_CLEAR(ssc, i+1);
1830 FALSE /* Already has been inverted */
1834 PERL_STATIC_INLINE void
1835 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1837 PERL_ARGS_ASSERT_SSC_UNION;
1839 assert(is_ANYOF_SYNTHETIC(ssc));
1841 _invlist_union_maybe_complement_2nd(ssc->invlist,
1847 PERL_STATIC_INLINE void
1848 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1850 const bool invert2nd)
1852 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1854 assert(is_ANYOF_SYNTHETIC(ssc));
1856 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1862 PERL_STATIC_INLINE void
1863 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1865 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1867 assert(is_ANYOF_SYNTHETIC(ssc));
1869 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1872 PERL_STATIC_INLINE void
1873 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1875 /* AND just the single code point 'cp' into the SSC 'ssc' */
1877 SV* cp_list = _new_invlist(2);
1879 PERL_ARGS_ASSERT_SSC_CP_AND;
1881 assert(is_ANYOF_SYNTHETIC(ssc));
1883 cp_list = add_cp_to_invlist(cp_list, cp);
1884 ssc_intersection(ssc, cp_list,
1885 FALSE /* Not inverted */
1887 SvREFCNT_dec_NN(cp_list);
1890 PERL_STATIC_INLINE void
1891 S_ssc_clear_locale(regnode_ssc *ssc)
1893 /* Set the SSC 'ssc' to not match any locale things */
1894 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1896 assert(is_ANYOF_SYNTHETIC(ssc));
1898 ANYOF_POSIXL_ZERO(ssc);
1899 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1902 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1905 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1907 /* The synthetic start class is used to hopefully quickly winnow down
1908 * places where a pattern could start a match in the target string. If it
1909 * doesn't really narrow things down that much, there isn't much point to
1910 * having the overhead of using it. This function uses some very crude
1911 * heuristics to decide if to use the ssc or not.
1913 * It returns TRUE if 'ssc' rules out more than half what it considers to
1914 * be the "likely" possible matches, but of course it doesn't know what the
1915 * actual things being matched are going to be; these are only guesses
1917 * For /l matches, it assumes that the only likely matches are going to be
1918 * in the 0-255 range, uniformly distributed, so half of that is 127
1919 * For /a and /d matches, it assumes that the likely matches will be just
1920 * the ASCII range, so half of that is 63
1921 * For /u and there isn't anything matching above the Latin1 range, it
1922 * assumes that that is the only range likely to be matched, and uses
1923 * half that as the cut-off: 127. If anything matches above Latin1,
1924 * it assumes that all of Unicode could match (uniformly), except for
1925 * non-Unicode code points and things in the General Category "Other"
1926 * (unassigned, private use, surrogates, controls and formats). This
1927 * is a much large number. */
1929 U32 count = 0; /* Running total of number of code points matched by
1931 UV start, end; /* Start and end points of current range in inversion
1933 const U32 max_code_points = (LOC)
1935 : (( ! UNI_SEMANTICS
1936 || invlist_highest(ssc->invlist) < 256)
1939 const U32 max_match = max_code_points / 2;
1941 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1943 invlist_iterinit(ssc->invlist);
1944 while (invlist_iternext(ssc->invlist, &start, &end)) {
1945 if (start >= max_code_points) {
1948 end = MIN(end, max_code_points - 1);
1949 count += end - start + 1;
1950 if (count >= max_match) {
1951 invlist_iterfinish(ssc->invlist);
1961 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1963 /* The inversion list in the SSC is marked mortal; now we need a more
1964 * permanent copy, which is stored the same way that is done in a regular
1965 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1968 SV* invlist = invlist_clone(ssc->invlist);
1970 PERL_ARGS_ASSERT_SSC_FINALIZE;
1972 assert(is_ANYOF_SYNTHETIC(ssc));
1974 /* The code in this file assumes that all but these flags aren't relevant
1975 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1976 * by the time we reach here */
1977 assert(! (ANYOF_FLAGS(ssc)
1978 & ~( ANYOF_COMMON_FLAGS
1979 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1980 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
1982 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1984 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1985 NULL, NULL, NULL, FALSE);
1987 /* Make sure is clone-safe */
1988 ssc->invlist = NULL;
1990 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1991 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1994 if (RExC_contains_locale) {
1998 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2001 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2002 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2003 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2004 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2005 ? (TRIE_LIST_CUR( idx ) - 1) \
2011 dump_trie(trie,widecharmap,revcharmap)
2012 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2013 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2015 These routines dump out a trie in a somewhat readable format.
2016 The _interim_ variants are used for debugging the interim
2017 tables that are used to generate the final compressed
2018 representation which is what dump_trie expects.
2020 Part of the reason for their existence is to provide a form
2021 of documentation as to how the different representations function.
2026 Dumps the final compressed table form of the trie to Perl_debug_log.
2027 Used for debugging make_trie().
2031 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2032 AV *revcharmap, U32 depth)
2035 SV *sv=sv_newmortal();
2036 int colwidth= widecharmap ? 6 : 4;
2038 GET_RE_DEBUG_FLAGS_DECL;
2040 PERL_ARGS_ASSERT_DUMP_TRIE;
2042 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2043 depth+1, "Match","Base","Ofs" );
2045 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2046 SV ** const tmp = av_fetch( revcharmap, state, 0);
2048 Perl_re_printf( aTHX_ "%*s",
2050 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2051 PL_colors[0], PL_colors[1],
2052 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2053 PERL_PV_ESCAPE_FIRSTCHAR
2058 Perl_re_printf( aTHX_ "\n");
2059 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2061 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2062 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2063 Perl_re_printf( aTHX_ "\n");
2065 for( state = 1 ; state < trie->statecount ; state++ ) {
2066 const U32 base = trie->states[ state ].trans.base;
2068 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2070 if ( trie->states[ state ].wordnum ) {
2071 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2073 Perl_re_printf( aTHX_ "%6s", "" );
2076 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2081 while( ( base + ofs < trie->uniquecharcount ) ||
2082 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2083 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2087 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2089 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2090 if ( ( base + ofs >= trie->uniquecharcount )
2091 && ( base + ofs - trie->uniquecharcount
2093 && trie->trans[ base + ofs
2094 - trie->uniquecharcount ].check == state )
2096 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2097 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2100 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
2104 Perl_re_printf( aTHX_ "]");
2107 Perl_re_printf( aTHX_ "\n" );
2109 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2111 for (word=1; word <= trie->wordcount; word++) {
2112 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2113 (int)word, (int)(trie->wordinfo[word].prev),
2114 (int)(trie->wordinfo[word].len));
2116 Perl_re_printf( aTHX_ "\n" );
2119 Dumps a fully constructed but uncompressed trie in list form.
2120 List tries normally only are used for construction when the number of
2121 possible chars (trie->uniquecharcount) is very high.
2122 Used for debugging make_trie().
2125 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2126 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2130 SV *sv=sv_newmortal();
2131 int colwidth= widecharmap ? 6 : 4;
2132 GET_RE_DEBUG_FLAGS_DECL;
2134 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2136 /* print out the table precompression. */
2137 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2139 Perl_re_indentf( aTHX_ "%s",
2140 depth+1, "------:-----+-----------------\n" );
2142 for( state=1 ; state < next_alloc ; state ++ ) {
2145 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2146 depth+1, (UV)state );
2147 if ( ! trie->states[ state ].wordnum ) {
2148 Perl_re_printf( aTHX_ "%5s| ","");
2150 Perl_re_printf( aTHX_ "W%4x| ",
2151 trie->states[ state ].wordnum
2154 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2155 SV ** const tmp = av_fetch( revcharmap,
2156 TRIE_LIST_ITEM(state,charid).forid, 0);
2158 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2160 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2162 PL_colors[0], PL_colors[1],
2163 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2164 | PERL_PV_ESCAPE_FIRSTCHAR
2166 TRIE_LIST_ITEM(state,charid).forid,
2167 (UV)TRIE_LIST_ITEM(state,charid).newstate
2170 Perl_re_printf( aTHX_ "\n%*s| ",
2171 (int)((depth * 2) + 14), "");
2174 Perl_re_printf( aTHX_ "\n");
2179 Dumps a fully constructed but uncompressed trie in table form.
2180 This is the normal DFA style state transition table, with a few
2181 twists to facilitate compression later.
2182 Used for debugging make_trie().
2185 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2186 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2191 SV *sv=sv_newmortal();
2192 int colwidth= widecharmap ? 6 : 4;
2193 GET_RE_DEBUG_FLAGS_DECL;
2195 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2198 print out the table precompression so that we can do a visual check
2199 that they are identical.
2202 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2204 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2205 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2207 Perl_re_printf( aTHX_ "%*s",
2209 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2210 PL_colors[0], PL_colors[1],
2211 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2212 PERL_PV_ESCAPE_FIRSTCHAR
2218 Perl_re_printf( aTHX_ "\n");
2219 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2221 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2222 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2225 Perl_re_printf( aTHX_ "\n" );
2227 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2229 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2231 (UV)TRIE_NODENUM( state ) );
2233 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2234 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2236 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2238 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2240 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2241 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2242 (UV)trie->trans[ state ].check );
2244 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2245 (UV)trie->trans[ state ].check,
2246 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2254 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2255 startbranch: the first branch in the whole branch sequence
2256 first : start branch of sequence of branch-exact nodes.
2257 May be the same as startbranch
2258 last : Thing following the last branch.
2259 May be the same as tail.
2260 tail : item following the branch sequence
2261 count : words in the sequence
2262 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2263 depth : indent depth
2265 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2267 A trie is an N'ary tree where the branches are determined by digital
2268 decomposition of the key. IE, at the root node you look up the 1st character and
2269 follow that branch repeat until you find the end of the branches. Nodes can be
2270 marked as "accepting" meaning they represent a complete word. Eg:
2274 would convert into the following structure. Numbers represent states, letters
2275 following numbers represent valid transitions on the letter from that state, if
2276 the number is in square brackets it represents an accepting state, otherwise it
2277 will be in parenthesis.
2279 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2283 (1) +-i->(6)-+-s->[7]
2285 +-s->(3)-+-h->(4)-+-e->[5]
2287 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2289 This shows that when matching against the string 'hers' we will begin at state 1
2290 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2291 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2292 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2293 single traverse. We store a mapping from accepting to state to which word was
2294 matched, and then when we have multiple possibilities we try to complete the
2295 rest of the regex in the order in which they occurred in the alternation.
2297 The only prior NFA like behaviour that would be changed by the TRIE support is
2298 the silent ignoring of duplicate alternations which are of the form:
2300 / (DUPE|DUPE) X? (?{ ... }) Y /x
2302 Thus EVAL blocks following a trie may be called a different number of times with
2303 and without the optimisation. With the optimisations dupes will be silently
2304 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2305 the following demonstrates:
2307 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2309 which prints out 'word' three times, but
2311 'words'=~/(word|word|word)(?{ print $1 })S/
2313 which doesnt print it out at all. This is due to other optimisations kicking in.
2315 Example of what happens on a structural level:
2317 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2319 1: CURLYM[1] {1,32767}(18)
2330 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2331 and should turn into:
2333 1: CURLYM[1] {1,32767}(18)
2335 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2343 Cases where tail != last would be like /(?foo|bar)baz/:
2353 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2354 and would end up looking like:
2357 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2364 d = uvchr_to_utf8_flags(d, uv, 0);
2366 is the recommended Unicode-aware way of saying
2371 #define TRIE_STORE_REVCHAR(val) \
2374 SV *zlopp = newSV(UTF8_MAXBYTES); \
2375 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2376 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2377 SvCUR_set(zlopp, kapow - flrbbbbb); \
2380 av_push(revcharmap, zlopp); \
2382 char ooooff = (char)val; \
2383 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2387 /* This gets the next character from the input, folding it if not already
2389 #define TRIE_READ_CHAR STMT_START { \
2392 /* if it is UTF then it is either already folded, or does not need \
2394 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2396 else if (folder == PL_fold_latin1) { \
2397 /* This folder implies Unicode rules, which in the range expressible \
2398 * by not UTF is the lower case, with the two exceptions, one of \
2399 * which should have been taken care of before calling this */ \
2400 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2401 uvc = toLOWER_L1(*uc); \
2402 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2405 /* raw data, will be folded later if needed */ \
2413 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2414 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2415 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2416 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2417 TRIE_LIST_LEN( state ) = ging; \
2419 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2420 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2421 TRIE_LIST_CUR( state )++; \
2424 #define TRIE_LIST_NEW(state) STMT_START { \
2425 Newx( trie->states[ state ].trans.list, \
2426 4, reg_trie_trans_le ); \
2427 TRIE_LIST_CUR( state ) = 1; \
2428 TRIE_LIST_LEN( state ) = 4; \
2431 #define TRIE_HANDLE_WORD(state) STMT_START { \
2432 U16 dupe= trie->states[ state ].wordnum; \
2433 regnode * const noper_next = regnext( noper ); \
2436 /* store the word for dumping */ \
2438 if (OP(noper) != NOTHING) \
2439 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2441 tmp = newSVpvn_utf8( "", 0, UTF ); \
2442 av_push( trie_words, tmp ); \
2446 trie->wordinfo[curword].prev = 0; \
2447 trie->wordinfo[curword].len = wordlen; \
2448 trie->wordinfo[curword].accept = state; \
2450 if ( noper_next < tail ) { \
2452 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2454 trie->jump[curword] = (U16)(noper_next - convert); \
2456 jumper = noper_next; \
2458 nextbranch= regnext(cur); \
2462 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2463 /* chain, so that when the bits of chain are later */\
2464 /* linked together, the dups appear in the chain */\
2465 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2466 trie->wordinfo[dupe].prev = curword; \
2468 /* we haven't inserted this word yet. */ \
2469 trie->states[ state ].wordnum = curword; \
2474 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2475 ( ( base + charid >= ucharcount \
2476 && base + charid < ubound \
2477 && state == trie->trans[ base - ucharcount + charid ].check \
2478 && trie->trans[ base - ucharcount + charid ].next ) \
2479 ? trie->trans[ base - ucharcount + charid ].next \
2480 : ( state==1 ? special : 0 ) \
2483 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2485 TRIE_BITMAP_SET(trie, uvc); \
2486 /* store the folded codepoint */ \
2488 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2491 /* store first byte of utf8 representation of */ \
2492 /* variant codepoints */ \
2493 if (! UVCHR_IS_INVARIANT(uvc)) { \
2494 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2499 #define MADE_JUMP_TRIE 2
2500 #define MADE_EXACT_TRIE 4
2503 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2504 regnode *first, regnode *last, regnode *tail,
2505 U32 word_count, U32 flags, U32 depth)
2507 /* first pass, loop through and scan words */
2508 reg_trie_data *trie;
2509 HV *widecharmap = NULL;
2510 AV *revcharmap = newAV();
2516 regnode *jumper = NULL;
2517 regnode *nextbranch = NULL;
2518 regnode *convert = NULL;
2519 U32 *prev_states; /* temp array mapping each state to previous one */
2520 /* we just use folder as a flag in utf8 */
2521 const U8 * folder = NULL;
2523 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2524 * which stands for one trie structure, one hash, optionally followed
2527 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2528 AV *trie_words = NULL;
2529 /* along with revcharmap, this only used during construction but both are
2530 * useful during debugging so we store them in the struct when debugging.
2533 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2534 STRLEN trie_charcount=0;
2536 SV *re_trie_maxbuff;
2537 GET_RE_DEBUG_FLAGS_DECL;
2539 PERL_ARGS_ASSERT_MAKE_TRIE;
2541 PERL_UNUSED_ARG(depth);
2545 case EXACT: case EXACTL: break;
2549 case EXACTFLU8: folder = PL_fold_latin1; break;
2550 case EXACTF: folder = PL_fold; break;
2551 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2554 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2556 trie->startstate = 1;
2557 trie->wordcount = word_count;
2558 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2559 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2560 if (flags == EXACT || flags == EXACTL)
2561 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2562 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2563 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2566 trie_words = newAV();
2569 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2570 assert(re_trie_maxbuff);
2571 if (!SvIOK(re_trie_maxbuff)) {
2572 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2574 DEBUG_TRIE_COMPILE_r({
2575 Perl_re_indentf( aTHX_
2576 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2578 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2579 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2582 /* Find the node we are going to overwrite */
2583 if ( first == startbranch && OP( last ) != BRANCH ) {
2584 /* whole branch chain */
2587 /* branch sub-chain */
2588 convert = NEXTOPER( first );
2591 /* -- First loop and Setup --
2593 We first traverse the branches and scan each word to determine if it
2594 contains widechars, and how many unique chars there are, this is
2595 important as we have to build a table with at least as many columns as we
2598 We use an array of integers to represent the character codes 0..255
2599 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2600 the native representation of the character value as the key and IV's for
2603 *TODO* If we keep track of how many times each character is used we can
2604 remap the columns so that the table compression later on is more
2605 efficient in terms of memory by ensuring the most common value is in the
2606 middle and the least common are on the outside. IMO this would be better
2607 than a most to least common mapping as theres a decent chance the most
2608 common letter will share a node with the least common, meaning the node
2609 will not be compressible. With a middle is most common approach the worst
2610 case is when we have the least common nodes twice.
2614 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2615 regnode *noper = NEXTOPER( cur );
2619 U32 wordlen = 0; /* required init */
2620 STRLEN minchars = 0;
2621 STRLEN maxchars = 0;
2622 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2625 if (OP(noper) == NOTHING) {
2626 /* skip past a NOTHING at the start of an alternation
2627 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2629 regnode *noper_next= regnext(noper);
2630 if (noper_next < tail)
2634 if ( noper < tail &&
2636 OP(noper) == flags ||
2639 OP(noper) == EXACTFU_SS
2643 uc= (U8*)STRING(noper);
2644 e= uc + STR_LEN(noper);
2651 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2652 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2653 regardless of encoding */
2654 if (OP( noper ) == EXACTFU_SS) {
2655 /* false positives are ok, so just set this */
2656 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2660 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2662 TRIE_CHARCOUNT(trie)++;
2665 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2666 * is in effect. Under /i, this character can match itself, or
2667 * anything that folds to it. If not under /i, it can match just
2668 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2669 * all fold to k, and all are single characters. But some folds
2670 * expand to more than one character, so for example LATIN SMALL
2671 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2672 * the string beginning at 'uc' is 'ffi', it could be matched by
2673 * three characters, or just by the one ligature character. (It
2674 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2675 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2676 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2677 * match.) The trie needs to know the minimum and maximum number
2678 * of characters that could match so that it can use size alone to
2679 * quickly reject many match attempts. The max is simple: it is
2680 * the number of folded characters in this branch (since a fold is
2681 * never shorter than what folds to it. */
2685 /* And the min is equal to the max if not under /i (indicated by
2686 * 'folder' being NULL), or there are no multi-character folds. If
2687 * there is a multi-character fold, the min is incremented just
2688 * once, for the character that folds to the sequence. Each
2689 * character in the sequence needs to be added to the list below of
2690 * characters in the trie, but we count only the first towards the
2691 * min number of characters needed. This is done through the
2692 * variable 'foldlen', which is returned by the macros that look
2693 * for these sequences as the number of bytes the sequence
2694 * occupies. Each time through the loop, we decrement 'foldlen' by
2695 * how many bytes the current char occupies. Only when it reaches
2696 * 0 do we increment 'minchars' or look for another multi-character
2698 if (folder == NULL) {
2701 else if (foldlen > 0) {
2702 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2707 /* See if *uc is the beginning of a multi-character fold. If
2708 * so, we decrement the length remaining to look at, to account
2709 * for the current character this iteration. (We can use 'uc'
2710 * instead of the fold returned by TRIE_READ_CHAR because for
2711 * non-UTF, the latin1_safe macro is smart enough to account
2712 * for all the unfolded characters, and because for UTF, the
2713 * string will already have been folded earlier in the
2714 * compilation process */
2716 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2717 foldlen -= UTF8SKIP(uc);
2720 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2725 /* The current character (and any potential folds) should be added
2726 * to the possible matching characters for this position in this
2730 U8 folded= folder[ (U8) uvc ];
2731 if ( !trie->charmap[ folded ] ) {
2732 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2733 TRIE_STORE_REVCHAR( folded );
2736 if ( !trie->charmap[ uvc ] ) {
2737 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2738 TRIE_STORE_REVCHAR( uvc );
2741 /* store the codepoint in the bitmap, and its folded
2743 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2744 set_bit = 0; /* We've done our bit :-) */
2748 /* XXX We could come up with the list of code points that fold
2749 * to this using PL_utf8_foldclosures, except not for
2750 * multi-char folds, as there may be multiple combinations
2751 * there that could work, which needs to wait until runtime to
2752 * resolve (The comment about LIGATURE FFI above is such an
2757 widecharmap = newHV();
2759 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2762 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2764 if ( !SvTRUE( *svpp ) ) {
2765 sv_setiv( *svpp, ++trie->uniquecharcount );
2766 TRIE_STORE_REVCHAR(uvc);
2769 } /* end loop through characters in this branch of the trie */
2771 /* We take the min and max for this branch and combine to find the min
2772 * and max for all branches processed so far */
2773 if( cur == first ) {
2774 trie->minlen = minchars;
2775 trie->maxlen = maxchars;
2776 } else if (minchars < trie->minlen) {
2777 trie->minlen = minchars;
2778 } else if (maxchars > trie->maxlen) {
2779 trie->maxlen = maxchars;
2781 } /* end first pass */
2782 DEBUG_TRIE_COMPILE_r(
2783 Perl_re_indentf( aTHX_
2784 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2786 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2787 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2788 (int)trie->minlen, (int)trie->maxlen )
2792 We now know what we are dealing with in terms of unique chars and
2793 string sizes so we can calculate how much memory a naive
2794 representation using a flat table will take. If it's over a reasonable
2795 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2796 conservative but potentially much slower representation using an array
2799 At the end we convert both representations into the same compressed
2800 form that will be used in regexec.c for matching with. The latter
2801 is a form that cannot be used to construct with but has memory
2802 properties similar to the list form and access properties similar
2803 to the table form making it both suitable for fast searches and
2804 small enough that its feasable to store for the duration of a program.
2806 See the comment in the code where the compressed table is produced
2807 inplace from the flat tabe representation for an explanation of how
2808 the compression works.
2813 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2816 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2817 > SvIV(re_trie_maxbuff) )
2820 Second Pass -- Array Of Lists Representation
2822 Each state will be represented by a list of charid:state records
2823 (reg_trie_trans_le) the first such element holds the CUR and LEN
2824 points of the allocated array. (See defines above).
2826 We build the initial structure using the lists, and then convert
2827 it into the compressed table form which allows faster lookups
2828 (but cant be modified once converted).
2831 STRLEN transcount = 1;
2833 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2836 trie->states = (reg_trie_state *)
2837 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2838 sizeof(reg_trie_state) );
2842 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2844 regnode *noper = NEXTOPER( cur );
2845 U32 state = 1; /* required init */
2846 U16 charid = 0; /* sanity init */
2847 U32 wordlen = 0; /* required init */
2849 if (OP(noper) == NOTHING) {
2850 regnode *noper_next= regnext(noper);
2851 if (noper_next < tail)
2855 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2856 const U8 *uc= (U8*)STRING(noper);
2857 const U8 *e= uc + STR_LEN(noper);
2859 for ( ; uc < e ; uc += len ) {
2864 charid = trie->charmap[ uvc ];
2866 SV** const svpp = hv_fetch( widecharmap,
2873 charid=(U16)SvIV( *svpp );
2876 /* charid is now 0 if we dont know the char read, or
2877 * nonzero if we do */
2884 if ( !trie->states[ state ].trans.list ) {
2885 TRIE_LIST_NEW( state );
2888 check <= TRIE_LIST_USED( state );
2891 if ( TRIE_LIST_ITEM( state, check ).forid
2894 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2899 newstate = next_alloc++;
2900 prev_states[newstate] = state;
2901 TRIE_LIST_PUSH( state, charid, newstate );
2906 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
2910 TRIE_HANDLE_WORD(state);
2912 } /* end second pass */
2914 /* next alloc is the NEXT state to be allocated */
2915 trie->statecount = next_alloc;
2916 trie->states = (reg_trie_state *)
2917 PerlMemShared_realloc( trie->states,
2919 * sizeof(reg_trie_state) );
2921 /* and now dump it out before we compress it */
2922 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2923 revcharmap, next_alloc,
2927 trie->trans = (reg_trie_trans *)
2928 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2935 for( state=1 ; state < next_alloc ; state ++ ) {
2939 DEBUG_TRIE_COMPILE_MORE_r(
2940 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
2944 if (trie->states[state].trans.list) {
2945 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2949 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2950 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2951 if ( forid < minid ) {
2953 } else if ( forid > maxid ) {
2957 if ( transcount < tp + maxid - minid + 1) {
2959 trie->trans = (reg_trie_trans *)
2960 PerlMemShared_realloc( trie->trans,
2962 * sizeof(reg_trie_trans) );
2963 Zero( trie->trans + (transcount / 2),
2967 base = trie->uniquecharcount + tp - minid;
2968 if ( maxid == minid ) {
2970 for ( ; zp < tp ; zp++ ) {
2971 if ( ! trie->trans[ zp ].next ) {
2972 base = trie->uniquecharcount + zp - minid;
2973 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2975 trie->trans[ zp ].check = state;
2981 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2983 trie->trans[ tp ].check = state;
2988 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2989 const U32 tid = base
2990 - trie->uniquecharcount
2991 + TRIE_LIST_ITEM( state, idx ).forid;
2992 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2994 trie->trans[ tid ].check = state;
2996 tp += ( maxid - minid + 1 );
2998 Safefree(trie->states[ state ].trans.list);
3001 DEBUG_TRIE_COMPILE_MORE_r(
3002 Perl_re_printf( aTHX_ " base: %d\n",base);
3005 trie->states[ state ].trans.base=base;
3007 trie->lasttrans = tp + 1;
3011 Second Pass -- Flat Table Representation.
3013 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3014 each. We know that we will need Charcount+1 trans at most to store
3015 the data (one row per char at worst case) So we preallocate both
3016 structures assuming worst case.
3018 We then construct the trie using only the .next slots of the entry
3021 We use the .check field of the first entry of the node temporarily
3022 to make compression both faster and easier by keeping track of how
3023 many non zero fields are in the node.
3025 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3028 There are two terms at use here: state as a TRIE_NODEIDX() which is
3029 a number representing the first entry of the node, and state as a
3030 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3031 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3032 if there are 2 entrys per node. eg:
3040 The table is internally in the right hand, idx form. However as we
3041 also have to deal with the states array which is indexed by nodenum
3042 we have to use TRIE_NODENUM() to convert.
3045 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3048 trie->trans = (reg_trie_trans *)
3049 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3050 * trie->uniquecharcount + 1,
3051 sizeof(reg_trie_trans) );
3052 trie->states = (reg_trie_state *)
3053 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3054 sizeof(reg_trie_state) );
3055 next_alloc = trie->uniquecharcount + 1;
3058 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3060 regnode *noper = NEXTOPER( cur );
3062 U32 state = 1; /* required init */
3064 U16 charid = 0; /* sanity init */
3065 U32 accept_state = 0; /* sanity init */
3067 U32 wordlen = 0; /* required init */
3069 if (OP(noper) == NOTHING) {
3070 regnode *noper_next= regnext(noper);
3071 if (noper_next < tail)
3075 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3076 const U8 *uc= (U8*)STRING(noper);
3077 const U8 *e= uc + STR_LEN(noper);
3079 for ( ; uc < e ; uc += len ) {
3084 charid = trie->charmap[ uvc ];
3086 SV* const * const svpp = hv_fetch( widecharmap,
3090 charid = svpp ? (U16)SvIV(*svpp) : 0;
3094 if ( !trie->trans[ state + charid ].next ) {
3095 trie->trans[ state + charid ].next = next_alloc;
3096 trie->trans[ state ].check++;
3097 prev_states[TRIE_NODENUM(next_alloc)]
3098 = TRIE_NODENUM(state);
3099 next_alloc += trie->uniquecharcount;
3101 state = trie->trans[ state + charid ].next;
3103 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3105 /* charid is now 0 if we dont know the char read, or
3106 * nonzero if we do */
3109 accept_state = TRIE_NODENUM( state );
3110 TRIE_HANDLE_WORD(accept_state);
3112 } /* end second pass */
3114 /* and now dump it out before we compress it */
3115 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3117 next_alloc, depth+1));
3121 * Inplace compress the table.*
3123 For sparse data sets the table constructed by the trie algorithm will
3124 be mostly 0/FAIL transitions or to put it another way mostly empty.
3125 (Note that leaf nodes will not contain any transitions.)
3127 This algorithm compresses the tables by eliminating most such
3128 transitions, at the cost of a modest bit of extra work during lookup:
3130 - Each states[] entry contains a .base field which indicates the
3131 index in the state[] array wheres its transition data is stored.
3133 - If .base is 0 there are no valid transitions from that node.
3135 - If .base is nonzero then charid is added to it to find an entry in
3138 -If trans[states[state].base+charid].check!=state then the
3139 transition is taken to be a 0/Fail transition. Thus if there are fail
3140 transitions at the front of the node then the .base offset will point
3141 somewhere inside the previous nodes data (or maybe even into a node
3142 even earlier), but the .check field determines if the transition is
3146 The following process inplace converts the table to the compressed
3147 table: We first do not compress the root node 1,and mark all its
3148 .check pointers as 1 and set its .base pointer as 1 as well. This
3149 allows us to do a DFA construction from the compressed table later,
3150 and ensures that any .base pointers we calculate later are greater
3153 - We set 'pos' to indicate the first entry of the second node.
3155 - We then iterate over the columns of the node, finding the first and
3156 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3157 and set the .check pointers accordingly, and advance pos
3158 appropriately and repreat for the next node. Note that when we copy
3159 the next pointers we have to convert them from the original
3160 NODEIDX form to NODENUM form as the former is not valid post
3163 - If a node has no transitions used we mark its base as 0 and do not
3164 advance the pos pointer.
3166 - If a node only has one transition we use a second pointer into the
3167 structure to fill in allocated fail transitions from other states.
3168 This pointer is independent of the main pointer and scans forward
3169 looking for null transitions that are allocated to a state. When it
3170 finds one it writes the single transition into the "hole". If the
3171 pointer doesnt find one the single transition is appended as normal.
3173 - Once compressed we can Renew/realloc the structures to release the
3176 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3177 specifically Fig 3.47 and the associated pseudocode.
3181 const U32 laststate = TRIE_NODENUM( next_alloc );
3184 trie->statecount = laststate;
3186 for ( state = 1 ; state < laststate ; state++ ) {
3188 const U32 stateidx = TRIE_NODEIDX( state );
3189 const U32 o_used = trie->trans[ stateidx ].check;
3190 U32 used = trie->trans[ stateidx ].check;
3191 trie->trans[ stateidx ].check = 0;
3194 used && charid < trie->uniquecharcount;
3197 if ( flag || trie->trans[ stateidx + charid ].next ) {
3198 if ( trie->trans[ stateidx + charid ].next ) {
3200 for ( ; zp < pos ; zp++ ) {
3201 if ( ! trie->trans[ zp ].next ) {
3205 trie->states[ state ].trans.base
3207 + trie->uniquecharcount
3209 trie->trans[ zp ].next
3210 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3212 trie->trans[ zp ].check = state;
3213 if ( ++zp > pos ) pos = zp;
3220 trie->states[ state ].trans.base
3221 = pos + trie->uniquecharcount - charid ;
3223 trie->trans[ pos ].next
3224 = SAFE_TRIE_NODENUM(
3225 trie->trans[ stateidx + charid ].next );
3226 trie->trans[ pos ].check = state;
3231 trie->lasttrans = pos + 1;
3232 trie->states = (reg_trie_state *)
3233 PerlMemShared_realloc( trie->states, laststate
3234 * sizeof(reg_trie_state) );
3235 DEBUG_TRIE_COMPILE_MORE_r(
3236 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3238 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3242 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3245 } /* end table compress */
3247 DEBUG_TRIE_COMPILE_MORE_r(
3248 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3250 (UV)trie->statecount,
3251 (UV)trie->lasttrans)
3253 /* resize the trans array to remove unused space */
3254 trie->trans = (reg_trie_trans *)
3255 PerlMemShared_realloc( trie->trans, trie->lasttrans
3256 * sizeof(reg_trie_trans) );
3258 { /* Modify the program and insert the new TRIE node */
3259 U8 nodetype =(U8)(flags & 0xFF);
3263 regnode *optimize = NULL;
3264 #ifdef RE_TRACK_PATTERN_OFFSETS
3267 U32 mjd_nodelen = 0;
3268 #endif /* RE_TRACK_PATTERN_OFFSETS */
3269 #endif /* DEBUGGING */
3271 This means we convert either the first branch or the first Exact,
3272 depending on whether the thing following (in 'last') is a branch
3273 or not and whther first is the startbranch (ie is it a sub part of
3274 the alternation or is it the whole thing.)
3275 Assuming its a sub part we convert the EXACT otherwise we convert
3276 the whole branch sequence, including the first.
3278 /* Find the node we are going to overwrite */
3279 if ( first != startbranch || OP( last ) == BRANCH ) {
3280 /* branch sub-chain */
3281 NEXT_OFF( first ) = (U16)(last - first);
3282 #ifdef RE_TRACK_PATTERN_OFFSETS
3284 mjd_offset= Node_Offset((convert));
3285 mjd_nodelen= Node_Length((convert));
3288 /* whole branch chain */
3290 #ifdef RE_TRACK_PATTERN_OFFSETS
3293 const regnode *nop = NEXTOPER( convert );
3294 mjd_offset= Node_Offset((nop));
3295 mjd_nodelen= Node_Length((nop));
3299 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3301 (UV)mjd_offset, (UV)mjd_nodelen)
3304 /* But first we check to see if there is a common prefix we can
3305 split out as an EXACT and put in front of the TRIE node. */
3306 trie->startstate= 1;
3307 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3308 /* we want to find the first state that has more than
3309 * one transition, if that state is not the first state
3310 * then we have a common prefix which we can remove.
3313 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3315 I32 first_ofs = -1; /* keeps track of the ofs of the first
3316 transition, -1 means none */
3318 const U32 base = trie->states[ state ].trans.base;
3320 /* does this state terminate an alternation? */
3321 if ( trie->states[state].wordnum )
3324 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3325 if ( ( base + ofs >= trie->uniquecharcount ) &&
3326 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3327 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3329 if ( ++count > 1 ) {
3330 /* we have more than one transition */
3333 /* if this is the first state there is no common prefix
3334 * to extract, so we can exit */
3335 if ( state == 1 ) break;
3336 tmp = av_fetch( revcharmap, ofs, 0);
3337 ch = (U8*)SvPV_nolen_const( *tmp );
3339 /* if we are on count 2 then we need to initialize the
3340 * bitmap, and store the previous char if there was one
3343 /* clear the bitmap */
3344 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3346 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3349 if (first_ofs >= 0) {
3350 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3351 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3353 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3355 Perl_re_printf( aTHX_ "%s", (char*)ch)
3359 /* store the current firstchar in the bitmap */
3360 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3361 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3367 /* This state has only one transition, its transition is part
3368 * of a common prefix - we need to concatenate the char it
3369 * represents to what we have so far. */
3370 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3372 char *ch = SvPV( *tmp, len );
3374 SV *sv=sv_newmortal();
3375 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3377 (UV)state, (UV)first_ofs,
3378 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3379 PL_colors[0], PL_colors[1],
3380 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3381 PERL_PV_ESCAPE_FIRSTCHAR
3386 OP( convert ) = nodetype;
3387 str=STRING(convert);
3390 STR_LEN(convert) += len;
3396 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3401 trie->prefixlen = (state-1);
3403 regnode *n = convert+NODE_SZ_STR(convert);
3404 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3405 trie->startstate = state;
3406 trie->minlen -= (state - 1);
3407 trie->maxlen -= (state - 1);
3409 /* At least the UNICOS C compiler choked on this
3410 * being argument to DEBUG_r(), so let's just have
3413 #ifdef PERL_EXT_RE_BUILD
3419 regnode *fix = convert;
3420 U32 word = trie->wordcount;
3422 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3423 while( ++fix < n ) {
3424 Set_Node_Offset_Length(fix, 0, 0);
3427 SV ** const tmp = av_fetch( trie_words, word, 0 );
3429 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3430 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3432 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3440 NEXT_OFF(convert) = (U16)(tail - convert);
3441 DEBUG_r(optimize= n);
3447 if ( trie->maxlen ) {
3448 NEXT_OFF( convert ) = (U16)(tail - convert);
3449 ARG_SET( convert, data_slot );
3450 /* Store the offset to the first unabsorbed branch in
3451 jump[0], which is otherwise unused by the jump logic.
3452 We use this when dumping a trie and during optimisation. */
3454 trie->jump[0] = (U16)(nextbranch - convert);
3456 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3457 * and there is a bitmap
3458 * and the first "jump target" node we found leaves enough room
3459 * then convert the TRIE node into a TRIEC node, with the bitmap
3460 * embedded inline in the opcode - this is hypothetically faster.
3462 if ( !trie->states[trie->startstate].wordnum
3464 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3466 OP( convert ) = TRIEC;
3467 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3468 PerlMemShared_free(trie->bitmap);
3471 OP( convert ) = TRIE;
3473 /* store the type in the flags */
3474 convert->flags = nodetype;
3478 + regarglen[ OP( convert ) ];
3480 /* XXX We really should free up the resource in trie now,
3481 as we won't use them - (which resources?) dmq */
3483 /* needed for dumping*/
3484 DEBUG_r(if (optimize) {
3485 regnode *opt = convert;
3487 while ( ++opt < optimize) {
3488 Set_Node_Offset_Length(opt,0,0);
3491 Try to clean up some of the debris left after the
3494 while( optimize < jumper ) {
3495 mjd_nodelen += Node_Length((optimize));
3496 OP( optimize ) = OPTIMIZED;
3497 Set_Node_Offset_Length(optimize,0,0);
3500 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3502 } /* end node insert */
3504 /* Finish populating the prev field of the wordinfo array. Walk back
3505 * from each accept state until we find another accept state, and if
3506 * so, point the first word's .prev field at the second word. If the
3507 * second already has a .prev field set, stop now. This will be the
3508 * case either if we've already processed that word's accept state,
3509 * or that state had multiple words, and the overspill words were
3510 * already linked up earlier.
3517 for (word=1; word <= trie->wordcount; word++) {
3519 if (trie->wordinfo[word].prev)
3521 state = trie->wordinfo[word].accept;
3523 state = prev_states[state];
3526 prev = trie->states[state].wordnum;
3530 trie->wordinfo[word].prev = prev;
3532 Safefree(prev_states);
3536 /* and now dump out the compressed format */
3537 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3539 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3541 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3542 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3544 SvREFCNT_dec_NN(revcharmap);
3548 : trie->startstate>1
3554 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3556 /* The Trie is constructed and compressed now so we can build a fail array if
3559 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3561 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3565 We find the fail state for each state in the trie, this state is the longest
3566 proper suffix of the current state's 'word' that is also a proper prefix of
3567 another word in our trie. State 1 represents the word '' and is thus the
3568 default fail state. This allows the DFA not to have to restart after its
3569 tried and failed a word at a given point, it simply continues as though it
3570 had been matching the other word in the first place.
3572 'abcdgu'=~/abcdefg|cdgu/
3573 When we get to 'd' we are still matching the first word, we would encounter
3574 'g' which would fail, which would bring us to the state representing 'd' in
3575 the second word where we would try 'g' and succeed, proceeding to match
3578 /* add a fail transition */
3579 const U32 trie_offset = ARG(source);
3580 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3582 const U32 ucharcount = trie->uniquecharcount;
3583 const U32 numstates = trie->statecount;
3584 const U32 ubound = trie->lasttrans + ucharcount;
3588 U32 base = trie->states[ 1 ].trans.base;
3591 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3593 GET_RE_DEBUG_FLAGS_DECL;
3595 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3596 PERL_UNUSED_CONTEXT;
3598 PERL_UNUSED_ARG(depth);
3601 if ( OP(source) == TRIE ) {
3602 struct regnode_1 *op = (struct regnode_1 *)
3603 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3604 StructCopy(source,op,struct regnode_1);
3605 stclass = (regnode *)op;
3607 struct regnode_charclass *op = (struct regnode_charclass *)
3608 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3609 StructCopy(source,op,struct regnode_charclass);
3610 stclass = (regnode *)op;
3612 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3614 ARG_SET( stclass, data_slot );
3615 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3616 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3617 aho->trie=trie_offset;
3618 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3619 Copy( trie->states, aho->states, numstates, reg_trie_state );
3620 Newx( q, numstates, U32);
3621 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3624 /* initialize fail[0..1] to be 1 so that we always have
3625 a valid final fail state */
3626 fail[ 0 ] = fail[ 1 ] = 1;
3628 for ( charid = 0; charid < ucharcount ; charid++ ) {
3629 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3631 q[ q_write ] = newstate;
3632 /* set to point at the root */
3633 fail[ q[ q_write++ ] ]=1;
3636 while ( q_read < q_write) {
3637 const U32 cur = q[ q_read++ % numstates ];
3638 base = trie->states[ cur ].trans.base;
3640 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3641 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3643 U32 fail_state = cur;
3646 fail_state = fail[ fail_state ];
3647 fail_base = aho->states[ fail_state ].trans.base;
3648 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3650 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3651 fail[ ch_state ] = fail_state;
3652 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3654 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3656 q[ q_write++ % numstates] = ch_state;
3660 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3661 when we fail in state 1, this allows us to use the
3662 charclass scan to find a valid start char. This is based on the principle
3663 that theres a good chance the string being searched contains lots of stuff
3664 that cant be a start char.
3666 fail[ 0 ] = fail[ 1 ] = 0;
3667 DEBUG_TRIE_COMPILE_r({
3668 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3669 depth, (UV)numstates
3671 for( q_read=1; q_read<numstates; q_read++ ) {
3672 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3674 Perl_re_printf( aTHX_ "\n");
3677 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3682 /* The below joins as many adjacent EXACTish nodes as possible into a single
3683 * one. The regop may be changed if the node(s) contain certain sequences that
3684 * require special handling. The joining is only done if:
3685 * 1) there is room in the current conglomerated node to entirely contain the
3687 * 2) they are the exact same node type
3689 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3690 * these get optimized out
3692 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3693 * as possible, even if that means splitting an existing node so that its first
3694 * part is moved to the preceeding node. This would maximise the efficiency of
3695 * memEQ during matching. Elsewhere in this file, khw proposes splitting
3696 * EXACTFish nodes into portions that don't change under folding vs those that
3697 * do. Those portions that don't change may be the only things in the pattern that
3698 * could be used to find fixed and floating strings.
3700 * If a node is to match under /i (folded), the number of characters it matches
3701 * can be different than its character length if it contains a multi-character
3702 * fold. *min_subtract is set to the total delta number of characters of the
3705 * And *unfolded_multi_char is set to indicate whether or not the node contains
3706 * an unfolded multi-char fold. This happens when whether the fold is valid or
3707 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3708 * SMALL LETTER SHARP S, as only if the target string being matched against
3709 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3710 * folding rules depend on the locale in force at runtime. (Multi-char folds
3711 * whose components are all above the Latin1 range are not run-time locale
3712 * dependent, and have already been folded by the time this function is
3715 * This is as good a place as any to discuss the design of handling these
3716 * multi-character fold sequences. It's been wrong in Perl for a very long
3717 * time. There are three code points in Unicode whose multi-character folds
3718 * were long ago discovered to mess things up. The previous designs for
3719 * dealing with these involved assigning a special node for them. This
3720 * approach doesn't always work, as evidenced by this example:
3721 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3722 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3723 * would match just the \xDF, it won't be able to handle the case where a
3724 * successful match would have to cross the node's boundary. The new approach
3725 * that hopefully generally solves the problem generates an EXACTFU_SS node
3726 * that is "sss" in this case.
3728 * It turns out that there are problems with all multi-character folds, and not
3729 * just these three. Now the code is general, for all such cases. The
3730 * approach taken is:
3731 * 1) This routine examines each EXACTFish node that could contain multi-
3732 * character folded sequences. Since a single character can fold into
3733 * such a sequence, the minimum match length for this node is less than
3734 * the number of characters in the node. This routine returns in
3735 * *min_subtract how many characters to subtract from the the actual
3736 * length of the string to get a real minimum match length; it is 0 if
3737 * there are no multi-char foldeds. This delta is used by the caller to
3738 * adjust the min length of the match, and the delta between min and max,
3739 * so that the optimizer doesn't reject these possibilities based on size
3741 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3742 * is used for an EXACTFU node that contains at least one "ss" sequence in
3743 * it. For non-UTF-8 patterns and strings, this is the only case where
3744 * there is a possible fold length change. That means that a regular
3745 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3746 * with length changes, and so can be processed faster. regexec.c takes
3747 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3748 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3749 * known until runtime). This saves effort in regex matching. However,
3750 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3751 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3752 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3753 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3754 * possibilities for the non-UTF8 patterns are quite simple, except for
3755 * the sharp s. All the ones that don't involve a UTF-8 target string are
3756 * members of a fold-pair, and arrays are set up for all of them so that
3757 * the other member of the pair can be found quickly. Code elsewhere in
3758 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3759 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3760 * described in the next item.
3761 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3762 * validity of the fold won't be known until runtime, and so must remain
3763 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3764 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3765 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3766 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3767 * The reason this is a problem is that the optimizer part of regexec.c
3768 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3769 * that a character in the pattern corresponds to at most a single
3770 * character in the target string. (And I do mean character, and not byte
3771 * here, unlike other parts of the documentation that have never been
3772 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3773 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3774 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3775 * nodes, violate the assumption, and they are the only instances where it
3776 * is violated. I'm reluctant to try to change the assumption, as the
3777 * code involved is impenetrable to me (khw), so instead the code here
3778 * punts. This routine examines EXACTFL nodes, and (when the pattern
3779 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3780 * boolean indicating whether or not the node contains such a fold. When
3781 * it is true, the caller sets a flag that later causes the optimizer in
3782 * this file to not set values for the floating and fixed string lengths,
3783 * and thus avoids the optimizer code in regexec.c that makes the invalid
3784 * assumption. Thus, there is no optimization based on string lengths for
3785 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3786 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3787 * assumption is wrong only in these cases is that all other non-UTF-8
3788 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3789 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3790 * EXACTF nodes because we don't know at compile time if it actually
3791 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3792 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3793 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3794 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3795 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3796 * string would require the pattern to be forced into UTF-8, the overhead
3797 * of which we want to avoid. Similarly the unfolded multi-char folds in
3798 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3801 * Similarly, the code that generates tries doesn't currently handle
3802 * not-already-folded multi-char folds, and it looks like a pain to change
3803 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3804 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3805 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3806 * using /iaa matching will be doing so almost entirely with ASCII
3807 * strings, so this should rarely be encountered in practice */
3809 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3810 if (PL_regkind[OP(scan)] == EXACT) \
3811 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3814 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3815 UV *min_subtract, bool *unfolded_multi_char,
3816 U32 flags,regnode *val, U32 depth)
3818 /* Merge several consecutive EXACTish nodes into one. */
3819 regnode *n = regnext(scan);
3821 regnode *next = scan + NODE_SZ_STR(scan);
3825 regnode *stop = scan;
3826 GET_RE_DEBUG_FLAGS_DECL;
3828 PERL_UNUSED_ARG(depth);
3831 PERL_ARGS_ASSERT_JOIN_EXACT;
3832 #ifndef EXPERIMENTAL_INPLACESCAN
3833 PERL_UNUSED_ARG(flags);
3834 PERL_UNUSED_ARG(val);
3836 DEBUG_PEEP("join", scan, depth, 0);
3838 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3839 * EXACT ones that are mergeable to the current one. */
3841 && (PL_regkind[OP(n)] == NOTHING
3842 || (stringok && OP(n) == OP(scan)))
3844 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3847 if (OP(n) == TAIL || n > next)
3849 if (PL_regkind[OP(n)] == NOTHING) {
3850 DEBUG_PEEP("skip:", n, depth, 0);
3851 NEXT_OFF(scan) += NEXT_OFF(n);
3852 next = n + NODE_STEP_REGNODE;
3859 else if (stringok) {
3860 const unsigned int oldl = STR_LEN(scan);
3861 regnode * const nnext = regnext(n);
3863 /* XXX I (khw) kind of doubt that this works on platforms (should
3864 * Perl ever run on one) where U8_MAX is above 255 because of lots
3865 * of other assumptions */
3866 /* Don't join if the sum can't fit into a single node */
3867 if (oldl + STR_LEN(n) > U8_MAX)
3870 DEBUG_PEEP("merg", n, depth, 0);
3873 NEXT_OFF(scan) += NEXT_OFF(n);
3874 STR_LEN(scan) += STR_LEN(n);
3875 next = n + NODE_SZ_STR(n);
3876 /* Now we can overwrite *n : */
3877 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3885 #ifdef EXPERIMENTAL_INPLACESCAN
3886 if (flags && !NEXT_OFF(n)) {
3887 DEBUG_PEEP("atch", val, depth, 0);
3888 if (reg_off_by_arg[OP(n)]) {
3889 ARG_SET(n, val - n);
3892 NEXT_OFF(n) = val - n;
3900 *unfolded_multi_char = FALSE;
3902 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3903 * can now analyze for sequences of problematic code points. (Prior to
3904 * this final joining, sequences could have been split over boundaries, and
3905 * hence missed). The sequences only happen in folding, hence for any
3906 * non-EXACT EXACTish node */
3907 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3908 U8* s0 = (U8*) STRING(scan);
3910 U8* s_end = s0 + STR_LEN(scan);
3912 int total_count_delta = 0; /* Total delta number of characters that
3913 multi-char folds expand to */
3915 /* One pass is made over the node's string looking for all the
3916 * possibilities. To avoid some tests in the loop, there are two main
3917 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3922 if (OP(scan) == EXACTFL) {
3925 /* An EXACTFL node would already have been changed to another
3926 * node type unless there is at least one character in it that
3927 * is problematic; likely a character whose fold definition
3928 * won't be known until runtime, and so has yet to be folded.
3929 * For all but the UTF-8 locale, folds are 1-1 in length, but
3930 * to handle the UTF-8 case, we need to create a temporary
3931 * folded copy using UTF-8 locale rules in order to analyze it.
3932 * This is because our macros that look to see if a sequence is
3933 * a multi-char fold assume everything is folded (otherwise the
3934 * tests in those macros would be too complicated and slow).
3935 * Note that here, the non-problematic folds will have already
3936 * been done, so we can just copy such characters. We actually
3937 * don't completely fold the EXACTFL string. We skip the
3938 * unfolded multi-char folds, as that would just create work
3939 * below to figure out the size they already are */
3941 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3944 STRLEN s_len = UTF8SKIP(s);
3945 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3946 Copy(s, d, s_len, U8);
3949 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3950 *unfolded_multi_char = TRUE;
3951 Copy(s, d, s_len, U8);
3954 else if (isASCII(*s)) {
3955 *(d++) = toFOLD(*s);
3959 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
3965 /* Point the remainder of the routine to look at our temporary
3969 } /* End of creating folded copy of EXACTFL string */
3971 /* Examine the string for a multi-character fold sequence. UTF-8
3972 * patterns have all characters pre-folded by the time this code is
3974 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3975 length sequence we are looking for is 2 */
3977 int count = 0; /* How many characters in a multi-char fold */
3978 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3979 if (! len) { /* Not a multi-char fold: get next char */
3984 /* Nodes with 'ss' require special handling, except for
3985 * EXACTFA-ish for which there is no multi-char fold to this */
3986 if (len == 2 && *s == 's' && *(s+1) == 's'
3987 && OP(scan) != EXACTFA
3988 && OP(scan) != EXACTFA_NO_TRIE)
3991 if (OP(scan) != EXACTFL) {
3992 OP(scan) = EXACTFU_SS;
3996 else { /* Here is a generic multi-char fold. */
3997 U8* multi_end = s + len;
3999 /* Count how many characters are in it. In the case of
4000 * /aa, no folds which contain ASCII code points are
4001 * allowed, so check for those, and skip if found. */
4002 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
4003 count = utf8_length(s, multi_end);
4007 while (s < multi_end) {
4010 goto next_iteration;
4020 /* The delta is how long the sequence is minus 1 (1 is how long
4021 * the character that folds to the sequence is) */
4022 total_count_delta += count - 1;
4026 /* We created a temporary folded copy of the string in EXACTFL
4027 * nodes. Therefore we need to be sure it doesn't go below zero,
4028 * as the real string could be shorter */
4029 if (OP(scan) == EXACTFL) {
4030 int total_chars = utf8_length((U8*) STRING(scan),
4031 (U8*) STRING(scan) + STR_LEN(scan));
4032 if (total_count_delta > total_chars) {
4033 total_count_delta = total_chars;
4037 *min_subtract += total_count_delta;
4040 else if (OP(scan) == EXACTFA) {
4042 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
4043 * fold to the ASCII range (and there are no existing ones in the
4044 * upper latin1 range). But, as outlined in the comments preceding
4045 * this function, we need to flag any occurrences of the sharp s.
4046 * This character forbids trie formation (because of added
4048 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4049 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4050 || UNICODE_DOT_DOT_VERSION > 0)
4052 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4053 OP(scan) = EXACTFA_NO_TRIE;
4054 *unfolded_multi_char = TRUE;
4062 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
4063 * folds that are all Latin1. As explained in the comments
4064 * preceding this function, we look also for the sharp s in EXACTF
4065 * and EXACTFL nodes; it can be in the final position. Otherwise
4066 * we can stop looking 1 byte earlier because have to find at least
4067 * two characters for a multi-fold */
4068 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4073 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4074 if (! len) { /* Not a multi-char fold. */
4075 if (*s == LATIN_SMALL_LETTER_SHARP_S
4076 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4078 *unfolded_multi_char = TRUE;
4085 && isALPHA_FOLD_EQ(*s, 's')
4086 && isALPHA_FOLD_EQ(*(s+1), 's'))
4089 /* EXACTF nodes need to know that the minimum length
4090 * changed so that a sharp s in the string can match this
4091 * ss in the pattern, but they remain EXACTF nodes, as they
4092 * won't match this unless the target string is is UTF-8,
4093 * which we don't know until runtime. EXACTFL nodes can't
4094 * transform into EXACTFU nodes */
4095 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4096 OP(scan) = EXACTFU_SS;
4100 *min_subtract += len - 1;
4108 /* Allow dumping but overwriting the collection of skipped
4109 * ops and/or strings with fake optimized ops */
4110 n = scan + NODE_SZ_STR(scan);
4118 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl", scan, depth, 0);});
4122 /* REx optimizer. Converts nodes into quicker variants "in place".
4123 Finds fixed substrings. */
4125 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4126 to the position after last scanned or to NULL. */
4128 #define INIT_AND_WITHP \
4129 assert(!and_withp); \
4130 Newx(and_withp,1, regnode_ssc); \
4131 SAVEFREEPV(and_withp)
4135 S_unwind_scan_frames(pTHX_ const void *p)
4137 scan_frame *f= (scan_frame *)p;
4139 scan_frame *n= f->next_frame;
4147 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4148 SSize_t *minlenp, SSize_t *deltap,
4153 regnode_ssc *and_withp,
4154 U32 flags, U32 depth)
4155 /* scanp: Start here (read-write). */
4156 /* deltap: Write maxlen-minlen here. */
4157 /* last: Stop before this one. */
4158 /* data: string data about the pattern */
4159 /* stopparen: treat close N as END */
4160 /* recursed: which subroutines have we recursed into */
4161 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4163 /* There must be at least this number of characters to match */
4166 regnode *scan = *scanp, *next;
4168 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4169 int is_inf_internal = 0; /* The studied chunk is infinite */
4170 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4171 scan_data_t data_fake;
4172 SV *re_trie_maxbuff = NULL;
4173 regnode *first_non_open = scan;
4174 SSize_t stopmin = SSize_t_MAX;
4175 scan_frame *frame = NULL;
4176 GET_RE_DEBUG_FLAGS_DECL;
4178 PERL_ARGS_ASSERT_STUDY_CHUNK;
4179 RExC_study_started= 1;
4183 while (first_non_open && OP(first_non_open) == OPEN)
4184 first_non_open=regnext(first_non_open);
4190 RExC_study_chunk_recursed_count++;
4192 DEBUG_OPTIMISE_MORE_r(
4194 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4195 depth, (long)stopparen,
4196 (unsigned long)RExC_study_chunk_recursed_count,
4197 (unsigned long)depth, (unsigned long)recursed_depth,
4200 if (recursed_depth) {
4203 for ( j = 0 ; j < recursed_depth ; j++ ) {
4204 for ( i = 0 ; i < (U32)RExC_npar ; i++ ) {
4206 PAREN_TEST(RExC_study_chunk_recursed +
4207 ( j * RExC_study_chunk_recursed_bytes), i )
4210 !PAREN_TEST(RExC_study_chunk_recursed +
4211 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
4214 Perl_re_printf( aTHX_ " %d",(int)i);
4218 if ( j + 1 < recursed_depth ) {
4219 Perl_re_printf( aTHX_ ",");
4223 Perl_re_printf( aTHX_ "\n");
4226 while ( scan && OP(scan) != END && scan < last ){
4227 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4228 node length to get a real minimum (because
4229 the folded version may be shorter) */
4230 bool unfolded_multi_char = FALSE;
4231 /* Peephole optimizer: */
4232 DEBUG_STUDYDATA("Peep", data, depth, is_inf);
4233 DEBUG_PEEP("Peep", scan, depth, flags);
4236 /* The reason we do this here is that we need to deal with things like
4237 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4238 * parsing code, as each (?:..) is handled by a different invocation of
4241 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
4243 /* Follow the next-chain of the current node and optimize
4244 away all the NOTHINGs from it. */
4245 if (OP(scan) != CURLYX) {
4246 const int max = (reg_off_by_arg[OP(scan)]
4248 /* I32 may be smaller than U16 on CRAYs! */
4249 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4250 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
4254 /* Skip NOTHING and LONGJMP. */
4255 while ((n = regnext(n))
4256 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4257 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
4258 && off + noff < max)
4260 if (reg_off_by_arg[OP(scan)])
4263 NEXT_OFF(scan) = off;
4266 /* The principal pseudo-switch. Cannot be a switch, since we
4267 look into several different things. */
4268 if ( OP(scan) == DEFINEP ) {
4270 SSize_t deltanext = 0;
4271 SSize_t fake_last_close = 0;
4272 I32 f = SCF_IN_DEFINE;
4274 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4275 scan = regnext(scan);
4276 assert( OP(scan) == IFTHEN );
4277 DEBUG_PEEP("expect IFTHEN", scan, depth, flags);
4279 data_fake.last_closep= &fake_last_close;
4281 next = regnext(scan);
4282 scan = NEXTOPER(NEXTOPER(scan));
4283 DEBUG_PEEP("scan", scan, depth, flags);
4284 DEBUG_PEEP("next", next, depth, flags);
4286 /* we suppose the run is continuous, last=next...
4287 * NOTE we dont use the return here! */
4288 (void)study_chunk(pRExC_state, &scan, &minlen,
4289 &deltanext, next, &data_fake, stopparen,
4290 recursed_depth, NULL, f, depth+1);
4295 OP(scan) == BRANCH ||
4296 OP(scan) == BRANCHJ ||
4299 next = regnext(scan);
4302 /* The op(next)==code check below is to see if we
4303 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4304 * IFTHEN is special as it might not appear in pairs.
4305 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4306 * we dont handle it cleanly. */
4307 if (OP(next) == code || code == IFTHEN) {
4308 /* NOTE - There is similar code to this block below for
4309 * handling TRIE nodes on a re-study. If you change stuff here
4310 * check there too. */
4311 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
4313 regnode * const startbranch=scan;
4315 if (flags & SCF_DO_SUBSTR) {
4316 /* Cannot merge strings after this. */
4317 scan_commit(pRExC_state, data, minlenp, is_inf);
4320 if (flags & SCF_DO_STCLASS)
4321 ssc_init_zero(pRExC_state, &accum);
4323 while (OP(scan) == code) {
4324 SSize_t deltanext, minnext, fake;
4326 regnode_ssc this_class;
4328 DEBUG_PEEP("Branch", scan, depth, flags);
4331 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4333 data_fake.whilem_c = data->whilem_c;
4334 data_fake.last_closep = data->last_closep;
4337 data_fake.last_closep = &fake;
4339 data_fake.pos_delta = delta;
4340 next = regnext(scan);
4342 scan = NEXTOPER(scan); /* everything */
4343 if (code != BRANCH) /* everything but BRANCH */
4344 scan = NEXTOPER(scan);
4346 if (flags & SCF_DO_STCLASS) {
4347 ssc_init(pRExC_state, &this_class);
4348 data_fake.start_class = &this_class;
4349 f = SCF_DO_STCLASS_AND;
4351 if (flags & SCF_WHILEM_VISITED_POS)
4352 f |= SCF_WHILEM_VISITED_POS;
4354 /* we suppose the run is continuous, last=next...*/
4355 minnext = study_chunk(pRExC_state, &scan, minlenp,
4356 &deltanext, next, &data_fake, stopparen,
4357 recursed_depth, NULL, f,depth+1);
4361 if (deltanext == SSize_t_MAX) {
4362 is_inf = is_inf_internal = 1;
4364 } else if (max1 < minnext + deltanext)
4365 max1 = minnext + deltanext;
4367 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4369 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4370 if ( stopmin > minnext)
4371 stopmin = min + min1;
4372 flags &= ~SCF_DO_SUBSTR;
4374 data->flags |= SCF_SEEN_ACCEPT;
4377 if (data_fake.flags & SF_HAS_EVAL)
4378 data->flags |= SF_HAS_EVAL;
4379 data->whilem_c = data_fake.whilem_c;
4381 if (flags & SCF_DO_STCLASS)
4382 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4384 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4386 if (flags & SCF_DO_SUBSTR) {
4387 data->pos_min += min1;
4388 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
4389 data->pos_delta = SSize_t_MAX;
4391 data->pos_delta += max1 - min1;
4392 if (max1 != min1 || is_inf)
4393 data->cur_is_floating = 1;
4396 if (delta == SSize_t_MAX
4397 || SSize_t_MAX - delta - (max1 - min1) < 0)
4398 delta = SSize_t_MAX;
4400 delta += max1 - min1;
4401 if (flags & SCF_DO_STCLASS_OR) {
4402 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4404 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4405 flags &= ~SCF_DO_STCLASS;
4408 else if (flags & SCF_DO_STCLASS_AND) {
4410 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4411 flags &= ~SCF_DO_STCLASS;
4414 /* Switch to OR mode: cache the old value of
4415 * data->start_class */
4417 StructCopy(data->start_class, and_withp, regnode_ssc);
4418 flags &= ~SCF_DO_STCLASS_AND;
4419 StructCopy(&accum, data->start_class, regnode_ssc);
4420 flags |= SCF_DO_STCLASS_OR;
4424 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4425 OP( startbranch ) == BRANCH )
4429 Assuming this was/is a branch we are dealing with: 'scan'
4430 now points at the item that follows the branch sequence,
4431 whatever it is. We now start at the beginning of the
4432 sequence and look for subsequences of
4438 which would be constructed from a pattern like
4441 If we can find such a subsequence we need to turn the first
4442 element into a trie and then add the subsequent branch exact
4443 strings to the trie.
4447 1. patterns where the whole set of branches can be
4450 2. patterns where only a subset can be converted.
4452 In case 1 we can replace the whole set with a single regop
4453 for the trie. In case 2 we need to keep the start and end
4456 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4457 becomes BRANCH TRIE; BRANCH X;
4459 There is an additional case, that being where there is a
4460 common prefix, which gets split out into an EXACT like node
4461 preceding the TRIE node.
4463 If x(1..n)==tail then we can do a simple trie, if not we make
4464 a "jump" trie, such that when we match the appropriate word
4465 we "jump" to the appropriate tail node. Essentially we turn
4466 a nested if into a case structure of sorts.
4471 if (!re_trie_maxbuff) {
4472 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4473 if (!SvIOK(re_trie_maxbuff))
4474 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4476 if ( SvIV(re_trie_maxbuff)>=0 ) {
4478 regnode *first = (regnode *)NULL;
4479 regnode *last = (regnode *)NULL;
4480 regnode *tail = scan;
4484 /* var tail is used because there may be a TAIL
4485 regop in the way. Ie, the exacts will point to the
4486 thing following the TAIL, but the last branch will
4487 point at the TAIL. So we advance tail. If we
4488 have nested (?:) we may have to move through several
4492 while ( OP( tail ) == TAIL ) {
4493 /* this is the TAIL generated by (?:) */
4494 tail = regnext( tail );
4498 DEBUG_TRIE_COMPILE_r({
4499 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4500 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4502 "Looking for TRIE'able sequences. Tail node is ",
4503 (UV)(tail - RExC_emit_start),
4504 SvPV_nolen_const( RExC_mysv )
4510 Step through the branches
4511 cur represents each branch,
4512 noper is the first thing to be matched as part
4514 noper_next is the regnext() of that node.
4516 We normally handle a case like this
4517 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4518 support building with NOJUMPTRIE, which restricts
4519 the trie logic to structures like /FOO|BAR/.
4521 If noper is a trieable nodetype then the branch is
4522 a possible optimization target. If we are building
4523 under NOJUMPTRIE then we require that noper_next is
4524 the same as scan (our current position in the regex
4527 Once we have two or more consecutive such branches
4528 we can create a trie of the EXACT's contents and
4529 stitch it in place into the program.
4531 If the sequence represents all of the branches in
4532 the alternation we replace the entire thing with a
4535 Otherwise when it is a subsequence we need to
4536 stitch it in place and replace only the relevant
4537 branches. This means the first branch has to remain
4538 as it is used by the alternation logic, and its
4539 next pointer, and needs to be repointed at the item
4540 on the branch chain following the last branch we
4541 have optimized away.
4543 This could be either a BRANCH, in which case the
4544 subsequence is internal, or it could be the item
4545 following the branch sequence in which case the
4546 subsequence is at the end (which does not
4547 necessarily mean the first node is the start of the
4550 TRIE_TYPE(X) is a define which maps the optype to a
4554 ----------------+-----------
4558 EXACTFU_SS | EXACTFU
4561 EXACTFLU8 | EXACTFLU8
4565 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4567 : ( EXACT == (X) ) \
4569 : ( EXACTFU == (X) || EXACTFU_SS == (X) ) \
4571 : ( EXACTFA == (X) ) \
4573 : ( EXACTL == (X) ) \
4575 : ( EXACTFLU8 == (X) ) \
4579 /* dont use tail as the end marker for this traverse */
4580 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4581 regnode * const noper = NEXTOPER( cur );
4582 U8 noper_type = OP( noper );
4583 U8 noper_trietype = TRIE_TYPE( noper_type );
4584 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4585 regnode * const noper_next = regnext( noper );
4586 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4587 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4590 DEBUG_TRIE_COMPILE_r({
4591 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4592 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4594 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4596 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4597 Perl_re_printf( aTHX_ " -> %d:%s",
4598 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
4601 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4602 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
4603 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
4605 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
4606 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4607 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4611 /* Is noper a trieable nodetype that can be merged
4612 * with the current trie (if there is one)? */
4616 ( noper_trietype == NOTHING )
4617 || ( trietype == NOTHING )
4618 || ( trietype == noper_trietype )
4621 && noper_next >= tail
4625 /* Handle mergable triable node Either we are
4626 * the first node in a new trieable sequence,
4627 * in which case we do some bookkeeping,
4628 * otherwise we update the end pointer. */
4631 if ( noper_trietype == NOTHING ) {
4632 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4633 regnode * const noper_next = regnext( noper );
4634 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4635 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4638 if ( noper_next_trietype ) {
4639 trietype = noper_next_trietype;
4640 } else if (noper_next_type) {
4641 /* a NOTHING regop is 1 regop wide.
4642 * We need at least two for a trie
4643 * so we can't merge this in */
4647 trietype = noper_trietype;
4650 if ( trietype == NOTHING )
4651 trietype = noper_trietype;
4656 } /* end handle mergable triable node */
4658 /* handle unmergable node -
4659 * noper may either be a triable node which can
4660 * not be tried together with the current trie,
4661 * or a non triable node */
4663 /* If last is set and trietype is not
4664 * NOTHING then we have found at least two
4665 * triable branch sequences in a row of a
4666 * similar trietype so we can turn them
4667 * into a trie. If/when we allow NOTHING to
4668 * start a trie sequence this condition
4669 * will be required, and it isn't expensive
4670 * so we leave it in for now. */
4671 if ( trietype && trietype != NOTHING )
4672 make_trie( pRExC_state,
4673 startbranch, first, cur, tail,
4674 count, trietype, depth+1 );
4675 last = NULL; /* note: we clear/update
4676 first, trietype etc below,
4677 so we dont do it here */
4681 && noper_next >= tail
4684 /* noper is triable, so we can start a new
4688 trietype = noper_trietype;
4690 /* if we already saw a first but the
4691 * current node is not triable then we have
4692 * to reset the first information. */
4697 } /* end handle unmergable node */
4698 } /* loop over branches */
4699 DEBUG_TRIE_COMPILE_r({
4700 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4701 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
4702 depth+1, SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4703 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
4704 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4705 PL_reg_name[trietype]
4709 if ( last && trietype ) {
4710 if ( trietype != NOTHING ) {
4711 /* the last branch of the sequence was part of
4712 * a trie, so we have to construct it here
4713 * outside of the loop */
4714 made= make_trie( pRExC_state, startbranch,
4715 first, scan, tail, count,
4716 trietype, depth+1 );
4717 #ifdef TRIE_STUDY_OPT
4718 if ( ((made == MADE_EXACT_TRIE &&
4719 startbranch == first)
4720 || ( first_non_open == first )) &&
4722 flags |= SCF_TRIE_RESTUDY;
4723 if ( startbranch == first
4726 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4731 /* at this point we know whatever we have is a
4732 * NOTHING sequence/branch AND if 'startbranch'
4733 * is 'first' then we can turn the whole thing
4736 if ( startbranch == first ) {
4738 /* the entire thing is a NOTHING sequence,
4739 * something like this: (?:|) So we can
4740 * turn it into a plain NOTHING op. */
4741 DEBUG_TRIE_COMPILE_r({
4742 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4743 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
4745 SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4748 OP(startbranch)= NOTHING;
4749 NEXT_OFF(startbranch)= tail - startbranch;
4750 for ( opt= startbranch + 1; opt < tail ; opt++ )
4754 } /* end if ( last) */
4755 } /* TRIE_MAXBUF is non zero */
4760 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4761 scan = NEXTOPER(NEXTOPER(scan));
4762 } else /* single branch is optimized. */
4763 scan = NEXTOPER(scan);
4765 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
4767 regnode *start = NULL;
4768 regnode *end = NULL;
4769 U32 my_recursed_depth= recursed_depth;
4771 if (OP(scan) != SUSPEND) { /* GOSUB */
4772 /* Do setup, note this code has side effects beyond
4773 * the rest of this block. Specifically setting
4774 * RExC_recurse[] must happen at least once during
4777 RExC_recurse[ARG2L(scan)] = scan;
4778 start = RExC_open_parens[paren];
4779 end = RExC_close_parens[paren];
4781 /* NOTE we MUST always execute the above code, even
4782 * if we do nothing with a GOSUB */
4784 ( flags & SCF_IN_DEFINE )
4787 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
4789 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
4792 /* no need to do anything here if we are in a define. */
4793 /* or we are after some kind of infinite construct
4794 * so we can skip recursing into this item.
4795 * Since it is infinite we will not change the maxlen
4796 * or delta, and if we miss something that might raise
4797 * the minlen it will merely pessimise a little.
4799 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
4800 * might result in a minlen of 1 and not of 4,
4801 * but this doesn't make us mismatch, just try a bit
4802 * harder than we should.
4804 scan= regnext(scan);
4811 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4813 /* it is quite possible that there are more efficient ways
4814 * to do this. We maintain a bitmap per level of recursion
4815 * of which patterns we have entered so we can detect if a
4816 * pattern creates a possible infinite loop. When we
4817 * recurse down a level we copy the previous levels bitmap
4818 * down. When we are at recursion level 0 we zero the top
4819 * level bitmap. It would be nice to implement a different
4820 * more efficient way of doing this. In particular the top
4821 * level bitmap may be unnecessary.
4823 if (!recursed_depth) {
4824 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4826 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4827 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4828 RExC_study_chunk_recursed_bytes, U8);
4830 /* we havent recursed into this paren yet, so recurse into it */
4831 DEBUG_STUDYDATA("gosub-set", data, depth, is_inf);
4832 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4833 my_recursed_depth= recursed_depth + 1;
4835 DEBUG_STUDYDATA("gosub-inf", data, depth, is_inf);
4836 /* some form of infinite recursion, assume infinite length
4838 if (flags & SCF_DO_SUBSTR) {
4839 scan_commit(pRExC_state, data, minlenp, is_inf);
4840 data->cur_is_floating = 1;
4842 is_inf = is_inf_internal = 1;
4843 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4844 ssc_anything(data->start_class);
4845 flags &= ~SCF_DO_STCLASS;
4847 start= NULL; /* reset start so we dont recurse later on. */
4852 end = regnext(scan);
4855 scan_frame *newframe;
4857 if (!RExC_frame_last) {
4858 Newxz(newframe, 1, scan_frame);
4859 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
4860 RExC_frame_head= newframe;
4862 } else if (!RExC_frame_last->next_frame) {
4863 Newxz(newframe,1,scan_frame);
4864 RExC_frame_last->next_frame= newframe;
4865 newframe->prev_frame= RExC_frame_last;
4868 newframe= RExC_frame_last->next_frame;
4870 RExC_frame_last= newframe;
4872 newframe->next_regnode = regnext(scan);
4873 newframe->last_regnode = last;
4874 newframe->stopparen = stopparen;
4875 newframe->prev_recursed_depth = recursed_depth;
4876 newframe->this_prev_frame= frame;
4878 DEBUG_STUDYDATA("frame-new", data, depth, is_inf);
4879 DEBUG_PEEP("fnew", scan, depth, flags);
4886 recursed_depth= my_recursed_depth;
4891 else if (OP(scan) == EXACT || OP(scan) == EXACTL) {
4892 SSize_t l = STR_LEN(scan);
4896 const U8 * const s = (U8*)STRING(scan);
4897 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4898 l = utf8_length(s, s + l);
4900 uc = *((U8*)STRING(scan));
4903 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4904 /* The code below prefers earlier match for fixed
4905 offset, later match for variable offset. */
4906 if (data->last_end == -1) { /* Update the start info. */
4907 data->last_start_min = data->pos_min;
4908 data->last_start_max = is_inf
4909 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4911 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4913 SvUTF8_on(data->last_found);
4915 SV * const sv = data->last_found;
4916 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4917 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4918 if (mg && mg->mg_len >= 0)
4919 mg->mg_len += utf8_length((U8*)STRING(scan),
4920 (U8*)STRING(scan)+STR_LEN(scan));
4922 data->last_end = data->pos_min + l;
4923 data->pos_min += l; /* As in the first entry. */
4924 data->flags &= ~SF_BEFORE_EOL;
4927 /* ANDing the code point leaves at most it, and not in locale, and
4928 * can't match null string */
4929 if (flags & SCF_DO_STCLASS_AND) {
4930 ssc_cp_and(data->start_class, uc);
4931 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4932 ssc_clear_locale(data->start_class);
4934 else if (flags & SCF_DO_STCLASS_OR) {
4935 ssc_add_cp(data->start_class, uc);
4936 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4938 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4939 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4941 flags &= ~SCF_DO_STCLASS;
4943 else if (PL_regkind[OP(scan)] == EXACT) {
4944 /* But OP != EXACT!, so is EXACTFish */
4945 SSize_t l = STR_LEN(scan);
4946 const U8 * s = (U8*)STRING(scan);
4948 /* Search for fixed substrings supports EXACT only. */
4949 if (flags & SCF_DO_SUBSTR) {
4951 scan_commit(pRExC_state, data, minlenp, is_inf);
4954 l = utf8_length(s, s + l);
4956 if (unfolded_multi_char) {
4957 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4959 min += l - min_subtract;
4961 delta += min_subtract;
4962 if (flags & SCF_DO_SUBSTR) {
4963 data->pos_min += l - min_subtract;
4964 if (data->pos_min < 0) {
4967 data->pos_delta += min_subtract;
4969 data->cur_is_floating = 1; /* float */
4973 if (flags & SCF_DO_STCLASS) {
4974 SV* EXACTF_invlist = _make_exactf_invlist(pRExC_state, scan);
4976 assert(EXACTF_invlist);
4977 if (flags & SCF_DO_STCLASS_AND) {
4978 if (OP(scan) != EXACTFL)
4979 ssc_clear_locale(data->start_class);
4980 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4981 ANYOF_POSIXL_ZERO(data->start_class);
4982 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4984 else { /* SCF_DO_STCLASS_OR */
4985 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4986 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4988 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4989 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4991 flags &= ~SCF_DO_STCLASS;
4992 SvREFCNT_dec(EXACTF_invlist);
4995 else if (REGNODE_VARIES(OP(scan))) {
4996 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4997 I32 fl = 0, f = flags;
4998 regnode * const oscan = scan;
4999 regnode_ssc this_class;
5000 regnode_ssc *oclass = NULL;
5001 I32 next_is_eval = 0;
5003 switch (PL_regkind[OP(scan)]) {
5004 case WHILEM: /* End of (?:...)* . */
5005 scan = NEXTOPER(scan);
5008 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
5009 next = NEXTOPER(scan);
5010 if (OP(next) == EXACT
5011 || OP(next) == EXACTL
5012 || (flags & SCF_DO_STCLASS))
5015 maxcount = REG_INFTY;
5016 next = regnext(scan);
5017 scan = NEXTOPER(scan);
5021 if (flags & SCF_DO_SUBSTR)
5026 if (flags & SCF_DO_STCLASS) {
5028 maxcount = REG_INFTY;
5029 next = regnext(scan);
5030 scan = NEXTOPER(scan);
5033 if (flags & SCF_DO_SUBSTR) {
5034 scan_commit(pRExC_state, data, minlenp, is_inf);
5035 /* Cannot extend fixed substrings */
5036 data->cur_is_floating = 1; /* float */
5038 is_inf = is_inf_internal = 1;
5039 scan = regnext(scan);
5040 goto optimize_curly_tail;
5042 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
5043 && (scan->flags == stopparen))
5048 mincount = ARG1(scan);
5049 maxcount = ARG2(scan);
5051 next = regnext(scan);
5052 if (OP(scan) == CURLYX) {
5053 I32 lp = (data ? *(data->last_closep) : 0);
5054 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
5056 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
5057 next_is_eval = (OP(scan) == EVAL);
5059 if (flags & SCF_DO_SUBSTR) {
5061 scan_commit(pRExC_state, data, minlenp, is_inf);
5062 /* Cannot extend fixed substrings */
5063 pos_before = data->pos_min;
5067 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
5069 data->flags |= SF_IS_INF;
5071 if (flags & SCF_DO_STCLASS) {
5072 ssc_init(pRExC_state, &this_class);
5073 oclass = data->start_class;
5074 data->start_class = &this_class;
5075 f |= SCF_DO_STCLASS_AND;
5076 f &= ~SCF_DO_STCLASS_OR;
5078 /* Exclude from super-linear cache processing any {n,m}
5079 regops for which the combination of input pos and regex
5080 pos is not enough information to determine if a match
5083 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5084 regex pos at the \s*, the prospects for a match depend not
5085 only on the input position but also on how many (bar\s*)
5086 repeats into the {4,8} we are. */
5087 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5088 f &= ~SCF_WHILEM_VISITED_POS;
5090 /* This will finish on WHILEM, setting scan, or on NULL: */
5091 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5092 last, data, stopparen, recursed_depth, NULL,
5094 ? (f & ~SCF_DO_SUBSTR)
5098 if (flags & SCF_DO_STCLASS)
5099 data->start_class = oclass;
5100 if (mincount == 0 || minnext == 0) {
5101 if (flags & SCF_DO_STCLASS_OR) {
5102 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5104 else if (flags & SCF_DO_STCLASS_AND) {
5105 /* Switch to OR mode: cache the old value of
5106 * data->start_class */
5108 StructCopy(data->start_class, and_withp, regnode_ssc);
5109 flags &= ~SCF_DO_STCLASS_AND;
5110 StructCopy(&this_class, data->start_class, regnode_ssc);
5111 flags |= SCF_DO_STCLASS_OR;
5112 ANYOF_FLAGS(data->start_class)
5113 |= SSC_MATCHES_EMPTY_STRING;
5115 } else { /* Non-zero len */
5116 if (flags & SCF_DO_STCLASS_OR) {
5117 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5118 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5120 else if (flags & SCF_DO_STCLASS_AND)
5121 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5122 flags &= ~SCF_DO_STCLASS;
5124 if (!scan) /* It was not CURLYX, but CURLY. */
5126 if (((flags & (SCF_TRIE_DOING_RESTUDY|SCF_DO_SUBSTR))==SCF_DO_SUBSTR)
5127 /* ? quantifier ok, except for (?{ ... }) */
5128 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5129 && (minnext == 0) && (deltanext == 0)
5130 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5131 && maxcount <= REG_INFTY/3) /* Complement check for big
5134 /* Fatal warnings may leak the regexp without this: */
5135 SAVEFREESV(RExC_rx_sv);
5136 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5137 "Quantifier unexpected on zero-length expression "
5138 "in regex m/%" UTF8f "/",
5139 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5141 (void)ReREFCNT_inc(RExC_rx_sv);
5144 min += minnext * mincount;
5145 is_inf_internal |= deltanext == SSize_t_MAX
5146 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5147 is_inf |= is_inf_internal;
5149 delta = SSize_t_MAX;
5151 delta += (minnext + deltanext) * maxcount
5152 - minnext * mincount;
5154 /* Try powerful optimization CURLYX => CURLYN. */
5155 if ( OP(oscan) == CURLYX && data
5156 && data->flags & SF_IN_PAR
5157 && !(data->flags & SF_HAS_EVAL)
5158 && !deltanext && minnext == 1 ) {
5159 /* Try to optimize to CURLYN. */
5160 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5161 regnode * const nxt1 = nxt;
5168 if (!REGNODE_SIMPLE(OP(nxt))
5169 && !(PL_regkind[OP(nxt)] == EXACT
5170 && STR_LEN(nxt) == 1))
5176 if (OP(nxt) != CLOSE)
5178 if (RExC_open_parens) {
5179 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5180 RExC_close_parens[ARG(nxt1)]=nxt+2; /*close->while*/
5182 /* Now we know that nxt2 is the only contents: */
5183 oscan->flags = (U8)ARG(nxt);
5185 OP(nxt1) = NOTHING; /* was OPEN. */
5188 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5189 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5190 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5191 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5192 OP(nxt + 1) = OPTIMIZED; /* was count. */
5193 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5198 /* Try optimization CURLYX => CURLYM. */
5199 if ( OP(oscan) == CURLYX && data
5200 && !(data->flags & SF_HAS_PAR)
5201 && !(data->flags & SF_HAS_EVAL)
5202 && !deltanext /* atom is fixed width */
5203 && minnext != 0 /* CURLYM can't handle zero width */
5205 /* Nor characters whose fold at run-time may be
5206 * multi-character */
5207 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5209 /* XXXX How to optimize if data == 0? */
5210 /* Optimize to a simpler form. */
5211 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5215 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5216 && (OP(nxt2) != WHILEM))
5218 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5219 /* Need to optimize away parenths. */
5220 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5221 /* Set the parenth number. */
5222 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5224 oscan->flags = (U8)ARG(nxt);
5225 if (RExC_open_parens) {
5226 RExC_open_parens[ARG(nxt1)]=oscan; /*open->CURLYM*/
5227 RExC_close_parens[ARG(nxt1)]=nxt2+1; /*close->NOTHING*/
5229 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5230 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5233 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5234 OP(nxt + 1) = OPTIMIZED; /* was count. */
5235 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5236 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5239 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5240 regnode *nnxt = regnext(nxt1);
5242 if (reg_off_by_arg[OP(nxt1)])
5243 ARG_SET(nxt1, nxt2 - nxt1);
5244 else if (nxt2 - nxt1 < U16_MAX)
5245 NEXT_OFF(nxt1) = nxt2 - nxt1;
5247 OP(nxt) = NOTHING; /* Cannot beautify */
5252 /* Optimize again: */
5253 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5254 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
5259 else if ((OP(oscan) == CURLYX)
5260 && (flags & SCF_WHILEM_VISITED_POS)
5261 /* See the comment on a similar expression above.
5262 However, this time it's not a subexpression
5263 we care about, but the expression itself. */
5264 && (maxcount == REG_INFTY)
5266 /* This stays as CURLYX, we can put the count/of pair. */
5267 /* Find WHILEM (as in regexec.c) */
5268 regnode *nxt = oscan + NEXT_OFF(oscan);
5270 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5272 nxt = PREVOPER(nxt);
5273 if (nxt->flags & 0xf) {
5274 /* we've already set whilem count on this node */
5275 } else if (++data->whilem_c < 16) {
5276 assert(data->whilem_c <= RExC_whilem_seen);
5277 nxt->flags = (U8)(data->whilem_c
5278 | (RExC_whilem_seen << 4)); /* On WHILEM */
5281 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5283 if (flags & SCF_DO_SUBSTR) {
5284 SV *last_str = NULL;
5285 STRLEN last_chrs = 0;
5286 int counted = mincount != 0;
5288 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5290 SSize_t b = pos_before >= data->last_start_min
5291 ? pos_before : data->last_start_min;
5293 const char * const s = SvPV_const(data->last_found, l);
5294 SSize_t old = b - data->last_start_min;
5297 old = utf8_hop((U8*)s, old) - (U8*)s;
5299 /* Get the added string: */
5300 last_str = newSVpvn_utf8(s + old, l, UTF);
5301 last_chrs = UTF ? utf8_length((U8*)(s + old),
5302 (U8*)(s + old + l)) : l;
5303 if (deltanext == 0 && pos_before == b) {
5304 /* What was added is a constant string */
5307 SvGROW(last_str, (mincount * l) + 1);
5308 repeatcpy(SvPVX(last_str) + l,
5309 SvPVX_const(last_str), l,
5311 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5312 /* Add additional parts. */
5313 SvCUR_set(data->last_found,
5314 SvCUR(data->last_found) - l);
5315 sv_catsv(data->last_found, last_str);
5317 SV * sv = data->last_found;
5319 SvUTF8(sv) && SvMAGICAL(sv) ?
5320 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5321 if (mg && mg->mg_len >= 0)
5322 mg->mg_len += last_chrs * (mincount-1);
5324 last_chrs *= mincount;
5325 data->last_end += l * (mincount - 1);
5328 /* start offset must point into the last copy */
5329 data->last_start_min += minnext * (mincount - 1);
5330 data->last_start_max =
5333 : data->last_start_max +
5334 (maxcount - 1) * (minnext + data->pos_delta);
5337 /* It is counted once already... */
5338 data->pos_min += minnext * (mincount - counted);
5340 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5341 " SSize_t_MAX=%" UVuf " minnext=%" UVuf
5342 " maxcount=%" UVuf " mincount=%" UVuf "\n",
5343 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
5345 if (deltanext != SSize_t_MAX)
5346 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5347 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5348 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
5350 if (deltanext == SSize_t_MAX
5351 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
5352 data->pos_delta = SSize_t_MAX;
5354 data->pos_delta += - counted * deltanext +
5355 (minnext + deltanext) * maxcount - minnext * mincount;
5356 if (mincount != maxcount) {
5357 /* Cannot extend fixed substrings found inside
5359 scan_commit(pRExC_state, data, minlenp, is_inf);
5360 if (mincount && last_str) {
5361 SV * const sv = data->last_found;
5362 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5363 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5367 sv_setsv(sv, last_str);
5368 data->last_end = data->pos_min;
5369 data->last_start_min = data->pos_min - last_chrs;
5370 data->last_start_max = is_inf
5372 : data->pos_min + data->pos_delta - last_chrs;
5374 data->cur_is_floating = 1; /* float */
5376 SvREFCNT_dec(last_str);
5378 if (data && (fl & SF_HAS_EVAL))
5379 data->flags |= SF_HAS_EVAL;
5380 optimize_curly_tail:
5381 if (OP(oscan) != CURLYX) {
5382 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
5384 NEXT_OFF(oscan) += NEXT_OFF(next);
5390 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5395 if (flags & SCF_DO_SUBSTR) {
5396 /* Cannot expect anything... */
5397 scan_commit(pRExC_state, data, minlenp, is_inf);
5398 data->cur_is_floating = 1; /* float */
5400 is_inf = is_inf_internal = 1;
5401 if (flags & SCF_DO_STCLASS_OR) {
5402 if (OP(scan) == CLUMP) {
5403 /* Actually is any start char, but very few code points
5404 * aren't start characters */
5405 ssc_match_all_cp(data->start_class);
5408 ssc_anything(data->start_class);
5411 flags &= ~SCF_DO_STCLASS;
5415 else if (OP(scan) == LNBREAK) {
5416 if (flags & SCF_DO_STCLASS) {
5417 if (flags & SCF_DO_STCLASS_AND) {
5418 ssc_intersection(data->start_class,
5419 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5420 ssc_clear_locale(data->start_class);
5421 ANYOF_FLAGS(data->start_class)
5422 &= ~SSC_MATCHES_EMPTY_STRING;
5424 else if (flags & SCF_DO_STCLASS_OR) {
5425 ssc_union(data->start_class,
5426 PL_XPosix_ptrs[_CC_VERTSPACE],
5428 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5430 /* See commit msg for
5431 * 749e076fceedeb708a624933726e7989f2302f6a */
5432 ANYOF_FLAGS(data->start_class)
5433 &= ~SSC_MATCHES_EMPTY_STRING;
5435 flags &= ~SCF_DO_STCLASS;
5438 if (delta != SSize_t_MAX)
5439 delta++; /* Because of the 2 char string cr-lf */
5440 if (flags & SCF_DO_SUBSTR) {
5441 /* Cannot expect anything... */
5442 scan_commit(pRExC_state, data, minlenp, is_inf);
5444 data->pos_delta += 1;
5445 data->cur_is_floating = 1; /* float */
5448 else if (REGNODE_SIMPLE(OP(scan))) {
5450 if (flags & SCF_DO_SUBSTR) {
5451 scan_commit(pRExC_state, data, minlenp, is_inf);
5455 if (flags & SCF_DO_STCLASS) {
5457 SV* my_invlist = NULL;
5460 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5461 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5463 /* Some of the logic below assumes that switching
5464 locale on will only add false positives. */
5469 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5473 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5474 ssc_match_all_cp(data->start_class);
5479 SV* REG_ANY_invlist = _new_invlist(2);
5480 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5482 if (flags & SCF_DO_STCLASS_OR) {
5483 ssc_union(data->start_class,
5485 TRUE /* TRUE => invert, hence all but \n
5489 else if (flags & SCF_DO_STCLASS_AND) {
5490 ssc_intersection(data->start_class,
5492 TRUE /* TRUE => invert */
5494 ssc_clear_locale(data->start_class);
5496 SvREFCNT_dec_NN(REG_ANY_invlist);
5503 if (flags & SCF_DO_STCLASS_AND)
5504 ssc_and(pRExC_state, data->start_class,
5505 (regnode_charclass *) scan);
5507 ssc_or(pRExC_state, data->start_class,
5508 (regnode_charclass *) scan);
5516 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5517 if (flags & SCF_DO_STCLASS_AND) {
5518 bool was_there = cBOOL(
5519 ANYOF_POSIXL_TEST(data->start_class,
5521 ANYOF_POSIXL_ZERO(data->start_class);
5522 if (was_there) { /* Do an AND */
5523 ANYOF_POSIXL_SET(data->start_class, namedclass);
5525 /* No individual code points can now match */
5526 data->start_class->invlist
5527 = sv_2mortal(_new_invlist(0));
5530 int complement = namedclass + ((invert) ? -1 : 1);
5532 assert(flags & SCF_DO_STCLASS_OR);
5534 /* If the complement of this class was already there,
5535 * the result is that they match all code points,
5536 * (\d + \D == everything). Remove the classes from
5537 * future consideration. Locale is not relevant in
5539 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5540 ssc_match_all_cp(data->start_class);
5541 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5542 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5544 else { /* The usual case; just add this class to the
5546 ANYOF_POSIXL_SET(data->start_class, namedclass);
5551 case NPOSIXA: /* For these, we always know the exact set of
5556 if (FLAGS(scan) == _CC_ASCII) {
5557 my_invlist = invlist_clone(PL_XPosix_ptrs[_CC_ASCII]);
5560 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
5561 PL_XPosix_ptrs[_CC_ASCII],
5572 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
5574 /* NPOSIXD matches all upper Latin1 code points unless the
5575 * target string being matched is UTF-8, which is
5576 * unknowable until match time. Since we are going to
5577 * invert, we want to get rid of all of them so that the
5578 * inversion will match all */
5579 if (OP(scan) == NPOSIXD) {
5580 _invlist_subtract(my_invlist, PL_UpperLatin1,
5586 if (flags & SCF_DO_STCLASS_AND) {
5587 ssc_intersection(data->start_class, my_invlist, invert);
5588 ssc_clear_locale(data->start_class);
5591 assert(flags & SCF_DO_STCLASS_OR);
5592 ssc_union(data->start_class, my_invlist, invert);
5594 SvREFCNT_dec(my_invlist);
5596 if (flags & SCF_DO_STCLASS_OR)
5597 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5598 flags &= ~SCF_DO_STCLASS;
5601 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5602 data->flags |= (OP(scan) == MEOL
5605 scan_commit(pRExC_state, data, minlenp, is_inf);
5608 else if ( PL_regkind[OP(scan)] == BRANCHJ
5609 /* Lookbehind, or need to calculate parens/evals/stclass: */
5610 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5611 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5613 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5614 || OP(scan) == UNLESSM )
5616 /* Negative Lookahead/lookbehind
5617 In this case we can't do fixed string optimisation.
5620 SSize_t deltanext, minnext, fake = 0;
5625 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5627 data_fake.whilem_c = data->whilem_c;
5628 data_fake.last_closep = data->last_closep;
5631 data_fake.last_closep = &fake;
5632 data_fake.pos_delta = delta;
5633 if ( flags & SCF_DO_STCLASS && !scan->flags
5634 && OP(scan) == IFMATCH ) { /* Lookahead */
5635 ssc_init(pRExC_state, &intrnl);
5636 data_fake.start_class = &intrnl;
5637 f |= SCF_DO_STCLASS_AND;
5639 if (flags & SCF_WHILEM_VISITED_POS)
5640 f |= SCF_WHILEM_VISITED_POS;
5641 next = regnext(scan);
5642 nscan = NEXTOPER(NEXTOPER(scan));
5643 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5644 last, &data_fake, stopparen,
5645 recursed_depth, NULL, f, depth+1);
5648 FAIL("Variable length lookbehind not implemented");
5650 else if (minnext > (I32)U8_MAX) {
5651 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5654 scan->flags = (U8)minnext;
5657 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5659 if (data_fake.flags & SF_HAS_EVAL)
5660 data->flags |= SF_HAS_EVAL;
5661 data->whilem_c = data_fake.whilem_c;
5663 if (f & SCF_DO_STCLASS_AND) {
5664 if (flags & SCF_DO_STCLASS_OR) {
5665 /* OR before, AND after: ideally we would recurse with
5666 * data_fake to get the AND applied by study of the
5667 * remainder of the pattern, and then derecurse;
5668 * *** HACK *** for now just treat as "no information".
5669 * See [perl #56690].
5671 ssc_init(pRExC_state, data->start_class);
5673 /* AND before and after: combine and continue. These
5674 * assertions are zero-length, so can match an EMPTY
5676 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5677 ANYOF_FLAGS(data->start_class)
5678 |= SSC_MATCHES_EMPTY_STRING;
5682 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5684 /* Positive Lookahead/lookbehind
5685 In this case we can do fixed string optimisation,
5686 but we must be careful about it. Note in the case of
5687 lookbehind the positions will be offset by the minimum
5688 length of the pattern, something we won't know about
5689 until after the recurse.
5691 SSize_t deltanext, fake = 0;
5695 /* We use SAVEFREEPV so that when the full compile
5696 is finished perl will clean up the allocated
5697 minlens when it's all done. This way we don't
5698 have to worry about freeing them when we know
5699 they wont be used, which would be a pain.
5702 Newx( minnextp, 1, SSize_t );
5703 SAVEFREEPV(minnextp);
5706 StructCopy(data, &data_fake, scan_data_t);
5707 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5710 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5711 data_fake.last_found=newSVsv(data->last_found);
5715 data_fake.last_closep = &fake;
5716 data_fake.flags = 0;
5717 data_fake.substrs[0].flags = 0;
5718 data_fake.substrs[1].flags = 0;
5719 data_fake.pos_delta = delta;
5721 data_fake.flags |= SF_IS_INF;
5722 if ( flags & SCF_DO_STCLASS && !scan->flags
5723 && OP(scan) == IFMATCH ) { /* Lookahead */
5724 ssc_init(pRExC_state, &intrnl);
5725 data_fake.start_class = &intrnl;
5726 f |= SCF_DO_STCLASS_AND;
5728 if (flags & SCF_WHILEM_VISITED_POS)
5729 f |= SCF_WHILEM_VISITED_POS;
5730 next = regnext(scan);
5731 nscan = NEXTOPER(NEXTOPER(scan));
5733 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5734 &deltanext, last, &data_fake,
5735 stopparen, recursed_depth, NULL,
5739 FAIL("Variable length lookbehind not implemented");
5741 else if (*minnextp > (I32)U8_MAX) {
5742 FAIL2("Lookbehind longer than %" UVuf " not implemented",
5745 scan->flags = (U8)*minnextp;
5750 if (f & SCF_DO_STCLASS_AND) {
5751 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5752 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
5755 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5757 if (data_fake.flags & SF_HAS_EVAL)
5758 data->flags |= SF_HAS_EVAL;
5759 data->whilem_c = data_fake.whilem_c;
5760 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5762 if (RExC_rx->minlen<*minnextp)
5763 RExC_rx->minlen=*minnextp;
5764 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5765 SvREFCNT_dec_NN(data_fake.last_found);
5767 for (i = 0; i < 2; i++) {
5768 if (data_fake.substrs[i].minlenp != minlenp) {
5769 data->substrs[i].min_offset =
5770 data_fake.substrs[i].min_offset;
5771 data->substrs[i].max_offset =
5772 data_fake.substrs[i].max_offset;
5773 data->substrs[i].minlenp =
5774 data_fake.substrs[i].minlenp;
5775 data->substrs[i].lookbehind += scan->flags;
5784 else if (OP(scan) == OPEN) {
5785 if (stopparen != (I32)ARG(scan))
5788 else if (OP(scan) == CLOSE) {
5789 if (stopparen == (I32)ARG(scan)) {
5792 if ((I32)ARG(scan) == is_par) {
5793 next = regnext(scan);
5795 if ( next && (OP(next) != WHILEM) && next < last)
5796 is_par = 0; /* Disable optimization */
5799 *(data->last_closep) = ARG(scan);
5801 else if (OP(scan) == EVAL) {
5803 data->flags |= SF_HAS_EVAL;
5805 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5806 if (flags & SCF_DO_SUBSTR) {
5807 scan_commit(pRExC_state, data, minlenp, is_inf);
5808 flags &= ~SCF_DO_SUBSTR;
5810 if (data && OP(scan)==ACCEPT) {
5811 data->flags |= SCF_SEEN_ACCEPT;
5816 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5818 if (flags & SCF_DO_SUBSTR) {
5819 scan_commit(pRExC_state, data, minlenp, is_inf);
5820 data->cur_is_floating = 1; /* float */
5822 is_inf = is_inf_internal = 1;
5823 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5824 ssc_anything(data->start_class);
5825 flags &= ~SCF_DO_STCLASS;
5827 else if (OP(scan) == GPOS) {
5828 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5829 !(delta || is_inf || (data && data->pos_delta)))
5831 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5832 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5833 if (RExC_rx->gofs < (STRLEN)min)
5834 RExC_rx->gofs = min;
5836 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5840 #ifdef TRIE_STUDY_OPT
5841 #ifdef FULL_TRIE_STUDY
5842 else if (PL_regkind[OP(scan)] == TRIE) {
5843 /* NOTE - There is similar code to this block above for handling
5844 BRANCH nodes on the initial study. If you change stuff here
5846 regnode *trie_node= scan;
5847 regnode *tail= regnext(scan);
5848 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5849 SSize_t max1 = 0, min1 = SSize_t_MAX;
5852 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5853 /* Cannot merge strings after this. */
5854 scan_commit(pRExC_state, data, minlenp, is_inf);
5856 if (flags & SCF_DO_STCLASS)
5857 ssc_init_zero(pRExC_state, &accum);
5863 const regnode *nextbranch= NULL;
5866 for ( word=1 ; word <= trie->wordcount ; word++)
5868 SSize_t deltanext=0, minnext=0, f = 0, fake;
5869 regnode_ssc this_class;
5871 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5873 data_fake.whilem_c = data->whilem_c;
5874 data_fake.last_closep = data->last_closep;
5877 data_fake.last_closep = &fake;
5878 data_fake.pos_delta = delta;
5879 if (flags & SCF_DO_STCLASS) {
5880 ssc_init(pRExC_state, &this_class);
5881 data_fake.start_class = &this_class;
5882 f = SCF_DO_STCLASS_AND;
5884 if (flags & SCF_WHILEM_VISITED_POS)
5885 f |= SCF_WHILEM_VISITED_POS;
5887 if (trie->jump[word]) {
5889 nextbranch = trie_node + trie->jump[0];
5890 scan= trie_node + trie->jump[word];
5891 /* We go from the jump point to the branch that follows
5892 it. Note this means we need the vestigal unused
5893 branches even though they arent otherwise used. */
5894 minnext = study_chunk(pRExC_state, &scan, minlenp,
5895 &deltanext, (regnode *)nextbranch, &data_fake,
5896 stopparen, recursed_depth, NULL, f,depth+1);
5898 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5899 nextbranch= regnext((regnode*)nextbranch);
5901 if (min1 > (SSize_t)(minnext + trie->minlen))
5902 min1 = minnext + trie->minlen;
5903 if (deltanext == SSize_t_MAX) {
5904 is_inf = is_inf_internal = 1;
5906 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5907 max1 = minnext + deltanext + trie->maxlen;
5909 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5911 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5912 if ( stopmin > min + min1)
5913 stopmin = min + min1;
5914 flags &= ~SCF_DO_SUBSTR;
5916 data->flags |= SCF_SEEN_ACCEPT;
5919 if (data_fake.flags & SF_HAS_EVAL)
5920 data->flags |= SF_HAS_EVAL;
5921 data->whilem_c = data_fake.whilem_c;
5923 if (flags & SCF_DO_STCLASS)
5924 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5927 if (flags & SCF_DO_SUBSTR) {
5928 data->pos_min += min1;
5929 data->pos_delta += max1 - min1;
5930 if (max1 != min1 || is_inf)
5931 data->cur_is_floating = 1; /* float */
5934 if (delta != SSize_t_MAX)
5935 delta += max1 - min1;
5936 if (flags & SCF_DO_STCLASS_OR) {
5937 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5939 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5940 flags &= ~SCF_DO_STCLASS;
5943 else if (flags & SCF_DO_STCLASS_AND) {
5945 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5946 flags &= ~SCF_DO_STCLASS;
5949 /* Switch to OR mode: cache the old value of
5950 * data->start_class */
5952 StructCopy(data->start_class, and_withp, regnode_ssc);
5953 flags &= ~SCF_DO_STCLASS_AND;
5954 StructCopy(&accum, data->start_class, regnode_ssc);
5955 flags |= SCF_DO_STCLASS_OR;
5962 else if (PL_regkind[OP(scan)] == TRIE) {
5963 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5966 min += trie->minlen;
5967 delta += (trie->maxlen - trie->minlen);
5968 flags &= ~SCF_DO_STCLASS; /* xxx */
5969 if (flags & SCF_DO_SUBSTR) {
5970 /* Cannot expect anything... */
5971 scan_commit(pRExC_state, data, minlenp, is_inf);
5972 data->pos_min += trie->minlen;
5973 data->pos_delta += (trie->maxlen - trie->minlen);
5974 if (trie->maxlen != trie->minlen)
5975 data->cur_is_floating = 1; /* float */
5977 if (trie->jump) /* no more substrings -- for now /grr*/
5978 flags &= ~SCF_DO_SUBSTR;
5980 #endif /* old or new */
5981 #endif /* TRIE_STUDY_OPT */
5983 /* Else: zero-length, ignore. */
5984 scan = regnext(scan);
5989 /* we need to unwind recursion. */
5992 DEBUG_STUDYDATA("frame-end", data, depth, is_inf);
5993 DEBUG_PEEP("fend", scan, depth, flags);
5995 /* restore previous context */
5996 last = frame->last_regnode;
5997 scan = frame->next_regnode;
5998 stopparen = frame->stopparen;
5999 recursed_depth = frame->prev_recursed_depth;
6001 RExC_frame_last = frame->prev_frame;
6002 frame = frame->this_prev_frame;
6003 goto fake_study_recurse;
6007 DEBUG_STUDYDATA("pre-fin", data, depth, is_inf);
6010 *deltap = is_inf_internal ? SSize_t_MAX : delta;
6012 if (flags & SCF_DO_SUBSTR && is_inf)
6013 data->pos_delta = SSize_t_MAX - data->pos_min;
6014 if (is_par > (I32)U8_MAX)
6016 if (is_par && pars==1 && data) {
6017 data->flags |= SF_IN_PAR;
6018 data->flags &= ~SF_HAS_PAR;
6020 else if (pars && data) {
6021 data->flags |= SF_HAS_PAR;
6022 data->flags &= ~SF_IN_PAR;
6024 if (flags & SCF_DO_STCLASS_OR)
6025 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6026 if (flags & SCF_TRIE_RESTUDY)
6027 data->flags |= SCF_TRIE_RESTUDY;
6029 DEBUG_STUDYDATA("post-fin", data, depth, is_inf);
6032 SSize_t final_minlen= min < stopmin ? min : stopmin;
6034 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
6035 if (final_minlen > SSize_t_MAX - delta)
6036 RExC_maxlen = SSize_t_MAX;
6037 else if (RExC_maxlen < final_minlen + delta)
6038 RExC_maxlen = final_minlen + delta;
6040 return final_minlen;
6042 NOT_REACHED; /* NOTREACHED */
6046 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
6048 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
6050 PERL_ARGS_ASSERT_ADD_DATA;
6052 Renewc(RExC_rxi->data,
6053 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6054 char, struct reg_data);
6056 Renew(RExC_rxi->data->what, count + n, U8);
6058 Newx(RExC_rxi->data->what, n, U8);
6059 RExC_rxi->data->count = count + n;
6060 Copy(s, RExC_rxi->data->what + count, n, U8);
6064 /*XXX: todo make this not included in a non debugging perl, but appears to be
6065 * used anyway there, in 'use re' */
6066 #ifndef PERL_IN_XSUB_RE
6068 Perl_reginitcolors(pTHX)
6070 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6072 char *t = savepv(s);
6076 t = strchr(t, '\t');
6082 PL_colors[i] = t = (char *)"";
6087 PL_colors[i++] = (char *)"";
6094 #ifdef TRIE_STUDY_OPT
6095 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6098 (data.flags & SCF_TRIE_RESTUDY) \
6106 #define CHECK_RESTUDY_GOTO_butfirst
6110 * pregcomp - compile a regular expression into internal code
6112 * Decides which engine's compiler to call based on the hint currently in
6116 #ifndef PERL_IN_XSUB_RE
6118 /* return the currently in-scope regex engine (or the default if none) */
6120 regexp_engine const *
6121 Perl_current_re_engine(pTHX)
6123 if (IN_PERL_COMPILETIME) {
6124 HV * const table = GvHV(PL_hintgv);
6127 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6128 return &PL_core_reg_engine;
6129 ptr = hv_fetchs(table, "regcomp", FALSE);
6130 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6131 return &PL_core_reg_engine;
6132 return INT2PTR(regexp_engine*,SvIV(*ptr));
6136 if (!PL_curcop->cop_hints_hash)
6137 return &PL_core_reg_engine;
6138 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6139 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6140 return &PL_core_reg_engine;
6141 return INT2PTR(regexp_engine*,SvIV(ptr));
6147 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6149 regexp_engine const *eng = current_re_engine();
6150 GET_RE_DEBUG_FLAGS_DECL;
6152 PERL_ARGS_ASSERT_PREGCOMP;
6154 /* Dispatch a request to compile a regexp to correct regexp engine. */
6156 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6159 return CALLREGCOMP_ENG(eng, pattern, flags);
6163 /* public(ish) entry point for the perl core's own regex compiling code.
6164 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6165 * pattern rather than a list of OPs, and uses the internal engine rather
6166 * than the current one */
6169 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6171 SV *pat = pattern; /* defeat constness! */
6172 PERL_ARGS_ASSERT_RE_COMPILE;
6173 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6174 #ifdef PERL_IN_XSUB_RE
6177 &PL_core_reg_engine,
6179 NULL, NULL, rx_flags, 0);
6184 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6188 if (--cbs->refcnt > 0)
6190 for (n = 0; n < cbs->count; n++) {
6191 REGEXP *rx = cbs->cb[n].src_regex;
6192 cbs->cb[n].src_regex = NULL;
6200 static struct reg_code_blocks *
6201 S_alloc_code_blocks(pTHX_ int ncode)
6203 struct reg_code_blocks *cbs;
6204 Newx(cbs, 1, struct reg_code_blocks);
6207 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6209 Newx(cbs->cb, ncode, struct reg_code_block);
6216 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6217 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6218 * point to the realloced string and length.
6220 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6224 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6225 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6227 U8 *const src = (U8*)*pat_p;
6232 GET_RE_DEBUG_FLAGS_DECL;
6234 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6235 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6237 Newx(dst, *plen_p * 2 + 1, U8);
6240 while (s < *plen_p) {
6241 append_utf8_from_native_byte(src[s], &d);
6243 if (n < num_code_blocks) {
6244 assert(pRExC_state->code_blocks);
6245 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6246 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6247 assert(*(d - 1) == '(');
6250 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6251 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6252 assert(*(d - 1) == ')');
6261 *pat_p = (char*) dst;
6263 RExC_orig_utf8 = RExC_utf8 = 1;
6268 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6269 * while recording any code block indices, and handling overloading,
6270 * nested qr// objects etc. If pat is null, it will allocate a new
6271 * string, or just return the first arg, if there's only one.
6273 * Returns the malloced/updated pat.
6274 * patternp and pat_count is the array of SVs to be concatted;
6275 * oplist is the optional list of ops that generated the SVs;
6276 * recompile_p is a pointer to a boolean that will be set if
6277 * the regex will need to be recompiled.
6278 * delim, if non-null is an SV that will be inserted between each element
6282 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6283 SV *pat, SV ** const patternp, int pat_count,
6284 OP *oplist, bool *recompile_p, SV *delim)
6288 bool use_delim = FALSE;
6289 bool alloced = FALSE;
6291 /* if we know we have at least two args, create an empty string,
6292 * then concatenate args to that. For no args, return an empty string */
6293 if (!pat && pat_count != 1) {
6299 for (svp = patternp; svp < patternp + pat_count; svp++) {
6302 STRLEN orig_patlen = 0;
6304 SV *msv = use_delim ? delim : *svp;
6305 if (!msv) msv = &PL_sv_undef;
6307 /* if we've got a delimiter, we go round the loop twice for each
6308 * svp slot (except the last), using the delimiter the second
6317 if (SvTYPE(msv) == SVt_PVAV) {
6318 /* we've encountered an interpolated array within
6319 * the pattern, e.g. /...@a..../. Expand the list of elements,
6320 * then recursively append elements.
6321 * The code in this block is based on S_pushav() */
6323 AV *const av = (AV*)msv;
6324 const SSize_t maxarg = AvFILL(av) + 1;
6328 assert(oplist->op_type == OP_PADAV
6329 || oplist->op_type == OP_RV2AV);
6330 oplist = OpSIBLING(oplist);
6333 if (SvRMAGICAL(av)) {
6336 Newx(array, maxarg, SV*);
6338 for (i=0; i < maxarg; i++) {
6339 SV ** const svp = av_fetch(av, i, FALSE);
6340 array[i] = svp ? *svp : &PL_sv_undef;
6344 array = AvARRAY(av);
6346 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6347 array, maxarg, NULL, recompile_p,
6349 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6355 /* we make the assumption here that each op in the list of
6356 * op_siblings maps to one SV pushed onto the stack,
6357 * except for code blocks, with have both an OP_NULL and
6359 * This allows us to match up the list of SVs against the
6360 * list of OPs to find the next code block.
6362 * Note that PUSHMARK PADSV PADSV ..
6364 * PADRANGE PADSV PADSV ..
6365 * so the alignment still works. */
6368 if (oplist->op_type == OP_NULL
6369 && (oplist->op_flags & OPf_SPECIAL))
6371 assert(n < pRExC_state->code_blocks->count);
6372 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6373 pRExC_state->code_blocks->cb[n].block = oplist;
6374 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6377 oplist = OpSIBLING(oplist); /* skip CONST */
6380 oplist = OpSIBLING(oplist);;
6383 /* apply magic and QR overloading to arg */
6386 if (SvROK(msv) && SvAMAGIC(msv)) {
6387 SV *sv = AMG_CALLunary(msv, regexp_amg);
6391 if (SvTYPE(sv) != SVt_REGEXP)
6392 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6397 /* try concatenation overload ... */
6398 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6399 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6402 /* overloading involved: all bets are off over literal
6403 * code. Pretend we haven't seen it */
6405 pRExC_state->code_blocks->count -= n;
6409 /* ... or failing that, try "" overload */
6410 while (SvAMAGIC(msv)
6411 && (sv = AMG_CALLunary(msv, string_amg))
6415 && SvRV(msv) == SvRV(sv))
6420 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6424 /* this is a partially unrolled
6425 * sv_catsv_nomg(pat, msv);
6426 * that allows us to adjust code block indices if
6429 char *dst = SvPV_force_nomg(pat, dlen);
6431 if (SvUTF8(msv) && !SvUTF8(pat)) {
6432 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6433 sv_setpvn(pat, dst, dlen);
6436 sv_catsv_nomg(pat, msv);
6440 /* We have only one SV to process, but we need to verify
6441 * it is properly null terminated or we will fail asserts
6442 * later. In theory we probably shouldn't get such SV's,
6443 * but if we do we should handle it gracefully. */
6444 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) ) {
6445 /* not a string, or a string with a trailing null */
6448 /* a string with no trailing null, we need to copy it
6449 * so it we have a trailing null */
6455 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
6458 /* extract any code blocks within any embedded qr//'s */
6459 if (rx && SvTYPE(rx) == SVt_REGEXP
6460 && RX_ENGINE((REGEXP*)rx)->op_comp)
6463 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6464 if (ri->code_blocks && ri->code_blocks->count) {
6466 /* the presence of an embedded qr// with code means
6467 * we should always recompile: the text of the
6468 * qr// may not have changed, but it may be a
6469 * different closure than last time */
6471 if (pRExC_state->code_blocks) {
6472 int new_count = pRExC_state->code_blocks->count
6473 + ri->code_blocks->count;
6474 Renew(pRExC_state->code_blocks->cb,
6475 new_count, struct reg_code_block);
6476 pRExC_state->code_blocks->count = new_count;
6479 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
6480 ri->code_blocks->count);
6482 for (i=0; i < ri->code_blocks->count; i++) {
6483 struct reg_code_block *src, *dst;
6484 STRLEN offset = orig_patlen
6485 + ReANY((REGEXP *)rx)->pre_prefix;
6486 assert(n < pRExC_state->code_blocks->count);
6487 src = &ri->code_blocks->cb[i];
6488 dst = &pRExC_state->code_blocks->cb[n];
6489 dst->start = src->start + offset;
6490 dst->end = src->end + offset;
6491 dst->block = src->block;
6492 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6501 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6510 /* see if there are any run-time code blocks in the pattern.
6511 * False positives are allowed */
6514 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6515 char *pat, STRLEN plen)
6520 PERL_UNUSED_CONTEXT;
6522 for (s = 0; s < plen; s++) {
6523 if ( pRExC_state->code_blocks
6524 && n < pRExC_state->code_blocks->count
6525 && s == pRExC_state->code_blocks->cb[n].start)
6527 s = pRExC_state->code_blocks->cb[n].end;
6531 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6533 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6535 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6542 /* Handle run-time code blocks. We will already have compiled any direct
6543 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6544 * copy of it, but with any literal code blocks blanked out and
6545 * appropriate chars escaped; then feed it into
6547 * eval "qr'modified_pattern'"
6551 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6555 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6557 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6558 * and merge them with any code blocks of the original regexp.
6560 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6561 * instead, just save the qr and return FALSE; this tells our caller that
6562 * the original pattern needs upgrading to utf8.
6566 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6567 char *pat, STRLEN plen)
6571 GET_RE_DEBUG_FLAGS_DECL;
6573 if (pRExC_state->runtime_code_qr) {
6574 /* this is the second time we've been called; this should
6575 * only happen if the main pattern got upgraded to utf8
6576 * during compilation; re-use the qr we compiled first time
6577 * round (which should be utf8 too)
6579 qr = pRExC_state->runtime_code_qr;
6580 pRExC_state->runtime_code_qr = NULL;
6581 assert(RExC_utf8 && SvUTF8(qr));
6587 int newlen = plen + 7; /* allow for "qr''xx\0" extra chars */
6591 /* determine how many extra chars we need for ' and \ escaping */
6592 for (s = 0; s < plen; s++) {
6593 if (pat[s] == '\'' || pat[s] == '\\')
6597 Newx(newpat, newlen, char);
6599 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6601 for (s = 0; s < plen; s++) {
6602 if ( pRExC_state->code_blocks
6603 && n < pRExC_state->code_blocks->count
6604 && s == pRExC_state->code_blocks->cb[n].start)
6606 /* blank out literal code block */
6607 assert(pat[s] == '(');
6608 while (s <= pRExC_state->code_blocks->cb[n].end) {
6616 if (pat[s] == '\'' || pat[s] == '\\')
6621 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
6623 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
6629 Perl_re_printf( aTHX_
6630 "%sre-parsing pattern for runtime code:%s %s\n",
6631 PL_colors[4],PL_colors[5],newpat);
6634 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
6640 PUSHSTACKi(PERLSI_REQUIRE);
6641 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
6642 * parsing qr''; normally only q'' does this. It also alters
6644 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
6645 SvREFCNT_dec_NN(sv);
6650 SV * const errsv = ERRSV;
6651 if (SvTRUE_NN(errsv))
6652 /* use croak_sv ? */
6653 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
6655 assert(SvROK(qr_ref));
6657 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
6658 /* the leaving below frees the tmp qr_ref.
6659 * Give qr a life of its own */
6667 if (!RExC_utf8 && SvUTF8(qr)) {
6668 /* first time through; the pattern got upgraded; save the
6669 * qr for the next time through */
6670 assert(!pRExC_state->runtime_code_qr);
6671 pRExC_state->runtime_code_qr = qr;
6676 /* extract any code blocks within the returned qr// */
6679 /* merge the main (r1) and run-time (r2) code blocks into one */
6681 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6682 struct reg_code_block *new_block, *dst;
6683 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6687 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
6689 SvREFCNT_dec_NN(qr);
6693 if (!r1->code_blocks)
6694 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
6696 r1c = r1->code_blocks->count;
6697 r2c = r2->code_blocks->count;
6699 Newx(new_block, r1c + r2c, struct reg_code_block);
6703 while (i1 < r1c || i2 < r2c) {
6704 struct reg_code_block *src;
6708 src = &r2->code_blocks->cb[i2++];
6712 src = &r1->code_blocks->cb[i1++];
6713 else if ( r1->code_blocks->cb[i1].start
6714 < r2->code_blocks->cb[i2].start)
6716 src = &r1->code_blocks->cb[i1++];
6717 assert(src->end < r2->code_blocks->cb[i2].start);
6720 assert( r1->code_blocks->cb[i1].start
6721 > r2->code_blocks->cb[i2].start);
6722 src = &r2->code_blocks->cb[i2++];
6724 assert(src->end < r1->code_blocks->cb[i1].start);
6727 assert(pat[src->start] == '(');
6728 assert(pat[src->end] == ')');
6729 dst->start = src->start;
6730 dst->end = src->end;
6731 dst->block = src->block;
6732 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6736 r1->code_blocks->count += r2c;
6737 Safefree(r1->code_blocks->cb);
6738 r1->code_blocks->cb = new_block;
6741 SvREFCNT_dec_NN(qr);
6747 S_setup_longest(pTHX_ RExC_state_t *pRExC_state,
6748 struct reg_substr_datum *rsd,
6749 struct scan_data_substrs *sub,
6750 STRLEN longest_length)
6752 /* This is the common code for setting up the floating and fixed length
6753 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6754 * as to whether succeeded or not */
6758 bool eol = cBOOL(sub->flags & SF_BEFORE_EOL);
6759 bool meol = cBOOL(sub->flags & SF_BEFORE_MEOL);
6761 if (! (longest_length
6762 || (eol /* Can't have SEOL and MULTI */
6763 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6765 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6766 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6771 /* copy the information about the longest from the reg_scan_data
6772 over to the program. */
6773 if (SvUTF8(sub->str)) {
6775 rsd->utf8_substr = sub->str;
6777 rsd->substr = sub->str;
6778 rsd->utf8_substr = NULL;
6780 /* end_shift is how many chars that must be matched that
6781 follow this item. We calculate it ahead of time as once the
6782 lookbehind offset is added in we lose the ability to correctly
6784 ml = sub->minlenp ? *(sub->minlenp) : (SSize_t)longest_length;
6785 rsd->end_shift = ml - sub->min_offset
6787 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
6789 + (SvTAIL(sub->str) != 0)
6793 t = (eol/* Can't have SEOL and MULTI */
6794 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6795 fbm_compile(sub->str, t ? FBMcf_TAIL : 0);
6801 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6802 * regular expression into internal code.
6803 * The pattern may be passed either as:
6804 * a list of SVs (patternp plus pat_count)
6805 * a list of OPs (expr)
6806 * If both are passed, the SV list is used, but the OP list indicates
6807 * which SVs are actually pre-compiled code blocks
6809 * The SVs in the list have magic and qr overloading applied to them (and
6810 * the list may be modified in-place with replacement SVs in the latter
6813 * If the pattern hasn't changed from old_re, then old_re will be
6816 * eng is the current engine. If that engine has an op_comp method, then
6817 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6818 * do the initial concatenation of arguments and pass on to the external
6821 * If is_bare_re is not null, set it to a boolean indicating whether the
6822 * arg list reduced (after overloading) to a single bare regex which has
6823 * been returned (i.e. /$qr/).
6825 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6827 * pm_flags contains the PMf_* flags, typically based on those from the
6828 * pm_flags field of the related PMOP. Currently we're only interested in
6829 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6831 * We can't allocate space until we know how big the compiled form will be,
6832 * but we can't compile it (and thus know how big it is) until we've got a
6833 * place to put the code. So we cheat: we compile it twice, once with code
6834 * generation turned off and size counting turned on, and once "for real".
6835 * This also means that we don't allocate space until we are sure that the
6836 * thing really will compile successfully, and we never have to move the
6837 * code and thus invalidate pointers into it. (Note that it has to be in
6838 * one piece because free() must be able to free it all.) [NB: not true in perl]
6840 * Beware that the optimization-preparation code in here knows about some
6841 * of the structure of the compiled regexp. [I'll say.]
6845 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6846 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6847 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6851 regexp_internal *ri;
6859 SV** new_patternp = patternp;
6861 /* these are all flags - maybe they should be turned
6862 * into a single int with different bit masks */
6863 I32 sawlookahead = 0;
6868 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6870 bool runtime_code = 0;
6872 RExC_state_t RExC_state;
6873 RExC_state_t * const pRExC_state = &RExC_state;
6874 #ifdef TRIE_STUDY_OPT
6876 RExC_state_t copyRExC_state;
6878 GET_RE_DEBUG_FLAGS_DECL;
6880 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6882 DEBUG_r(if (!PL_colorset) reginitcolors());
6884 /* Initialize these here instead of as-needed, as is quick and avoids
6885 * having to test them each time otherwise */
6886 if (! PL_AboveLatin1) {
6888 char * dump_len_string;
6891 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6892 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6893 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6894 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6895 PL_HasMultiCharFold =
6896 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6898 /* This is calculated here, because the Perl program that generates the
6899 * static global ones doesn't currently have access to
6900 * NUM_ANYOF_CODE_POINTS */
6901 PL_InBitmap = _new_invlist(2);
6902 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
6903 NUM_ANYOF_CODE_POINTS - 1);
6905 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
6906 if ( ! dump_len_string
6907 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
6909 PL_dump_re_max_len = 60; /* A reasonable default */
6914 pRExC_state->warn_text = NULL;
6915 pRExC_state->code_blocks = NULL;
6918 *is_bare_re = FALSE;
6920 if (expr && (expr->op_type == OP_LIST ||
6921 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6922 /* allocate code_blocks if needed */
6926 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
6927 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6928 ncode++; /* count of DO blocks */
6931 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
6935 /* compile-time pattern with just OP_CONSTs and DO blocks */
6940 /* find how many CONSTs there are */
6943 if (expr->op_type == OP_CONST)
6946 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6947 if (o->op_type == OP_CONST)
6951 /* fake up an SV array */
6953 assert(!new_patternp);
6954 Newx(new_patternp, n, SV*);
6955 SAVEFREEPV(new_patternp);
6959 if (expr->op_type == OP_CONST)
6960 new_patternp[n] = cSVOPx_sv(expr);
6962 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6963 if (o->op_type == OP_CONST)
6964 new_patternp[n++] = cSVOPo_sv;
6969 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6970 "Assembling pattern from %d elements%s\n", pat_count,
6971 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6973 /* set expr to the first arg op */
6975 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
6976 && expr->op_type != OP_CONST)
6978 expr = cLISTOPx(expr)->op_first;
6979 assert( expr->op_type == OP_PUSHMARK
6980 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6981 || expr->op_type == OP_PADRANGE);
6982 expr = OpSIBLING(expr);
6985 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6986 expr, &recompile, NULL);
6988 /* handle bare (possibly after overloading) regex: foo =~ $re */
6993 if (SvTYPE(re) == SVt_REGEXP) {
6997 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6998 "Precompiled pattern%s\n",
6999 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7005 exp = SvPV_nomg(pat, plen);
7007 if (!eng->op_comp) {
7008 if ((SvUTF8(pat) && IN_BYTES)
7009 || SvGMAGICAL(pat) || SvAMAGIC(pat))
7011 /* make a temporary copy; either to convert to bytes,
7012 * or to avoid repeating get-magic / overloaded stringify */
7013 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
7014 (IN_BYTES ? 0 : SvUTF8(pat)));
7016 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
7019 /* ignore the utf8ness if the pattern is 0 length */
7020 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
7022 RExC_uni_semantics = 0;
7023 RExC_seen_unfolded_sharp_s = 0;
7024 RExC_contains_locale = 0;
7025 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
7026 RExC_study_started = 0;
7027 pRExC_state->runtime_code_qr = NULL;
7028 RExC_frame_head= NULL;
7029 RExC_frame_last= NULL;
7030 RExC_frame_count= 0;
7033 RExC_mysv1= sv_newmortal();
7034 RExC_mysv2= sv_newmortal();
7037 SV *dsv= sv_newmortal();
7038 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7039 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
7040 PL_colors[4],PL_colors[5],s);
7044 /* we jump here if we have to recompile, e.g., from upgrading the pattern
7047 if ((pm_flags & PMf_USE_RE_EVAL)
7048 /* this second condition covers the non-regex literal case,
7049 * i.e. $foo =~ '(?{})'. */
7050 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
7052 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
7054 /* return old regex if pattern hasn't changed */
7055 /* XXX: note in the below we have to check the flags as well as the
7058 * Things get a touch tricky as we have to compare the utf8 flag
7059 * independently from the compile flags. */
7063 && !!RX_UTF8(old_re) == !!RExC_utf8
7064 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7065 && RX_PRECOMP(old_re)
7066 && RX_PRELEN(old_re) == plen
7067 && memEQ(RX_PRECOMP(old_re), exp, plen)
7068 && !runtime_code /* with runtime code, always recompile */ )
7073 rx_flags = orig_rx_flags;
7075 if ( initial_charset == REGEX_DEPENDS_CHARSET
7076 && (RExC_utf8 ||RExC_uni_semantics))
7079 /* Set to use unicode semantics if the pattern is in utf8 and has the
7080 * 'depends' charset specified, as it means unicode when utf8 */
7081 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7085 RExC_precomp_adj = 0;
7086 RExC_flags = rx_flags;
7087 RExC_pm_flags = pm_flags;
7090 assert(TAINTING_get || !TAINT_get);
7092 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7094 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7095 /* whoops, we have a non-utf8 pattern, whilst run-time code
7096 * got compiled as utf8. Try again with a utf8 pattern */
7097 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7098 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7099 goto redo_first_pass;
7102 assert(!pRExC_state->runtime_code_qr);
7108 RExC_in_lookbehind = 0;
7109 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7112 RExC_recode_x_to_native = 0;
7114 RExC_in_multi_char_class = 0;
7116 /* First pass: determine size, legality. */
7118 RExC_start = RExC_adjusted_start = exp;
7119 RExC_end = exp + plen;
7120 RExC_precomp_end = RExC_end;
7125 RExC_emit = (regnode *) &RExC_emit_dummy;
7126 RExC_whilem_seen = 0;
7127 RExC_open_parens = NULL;
7128 RExC_close_parens = NULL;
7130 RExC_paren_names = NULL;
7132 RExC_paren_name_list = NULL;
7134 RExC_recurse = NULL;
7135 RExC_study_chunk_recursed = NULL;
7136 RExC_study_chunk_recursed_bytes= 0;
7137 RExC_recurse_count = 0;
7138 pRExC_state->code_index = 0;
7140 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7141 * code makes sure the final byte is an uncounted NUL. But should this
7142 * ever not be the case, lots of things could read beyond the end of the
7143 * buffer: loops like
7144 * while(isFOO(*RExC_parse)) RExC_parse++;
7145 * strchr(RExC_parse, "foo");
7146 * etc. So it is worth noting. */
7147 assert(*RExC_end == '\0');
7150 Perl_re_printf( aTHX_ "Starting first pass (sizing)\n");
7152 RExC_lastparse=NULL;
7155 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7156 /* It's possible to write a regexp in ascii that represents Unicode
7157 codepoints outside of the byte range, such as via \x{100}. If we
7158 detect such a sequence we have to convert the entire pattern to utf8
7159 and then recompile, as our sizing calculation will have been based
7160 on 1 byte == 1 character, but we will need to use utf8 to encode
7161 at least some part of the pattern, and therefore must convert the whole
7164 if (flags & RESTART_PASS1) {
7165 if (flags & NEED_UTF8) {
7166 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7167 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7170 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7171 "Need to redo pass 1\n"));
7174 goto redo_first_pass;
7176 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#" UVxf, (UV) flags);
7180 Perl_re_printf( aTHX_
7181 "Required size %" IVdf " nodes\n"
7182 "Starting second pass (creation)\n",
7185 RExC_lastparse=NULL;
7188 /* The first pass could have found things that force Unicode semantics */
7189 if ((RExC_utf8 || RExC_uni_semantics)
7190 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
7192 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7195 /* Small enough for pointer-storage convention?
7196 If extralen==0, this means that we will not need long jumps. */
7197 if (RExC_size >= 0x10000L && RExC_extralen)
7198 RExC_size += RExC_extralen;
7201 if (RExC_whilem_seen > 15)
7202 RExC_whilem_seen = 15;
7204 /* Allocate space and zero-initialize. Note, the two step process
7205 of zeroing when in debug mode, thus anything assigned has to
7206 happen after that */
7207 rx = (REGEXP*) newSV_type(SVt_REGEXP);
7209 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7210 char, regexp_internal);
7211 if ( r == NULL || ri == NULL )
7212 FAIL("Regexp out of space");
7214 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
7215 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
7218 /* bulk initialize base fields with 0. */
7219 Zero(ri, sizeof(regexp_internal), char);
7222 /* non-zero initialization begins here */
7225 r->extflags = rx_flags;
7226 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7228 if (pm_flags & PMf_IS_QR) {
7229 ri->code_blocks = pRExC_state->code_blocks;
7230 if (ri->code_blocks)
7231 ri->code_blocks->refcnt++;
7235 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7236 bool has_charset = (get_regex_charset(r->extflags)
7237 != REGEX_DEPENDS_CHARSET);
7239 /* The caret is output if there are any defaults: if not all the STD
7240 * flags are set, or if no character set specifier is needed */
7242 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7244 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7245 == REG_RUN_ON_COMMENT_SEEN);
7246 U8 reganch = (U8)((r->extflags & RXf_PMf_STD_PMMOD)
7247 >> RXf_PMf_STD_PMMOD_SHIFT);
7248 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7251 /* We output all the necessary flags; we never output a minus, as all
7252 * those are defaults, so are
7253 * covered by the caret */
7254 const STRLEN wraplen = plen + has_p + has_runon
7255 + has_default /* If needs a caret */
7256 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7258 /* If needs a character set specifier */
7259 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7260 + (sizeof("(?:)") - 1);
7262 /* make sure PL_bitcount bounds not exceeded */
7263 assert(sizeof(STD_PAT_MODS) <= 8);
7265 p = sv_grow(MUTABLE_SV(rx), wraplen + 1); /* +1 for the ending NUL */
7268 SvFLAGS(rx) |= SVf_UTF8;
7271 /* If a default, cover it using the caret */
7273 *p++= DEFAULT_PAT_MOD;
7277 const char* const name = get_regex_charset_name(r->extflags, &len);
7278 Copy(name, p, len, char);
7282 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7285 while((ch = *fptr++)) {
7293 Copy(RExC_precomp, p, plen, char);
7294 assert ((RX_WRAPPED(rx) - p) < 16);
7295 r->pre_prefix = p - RX_WRAPPED(rx);
7301 SvCUR_set(rx, p - RX_WRAPPED(rx));
7305 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
7307 /* Useful during FAIL. */
7308 #ifdef RE_TRACK_PATTERN_OFFSETS
7309 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
7310 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7311 "%s %" UVuf " bytes for offset annotations.\n",
7312 ri->u.offsets ? "Got" : "Couldn't get",
7313 (UV)((2*RExC_size+1) * sizeof(U32))));
7315 SetProgLen(ri,RExC_size);
7320 /* Second pass: emit code. */
7321 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7322 RExC_pm_flags = pm_flags;
7324 RExC_end = exp + plen;
7326 RExC_emit_start = ri->program;
7327 RExC_emit = ri->program;
7328 RExC_emit_bound = ri->program + RExC_size + 1;
7329 pRExC_state->code_index = 0;
7331 *((char*) RExC_emit++) = (char) REG_MAGIC;
7332 /* setup various meta data about recursion, this all requires
7333 * RExC_npar to be correctly set, and a bit later on we clear it */
7334 if (RExC_seen & REG_RECURSE_SEEN) {
7335 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
7336 "%*s%*s Setting up open/close parens\n",
7337 22, "| |", (int)(0 * 2 + 1), ""));
7339 /* setup RExC_open_parens, which holds the address of each
7340 * OPEN tag, and to make things simpler for the 0 index
7341 * the start of the program - this is used later for offsets */
7342 Newxz(RExC_open_parens, RExC_npar,regnode *);
7343 SAVEFREEPV(RExC_open_parens);
7344 RExC_open_parens[0] = RExC_emit;
7346 /* setup RExC_close_parens, which holds the address of each
7347 * CLOSE tag, and to make things simpler for the 0 index
7348 * the end of the program - this is used later for offsets */
7349 Newxz(RExC_close_parens, RExC_npar,regnode *);
7350 SAVEFREEPV(RExC_close_parens);
7351 /* we dont know where end op starts yet, so we dont
7352 * need to set RExC_close_parens[0] like we do RExC_open_parens[0] above */
7354 /* Note, RExC_npar is 1 + the number of parens in a pattern.
7355 * So its 1 if there are no parens. */
7356 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
7357 ((RExC_npar & 0x07) != 0);
7358 Newx(RExC_study_chunk_recursed,
7359 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7360 SAVEFREEPV(RExC_study_chunk_recursed);
7363 if (reg(pRExC_state, 0, &flags,1) == NULL) {
7365 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#" UVxf, (UV) flags);
7368 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7371 /* XXXX To minimize changes to RE engine we always allocate
7372 3-units-long substrs field. */
7373 Newx(r->substrs, 1, struct reg_substr_data);
7374 if (RExC_recurse_count) {
7375 Newx(RExC_recurse,RExC_recurse_count,regnode *);
7376 SAVEFREEPV(RExC_recurse);
7380 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
7382 RExC_study_chunk_recursed_count= 0;
7384 Zero(r->substrs, 1, struct reg_substr_data);
7385 if (RExC_study_chunk_recursed) {
7386 Zero(RExC_study_chunk_recursed,
7387 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
7391 #ifdef TRIE_STUDY_OPT
7393 StructCopy(&zero_scan_data, &data, scan_data_t);
7394 copyRExC_state = RExC_state;
7397 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
7399 RExC_state = copyRExC_state;
7400 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
7401 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
7403 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
7404 StructCopy(&zero_scan_data, &data, scan_data_t);
7407 StructCopy(&zero_scan_data, &data, scan_data_t);
7410 /* Dig out information for optimizations. */
7411 r->extflags = RExC_flags; /* was pm_op */
7412 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
7415 SvUTF8_on(rx); /* Unicode in it? */
7416 ri->regstclass = NULL;
7417 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
7418 r->intflags |= PREGf_NAUGHTY;
7419 scan = ri->program + 1; /* First BRANCH. */
7421 /* testing for BRANCH here tells us whether there is "must appear"
7422 data in the pattern. If there is then we can use it for optimisations */
7423 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
7426 STRLEN longest_length[2];
7427 regnode_ssc ch_class; /* pointed to by data */
7429 SSize_t last_close = 0; /* pointed to by data */
7430 regnode *first= scan;
7431 regnode *first_next= regnext(first);
7435 * Skip introductions and multiplicators >= 1
7436 * so that we can extract the 'meat' of the pattern that must
7437 * match in the large if() sequence following.
7438 * NOTE that EXACT is NOT covered here, as it is normally
7439 * picked up by the optimiser separately.
7441 * This is unfortunate as the optimiser isnt handling lookahead
7442 * properly currently.
7445 while ((OP(first) == OPEN && (sawopen = 1)) ||
7446 /* An OR of *one* alternative - should not happen now. */
7447 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
7448 /* for now we can't handle lookbehind IFMATCH*/
7449 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
7450 (OP(first) == PLUS) ||
7451 (OP(first) == MINMOD) ||
7452 /* An {n,m} with n>0 */
7453 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
7454 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
7457 * the only op that could be a regnode is PLUS, all the rest
7458 * will be regnode_1 or regnode_2.
7460 * (yves doesn't think this is true)
7462 if (OP(first) == PLUS)
7465 if (OP(first) == MINMOD)
7467 first += regarglen[OP(first)];
7469 first = NEXTOPER(first);
7470 first_next= regnext(first);
7473 /* Starting-point info. */
7475 DEBUG_PEEP("first:", first, 0, 0);
7476 /* Ignore EXACT as we deal with it later. */
7477 if (PL_regkind[OP(first)] == EXACT) {
7478 if (OP(first) == EXACT || OP(first) == EXACTL)
7479 NOOP; /* Empty, get anchored substr later. */
7481 ri->regstclass = first;
7484 else if (PL_regkind[OP(first)] == TRIE &&
7485 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
7487 /* this can happen only on restudy */
7488 ri->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
7491 else if (REGNODE_SIMPLE(OP(first)))
7492 ri->regstclass = first;
7493 else if (PL_regkind[OP(first)] == BOUND ||
7494 PL_regkind[OP(first)] == NBOUND)
7495 ri->regstclass = first;
7496 else if (PL_regkind[OP(first)] == BOL) {
7497 r->intflags |= (OP(first) == MBOL
7500 first = NEXTOPER(first);
7503 else if (OP(first) == GPOS) {
7504 r->intflags |= PREGf_ANCH_GPOS;
7505 first = NEXTOPER(first);
7508 else if ((!sawopen || !RExC_sawback) &&
7510 (OP(first) == STAR &&
7511 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
7512 !(r->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
7514 /* turn .* into ^.* with an implied $*=1 */
7516 (OP(NEXTOPER(first)) == REG_ANY)
7519 r->intflags |= (type | PREGf_IMPLICIT);
7520 first = NEXTOPER(first);
7523 if (sawplus && !sawminmod && !sawlookahead
7524 && (!sawopen || !RExC_sawback)
7525 && !pRExC_state->code_blocks) /* May examine pos and $& */
7526 /* x+ must match at the 1st pos of run of x's */
7527 r->intflags |= PREGf_SKIP;
7529 /* Scan is after the zeroth branch, first is atomic matcher. */
7530 #ifdef TRIE_STUDY_OPT
7533 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7534 (IV)(first - scan + 1))
7538 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
7539 (IV)(first - scan + 1))
7545 * If there's something expensive in the r.e., find the
7546 * longest literal string that must appear and make it the
7547 * regmust. Resolve ties in favor of later strings, since
7548 * the regstart check works with the beginning of the r.e.
7549 * and avoiding duplication strengthens checking. Not a
7550 * strong reason, but sufficient in the absence of others.
7551 * [Now we resolve ties in favor of the earlier string if
7552 * it happens that c_offset_min has been invalidated, since the
7553 * earlier string may buy us something the later one won't.]
7556 data.substrs[0].str = newSVpvs("");
7557 data.substrs[1].str = newSVpvs("");
7558 data.last_found = newSVpvs("");
7559 data.cur_is_floating = 0; /* initially any found substring is fixed */
7560 ENTER_with_name("study_chunk");
7561 SAVEFREESV(data.substrs[0].str);
7562 SAVEFREESV(data.substrs[1].str);
7563 SAVEFREESV(data.last_found);
7565 if (!ri->regstclass) {
7566 ssc_init(pRExC_state, &ch_class);
7567 data.start_class = &ch_class;
7568 stclass_flag = SCF_DO_STCLASS_AND;
7569 } else /* XXXX Check for BOUND? */
7571 data.last_closep = &last_close;
7574 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
7575 scan + RExC_size, /* Up to end */
7577 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
7578 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
7582 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
7585 if ( RExC_npar == 1 && !data.cur_is_floating
7586 && data.last_start_min == 0 && data.last_end > 0
7587 && !RExC_seen_zerolen
7588 && !(RExC_seen & REG_VERBARG_SEEN)
7589 && !(RExC_seen & REG_GPOS_SEEN)
7591 r->extflags |= RXf_CHECK_ALL;
7593 scan_commit(pRExC_state, &data,&minlen,0);
7596 /* XXX this is done in reverse order because that's the way the
7597 * code was before it was parameterised. Don't know whether it
7598 * actually needs doing in reverse order. DAPM */
7599 for (i = 1; i >= 0; i--) {
7600 longest_length[i] = CHR_SVLEN(data.substrs[i].str);
7603 && SvCUR(data.substrs[0].str) /* ok to leave SvCUR */
7604 && data.substrs[0].min_offset
7605 == data.substrs[1].min_offset
7606 && SvCUR(data.substrs[0].str)
7607 == SvCUR(data.substrs[1].str)
7609 && S_setup_longest (aTHX_ pRExC_state,
7610 &(r->substrs->data[i]),
7614 r->substrs->data[i].min_offset =
7615 data.substrs[i].min_offset - data.substrs[i].lookbehind;
7617 r->substrs->data[i].max_offset = data.substrs[i].max_offset;
7618 /* Don't offset infinity */
7619 if (data.substrs[i].max_offset < SSize_t_MAX)
7620 r->substrs->data[i].max_offset -= data.substrs[i].lookbehind;
7621 SvREFCNT_inc_simple_void_NN(data.substrs[i].str);
7624 r->substrs->data[i].substr = NULL;
7625 r->substrs->data[i].utf8_substr = NULL;
7626 longest_length[i] = 0;
7630 LEAVE_with_name("study_chunk");
7633 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
7634 ri->regstclass = NULL;
7636 if ((!(r->substrs->data[0].substr || r->substrs->data[0].utf8_substr)
7637 || r->substrs->data[0].min_offset)
7639 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7640 && is_ssc_worth_it(pRExC_state, data.start_class))
7642 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7644 ssc_finalize(pRExC_state, data.start_class);
7646 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7647 StructCopy(data.start_class,
7648 (regnode_ssc*)RExC_rxi->data->data[n],
7650 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7651 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7652 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
7653 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7654 Perl_re_printf( aTHX_
7655 "synthetic stclass \"%s\".\n",
7656 SvPVX_const(sv));});
7657 data.start_class = NULL;
7660 /* A temporary algorithm prefers floated substr to fixed one of
7661 * same length to dig more info. */
7662 i = (longest_length[0] <= longest_length[1]);
7663 r->substrs->check_ix = i;
7664 r->check_end_shift = r->substrs->data[i].end_shift;
7665 r->check_substr = r->substrs->data[i].substr;
7666 r->check_utf8 = r->substrs->data[i].utf8_substr;
7667 r->check_offset_min = r->substrs->data[i].min_offset;
7668 r->check_offset_max = r->substrs->data[i].max_offset;
7669 if (!i && (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS)))
7670 r->intflags |= PREGf_NOSCAN;
7672 if ((r->check_substr || r->check_utf8) ) {
7673 r->extflags |= RXf_USE_INTUIT;
7674 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
7675 r->extflags |= RXf_INTUIT_TAIL;
7678 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
7679 if ( (STRLEN)minlen < longest_length[1] )
7680 minlen= longest_length[1];
7681 if ( (STRLEN)minlen < longest_length[0] )
7682 minlen= longest_length[0];
7686 /* Several toplevels. Best we can is to set minlen. */
7688 regnode_ssc ch_class;
7689 SSize_t last_close = 0;
7691 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
7693 scan = ri->program + 1;
7694 ssc_init(pRExC_state, &ch_class);
7695 data.start_class = &ch_class;
7696 data.last_closep = &last_close;
7699 minlen = study_chunk(pRExC_state,
7700 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7701 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7702 ? SCF_TRIE_DOING_RESTUDY
7706 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7708 r->check_substr = NULL;
7709 r->check_utf8 = NULL;
7710 r->substrs->data[0].substr = NULL;
7711 r->substrs->data[0].utf8_substr = NULL;
7712 r->substrs->data[1].substr = NULL;
7713 r->substrs->data[1].utf8_substr = NULL;
7715 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7716 && is_ssc_worth_it(pRExC_state, data.start_class))
7718 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7720 ssc_finalize(pRExC_state, data.start_class);
7722 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7723 StructCopy(data.start_class,
7724 (regnode_ssc*)RExC_rxi->data->data[n],
7726 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7727 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7728 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7729 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7730 Perl_re_printf( aTHX_
7731 "synthetic stclass \"%s\".\n",
7732 SvPVX_const(sv));});
7733 data.start_class = NULL;
7737 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7738 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7739 r->maxlen = REG_INFTY;
7742 r->maxlen = RExC_maxlen;
7745 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7746 the "real" pattern. */
7748 Perl_re_printf( aTHX_ "minlen: %" IVdf " r->minlen:%" IVdf " maxlen:%" IVdf "\n",
7749 (IV)minlen, (IV)r->minlen, (IV)RExC_maxlen);
7751 r->minlenret = minlen;
7752 if (r->minlen < minlen)
7755 if (RExC_seen & REG_RECURSE_SEEN ) {
7756 r->intflags |= PREGf_RECURSE_SEEN;
7757 Newx(r->recurse_locinput, r->nparens + 1, char *);
7759 if (RExC_seen & REG_GPOS_SEEN)
7760 r->intflags |= PREGf_GPOS_SEEN;
7761 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7762 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7764 if (pRExC_state->code_blocks)
7765 r->extflags |= RXf_EVAL_SEEN;
7766 if (RExC_seen & REG_VERBARG_SEEN)
7768 r->intflags |= PREGf_VERBARG_SEEN;
7769 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7771 if (RExC_seen & REG_CUTGROUP_SEEN)
7772 r->intflags |= PREGf_CUTGROUP_SEEN;
7773 if (pm_flags & PMf_USE_RE_EVAL)
7774 r->intflags |= PREGf_USE_RE_EVAL;
7775 if (RExC_paren_names)
7776 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7778 RXp_PAREN_NAMES(r) = NULL;
7780 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7781 * so it can be used in pp.c */
7782 if (r->intflags & PREGf_ANCH)
7783 r->extflags |= RXf_IS_ANCHORED;
7787 /* this is used to identify "special" patterns that might result
7788 * in Perl NOT calling the regex engine and instead doing the match "itself",
7789 * particularly special cases in split//. By having the regex compiler
7790 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7791 * we avoid weird issues with equivalent patterns resulting in different behavior,
7792 * AND we allow non Perl engines to get the same optimizations by the setting the
7793 * flags appropriately - Yves */
7794 regnode *first = ri->program + 1;
7796 regnode *next = regnext(first);
7799 if (PL_regkind[fop] == NOTHING && nop == END)
7800 r->extflags |= RXf_NULL;
7801 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
7802 /* when fop is SBOL first->flags will be true only when it was
7803 * produced by parsing /\A/, and not when parsing /^/. This is
7804 * very important for the split code as there we want to
7805 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
7806 * See rt #122761 for more details. -- Yves */
7807 r->extflags |= RXf_START_ONLY;
7808 else if (fop == PLUS
7809 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7811 r->extflags |= RXf_WHITE;
7812 else if ( r->extflags & RXf_SPLIT
7813 && (fop == EXACT || fop == EXACTL)
7814 && STR_LEN(first) == 1
7815 && *(STRING(first)) == ' '
7817 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7821 if (RExC_contains_locale) {
7822 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7826 if (RExC_paren_names) {
7827 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7828 ri->data->data[ri->name_list_idx]
7829 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7832 ri->name_list_idx = 0;
7834 while ( RExC_recurse_count > 0 ) {
7835 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
7837 * This data structure is set up in study_chunk() and is used
7838 * to calculate the distance between a GOSUB regopcode and
7839 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
7842 * If for some reason someone writes code that optimises
7843 * away a GOSUB opcode then the assert should be changed to
7844 * an if(scan) to guard the ARG2L_SET() - Yves
7847 assert(scan && OP(scan) == GOSUB);
7848 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - scan );
7851 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7852 /* assume we don't need to swap parens around before we match */
7854 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
7855 (unsigned long)RExC_study_chunk_recursed_count);
7859 Perl_re_printf( aTHX_ "Final program:\n");
7862 #ifdef RE_TRACK_PATTERN_OFFSETS
7863 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7864 const STRLEN len = ri->u.offsets[0];
7866 GET_RE_DEBUG_FLAGS_DECL;
7867 Perl_re_printf( aTHX_
7868 "Offsets: [%" UVuf "]\n\t", (UV)ri->u.offsets[0]);
7869 for (i = 1; i <= len; i++) {
7870 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7871 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7872 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7874 Perl_re_printf( aTHX_ "\n");
7879 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7880 * by setting the regexp SV to readonly-only instead. If the
7881 * pattern's been recompiled, the USEDness should remain. */
7882 if (old_re && SvREADONLY(old_re))
7890 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7893 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7895 PERL_UNUSED_ARG(value);
7897 if (flags & RXapif_FETCH) {
7898 return reg_named_buff_fetch(rx, key, flags);
7899 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7900 Perl_croak_no_modify();
7902 } else if (flags & RXapif_EXISTS) {
7903 return reg_named_buff_exists(rx, key, flags)
7906 } else if (flags & RXapif_REGNAMES) {
7907 return reg_named_buff_all(rx, flags);
7908 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7909 return reg_named_buff_scalar(rx, flags);
7911 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7917 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7920 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7921 PERL_UNUSED_ARG(lastkey);
7923 if (flags & RXapif_FIRSTKEY)
7924 return reg_named_buff_firstkey(rx, flags);
7925 else if (flags & RXapif_NEXTKEY)
7926 return reg_named_buff_nextkey(rx, flags);
7928 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7935 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7939 struct regexp *const rx = ReANY(r);
7941 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7943 if (rx && RXp_PAREN_NAMES(rx)) {
7944 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7947 SV* sv_dat=HeVAL(he_str);
7948 I32 *nums=(I32*)SvPVX(sv_dat);
7949 AV * const retarray = (flags & RXapif_ALL) ? newAV() : NULL;
7950 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7951 if ((I32)(rx->nparens) >= nums[i]
7952 && rx->offs[nums[i]].start != -1
7953 && rx->offs[nums[i]].end != -1)
7956 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7961 ret = newSVsv(&PL_sv_undef);
7964 av_push(retarray, ret);
7967 return newRV_noinc(MUTABLE_SV(retarray));
7974 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7977 struct regexp *const rx = ReANY(r);
7979 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7981 if (rx && RXp_PAREN_NAMES(rx)) {
7982 if (flags & RXapif_ALL) {
7983 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7985 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7987 SvREFCNT_dec_NN(sv);
7999 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
8001 struct regexp *const rx = ReANY(r);
8003 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
8005 if ( rx && RXp_PAREN_NAMES(rx) ) {
8006 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
8008 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
8015 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
8017 struct regexp *const rx = ReANY(r);
8018 GET_RE_DEBUG_FLAGS_DECL;
8020 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
8022 if (rx && RXp_PAREN_NAMES(rx)) {
8023 HV *hv = RXp_PAREN_NAMES(rx);
8025 while ( (temphe = hv_iternext_flags(hv,0)) ) {
8028 SV* sv_dat = HeVAL(temphe);
8029 I32 *nums = (I32*)SvPVX(sv_dat);
8030 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8031 if ((I32)(rx->lastparen) >= nums[i] &&
8032 rx->offs[nums[i]].start != -1 &&
8033 rx->offs[nums[i]].end != -1)
8039 if (parno || flags & RXapif_ALL) {
8040 return newSVhek(HeKEY_hek(temphe));
8048 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8053 struct regexp *const rx = ReANY(r);
8055 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8057 if (rx && RXp_PAREN_NAMES(rx)) {
8058 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8059 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8060 } else if (flags & RXapif_ONE) {
8061 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8062 av = MUTABLE_AV(SvRV(ret));
8063 length = av_tindex(av);
8064 SvREFCNT_dec_NN(ret);
8065 return newSViv(length + 1);
8067 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8072 return &PL_sv_undef;
8076 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8078 struct regexp *const rx = ReANY(r);
8081 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8083 if (rx && RXp_PAREN_NAMES(rx)) {
8084 HV *hv= RXp_PAREN_NAMES(rx);
8086 (void)hv_iterinit(hv);
8087 while ( (temphe = hv_iternext_flags(hv,0)) ) {
8090 SV* sv_dat = HeVAL(temphe);
8091 I32 *nums = (I32*)SvPVX(sv_dat);
8092 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8093 if ((I32)(rx->lastparen) >= nums[i] &&
8094 rx->offs[nums[i]].start != -1 &&
8095 rx->offs[nums[i]].end != -1)
8101 if (parno || flags & RXapif_ALL) {
8102 av_push(av, newSVhek(HeKEY_hek(temphe)));
8107 return newRV_noinc(MUTABLE_SV(av));
8111 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8114 struct regexp *const rx = ReANY(r);
8120 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8122 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8123 || n == RX_BUFF_IDX_CARET_FULLMATCH
8124 || n == RX_BUFF_IDX_CARET_POSTMATCH
8127 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8129 /* on something like
8132 * the KEEPCOPY is set on the PMOP rather than the regex */
8133 if (PL_curpm && r == PM_GETRE(PL_curpm))
8134 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8143 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8144 /* no need to distinguish between them any more */
8145 n = RX_BUFF_IDX_FULLMATCH;
8147 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8148 && rx->offs[0].start != -1)
8150 /* $`, ${^PREMATCH} */
8151 i = rx->offs[0].start;
8155 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8156 && rx->offs[0].end != -1)
8158 /* $', ${^POSTMATCH} */
8159 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8160 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8163 if ( 0 <= n && n <= (I32)rx->nparens &&
8164 (s1 = rx->offs[n].start) != -1 &&
8165 (t1 = rx->offs[n].end) != -1)
8167 /* $&, ${^MATCH}, $1 ... */
8169 s = rx->subbeg + s1 - rx->suboffset;
8174 assert(s >= rx->subbeg);
8175 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8177 #ifdef NO_TAINT_SUPPORT
8178 sv_setpvn(sv, s, i);
8180 const int oldtainted = TAINT_get;
8182 sv_setpvn(sv, s, i);
8183 TAINT_set(oldtainted);
8185 if (RXp_MATCH_UTF8(rx))
8190 if (RXp_MATCH_TAINTED(rx)) {
8191 if (SvTYPE(sv) >= SVt_PVMG) {
8192 MAGIC* const mg = SvMAGIC(sv);
8195 SvMAGIC_set(sv, mg->mg_moremagic);
8197 if ((mgt = SvMAGIC(sv))) {
8198 mg->mg_moremagic = mgt;
8199 SvMAGIC_set(sv, mg);
8216 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8217 SV const * const value)
8219 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8221 PERL_UNUSED_ARG(rx);
8222 PERL_UNUSED_ARG(paren);
8223 PERL_UNUSED_ARG(value);
8226 Perl_croak_no_modify();
8230 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8233 struct regexp *const rx = ReANY(r);
8237 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8239 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8240 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8241 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8244 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8246 /* on something like
8249 * the KEEPCOPY is set on the PMOP rather than the regex */
8250 if (PL_curpm && r == PM_GETRE(PL_curpm))
8251 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8257 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8259 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8260 case RX_BUFF_IDX_PREMATCH: /* $` */
8261 if (rx->offs[0].start != -1) {
8262 i = rx->offs[0].start;
8271 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8272 case RX_BUFF_IDX_POSTMATCH: /* $' */
8273 if (rx->offs[0].end != -1) {
8274 i = rx->sublen - rx->offs[0].end;
8276 s1 = rx->offs[0].end;
8283 default: /* $& / ${^MATCH}, $1, $2, ... */
8284 if (paren <= (I32)rx->nparens &&
8285 (s1 = rx->offs[paren].start) != -1 &&
8286 (t1 = rx->offs[paren].end) != -1)
8292 if (ckWARN(WARN_UNINITIALIZED))
8293 report_uninit((const SV *)sv);
8298 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8299 const char * const s = rx->subbeg - rx->suboffset + s1;
8304 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8311 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8313 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8314 PERL_UNUSED_ARG(rx);
8318 return newSVpvs("Regexp");
8321 /* Scans the name of a named buffer from the pattern.
8322 * If flags is REG_RSN_RETURN_NULL returns null.
8323 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8324 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8325 * to the parsed name as looked up in the RExC_paren_names hash.
8326 * If there is an error throws a vFAIL().. type exception.
8329 #define REG_RSN_RETURN_NULL 0
8330 #define REG_RSN_RETURN_NAME 1
8331 #define REG_RSN_RETURN_DATA 2
8334 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8336 char *name_start = RExC_parse;
8338 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8340 assert (RExC_parse <= RExC_end);
8341 if (RExC_parse == RExC_end) NOOP;
8342 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8343 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8344 * using do...while */
8347 RExC_parse += UTF8SKIP(RExC_parse);
8348 } while ( RExC_parse < RExC_end
8349 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8353 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8355 RExC_parse++; /* so the <- from the vFAIL is after the offending
8357 vFAIL("Group name must start with a non-digit word character");
8361 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8362 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8363 if ( flags == REG_RSN_RETURN_NAME)
8365 else if (flags==REG_RSN_RETURN_DATA) {
8368 if ( ! sv_name ) /* should not happen*/
8369 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8370 if (RExC_paren_names)
8371 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8373 sv_dat = HeVAL(he_str);
8375 vFAIL("Reference to nonexistent named group");
8379 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
8380 (unsigned long) flags);
8382 NOT_REACHED; /* NOTREACHED */
8387 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
8389 if (RExC_lastparse!=RExC_parse) { \
8390 Perl_re_printf( aTHX_ "%s", \
8391 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
8392 RExC_end - RExC_parse, 16, \
8394 PERL_PV_ESCAPE_UNI_DETECT | \
8395 PERL_PV_PRETTY_ELLIPSES | \
8396 PERL_PV_PRETTY_LTGT | \
8397 PERL_PV_ESCAPE_RE | \
8398 PERL_PV_PRETTY_EXACTSIZE \
8402 Perl_re_printf( aTHX_ "%16s",""); \
8405 num = RExC_size + 1; \
8407 num=REG_NODE_NUM(RExC_emit); \
8408 if (RExC_lastnum!=num) \
8409 Perl_re_printf( aTHX_ "|%4d",num); \
8411 Perl_re_printf( aTHX_ "|%4s",""); \
8412 Perl_re_printf( aTHX_ "|%*s%-4s", \
8413 (int)((depth*2)), "", \
8417 RExC_lastparse=RExC_parse; \
8422 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8423 DEBUG_PARSE_MSG((funcname)); \
8424 Perl_re_printf( aTHX_ "%4s","\n"); \
8426 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
8427 DEBUG_PARSE_MSG((funcname)); \
8428 Perl_re_printf( aTHX_ fmt "\n",args); \
8431 /* This section of code defines the inversion list object and its methods. The
8432 * interfaces are highly subject to change, so as much as possible is static to
8433 * this file. An inversion list is here implemented as a malloc'd C UV array
8434 * as an SVt_INVLIST scalar.
8436 * An inversion list for Unicode is an array of code points, sorted by ordinal
8437 * number. Each element gives the code point that begins a range that extends
8438 * up-to but not including the code point given by the next element. The final
8439 * element gives the first code point of a range that extends to the platform's
8440 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
8441 * ...) give ranges whose code points are all in the inversion list. We say
8442 * that those ranges are in the set. The odd-numbered elements give ranges
8443 * whose code points are not in the inversion list, and hence not in the set.
8444 * Thus, element [0] is the first code point in the list. Element [1]
8445 * is the first code point beyond that not in the list; and element [2] is the
8446 * first code point beyond that that is in the list. In other words, the first
8447 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
8448 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
8449 * all code points in that range are not in the inversion list. The third
8450 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
8451 * list, and so forth. Thus every element whose index is divisible by two
8452 * gives the beginning of a range that is in the list, and every element whose
8453 * index is not divisible by two gives the beginning of a range not in the
8454 * list. If the final element's index is divisible by two, the inversion list
8455 * extends to the platform's infinity; otherwise the highest code point in the
8456 * inversion list is the contents of that element minus 1.
8458 * A range that contains just a single code point N will look like
8460 * invlist[i+1] == N+1
8462 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
8463 * impossible to represent, so element [i+1] is omitted. The single element
8465 * invlist[0] == UV_MAX
8466 * contains just UV_MAX, but is interpreted as matching to infinity.
8468 * Taking the complement (inverting) an inversion list is quite simple, if the
8469 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8470 * This implementation reserves an element at the beginning of each inversion
8471 * list to always contain 0; there is an additional flag in the header which
8472 * indicates if the list begins at the 0, or is offset to begin at the next
8473 * element. This means that the inversion list can be inverted without any
8474 * copying; just flip the flag.
8476 * More about inversion lists can be found in "Unicode Demystified"
8477 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8479 * The inversion list data structure is currently implemented as an SV pointing
8480 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8481 * array of UV whose memory management is automatically handled by the existing
8482 * facilities for SV's.
8484 * Some of the methods should always be private to the implementation, and some
8485 * should eventually be made public */
8487 /* The header definitions are in F<invlist_inline.h> */
8489 #ifndef PERL_IN_XSUB_RE
8491 PERL_STATIC_INLINE UV*
8492 S__invlist_array_init(SV* const invlist, const bool will_have_0)
8494 /* Returns a pointer to the first element in the inversion list's array.
8495 * This is called upon initialization of an inversion list. Where the
8496 * array begins depends on whether the list has the code point U+0000 in it
8497 * or not. The other parameter tells it whether the code that follows this
8498 * call is about to put a 0 in the inversion list or not. The first
8499 * element is either the element reserved for 0, if TRUE, or the element
8500 * after it, if FALSE */
8502 bool* offset = get_invlist_offset_addr(invlist);
8503 UV* zero_addr = (UV *) SvPVX(invlist);
8505 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
8508 assert(! _invlist_len(invlist));
8512 /* 1^1 = 0; 1^0 = 1 */
8513 *offset = 1 ^ will_have_0;
8514 return zero_addr + *offset;
8519 PERL_STATIC_INLINE void
8520 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8522 /* Sets the current number of elements stored in the inversion list.
8523 * Updates SvCUR correspondingly */
8524 PERL_UNUSED_CONTEXT;
8525 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8527 assert(SvTYPE(invlist) == SVt_INVLIST);
8532 : TO_INTERNAL_SIZE(len + offset));
8533 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
8536 #ifndef PERL_IN_XSUB_RE
8539 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
8541 /* Replaces the inversion list in 'dest' with the one from 'src'. It
8542 * steals the list from 'src', so 'src' is made to have a NULL list. This
8543 * is similar to what SvSetMagicSV() would do, if it were implemented on
8544 * inversion lists, though this routine avoids a copy */
8546 const UV src_len = _invlist_len(src);
8547 const bool src_offset = *get_invlist_offset_addr(src);
8548 const STRLEN src_byte_len = SvLEN(src);
8549 char * array = SvPVX(src);
8551 const int oldtainted = TAINT_get;
8553 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
8555 assert(SvTYPE(src) == SVt_INVLIST);
8556 assert(SvTYPE(dest) == SVt_INVLIST);
8557 assert(! invlist_is_iterating(src));
8558 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
8560 /* Make sure it ends in the right place with a NUL, as our inversion list
8561 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
8563 array[src_byte_len - 1] = '\0';
8565 TAINT_NOT; /* Otherwise it breaks */
8566 sv_usepvn_flags(dest,
8570 /* This flag is documented to cause a copy to be avoided */
8571 SV_HAS_TRAILING_NUL);
8572 TAINT_set(oldtainted);
8577 /* Finish up copying over the other fields in an inversion list */
8578 *get_invlist_offset_addr(dest) = src_offset;
8579 invlist_set_len(dest, src_len, src_offset);
8580 *get_invlist_previous_index_addr(dest) = 0;
8581 invlist_iterfinish(dest);
8584 PERL_STATIC_INLINE IV*
8585 S_get_invlist_previous_index_addr(SV* invlist)
8587 /* Return the address of the IV that is reserved to hold the cached index
8589 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
8591 assert(SvTYPE(invlist) == SVt_INVLIST);
8593 return &(((XINVLIST*) SvANY(invlist))->prev_index);
8596 PERL_STATIC_INLINE IV
8597 S_invlist_previous_index(SV* const invlist)
8599 /* Returns cached index of previous search */
8601 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
8603 return *get_invlist_previous_index_addr(invlist);
8606 PERL_STATIC_INLINE void
8607 S_invlist_set_previous_index(SV* const invlist, const IV index)
8609 /* Caches <index> for later retrieval */
8611 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
8613 assert(index == 0 || index < (int) _invlist_len(invlist));
8615 *get_invlist_previous_index_addr(invlist) = index;
8618 PERL_STATIC_INLINE void
8619 S_invlist_trim(SV* invlist)
8621 /* Free the not currently-being-used space in an inversion list */
8623 /* But don't free up the space needed for the 0 UV that is always at the
8624 * beginning of the list, nor the trailing NUL */
8625 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
8627 PERL_ARGS_ASSERT_INVLIST_TRIM;
8629 assert(SvTYPE(invlist) == SVt_INVLIST);
8631 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
8634 PERL_STATIC_INLINE void
8635 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
8637 PERL_ARGS_ASSERT_INVLIST_CLEAR;
8639 assert(SvTYPE(invlist) == SVt_INVLIST);
8641 invlist_set_len(invlist, 0, 0);
8642 invlist_trim(invlist);
8645 #endif /* ifndef PERL_IN_XSUB_RE */
8647 PERL_STATIC_INLINE bool
8648 S_invlist_is_iterating(SV* const invlist)
8650 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8652 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8655 #ifndef PERL_IN_XSUB_RE
8657 PERL_STATIC_INLINE UV
8658 S_invlist_max(SV* const invlist)
8660 /* Returns the maximum number of elements storable in the inversion list's
8661 * array, without having to realloc() */
8663 PERL_ARGS_ASSERT_INVLIST_MAX;
8665 assert(SvTYPE(invlist) == SVt_INVLIST);
8667 /* Assumes worst case, in which the 0 element is not counted in the
8668 * inversion list, so subtracts 1 for that */
8669 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
8670 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
8671 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
8674 Perl__new_invlist(pTHX_ IV initial_size)
8677 /* Return a pointer to a newly constructed inversion list, with enough
8678 * space to store 'initial_size' elements. If that number is negative, a
8679 * system default is used instead */
8683 if (initial_size < 0) {
8687 /* Allocate the initial space */
8688 new_list = newSV_type(SVt_INVLIST);
8690 /* First 1 is in case the zero element isn't in the list; second 1 is for
8692 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
8693 invlist_set_len(new_list, 0, 0);
8695 /* Force iterinit() to be used to get iteration to work */
8696 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
8698 *get_invlist_previous_index_addr(new_list) = 0;
8704 Perl__new_invlist_C_array(pTHX_ const UV* const list)
8706 /* Return a pointer to a newly constructed inversion list, initialized to
8707 * point to <list>, which has to be in the exact correct inversion list
8708 * form, including internal fields. Thus this is a dangerous routine that
8709 * should not be used in the wrong hands. The passed in 'list' contains
8710 * several header fields at the beginning that are not part of the
8711 * inversion list body proper */
8713 const STRLEN length = (STRLEN) list[0];
8714 const UV version_id = list[1];
8715 const bool offset = cBOOL(list[2]);
8716 #define HEADER_LENGTH 3
8717 /* If any of the above changes in any way, you must change HEADER_LENGTH
8718 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
8719 * perl -E 'say int(rand 2**31-1)'
8721 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
8722 data structure type, so that one being
8723 passed in can be validated to be an
8724 inversion list of the correct vintage.
8727 SV* invlist = newSV_type(SVt_INVLIST);
8729 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8731 if (version_id != INVLIST_VERSION_ID) {
8732 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8735 /* The generated array passed in includes header elements that aren't part
8736 * of the list proper, so start it just after them */
8737 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8739 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8740 shouldn't touch it */
8742 *(get_invlist_offset_addr(invlist)) = offset;
8744 /* The 'length' passed to us is the physical number of elements in the
8745 * inversion list. But if there is an offset the logical number is one
8747 invlist_set_len(invlist, length - offset, offset);
8749 invlist_set_previous_index(invlist, 0);
8751 /* Initialize the iteration pointer. */
8752 invlist_iterfinish(invlist);
8754 SvREADONLY_on(invlist);
8760 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8762 /* Grow the maximum size of an inversion list */
8764 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8766 assert(SvTYPE(invlist) == SVt_INVLIST);
8768 /* Add one to account for the zero element at the beginning which may not
8769 * be counted by the calling parameters */
8770 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8774 S__append_range_to_invlist(pTHX_ SV* const invlist,
8775 const UV start, const UV end)
8777 /* Subject to change or removal. Append the range from 'start' to 'end' at
8778 * the end of the inversion list. The range must be above any existing
8782 UV max = invlist_max(invlist);
8783 UV len = _invlist_len(invlist);
8786 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8788 if (len == 0) { /* Empty lists must be initialized */
8789 offset = start != 0;
8790 array = _invlist_array_init(invlist, ! offset);
8793 /* Here, the existing list is non-empty. The current max entry in the
8794 * list is generally the first value not in the set, except when the
8795 * set extends to the end of permissible values, in which case it is
8796 * the first entry in that final set, and so this call is an attempt to
8797 * append out-of-order */
8799 UV final_element = len - 1;
8800 array = invlist_array(invlist);
8801 if ( array[final_element] > start
8802 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8804 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",
8805 array[final_element], start,
8806 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8809 /* Here, it is a legal append. If the new range begins 1 above the end
8810 * of the range below it, it is extending the range below it, so the
8811 * new first value not in the set is one greater than the newly
8812 * extended range. */
8813 offset = *get_invlist_offset_addr(invlist);
8814 if (array[final_element] == start) {
8815 if (end != UV_MAX) {
8816 array[final_element] = end + 1;
8819 /* But if the end is the maximum representable on the machine,
8820 * assume that infinity was actually what was meant. Just let
8821 * the range that this would extend to have no end */
8822 invlist_set_len(invlist, len - 1, offset);
8828 /* Here the new range doesn't extend any existing set. Add it */
8830 len += 2; /* Includes an element each for the start and end of range */
8832 /* If wll overflow the existing space, extend, which may cause the array to
8835 invlist_extend(invlist, len);
8837 /* Have to set len here to avoid assert failure in invlist_array() */
8838 invlist_set_len(invlist, len, offset);
8840 array = invlist_array(invlist);
8843 invlist_set_len(invlist, len, offset);
8846 /* The next item on the list starts the range, the one after that is
8847 * one past the new range. */
8848 array[len - 2] = start;
8849 if (end != UV_MAX) {
8850 array[len - 1] = end + 1;
8853 /* But if the end is the maximum representable on the machine, just let
8854 * the range have no end */
8855 invlist_set_len(invlist, len - 1, offset);
8860 Perl__invlist_search(SV* const invlist, const UV cp)
8862 /* Searches the inversion list for the entry that contains the input code
8863 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8864 * return value is the index into the list's array of the range that
8865 * contains <cp>, that is, 'i' such that
8866 * array[i] <= cp < array[i+1]
8871 IV high = _invlist_len(invlist);
8872 const IV highest_element = high - 1;
8875 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8877 /* If list is empty, return failure. */
8882 /* (We can't get the array unless we know the list is non-empty) */
8883 array = invlist_array(invlist);
8885 mid = invlist_previous_index(invlist);
8887 if (mid > highest_element) {
8888 mid = highest_element;
8891 /* <mid> contains the cache of the result of the previous call to this
8892 * function (0 the first time). See if this call is for the same result,
8893 * or if it is for mid-1. This is under the theory that calls to this
8894 * function will often be for related code points that are near each other.
8895 * And benchmarks show that caching gives better results. We also test
8896 * here if the code point is within the bounds of the list. These tests
8897 * replace others that would have had to be made anyway to make sure that
8898 * the array bounds were not exceeded, and these give us extra information
8899 * at the same time */
8900 if (cp >= array[mid]) {
8901 if (cp >= array[highest_element]) {
8902 return highest_element;
8905 /* Here, array[mid] <= cp < array[highest_element]. This means that
8906 * the final element is not the answer, so can exclude it; it also
8907 * means that <mid> is not the final element, so can refer to 'mid + 1'
8909 if (cp < array[mid + 1]) {
8915 else { /* cp < aray[mid] */
8916 if (cp < array[0]) { /* Fail if outside the array */
8920 if (cp >= array[mid - 1]) {
8925 /* Binary search. What we are looking for is <i> such that
8926 * array[i] <= cp < array[i+1]
8927 * The loop below converges on the i+1. Note that there may not be an
8928 * (i+1)th element in the array, and things work nonetheless */
8929 while (low < high) {
8930 mid = (low + high) / 2;
8931 assert(mid <= highest_element);
8932 if (array[mid] <= cp) { /* cp >= array[mid] */
8935 /* We could do this extra test to exit the loop early.
8936 if (cp < array[low]) {
8941 else { /* cp < array[mid] */
8948 invlist_set_previous_index(invlist, high);
8953 Perl__invlist_populate_swatch(SV* const invlist,
8954 const UV start, const UV end, U8* swatch)
8956 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8957 * but is used when the swash has an inversion list. This makes this much
8958 * faster, as it uses a binary search instead of a linear one. This is
8959 * intimately tied to that function, and perhaps should be in utf8.c,
8960 * except it is intimately tied to inversion lists as well. It assumes
8961 * that <swatch> is all 0's on input */
8964 const IV len = _invlist_len(invlist);
8968 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8970 if (len == 0) { /* Empty inversion list */
8974 array = invlist_array(invlist);
8976 /* Find which element it is */
8977 i = _invlist_search(invlist, start);
8979 /* We populate from <start> to <end> */
8980 while (current < end) {
8983 /* The inversion list gives the results for every possible code point
8984 * after the first one in the list. Only those ranges whose index is
8985 * even are ones that the inversion list matches. For the odd ones,
8986 * and if the initial code point is not in the list, we have to skip
8987 * forward to the next element */
8988 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8990 if (i >= len) { /* Finished if beyond the end of the array */
8994 if (current >= end) { /* Finished if beyond the end of what we
8996 if (LIKELY(end < UV_MAX)) {
9000 /* We get here when the upper bound is the maximum
9001 * representable on the machine, and we are looking for just
9002 * that code point. Have to special case it */
9004 goto join_end_of_list;
9007 assert(current >= start);
9009 /* The current range ends one below the next one, except don't go past
9012 upper = (i < len && array[i] < end) ? array[i] : end;
9014 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
9015 * for each code point in it */
9016 for (; current < upper; current++) {
9017 const STRLEN offset = (STRLEN)(current - start);
9018 swatch[offset >> 3] |= 1 << (offset & 7);
9023 /* Quit if at the end of the list */
9026 /* But first, have to deal with the highest possible code point on
9027 * the platform. The previous code assumes that <end> is one
9028 * beyond where we want to populate, but that is impossible at the
9029 * platform's infinity, so have to handle it specially */
9030 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
9032 const STRLEN offset = (STRLEN)(end - start);
9033 swatch[offset >> 3] |= 1 << (offset & 7);
9038 /* Advance to the next range, which will be for code points not in the
9047 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9048 const bool complement_b, SV** output)
9050 /* Take the union of two inversion lists and point '*output' to it. On
9051 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9052 * even 'a' or 'b'). If to an inversion list, the contents of the original
9053 * list will be replaced by the union. The first list, 'a', may be
9054 * NULL, in which case a copy of the second list is placed in '*output'.
9055 * If 'complement_b' is TRUE, the union is taken of the complement
9056 * (inversion) of 'b' instead of b itself.
9058 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9059 * Richard Gillam, published by Addison-Wesley, and explained at some
9060 * length there. The preface says to incorporate its examples into your
9061 * code at your own risk.
9063 * The algorithm is like a merge sort. */
9065 const UV* array_a; /* a's array */
9067 UV len_a; /* length of a's array */
9070 SV* u; /* the resulting union */
9074 UV i_a = 0; /* current index into a's array */
9078 /* running count, as explained in the algorithm source book; items are
9079 * stopped accumulating and are output when the count changes to/from 0.
9080 * The count is incremented when we start a range that's in an input's set,
9081 * and decremented when we start a range that's not in a set. So this
9082 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9083 * and hence nothing goes into the union; 1, just one of the inputs is in
9084 * its set (and its current range gets added to the union); and 2 when both
9085 * inputs are in their sets. */
9088 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9090 assert(*output == NULL || SvTYPE(*output) == SVt_INVLIST);
9092 len_b = _invlist_len(b);
9095 /* Here, 'b' is empty, hence it's complement is all possible code
9096 * points. So if the union includes the complement of 'b', it includes
9097 * everything, and we need not even look at 'a'. It's easiest to
9098 * create a new inversion list that matches everything. */
9100 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9102 if (*output == NULL) { /* If the output didn't exist, just point it
9104 *output = everything;
9106 else { /* Otherwise, replace its contents with the new list */
9107 invlist_replace_list_destroys_src(*output, everything);
9108 SvREFCNT_dec_NN(everything);
9114 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9115 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9116 * output will be empty */
9118 if (a == NULL || _invlist_len(a) == 0) {
9119 if (*output == NULL) {
9120 *output = _new_invlist(0);
9123 invlist_clear(*output);
9128 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9129 * union. We can just return a copy of 'a' if '*output' doesn't point
9130 * to an existing list */
9131 if (*output == NULL) {
9132 *output = invlist_clone(a);
9136 /* If the output is to overwrite 'a', we have a no-op, as it's
9142 /* Here, '*output' is to be overwritten by 'a' */
9143 u = invlist_clone(a);
9144 invlist_replace_list_destroys_src(*output, u);
9150 /* Here 'b' is not empty. See about 'a' */
9152 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9154 /* Here, 'a' is empty (and b is not). That means the union will come
9155 * entirely from 'b'. If '*output' is NULL, we can directly return a
9156 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9159 SV ** dest = (*output == NULL) ? output : &u;
9160 *dest = invlist_clone(b);
9162 _invlist_invert(*dest);
9166 invlist_replace_list_destroys_src(*output, u);
9173 /* Here both lists exist and are non-empty */
9174 array_a = invlist_array(a);
9175 array_b = invlist_array(b);
9177 /* If are to take the union of 'a' with the complement of b, set it
9178 * up so are looking at b's complement. */
9181 /* To complement, we invert: if the first element is 0, remove it. To
9182 * do this, we just pretend the array starts one later */
9183 if (array_b[0] == 0) {
9189 /* But if the first element is not zero, we pretend the list starts
9190 * at the 0 that is always stored immediately before the array. */
9196 /* Size the union for the worst case: that the sets are completely
9198 u = _new_invlist(len_a + len_b);
9200 /* Will contain U+0000 if either component does */
9201 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9202 || (len_b > 0 && array_b[0] == 0));
9204 /* Go through each input list item by item, stopping when have exhausted
9206 while (i_a < len_a && i_b < len_b) {
9207 UV cp; /* The element to potentially add to the union's array */
9208 bool cp_in_set; /* is it in the the input list's set or not */
9210 /* We need to take one or the other of the two inputs for the union.
9211 * Since we are merging two sorted lists, we take the smaller of the
9212 * next items. In case of a tie, we take first the one that is in its
9213 * set. If we first took the one not in its set, it would decrement
9214 * the count, possibly to 0 which would cause it to be output as ending
9215 * the range, and the next time through we would take the same number,
9216 * and output it again as beginning the next range. By doing it the
9217 * opposite way, there is no possibility that the count will be
9218 * momentarily decremented to 0, and thus the two adjoining ranges will
9219 * be seamlessly merged. (In a tie and both are in the set or both not
9220 * in the set, it doesn't matter which we take first.) */
9221 if ( array_a[i_a] < array_b[i_b]
9222 || ( array_a[i_a] == array_b[i_b]
9223 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9225 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9226 cp = array_a[i_a++];
9229 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9230 cp = array_b[i_b++];
9233 /* Here, have chosen which of the two inputs to look at. Only output
9234 * if the running count changes to/from 0, which marks the
9235 * beginning/end of a range that's in the set */
9238 array_u[i_u++] = cp;
9245 array_u[i_u++] = cp;
9251 /* The loop above increments the index into exactly one of the input lists
9252 * each iteration, and ends when either index gets to its list end. That
9253 * means the other index is lower than its end, and so something is
9254 * remaining in that one. We decrement 'count', as explained below, if
9255 * that list is in its set. (i_a and i_b each currently index the element
9256 * beyond the one we care about.) */
9257 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9258 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9263 /* Above we decremented 'count' if the list that had unexamined elements in
9264 * it was in its set. This has made it so that 'count' being non-zero
9265 * means there isn't anything left to output; and 'count' equal to 0 means
9266 * that what is left to output is precisely that which is left in the
9267 * non-exhausted input list.
9269 * To see why, note first that the exhausted input obviously has nothing
9270 * left to add to the union. If it was in its set at its end, that means
9271 * the set extends from here to the platform's infinity, and hence so does
9272 * the union and the non-exhausted set is irrelevant. The exhausted set
9273 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9274 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9275 * 'count' remains at 1. This is consistent with the decremented 'count'
9276 * != 0 meaning there's nothing left to add to the union.
9278 * But if the exhausted input wasn't in its set, it contributed 0 to
9279 * 'count', and the rest of the union will be whatever the other input is.
9280 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9281 * otherwise it gets decremented to 0. This is consistent with 'count'
9282 * == 0 meaning the remainder of the union is whatever is left in the
9283 * non-exhausted list. */
9288 IV copy_count = len_a - i_a;
9289 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9290 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9292 else { /* The non-exhausted input is b */
9293 copy_count = len_b - i_b;
9294 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9296 len_u = i_u + copy_count;
9299 /* Set the result to the final length, which can change the pointer to
9300 * array_u, so re-find it. (Note that it is unlikely that this will
9301 * change, as we are shrinking the space, not enlarging it) */
9302 if (len_u != _invlist_len(u)) {
9303 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9305 array_u = invlist_array(u);
9308 if (*output == NULL) { /* Simply return the new inversion list */
9312 /* Otherwise, overwrite the inversion list that was in '*output'. We
9313 * could instead free '*output', and then set it to 'u', but experience
9314 * has shown [perl #127392] that if the input is a mortal, we can get a
9315 * huge build-up of these during regex compilation before they get
9317 invlist_replace_list_destroys_src(*output, u);
9325 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9326 const bool complement_b, SV** i)
9328 /* Take the intersection of two inversion lists and point '*i' to it. On
9329 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9330 * even 'a' or 'b'). If to an inversion list, the contents of the original
9331 * list will be replaced by the intersection. The first list, 'a', may be
9332 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9333 * TRUE, the result will be the intersection of 'a' and the complement (or
9334 * inversion) of 'b' instead of 'b' directly.
9336 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9337 * Richard Gillam, published by Addison-Wesley, and explained at some
9338 * length there. The preface says to incorporate its examples into your
9339 * code at your own risk. In fact, it had bugs
9341 * The algorithm is like a merge sort, and is essentially the same as the
9345 const UV* array_a; /* a's array */
9347 UV len_a; /* length of a's array */
9350 SV* r; /* the resulting intersection */
9354 UV i_a = 0; /* current index into a's array */
9358 /* running count of how many of the two inputs are postitioned at ranges
9359 * that are in their sets. As explained in the algorithm source book,
9360 * items are stopped accumulating and are output when the count changes
9361 * to/from 2. The count is incremented when we start a range that's in an
9362 * input's set, and decremented when we start a range that's not in a set.
9363 * Only when it is 2 are we in the intersection. */
9366 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9368 assert(*i == NULL || SvTYPE(*i) == SVt_INVLIST);
9370 /* Special case if either one is empty */
9371 len_a = (a == NULL) ? 0 : _invlist_len(a);
9372 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9373 if (len_a != 0 && complement_b) {
9375 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9376 * must be empty. Here, also we are using 'b's complement, which
9377 * hence must be every possible code point. Thus the intersection
9380 if (*i == a) { /* No-op */
9385 *i = invlist_clone(a);
9389 r = invlist_clone(a);
9390 invlist_replace_list_destroys_src(*i, r);
9395 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9396 * intersection must be empty */
9398 *i = _new_invlist(0);
9406 /* Here both lists exist and are non-empty */
9407 array_a = invlist_array(a);
9408 array_b = invlist_array(b);
9410 /* If are to take the intersection of 'a' with the complement of b, set it
9411 * up so are looking at b's complement. */
9414 /* To complement, we invert: if the first element is 0, remove it. To
9415 * do this, we just pretend the array starts one later */
9416 if (array_b[0] == 0) {
9422 /* But if the first element is not zero, we pretend the list starts
9423 * at the 0 that is always stored immediately before the array. */
9429 /* Size the intersection for the worst case: that the intersection ends up
9430 * fragmenting everything to be completely disjoint */
9431 r= _new_invlist(len_a + len_b);
9433 /* Will contain U+0000 iff both components do */
9434 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9435 && len_b > 0 && array_b[0] == 0);
9437 /* Go through each list item by item, stopping when have exhausted one of
9439 while (i_a < len_a && i_b < len_b) {
9440 UV cp; /* The element to potentially add to the intersection's
9442 bool cp_in_set; /* Is it in the input list's set or not */
9444 /* We need to take one or the other of the two inputs for the
9445 * intersection. Since we are merging two sorted lists, we take the
9446 * smaller of the next items. In case of a tie, we take first the one
9447 * that is not in its set (a difference from the union algorithm). If
9448 * we first took the one in its set, it would increment the count,
9449 * possibly to 2 which would cause it to be output as starting a range
9450 * in the intersection, and the next time through we would take that
9451 * same number, and output it again as ending the set. By doing the
9452 * opposite of this, there is no possibility that the count will be
9453 * momentarily incremented to 2. (In a tie and both are in the set or
9454 * both not in the set, it doesn't matter which we take first.) */
9455 if ( array_a[i_a] < array_b[i_b]
9456 || ( array_a[i_a] == array_b[i_b]
9457 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9459 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9460 cp = array_a[i_a++];
9463 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9467 /* Here, have chosen which of the two inputs to look at. Only output
9468 * if the running count changes to/from 2, which marks the
9469 * beginning/end of a range that's in the intersection */
9473 array_r[i_r++] = cp;
9478 array_r[i_r++] = cp;
9485 /* The loop above increments the index into exactly one of the input lists
9486 * each iteration, and ends when either index gets to its list end. That
9487 * means the other index is lower than its end, and so something is
9488 * remaining in that one. We increment 'count', as explained below, if the
9489 * exhausted list was in its set. (i_a and i_b each currently index the
9490 * element beyond the one we care about.) */
9491 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9492 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9497 /* Above we incremented 'count' if the exhausted list was in its set. This
9498 * has made it so that 'count' being below 2 means there is nothing left to
9499 * output; otheriwse what's left to add to the intersection is precisely
9500 * that which is left in the non-exhausted input list.
9502 * To see why, note first that the exhausted input obviously has nothing
9503 * left to affect the intersection. If it was in its set at its end, that
9504 * means the set extends from here to the platform's infinity, and hence
9505 * anything in the non-exhausted's list will be in the intersection, and
9506 * anything not in it won't be. Hence, the rest of the intersection is
9507 * precisely what's in the non-exhausted list The exhausted set also
9508 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9509 * it means 'count' is now at least 2. This is consistent with the
9510 * incremented 'count' being >= 2 means to add the non-exhausted list to
9513 * But if the exhausted input wasn't in its set, it contributed 0 to
9514 * 'count', and the intersection can't include anything further; the
9515 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9516 * incremented. This is consistent with 'count' being < 2 meaning nothing
9517 * further to add to the intersection. */
9518 if (count < 2) { /* Nothing left to put in the intersection. */
9521 else { /* copy the non-exhausted list, unchanged. */
9522 IV copy_count = len_a - i_a;
9523 if (copy_count > 0) { /* a is the one with stuff left */
9524 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9526 else { /* b is the one with stuff left */
9527 copy_count = len_b - i_b;
9528 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9530 len_r = i_r + copy_count;
9533 /* Set the result to the final length, which can change the pointer to
9534 * array_r, so re-find it. (Note that it is unlikely that this will
9535 * change, as we are shrinking the space, not enlarging it) */
9536 if (len_r != _invlist_len(r)) {
9537 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9539 array_r = invlist_array(r);
9542 if (*i == NULL) { /* Simply return the calculated intersection */
9545 else { /* Otherwise, replace the existing inversion list in '*i'. We could
9546 instead free '*i', and then set it to 'r', but experience has
9547 shown [perl #127392] that if the input is a mortal, we can get a
9548 huge build-up of these during regex compilation before they get
9551 invlist_replace_list_destroys_src(*i, r);
9563 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9565 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9566 * set. A pointer to the inversion list is returned. This may actually be
9567 * a new list, in which case the passed in one has been destroyed. The
9568 * passed-in inversion list can be NULL, in which case a new one is created
9569 * with just the one range in it. The new list is not necessarily
9570 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9571 * result of this function. The gain would not be large, and in many
9572 * cases, this is called multiple times on a single inversion list, so
9573 * anything freed may almost immediately be needed again.
9575 * This used to mostly call the 'union' routine, but that is much more
9576 * heavyweight than really needed for a single range addition */
9578 UV* array; /* The array implementing the inversion list */
9579 UV len; /* How many elements in 'array' */
9580 SSize_t i_s; /* index into the invlist array where 'start'
9582 SSize_t i_e = 0; /* And the index where 'end' should go */
9583 UV cur_highest; /* The highest code point in the inversion list
9584 upon entry to this function */
9586 /* This range becomes the whole inversion list if none already existed */
9587 if (invlist == NULL) {
9588 invlist = _new_invlist(2);
9589 _append_range_to_invlist(invlist, start, end);
9593 /* Likewise, if the inversion list is currently empty */
9594 len = _invlist_len(invlist);
9596 _append_range_to_invlist(invlist, start, end);
9600 /* Starting here, we have to know the internals of the list */
9601 array = invlist_array(invlist);
9603 /* If the new range ends higher than the current highest ... */
9604 cur_highest = invlist_highest(invlist);
9605 if (end > cur_highest) {
9607 /* If the whole range is higher, we can just append it */
9608 if (start > cur_highest) {
9609 _append_range_to_invlist(invlist, start, end);
9613 /* Otherwise, add the portion that is higher ... */
9614 _append_range_to_invlist(invlist, cur_highest + 1, end);
9616 /* ... and continue on below to handle the rest. As a result of the
9617 * above append, we know that the index of the end of the range is the
9618 * final even numbered one of the array. Recall that the final element
9619 * always starts a range that extends to infinity. If that range is in
9620 * the set (meaning the set goes from here to infinity), it will be an
9621 * even index, but if it isn't in the set, it's odd, and the final
9622 * range in the set is one less, which is even. */
9623 if (end == UV_MAX) {
9631 /* We have dealt with appending, now see about prepending. If the new
9632 * range starts lower than the current lowest ... */
9633 if (start < array[0]) {
9635 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
9636 * Let the union code handle it, rather than having to know the
9637 * trickiness in two code places. */
9638 if (UNLIKELY(start == 0)) {
9641 range_invlist = _new_invlist(2);
9642 _append_range_to_invlist(range_invlist, start, end);
9644 _invlist_union(invlist, range_invlist, &invlist);
9646 SvREFCNT_dec_NN(range_invlist);
9651 /* If the whole new range comes before the first entry, and doesn't
9652 * extend it, we have to insert it as an additional range */
9653 if (end < array[0] - 1) {
9655 goto splice_in_new_range;
9658 /* Here the new range adjoins the existing first range, extending it
9662 /* And continue on below to handle the rest. We know that the index of
9663 * the beginning of the range is the first one of the array */
9666 else { /* Not prepending any part of the new range to the existing list.
9667 * Find where in the list it should go. This finds i_s, such that:
9668 * invlist[i_s] <= start < array[i_s+1]
9670 i_s = _invlist_search(invlist, start);
9673 /* At this point, any extending before the beginning of the inversion list
9674 * and/or after the end has been done. This has made it so that, in the
9675 * code below, each endpoint of the new range is either in a range that is
9676 * in the set, or is in a gap between two ranges that are. This means we
9677 * don't have to worry about exceeding the array bounds.
9679 * Find where in the list the new range ends (but we can skip this if we
9680 * have already determined what it is, or if it will be the same as i_s,
9681 * which we already have computed) */
9683 i_e = (start == end)
9685 : _invlist_search(invlist, end);
9688 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
9689 * is a range that goes to infinity there is no element at invlist[i_e+1],
9690 * so only the first relation holds. */
9692 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9694 /* Here, the ranges on either side of the beginning of the new range
9695 * are in the set, and this range starts in the gap between them.
9697 * The new range extends the range above it downwards if the new range
9698 * ends at or above that range's start */
9699 const bool extends_the_range_above = ( end == UV_MAX
9700 || end + 1 >= array[i_s+1]);
9702 /* The new range extends the range below it upwards if it begins just
9703 * after where that range ends */
9704 if (start == array[i_s]) {
9706 /* If the new range fills the entire gap between the other ranges,
9707 * they will get merged together. Other ranges may also get
9708 * merged, depending on how many of them the new range spans. In
9709 * the general case, we do the merge later, just once, after we
9710 * figure out how many to merge. But in the case where the new
9711 * range exactly spans just this one gap (possibly extending into
9712 * the one above), we do the merge here, and an early exit. This
9713 * is done here to avoid having to special case later. */
9714 if (i_e - i_s <= 1) {
9716 /* If i_e - i_s == 1, it means that the new range terminates
9717 * within the range above, and hence 'extends_the_range_above'
9718 * must be true. (If the range above it extends to infinity,
9719 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
9720 * will be 0, so no harm done.) */
9721 if (extends_the_range_above) {
9722 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
9723 invlist_set_len(invlist,
9725 *(get_invlist_offset_addr(invlist)));
9729 /* Here, i_e must == i_s. We keep them in sync, as they apply
9730 * to the same range, and below we are about to decrement i_s
9735 /* Here, the new range is adjacent to the one below. (It may also
9736 * span beyond the range above, but that will get resolved later.)
9737 * Extend the range below to include this one. */
9738 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
9742 else if (extends_the_range_above) {
9744 /* Here the new range only extends the range above it, but not the
9745 * one below. It merges with the one above. Again, we keep i_e
9746 * and i_s in sync if they point to the same range */
9755 /* Here, we've dealt with the new range start extending any adjoining
9758 * If the new range extends to infinity, it is now the final one,
9759 * regardless of what was there before */
9760 if (UNLIKELY(end == UV_MAX)) {
9761 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
9765 /* If i_e started as == i_s, it has also been dealt with,
9766 * and been updated to the new i_s, which will fail the following if */
9767 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
9769 /* Here, the ranges on either side of the end of the new range are in
9770 * the set, and this range ends in the gap between them.
9772 * If this range is adjacent to (hence extends) the range above it, it
9773 * becomes part of that range; likewise if it extends the range below,
9774 * it becomes part of that range */
9775 if (end + 1 == array[i_e+1]) {
9779 else if (start <= array[i_e]) {
9780 array[i_e] = end + 1;
9787 /* If the range fits entirely in an existing range (as possibly already
9788 * extended above), it doesn't add anything new */
9789 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
9793 /* Here, no part of the range is in the list. Must add it. It will
9794 * occupy 2 more slots */
9795 splice_in_new_range:
9797 invlist_extend(invlist, len + 2);
9798 array = invlist_array(invlist);
9799 /* Move the rest of the array down two slots. Don't include any
9801 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
9803 /* Do the actual splice */
9804 array[i_e+1] = start;
9805 array[i_e+2] = end + 1;
9806 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
9810 /* Here the new range crossed the boundaries of a pre-existing range. The
9811 * code above has adjusted things so that both ends are in ranges that are
9812 * in the set. This means everything in between must also be in the set.
9813 * Just squash things together */
9814 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
9815 invlist_set_len(invlist,
9817 *(get_invlist_offset_addr(invlist)));
9823 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9824 UV** other_elements_ptr)
9826 /* Create and return an inversion list whose contents are to be populated
9827 * by the caller. The caller gives the number of elements (in 'size') and
9828 * the very first element ('element0'). This function will set
9829 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9832 * Obviously there is some trust involved that the caller will properly
9833 * fill in the other elements of the array.
9835 * (The first element needs to be passed in, as the underlying code does
9836 * things differently depending on whether it is zero or non-zero) */
9838 SV* invlist = _new_invlist(size);
9841 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9843 invlist = add_cp_to_invlist(invlist, element0);
9844 offset = *get_invlist_offset_addr(invlist);
9846 invlist_set_len(invlist, size, offset);
9847 *other_elements_ptr = invlist_array(invlist) + 1;
9853 PERL_STATIC_INLINE SV*
9854 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9855 return _add_range_to_invlist(invlist, cp, cp);
9858 #ifndef PERL_IN_XSUB_RE
9860 Perl__invlist_invert(pTHX_ SV* const invlist)
9862 /* Complement the input inversion list. This adds a 0 if the list didn't
9863 * have a zero; removes it otherwise. As described above, the data
9864 * structure is set up so that this is very efficient */
9866 PERL_ARGS_ASSERT__INVLIST_INVERT;
9868 assert(! invlist_is_iterating(invlist));
9870 /* The inverse of matching nothing is matching everything */
9871 if (_invlist_len(invlist) == 0) {
9872 _append_range_to_invlist(invlist, 0, UV_MAX);
9876 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9881 PERL_STATIC_INLINE SV*
9882 S_invlist_clone(pTHX_ SV* const invlist)
9885 /* Return a new inversion list that is a copy of the input one, which is
9886 * unchanged. The new list will not be mortal even if the old one was. */
9888 /* Need to allocate extra space to accommodate Perl's addition of a
9889 * trailing NUL to SvPV's, since it thinks they are always strings */
9890 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9891 STRLEN physical_length = SvCUR(invlist);
9892 bool offset = *(get_invlist_offset_addr(invlist));
9894 PERL_ARGS_ASSERT_INVLIST_CLONE;
9896 *(get_invlist_offset_addr(new_invlist)) = offset;
9897 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9898 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9903 PERL_STATIC_INLINE STRLEN*
9904 S_get_invlist_iter_addr(SV* invlist)
9906 /* Return the address of the UV that contains the current iteration
9909 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9911 assert(SvTYPE(invlist) == SVt_INVLIST);
9913 return &(((XINVLIST*) SvANY(invlist))->iterator);
9916 PERL_STATIC_INLINE void
9917 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9919 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9921 *get_invlist_iter_addr(invlist) = 0;
9924 PERL_STATIC_INLINE void
9925 S_invlist_iterfinish(SV* invlist)
9927 /* Terminate iterator for invlist. This is to catch development errors.
9928 * Any iteration that is interrupted before completed should call this
9929 * function. Functions that add code points anywhere else but to the end
9930 * of an inversion list assert that they are not in the middle of an
9931 * iteration. If they were, the addition would make the iteration
9932 * problematical: if the iteration hadn't reached the place where things
9933 * were being added, it would be ok */
9935 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
9937 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
9941 S_invlist_iternext(SV* invlist, UV* start, UV* end)
9943 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
9944 * This call sets in <*start> and <*end>, the next range in <invlist>.
9945 * Returns <TRUE> if successful and the next call will return the next
9946 * range; <FALSE> if was already at the end of the list. If the latter,
9947 * <*start> and <*end> are unchanged, and the next call to this function
9948 * will start over at the beginning of the list */
9950 STRLEN* pos = get_invlist_iter_addr(invlist);
9951 UV len = _invlist_len(invlist);
9954 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9957 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9961 array = invlist_array(invlist);
9963 *start = array[(*pos)++];
9969 *end = array[(*pos)++] - 1;
9975 PERL_STATIC_INLINE UV
9976 S_invlist_highest(SV* const invlist)
9978 /* Returns the highest code point that matches an inversion list. This API
9979 * has an ambiguity, as it returns 0 under either the highest is actually
9980 * 0, or if the list is empty. If this distinction matters to you, check
9981 * for emptiness before calling this function */
9983 UV len = _invlist_len(invlist);
9986 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
9992 array = invlist_array(invlist);
9994 /* The last element in the array in the inversion list always starts a
9995 * range that goes to infinity. That range may be for code points that are
9996 * matched in the inversion list, or it may be for ones that aren't
9997 * matched. In the latter case, the highest code point in the set is one
9998 * less than the beginning of this range; otherwise it is the final element
9999 * of this range: infinity */
10000 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
10002 : array[len - 1] - 1;
10006 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
10008 /* Get the contents of an inversion list into a string SV so that they can
10009 * be printed out. If 'traditional_style' is TRUE, it uses the format
10010 * traditionally done for debug tracing; otherwise it uses a format
10011 * suitable for just copying to the output, with blanks between ranges and
10012 * a dash between range components */
10016 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
10017 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
10019 if (traditional_style) {
10020 output = newSVpvs("\n");
10023 output = newSVpvs("");
10026 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
10028 assert(! invlist_is_iterating(invlist));
10030 invlist_iterinit(invlist);
10031 while (invlist_iternext(invlist, &start, &end)) {
10032 if (end == UV_MAX) {
10033 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFINITY%c",
10034 start, intra_range_delimiter,
10035 inter_range_delimiter);
10037 else if (end != start) {
10038 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
10040 intra_range_delimiter,
10041 end, inter_range_delimiter);
10044 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10045 start, inter_range_delimiter);
10049 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10050 SvCUR_set(output, SvCUR(output) - 1);
10056 #ifndef PERL_IN_XSUB_RE
10058 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10059 const char * const indent, SV* const invlist)
10061 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10062 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10063 * the string 'indent'. The output looks like this:
10064 [0] 0x000A .. 0x000D
10066 [4] 0x2028 .. 0x2029
10067 [6] 0x3104 .. INFINITY
10068 * This means that the first range of code points matched by the list are
10069 * 0xA through 0xD; the second range contains only the single code point
10070 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10071 * are used to define each range (except if the final range extends to
10072 * infinity, only a single element is needed). The array index of the
10073 * first element for the corresponding range is given in brackets. */
10078 PERL_ARGS_ASSERT__INVLIST_DUMP;
10080 if (invlist_is_iterating(invlist)) {
10081 Perl_dump_indent(aTHX_ level, file,
10082 "%sCan't dump inversion list because is in middle of iterating\n",
10087 invlist_iterinit(invlist);
10088 while (invlist_iternext(invlist, &start, &end)) {
10089 if (end == UV_MAX) {
10090 Perl_dump_indent(aTHX_ level, file,
10091 "%s[%" UVuf "] 0x%04" UVXf " .. INFINITY\n",
10092 indent, (UV)count, start);
10094 else if (end != start) {
10095 Perl_dump_indent(aTHX_ level, file,
10096 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10097 indent, (UV)count, start, end);
10100 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10101 indent, (UV)count, start);
10108 Perl__load_PL_utf8_foldclosures (pTHX)
10110 assert(! PL_utf8_foldclosures);
10112 /* If the folds haven't been read in, call a fold function
10114 if (! PL_utf8_tofold) {
10115 U8 dummy[UTF8_MAXBYTES_CASE+1];
10116 const U8 hyphen[] = HYPHEN_UTF8;
10118 /* This string is just a short named one above \xff */
10119 toFOLD_utf8_safe(hyphen, hyphen + sizeof(hyphen) - 1, dummy, NULL);
10120 assert(PL_utf8_tofold); /* Verify that worked */
10122 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
10126 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10128 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10130 /* Return a boolean as to if the two passed in inversion lists are
10131 * identical. The final argument, if TRUE, says to take the complement of
10132 * the second inversion list before doing the comparison */
10134 const UV* array_a = invlist_array(a);
10135 const UV* array_b = invlist_array(b);
10136 UV len_a = _invlist_len(a);
10137 UV len_b = _invlist_len(b);
10139 PERL_ARGS_ASSERT__INVLISTEQ;
10141 /* If are to compare 'a' with the complement of b, set it
10142 * up so are looking at b's complement. */
10143 if (complement_b) {
10145 /* The complement of nothing is everything, so <a> would have to have
10146 * just one element, starting at zero (ending at infinity) */
10148 return (len_a == 1 && array_a[0] == 0);
10150 else if (array_b[0] == 0) {
10152 /* Otherwise, to complement, we invert. Here, the first element is
10153 * 0, just remove it. To do this, we just pretend the array starts
10161 /* But if the first element is not zero, we pretend the list starts
10162 * at the 0 that is always stored immediately before the array. */
10168 return len_a == len_b
10169 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10175 * As best we can, determine the characters that can match the start of
10176 * the given EXACTF-ish node.
10178 * Returns the invlist as a new SV*; it is the caller's responsibility to
10179 * call SvREFCNT_dec() when done with it.
10182 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10184 const U8 * s = (U8*)STRING(node);
10185 SSize_t bytelen = STR_LEN(node);
10187 /* Start out big enough for 2 separate code points */
10188 SV* invlist = _new_invlist(4);
10190 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
10195 /* We punt and assume can match anything if the node begins
10196 * with a multi-character fold. Things are complicated. For
10197 * example, /ffi/i could match any of:
10198 * "\N{LATIN SMALL LIGATURE FFI}"
10199 * "\N{LATIN SMALL LIGATURE FF}I"
10200 * "F\N{LATIN SMALL LIGATURE FI}"
10201 * plus several other things; and making sure we have all the
10202 * possibilities is hard. */
10203 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10204 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10207 /* Any Latin1 range character can potentially match any
10208 * other depending on the locale */
10209 if (OP(node) == EXACTFL) {
10210 _invlist_union(invlist, PL_Latin1, &invlist);
10213 /* But otherwise, it matches at least itself. We can
10214 * quickly tell if it has a distinct fold, and if so,
10215 * it matches that as well */
10216 invlist = add_cp_to_invlist(invlist, uc);
10217 if (IS_IN_SOME_FOLD_L1(uc))
10218 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10221 /* Some characters match above-Latin1 ones under /i. This
10222 * is true of EXACTFL ones when the locale is UTF-8 */
10223 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10224 && (! isASCII(uc) || (OP(node) != EXACTFA
10225 && OP(node) != EXACTFA_NO_TRIE)))
10227 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10231 else { /* Pattern is UTF-8 */
10232 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10233 STRLEN foldlen = UTF8SKIP(s);
10234 const U8* e = s + bytelen;
10237 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10239 /* The only code points that aren't folded in a UTF EXACTFish
10240 * node are are the problematic ones in EXACTFL nodes */
10241 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10242 /* We need to check for the possibility that this EXACTFL
10243 * node begins with a multi-char fold. Therefore we fold
10244 * the first few characters of it so that we can make that
10249 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10251 *(d++) = (U8) toFOLD(*s);
10256 toFOLD_utf8_safe(s, e, d, &len);
10262 /* And set up so the code below that looks in this folded
10263 * buffer instead of the node's string */
10265 foldlen = UTF8SKIP(folded);
10269 /* When we reach here 's' points to the fold of the first
10270 * character(s) of the node; and 'e' points to far enough along
10271 * the folded string to be just past any possible multi-char
10272 * fold. 'foldlen' is the length in bytes of the first
10275 * Unlike the non-UTF-8 case, the macro for determining if a
10276 * string is a multi-char fold requires all the characters to
10277 * already be folded. This is because of all the complications
10278 * if not. Note that they are folded anyway, except in EXACTFL
10279 * nodes. Like the non-UTF case above, we punt if the node
10280 * begins with a multi-char fold */
10282 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10283 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10285 else { /* Single char fold */
10287 /* It matches all the things that fold to it, which are
10288 * found in PL_utf8_foldclosures (including itself) */
10289 invlist = add_cp_to_invlist(invlist, uc);
10290 if (! PL_utf8_foldclosures)
10291 _load_PL_utf8_foldclosures();
10292 if ((listp = hv_fetch(PL_utf8_foldclosures,
10293 (char *) s, foldlen, FALSE)))
10295 AV* list = (AV*) *listp;
10297 for (k = 0; k <= av_tindex_skip_len_mg(list); k++) {
10298 SV** c_p = av_fetch(list, k, FALSE);
10304 /* /aa doesn't allow folds between ASCII and non- */
10305 if ((OP(node) == EXACTFA || OP(node) == EXACTFA_NO_TRIE)
10306 && isASCII(c) != isASCII(uc))
10311 invlist = add_cp_to_invlist(invlist, c);
10320 #undef HEADER_LENGTH
10321 #undef TO_INTERNAL_SIZE
10322 #undef FROM_INTERNAL_SIZE
10323 #undef INVLIST_VERSION_ID
10325 /* End of inversion list object */
10328 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10330 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10331 * constructs, and updates RExC_flags with them. On input, RExC_parse
10332 * should point to the first flag; it is updated on output to point to the
10333 * final ')' or ':'. There needs to be at least one flag, or this will
10336 /* for (?g), (?gc), and (?o) warnings; warning
10337 about (?c) will warn about (?g) -- japhy */
10339 #define WASTED_O 0x01
10340 #define WASTED_G 0x02
10341 #define WASTED_C 0x04
10342 #define WASTED_GC (WASTED_G|WASTED_C)
10343 I32 wastedflags = 0x00;
10344 U32 posflags = 0, negflags = 0;
10345 U32 *flagsp = &posflags;
10346 char has_charset_modifier = '\0';
10348 bool has_use_defaults = FALSE;
10349 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10350 int x_mod_count = 0;
10352 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10354 /* '^' as an initial flag sets certain defaults */
10355 if (UCHARAT(RExC_parse) == '^') {
10357 has_use_defaults = TRUE;
10358 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10359 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
10360 ? REGEX_UNICODE_CHARSET
10361 : REGEX_DEPENDS_CHARSET);
10364 cs = get_regex_charset(RExC_flags);
10365 if (cs == REGEX_DEPENDS_CHARSET
10366 && (RExC_utf8 || RExC_uni_semantics))
10368 cs = REGEX_UNICODE_CHARSET;
10371 while (RExC_parse < RExC_end) {
10372 /* && strchr("iogcmsx", *RExC_parse) */
10373 /* (?g), (?gc) and (?o) are useless here
10374 and must be globally applied -- japhy */
10375 switch (*RExC_parse) {
10377 /* Code for the imsxn flags */
10378 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10380 case LOCALE_PAT_MOD:
10381 if (has_charset_modifier) {
10382 goto excess_modifier;
10384 else if (flagsp == &negflags) {
10387 cs = REGEX_LOCALE_CHARSET;
10388 has_charset_modifier = LOCALE_PAT_MOD;
10390 case UNICODE_PAT_MOD:
10391 if (has_charset_modifier) {
10392 goto excess_modifier;
10394 else if (flagsp == &negflags) {
10397 cs = REGEX_UNICODE_CHARSET;
10398 has_charset_modifier = UNICODE_PAT_MOD;
10400 case ASCII_RESTRICT_PAT_MOD:
10401 if (flagsp == &negflags) {
10404 if (has_charset_modifier) {
10405 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10406 goto excess_modifier;
10408 /* Doubled modifier implies more restricted */
10409 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10412 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10414 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10416 case DEPENDS_PAT_MOD:
10417 if (has_use_defaults) {
10418 goto fail_modifiers;
10420 else if (flagsp == &negflags) {
10423 else if (has_charset_modifier) {
10424 goto excess_modifier;
10427 /* The dual charset means unicode semantics if the
10428 * pattern (or target, not known until runtime) are
10429 * utf8, or something in the pattern indicates unicode
10431 cs = (RExC_utf8 || RExC_uni_semantics)
10432 ? REGEX_UNICODE_CHARSET
10433 : REGEX_DEPENDS_CHARSET;
10434 has_charset_modifier = DEPENDS_PAT_MOD;
10438 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10439 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10441 else if (has_charset_modifier == *(RExC_parse - 1)) {
10442 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10443 *(RExC_parse - 1));
10446 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10448 NOT_REACHED; /*NOTREACHED*/
10451 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10452 *(RExC_parse - 1));
10453 NOT_REACHED; /*NOTREACHED*/
10454 case ONCE_PAT_MOD: /* 'o' */
10455 case GLOBAL_PAT_MOD: /* 'g' */
10456 if (PASS2 && ckWARN(WARN_REGEXP)) {
10457 const I32 wflagbit = *RExC_parse == 'o'
10460 if (! (wastedflags & wflagbit) ) {
10461 wastedflags |= wflagbit;
10462 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10465 "Useless (%s%c) - %suse /%c modifier",
10466 flagsp == &negflags ? "?-" : "?",
10468 flagsp == &negflags ? "don't " : "",
10475 case CONTINUE_PAT_MOD: /* 'c' */
10476 if (PASS2 && ckWARN(WARN_REGEXP)) {
10477 if (! (wastedflags & WASTED_C) ) {
10478 wastedflags |= WASTED_GC;
10479 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10482 "Useless (%sc) - %suse /gc modifier",
10483 flagsp == &negflags ? "?-" : "?",
10484 flagsp == &negflags ? "don't " : ""
10489 case KEEPCOPY_PAT_MOD: /* 'p' */
10490 if (flagsp == &negflags) {
10492 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10494 *flagsp |= RXf_PMf_KEEPCOPY;
10498 /* A flag is a default iff it is following a minus, so
10499 * if there is a minus, it means will be trying to
10500 * re-specify a default which is an error */
10501 if (has_use_defaults || flagsp == &negflags) {
10502 goto fail_modifiers;
10504 flagsp = &negflags;
10505 wastedflags = 0; /* reset so (?g-c) warns twice */
10511 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
10512 negflags |= RXf_PMf_EXTENDED_MORE;
10514 RExC_flags |= posflags;
10516 if (negflags & RXf_PMf_EXTENDED) {
10517 negflags |= RXf_PMf_EXTENDED_MORE;
10519 RExC_flags &= ~negflags;
10520 set_regex_charset(&RExC_flags, cs);
10525 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10526 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10527 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
10528 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10529 NOT_REACHED; /*NOTREACHED*/
10532 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10535 vFAIL("Sequence (?... not terminated");
10539 - reg - regular expression, i.e. main body or parenthesized thing
10541 * Caller must absorb opening parenthesis.
10543 * Combining parenthesis handling with the base level of regular expression
10544 * is a trifle forced, but the need to tie the tails of the branches to what
10545 * follows makes it hard to avoid.
10547 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10549 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10551 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10554 PERL_STATIC_INLINE regnode *
10555 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10557 char * parse_start,
10562 char* name_start = RExC_parse;
10564 SV *sv_dat = reg_scan_name(pRExC_state, SIZE_ONLY
10565 ? REG_RSN_RETURN_NULL
10566 : REG_RSN_RETURN_DATA);
10567 GET_RE_DEBUG_FLAGS_DECL;
10569 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10571 if (RExC_parse == name_start || *RExC_parse != ch) {
10572 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10573 vFAIL2("Sequence %.3s... not terminated",parse_start);
10577 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10578 RExC_rxi->data->data[num]=(void*)sv_dat;
10579 SvREFCNT_inc_simple_void(sv_dat);
10582 ret = reganode(pRExC_state,
10585 : (ASCII_FOLD_RESTRICTED)
10587 : (AT_LEAST_UNI_SEMANTICS)
10593 *flagp |= HASWIDTH;
10595 Set_Node_Offset(ret, parse_start+1);
10596 Set_Node_Cur_Length(ret, parse_start);
10598 nextchar(pRExC_state);
10602 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
10603 flags. Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan
10604 needs to be restarted, or'd with NEED_UTF8 if the pattern needs to be
10605 upgraded to UTF-8. Otherwise would only return NULL if regbranch() returns
10606 NULL, which cannot happen. */
10608 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
10609 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
10610 * 2 is like 1, but indicates that nextchar() has been called to advance
10611 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
10612 * this flag alerts us to the need to check for that */
10614 regnode *ret; /* Will be the head of the group. */
10617 regnode *ender = NULL;
10620 U32 oregflags = RExC_flags;
10621 bool have_branch = 0;
10623 I32 freeze_paren = 0;
10624 I32 after_freeze = 0;
10625 I32 num; /* numeric backreferences */
10627 char * parse_start = RExC_parse; /* MJD */
10628 char * const oregcomp_parse = RExC_parse;
10630 GET_RE_DEBUG_FLAGS_DECL;
10632 PERL_ARGS_ASSERT_REG;
10633 DEBUG_PARSE("reg ");
10635 *flagp = 0; /* Tentatively. */
10637 /* Having this true makes it feasible to have a lot fewer tests for the
10638 * parse pointer being in scope. For example, we can write
10639 * while(isFOO(*RExC_parse)) RExC_parse++;
10641 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
10643 assert(*RExC_end == '\0');
10645 /* Make an OPEN node, if parenthesized. */
10648 /* Under /x, space and comments can be gobbled up between the '(' and
10649 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
10650 * intervening space, as the sequence is a token, and a token should be
10652 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
10654 if (RExC_parse >= RExC_end) {
10655 vFAIL("Unmatched (");
10658 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
10659 char *start_verb = RExC_parse + 1;
10661 char *start_arg = NULL;
10662 unsigned char op = 0;
10663 int arg_required = 0;
10664 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
10666 if (has_intervening_patws) {
10667 RExC_parse++; /* past the '*' */
10668 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
10670 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
10671 if ( *RExC_parse == ':' ) {
10672 start_arg = RExC_parse + 1;
10675 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10677 verb_len = RExC_parse - start_verb;
10679 if (RExC_parse >= RExC_end) {
10680 goto unterminated_verb_pattern;
10682 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10683 while ( RExC_parse < RExC_end && *RExC_parse != ')' )
10684 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10685 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10686 unterminated_verb_pattern:
10687 vFAIL("Unterminated verb pattern argument");
10688 if ( RExC_parse == start_arg )
10691 if ( RExC_parse >= RExC_end || *RExC_parse != ')' )
10692 vFAIL("Unterminated verb pattern");
10695 /* Here, we know that RExC_parse < RExC_end */
10697 switch ( *start_verb ) {
10698 case 'A': /* (*ACCEPT) */
10699 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
10701 internal_argval = RExC_nestroot;
10704 case 'C': /* (*COMMIT) */
10705 if ( memEQs(start_verb,verb_len,"COMMIT") )
10708 case 'F': /* (*FAIL) */
10709 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
10713 case ':': /* (*:NAME) */
10714 case 'M': /* (*MARK:NAME) */
10715 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
10720 case 'P': /* (*PRUNE) */
10721 if ( memEQs(start_verb,verb_len,"PRUNE") )
10724 case 'S': /* (*SKIP) */
10725 if ( memEQs(start_verb,verb_len,"SKIP") )
10728 case 'T': /* (*THEN) */
10729 /* [19:06] <TimToady> :: is then */
10730 if ( memEQs(start_verb,verb_len,"THEN") ) {
10732 RExC_seen |= REG_CUTGROUP_SEEN;
10737 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10739 "Unknown verb pattern '%" UTF8f "'",
10740 UTF8fARG(UTF, verb_len, start_verb));
10742 if ( arg_required && !start_arg ) {
10743 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
10744 verb_len, start_verb);
10746 if (internal_argval == -1) {
10747 ret = reganode(pRExC_state, op, 0);
10749 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
10751 RExC_seen |= REG_VERBARG_SEEN;
10752 if ( ! SIZE_ONLY ) {
10754 SV *sv = newSVpvn( start_arg,
10755 RExC_parse - start_arg);
10756 ARG(ret) = add_data( pRExC_state,
10757 STR_WITH_LEN("S"));
10758 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
10763 if ( internal_argval != -1 )
10764 ARG2L_SET(ret, internal_argval);
10766 nextchar(pRExC_state);
10769 else if (*RExC_parse == '?') { /* (?...) */
10770 bool is_logical = 0;
10771 const char * const seqstart = RExC_parse;
10772 const char * endptr;
10773 if (has_intervening_patws) {
10775 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
10778 RExC_parse++; /* past the '?' */
10779 paren = *RExC_parse; /* might be a trailing NUL, if not
10781 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10782 if (RExC_parse > RExC_end) {
10785 ret = NULL; /* For look-ahead/behind. */
10788 case 'P': /* (?P...) variants for those used to PCRE/Python */
10789 paren = *RExC_parse;
10790 if ( paren == '<') { /* (?P<...>) named capture */
10792 if (RExC_parse >= RExC_end) {
10793 vFAIL("Sequence (?P<... not terminated");
10795 goto named_capture;
10797 else if (paren == '>') { /* (?P>name) named recursion */
10799 if (RExC_parse >= RExC_end) {
10800 vFAIL("Sequence (?P>... not terminated");
10802 goto named_recursion;
10804 else if (paren == '=') { /* (?P=...) named backref */
10806 return handle_named_backref(pRExC_state, flagp,
10809 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10810 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10811 vFAIL3("Sequence (%.*s...) not recognized",
10812 RExC_parse-seqstart, seqstart);
10813 NOT_REACHED; /*NOTREACHED*/
10814 case '<': /* (?<...) */
10815 if (*RExC_parse == '!')
10817 else if (*RExC_parse != '=')
10824 case '\'': /* (?'...') */
10825 name_start = RExC_parse;
10826 svname = reg_scan_name(pRExC_state,
10827 SIZE_ONLY /* reverse test from the others */
10828 ? REG_RSN_RETURN_NAME
10829 : REG_RSN_RETURN_NULL);
10830 if ( RExC_parse == name_start
10831 || RExC_parse >= RExC_end
10832 || *RExC_parse != paren)
10834 vFAIL2("Sequence (?%c... not terminated",
10835 paren=='>' ? '<' : paren);
10840 if (!svname) /* shouldn't happen */
10842 "panic: reg_scan_name returned NULL");
10843 if (!RExC_paren_names) {
10844 RExC_paren_names= newHV();
10845 sv_2mortal(MUTABLE_SV(RExC_paren_names));
10847 RExC_paren_name_list= newAV();
10848 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
10851 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
10853 sv_dat = HeVAL(he_str);
10855 /* croak baby croak */
10857 "panic: paren_name hash element allocation failed");
10858 } else if ( SvPOK(sv_dat) ) {
10859 /* (?|...) can mean we have dupes so scan to check
10860 its already been stored. Maybe a flag indicating
10861 we are inside such a construct would be useful,
10862 but the arrays are likely to be quite small, so
10863 for now we punt -- dmq */
10864 IV count = SvIV(sv_dat);
10865 I32 *pv = (I32*)SvPVX(sv_dat);
10867 for ( i = 0 ; i < count ; i++ ) {
10868 if ( pv[i] == RExC_npar ) {
10874 pv = (I32*)SvGROW(sv_dat,
10875 SvCUR(sv_dat) + sizeof(I32)+1);
10876 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
10877 pv[count] = RExC_npar;
10878 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
10881 (void)SvUPGRADE(sv_dat,SVt_PVNV);
10882 sv_setpvn(sv_dat, (char *)&(RExC_npar),
10885 SvIV_set(sv_dat, 1);
10888 /* Yes this does cause a memory leak in debugging Perls
10890 if (!av_store(RExC_paren_name_list,
10891 RExC_npar, SvREFCNT_inc(svname)))
10892 SvREFCNT_dec_NN(svname);
10895 /*sv_dump(sv_dat);*/
10897 nextchar(pRExC_state);
10899 goto capturing_parens;
10901 RExC_seen |= REG_LOOKBEHIND_SEEN;
10902 RExC_in_lookbehind++;
10904 if (RExC_parse >= RExC_end) {
10905 vFAIL("Sequence (?... not terminated");
10909 case '=': /* (?=...) */
10910 RExC_seen_zerolen++;
10912 case '!': /* (?!...) */
10913 RExC_seen_zerolen++;
10914 /* check if we're really just a "FAIL" assertion */
10915 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
10916 FALSE /* Don't force to /x */ );
10917 if (*RExC_parse == ')') {
10918 ret=reganode(pRExC_state, OPFAIL, 0);
10919 nextchar(pRExC_state);
10923 case '|': /* (?|...) */
10924 /* branch reset, behave like a (?:...) except that
10925 buffers in alternations share the same numbers */
10927 after_freeze = freeze_paren = RExC_npar;
10929 case ':': /* (?:...) */
10930 case '>': /* (?>...) */
10932 case '$': /* (?$...) */
10933 case '@': /* (?@...) */
10934 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
10936 case '0' : /* (?0) */
10937 case 'R' : /* (?R) */
10938 if (RExC_parse == RExC_end || *RExC_parse != ')')
10939 FAIL("Sequence (?R) not terminated");
10941 RExC_seen |= REG_RECURSE_SEEN;
10942 *flagp |= POSTPONED;
10943 goto gen_recurse_regop;
10945 /* named and numeric backreferences */
10946 case '&': /* (?&NAME) */
10947 parse_start = RExC_parse - 1;
10950 SV *sv_dat = reg_scan_name(pRExC_state,
10951 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10952 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10954 if (RExC_parse >= RExC_end || *RExC_parse != ')')
10955 vFAIL("Sequence (?&... not terminated");
10956 goto gen_recurse_regop;
10959 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10961 vFAIL("Illegal pattern");
10963 goto parse_recursion;
10965 case '-': /* (?-1) */
10966 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10967 RExC_parse--; /* rewind to let it be handled later */
10971 case '1': case '2': case '3': case '4': /* (?1) */
10972 case '5': case '6': case '7': case '8': case '9':
10973 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
10976 bool is_neg = FALSE;
10978 parse_start = RExC_parse - 1; /* MJD */
10979 if (*RExC_parse == '-') {
10983 if (grok_atoUV(RExC_parse, &unum, &endptr)
10987 RExC_parse = (char*)endptr;
10991 /* Some limit for num? */
10995 if (*RExC_parse!=')')
10996 vFAIL("Expecting close bracket");
10999 if ( paren == '-' ) {
11001 Diagram of capture buffer numbering.
11002 Top line is the normal capture buffer numbers
11003 Bottom line is the negative indexing as from
11007 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
11011 num = RExC_npar + num;
11014 vFAIL("Reference to nonexistent group");
11016 } else if ( paren == '+' ) {
11017 num = RExC_npar + num - 1;
11019 /* We keep track how many GOSUB items we have produced.
11020 To start off the ARG2L() of the GOSUB holds its "id",
11021 which is used later in conjunction with RExC_recurse
11022 to calculate the offset we need to jump for the GOSUB,
11023 which it will store in the final representation.
11024 We have to defer the actual calculation until much later
11025 as the regop may move.
11028 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
11030 if (num > (I32)RExC_rx->nparens) {
11032 vFAIL("Reference to nonexistent group");
11034 RExC_recurse_count++;
11035 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11036 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
11037 22, "| |", (int)(depth * 2 + 1), "",
11038 (UV)ARG(ret), (IV)ARG2L(ret)));
11040 RExC_seen |= REG_RECURSE_SEEN;
11042 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
11043 Set_Node_Offset(ret, parse_start); /* MJD */
11045 *flagp |= POSTPONED;
11046 assert(*RExC_parse == ')');
11047 nextchar(pRExC_state);
11052 case '?': /* (??...) */
11054 if (*RExC_parse != '{') {
11055 RExC_parse += SKIP_IF_CHAR(RExC_parse);
11056 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11058 "Sequence (%" UTF8f "...) not recognized",
11059 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11060 NOT_REACHED; /*NOTREACHED*/
11062 *flagp |= POSTPONED;
11066 case '{': /* (?{...}) */
11069 struct reg_code_block *cb;
11071 RExC_seen_zerolen++;
11073 if ( !pRExC_state->code_blocks
11074 || pRExC_state->code_index
11075 >= pRExC_state->code_blocks->count
11076 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11077 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11080 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11081 FAIL("panic: Sequence (?{...}): no code block found\n");
11082 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11084 /* this is a pre-compiled code block (?{...}) */
11085 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11086 RExC_parse = RExC_start + cb->end;
11089 if (cb->src_regex) {
11090 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11091 RExC_rxi->data->data[n] =
11092 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11093 RExC_rxi->data->data[n+1] = (void*)o;
11096 n = add_data(pRExC_state,
11097 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11098 RExC_rxi->data->data[n] = (void*)o;
11101 pRExC_state->code_index++;
11102 nextchar(pRExC_state);
11106 ret = reg_node(pRExC_state, LOGICAL);
11108 eval = reg2Lanode(pRExC_state, EVAL,
11111 /* for later propagation into (??{})
11113 RExC_flags & RXf_PMf_COMPILETIME
11118 REGTAIL(pRExC_state, ret, eval);
11119 /* deal with the length of this later - MJD */
11122 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11123 Set_Node_Length(ret, RExC_parse - parse_start + 1);
11124 Set_Node_Offset(ret, parse_start);
11127 case '(': /* (?(?{...})...) and (?(?=...)...) */
11130 const int DEFINE_len = sizeof("DEFINE") - 1;
11131 if (RExC_parse[0] == '?') { /* (?(?...)) */
11132 if ( RExC_parse < RExC_end - 1
11133 && ( RExC_parse[1] == '='
11134 || RExC_parse[1] == '!'
11135 || RExC_parse[1] == '<'
11136 || RExC_parse[1] == '{')
11137 ) { /* Lookahead or eval. */
11141 ret = reg_node(pRExC_state, LOGICAL);
11145 tail = reg(pRExC_state, 1, &flag, depth+1);
11146 if (flag & (RESTART_PASS1|NEED_UTF8)) {
11147 *flagp = flag & (RESTART_PASS1|NEED_UTF8);
11150 REGTAIL(pRExC_state, ret, tail);
11153 /* Fall through to ‘Unknown switch condition’ at the
11154 end of the if/else chain. */
11156 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11157 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11159 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11160 char *name_start= RExC_parse++;
11162 SV *sv_dat=reg_scan_name(pRExC_state,
11163 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11164 if ( RExC_parse == name_start
11165 || RExC_parse >= RExC_end
11166 || *RExC_parse != ch)
11168 vFAIL2("Sequence (?(%c... not terminated",
11169 (ch == '>' ? '<' : ch));
11173 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11174 RExC_rxi->data->data[num]=(void*)sv_dat;
11175 SvREFCNT_inc_simple_void(sv_dat);
11177 ret = reganode(pRExC_state,NGROUPP,num);
11178 goto insert_if_check_paren;
11180 else if (memBEGINs(RExC_parse,
11181 (STRLEN) (RExC_end - RExC_parse),
11184 ret = reganode(pRExC_state,DEFINEP,0);
11185 RExC_parse += DEFINE_len;
11187 goto insert_if_check_paren;
11189 else if (RExC_parse[0] == 'R') {
11191 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11192 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11193 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11196 if (RExC_parse[0] == '0') {
11200 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11202 if (grok_atoUV(RExC_parse, &uv, &endptr)
11205 parno = (I32)uv + 1;
11206 RExC_parse = (char*)endptr;
11208 /* else "Switch condition not recognized" below */
11209 } else if (RExC_parse[0] == '&') {
11212 sv_dat = reg_scan_name(pRExC_state,
11214 ? REG_RSN_RETURN_NULL
11215 : REG_RSN_RETURN_DATA);
11217 /* we should only have a false sv_dat when
11218 * SIZE_ONLY is true, and we always have false
11219 * sv_dat when SIZE_ONLY is true.
11220 * reg_scan_name() will VFAIL() if the name is
11221 * unknown when SIZE_ONLY is false, and otherwise
11222 * will return something, and when SIZE_ONLY is
11223 * true, reg_scan_name() just parses the string,
11224 * and doesnt return anything. (in theory) */
11225 assert(SIZE_ONLY ? !sv_dat : !!sv_dat);
11228 parno = 1 + *((I32 *)SvPVX(sv_dat));
11230 ret = reganode(pRExC_state,INSUBP,parno);
11231 goto insert_if_check_paren;
11233 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
11237 if (grok_atoUV(RExC_parse, &uv, &endptr)
11241 RExC_parse = (char*)endptr;
11244 vFAIL("panic: grok_atoUV returned FALSE");
11246 ret = reganode(pRExC_state, GROUPP, parno);
11248 insert_if_check_paren:
11249 if (UCHARAT(RExC_parse) != ')') {
11250 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11251 vFAIL("Switch condition not recognized");
11253 nextchar(pRExC_state);
11255 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
11256 br = regbranch(pRExC_state, &flags, 1,depth+1);
11258 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11259 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11262 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11265 REGTAIL(pRExC_state, br, reganode(pRExC_state,
11267 c = UCHARAT(RExC_parse);
11268 nextchar(pRExC_state);
11269 if (flags&HASWIDTH)
11270 *flagp |= HASWIDTH;
11273 vFAIL("(?(DEFINE)....) does not allow branches");
11275 /* Fake one for optimizer. */
11276 lastbr = reganode(pRExC_state, IFTHEN, 0);
11278 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
11279 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11280 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11283 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf,
11286 REGTAIL(pRExC_state, ret, lastbr);
11287 if (flags&HASWIDTH)
11288 *flagp |= HASWIDTH;
11289 c = UCHARAT(RExC_parse);
11290 nextchar(pRExC_state);
11295 if (RExC_parse >= RExC_end)
11296 vFAIL("Switch (?(condition)... not terminated");
11298 vFAIL("Switch (?(condition)... contains too many branches");
11300 ender = reg_node(pRExC_state, TAIL);
11301 REGTAIL(pRExC_state, br, ender);
11303 REGTAIL(pRExC_state, lastbr, ender);
11304 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11307 REGTAIL(pRExC_state, ret, ender);
11308 RExC_size++; /* XXX WHY do we need this?!!
11309 For large programs it seems to be required
11310 but I can't figure out why. -- dmq*/
11313 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11314 vFAIL("Unknown switch condition (?(...))");
11316 case '[': /* (?[ ... ]) */
11317 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
11319 case 0: /* A NUL */
11320 RExC_parse--; /* for vFAIL to print correctly */
11321 vFAIL("Sequence (? incomplete");
11323 default: /* e.g., (?i) */
11324 RExC_parse = (char *) seqstart + 1;
11326 parse_lparen_question_flags(pRExC_state);
11327 if (UCHARAT(RExC_parse) != ':') {
11328 if (RExC_parse < RExC_end)
11329 nextchar(pRExC_state);
11334 nextchar(pRExC_state);
11339 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
11344 ret = reganode(pRExC_state, OPEN, parno);
11346 if (!RExC_nestroot)
11347 RExC_nestroot = parno;
11348 if (RExC_open_parens && !RExC_open_parens[parno])
11350 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11351 "%*s%*s Setting open paren #%" IVdf " to %d\n",
11352 22, "| |", (int)(depth * 2 + 1), "",
11353 (IV)parno, REG_NODE_NUM(ret)));
11354 RExC_open_parens[parno]= ret;
11357 Set_Node_Length(ret, 1); /* MJD */
11358 Set_Node_Offset(ret, RExC_parse); /* MJD */
11361 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
11370 /* Pick up the branches, linking them together. */
11371 parse_start = RExC_parse; /* MJD */
11372 br = regbranch(pRExC_state, &flags, 1,depth+1);
11374 /* branch_len = (paren != 0); */
11377 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11378 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11381 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11383 if (*RExC_parse == '|') {
11384 if (!SIZE_ONLY && RExC_extralen) {
11385 reginsert(pRExC_state, BRANCHJ, br, depth+1);
11388 reginsert(pRExC_state, BRANCH, br, depth+1);
11389 Set_Node_Length(br, paren != 0);
11390 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
11394 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
11396 else if (paren == ':') {
11397 *flagp |= flags&SIMPLE;
11399 if (is_open) { /* Starts with OPEN. */
11400 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
11402 else if (paren != '?') /* Not Conditional */
11404 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11406 while (*RExC_parse == '|') {
11407 if (!SIZE_ONLY && RExC_extralen) {
11408 ender = reganode(pRExC_state, LONGJMP,0);
11410 /* Append to the previous. */
11411 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
11414 RExC_extralen += 2; /* Account for LONGJMP. */
11415 nextchar(pRExC_state);
11416 if (freeze_paren) {
11417 if (RExC_npar > after_freeze)
11418 after_freeze = RExC_npar;
11419 RExC_npar = freeze_paren;
11421 br = regbranch(pRExC_state, &flags, 0, depth+1);
11424 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11425 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11428 FAIL2("panic: regbranch returned NULL, flags=%#" UVxf, (UV) flags);
11430 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
11432 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
11435 if (have_branch || paren != ':') {
11436 /* Make a closing node, and hook it on the end. */
11439 ender = reg_node(pRExC_state, TAIL);
11442 ender = reganode(pRExC_state, CLOSE, parno);
11443 if ( RExC_close_parens ) {
11444 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11445 "%*s%*s Setting close paren #%" IVdf " to %d\n",
11446 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
11447 RExC_close_parens[parno]= ender;
11448 if (RExC_nestroot == parno)
11451 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
11452 Set_Node_Length(ender,1); /* MJD */
11458 *flagp &= ~HASWIDTH;
11461 ender = reg_node(pRExC_state, SUCCEED);
11464 ender = reg_node(pRExC_state, END);
11466 assert(!RExC_end_op); /* there can only be one! */
11467 RExC_end_op = ender;
11468 if (RExC_close_parens) {
11469 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11470 "%*s%*s Setting close paren #0 (END) to %d\n",
11471 22, "| |", (int)(depth * 2 + 1), "", REG_NODE_NUM(ender)));
11473 RExC_close_parens[0]= ender;
11478 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11479 DEBUG_PARSE_MSG("lsbr");
11480 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
11481 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11482 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11483 SvPV_nolen_const(RExC_mysv1),
11484 (IV)REG_NODE_NUM(lastbr),
11485 SvPV_nolen_const(RExC_mysv2),
11486 (IV)REG_NODE_NUM(ender),
11487 (IV)(ender - lastbr)
11490 REGTAIL(pRExC_state, lastbr, ender);
11492 if (have_branch && !SIZE_ONLY) {
11493 char is_nothing= 1;
11495 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
11497 /* Hook the tails of the branches to the closing node. */
11498 for (br = ret; br; br = regnext(br)) {
11499 const U8 op = PL_regkind[OP(br)];
11500 if (op == BRANCH) {
11501 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
11502 if ( OP(NEXTOPER(br)) != NOTHING
11503 || regnext(NEXTOPER(br)) != ender)
11506 else if (op == BRANCHJ) {
11507 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
11508 /* for now we always disable this optimisation * /
11509 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
11510 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
11516 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
11517 DEBUG_PARSE_r(if (!SIZE_ONLY) {
11518 DEBUG_PARSE_MSG("NADA");
11519 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
11520 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
11521 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
11522 SvPV_nolen_const(RExC_mysv1),
11523 (IV)REG_NODE_NUM(ret),
11524 SvPV_nolen_const(RExC_mysv2),
11525 (IV)REG_NODE_NUM(ender),
11530 if (OP(ender) == TAIL) {
11535 for ( opt= br + 1; opt < ender ; opt++ )
11536 OP(opt)= OPTIMIZED;
11537 NEXT_OFF(br)= ender - br;
11545 static const char parens[] = "=!<,>";
11547 if (paren && (p = strchr(parens, paren))) {
11548 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
11549 int flag = (p - parens) > 1;
11552 node = SUSPEND, flag = 0;
11553 reginsert(pRExC_state, node,ret, depth+1);
11554 Set_Node_Cur_Length(ret, parse_start);
11555 Set_Node_Offset(ret, parse_start + 1);
11557 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
11561 /* Check for proper termination. */
11563 /* restore original flags, but keep (?p) and, if we've changed from /d
11564 * rules to /u, keep the /u */
11565 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
11566 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
11567 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
11569 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
11570 RExC_parse = oregcomp_parse;
11571 vFAIL("Unmatched (");
11573 nextchar(pRExC_state);
11575 else if (!paren && RExC_parse < RExC_end) {
11576 if (*RExC_parse == ')') {
11578 vFAIL("Unmatched )");
11581 FAIL("Junk on end of regexp"); /* "Can't happen". */
11582 NOT_REACHED; /* NOTREACHED */
11585 if (RExC_in_lookbehind) {
11586 RExC_in_lookbehind--;
11588 if (after_freeze > RExC_npar)
11589 RExC_npar = after_freeze;
11594 - regbranch - one alternative of an | operator
11596 * Implements the concatenation operator.
11598 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11599 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11602 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
11605 regnode *chain = NULL;
11607 I32 flags = 0, c = 0;
11608 GET_RE_DEBUG_FLAGS_DECL;
11610 PERL_ARGS_ASSERT_REGBRANCH;
11612 DEBUG_PARSE("brnc");
11617 if (!SIZE_ONLY && RExC_extralen)
11618 ret = reganode(pRExC_state, BRANCHJ,0);
11620 ret = reg_node(pRExC_state, BRANCH);
11621 Set_Node_Length(ret, 1);
11625 if (!first && SIZE_ONLY)
11626 RExC_extralen += 1; /* BRANCHJ */
11628 *flagp = WORST; /* Tentatively. */
11630 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11631 FALSE /* Don't force to /x */ );
11632 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
11633 flags &= ~TRYAGAIN;
11634 latest = regpiece(pRExC_state, &flags,depth+1);
11635 if (latest == NULL) {
11636 if (flags & TRYAGAIN)
11638 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11639 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11642 FAIL2("panic: regpiece returned NULL, flags=%#" UVxf, (UV) flags);
11644 else if (ret == NULL)
11646 *flagp |= flags&(HASWIDTH|POSTPONED);
11647 if (chain == NULL) /* First piece. */
11648 *flagp |= flags&SPSTART;
11650 /* FIXME adding one for every branch after the first is probably
11651 * excessive now we have TRIE support. (hv) */
11653 REGTAIL(pRExC_state, chain, latest);
11658 if (chain == NULL) { /* Loop ran zero times. */
11659 chain = reg_node(pRExC_state, NOTHING);
11664 *flagp |= flags&SIMPLE;
11671 - regpiece - something followed by possible quantifier * + ? {n,m}
11673 * Note that the branching code sequences used for ? and the general cases
11674 * of * and + are somewhat optimized: they use the same NOTHING node as
11675 * both the endmarker for their branch list and the body of the last branch.
11676 * It might seem that this node could be dispensed with entirely, but the
11677 * endmarker role is not redundant.
11679 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
11681 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11682 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11685 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11691 const char * const origparse = RExC_parse;
11693 I32 max = REG_INFTY;
11694 #ifdef RE_TRACK_PATTERN_OFFSETS
11697 const char *maxpos = NULL;
11700 /* Save the original in case we change the emitted regop to a FAIL. */
11701 regnode * const orig_emit = RExC_emit;
11703 GET_RE_DEBUG_FLAGS_DECL;
11705 PERL_ARGS_ASSERT_REGPIECE;
11707 DEBUG_PARSE("piec");
11709 ret = regatom(pRExC_state, &flags,depth+1);
11711 if (flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8))
11712 *flagp |= flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8);
11714 FAIL2("panic: regatom returned NULL, flags=%#" UVxf, (UV) flags);
11720 if (op == '{' && regcurly(RExC_parse)) {
11722 #ifdef RE_TRACK_PATTERN_OFFSETS
11723 parse_start = RExC_parse; /* MJD */
11725 next = RExC_parse + 1;
11726 while (isDIGIT(*next) || *next == ',') {
11727 if (*next == ',') {
11735 if (*next == '}') { /* got one */
11736 const char* endptr;
11740 if (isDIGIT(*RExC_parse)) {
11741 if (!grok_atoUV(RExC_parse, &uv, &endptr))
11742 vFAIL("Invalid quantifier in {,}");
11743 if (uv >= REG_INFTY)
11744 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11749 if (*maxpos == ',')
11752 maxpos = RExC_parse;
11753 if (isDIGIT(*maxpos)) {
11754 if (!grok_atoUV(maxpos, &uv, &endptr))
11755 vFAIL("Invalid quantifier in {,}");
11756 if (uv >= REG_INFTY)
11757 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
11760 max = REG_INFTY; /* meaning "infinity" */
11763 nextchar(pRExC_state);
11764 if (max < min) { /* If can't match, warn and optimize to fail
11766 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
11768 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
11769 NEXT_OFF(orig_emit)= regarglen[OPFAIL] + NODE_STEP_REGNODE;
11773 else if (min == max && *RExC_parse == '?')
11776 ckWARN2reg(RExC_parse + 1,
11777 "Useless use of greediness modifier '%c'",
11783 if ((flags&SIMPLE)) {
11784 if (min == 0 && max == REG_INFTY) {
11785 reginsert(pRExC_state, STAR, ret, depth+1);
11787 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11790 if (min == 1 && max == REG_INFTY) {
11791 reginsert(pRExC_state, PLUS, ret, depth+1);
11793 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11796 MARK_NAUGHTY_EXP(2, 2);
11797 reginsert(pRExC_state, CURLY, ret, depth+1);
11798 Set_Node_Offset(ret, parse_start+1); /* MJD */
11799 Set_Node_Cur_Length(ret, parse_start);
11802 regnode * const w = reg_node(pRExC_state, WHILEM);
11805 REGTAIL(pRExC_state, ret, w);
11806 if (!SIZE_ONLY && RExC_extralen) {
11807 reginsert(pRExC_state, LONGJMP,ret, depth+1);
11808 reginsert(pRExC_state, NOTHING,ret, depth+1);
11809 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
11811 reginsert(pRExC_state, CURLYX,ret, depth+1);
11813 Set_Node_Offset(ret, parse_start+1);
11814 Set_Node_Length(ret,
11815 op == '{' ? (RExC_parse - parse_start) : 1);
11817 if (!SIZE_ONLY && RExC_extralen)
11818 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
11819 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
11821 RExC_whilem_seen++, RExC_extralen += 3;
11822 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
11829 *flagp |= HASWIDTH;
11831 ARG1_SET(ret, (U16)min);
11832 ARG2_SET(ret, (U16)max);
11834 if (max == REG_INFTY)
11835 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11841 if (!ISMULT1(op)) {
11846 #if 0 /* Now runtime fix should be reliable. */
11848 /* if this is reinstated, don't forget to put this back into perldiag:
11850 =item Regexp *+ operand could be empty at {#} in regex m/%s/
11852 (F) The part of the regexp subject to either the * or + quantifier
11853 could match an empty string. The {#} shows in the regular
11854 expression about where the problem was discovered.
11858 if (!(flags&HASWIDTH) && op != '?')
11859 vFAIL("Regexp *+ operand could be empty");
11862 #ifdef RE_TRACK_PATTERN_OFFSETS
11863 parse_start = RExC_parse;
11865 nextchar(pRExC_state);
11867 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
11873 else if (op == '+') {
11877 else if (op == '?') {
11882 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
11883 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
11884 ckWARN2reg(RExC_parse,
11885 "%" UTF8f " matches null string many times",
11886 UTF8fARG(UTF, (RExC_parse >= origparse
11887 ? RExC_parse - origparse
11890 (void)ReREFCNT_inc(RExC_rx_sv);
11893 if (*RExC_parse == '?') {
11894 nextchar(pRExC_state);
11895 reginsert(pRExC_state, MINMOD, ret, depth+1);
11896 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
11898 else if (*RExC_parse == '+') {
11900 nextchar(pRExC_state);
11901 ender = reg_node(pRExC_state, SUCCEED);
11902 REGTAIL(pRExC_state, ret, ender);
11903 reginsert(pRExC_state, SUSPEND, ret, depth+1);
11904 ender = reg_node(pRExC_state, TAIL);
11905 REGTAIL(pRExC_state, ret, ender);
11908 if (ISMULT2(RExC_parse)) {
11910 vFAIL("Nested quantifiers");
11917 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
11926 /* This routine teases apart the various meanings of \N and returns
11927 * accordingly. The input parameters constrain which meaning(s) is/are valid
11928 * in the current context.
11930 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
11932 * If <code_point_p> is not NULL, the context is expecting the result to be a
11933 * single code point. If this \N instance turns out to a single code point,
11934 * the function returns TRUE and sets *code_point_p to that code point.
11936 * If <node_p> is not NULL, the context is expecting the result to be one of
11937 * the things representable by a regnode. If this \N instance turns out to be
11938 * one such, the function generates the regnode, returns TRUE and sets *node_p
11939 * to point to that regnode.
11941 * If this instance of \N isn't legal in any context, this function will
11942 * generate a fatal error and not return.
11944 * On input, RExC_parse should point to the first char following the \N at the
11945 * time of the call. On successful return, RExC_parse will have been updated
11946 * to point to just after the sequence identified by this routine. Also
11947 * *flagp has been updated as needed.
11949 * When there is some problem with the current context and this \N instance,
11950 * the function returns FALSE, without advancing RExC_parse, nor setting
11951 * *node_p, nor *code_point_p, nor *flagp.
11953 * If <cp_count> is not NULL, the caller wants to know the length (in code
11954 * points) that this \N sequence matches. This is set even if the function
11955 * returns FALSE, as detailed below.
11957 * There are 5 possibilities here, as detailed in the next 5 paragraphs.
11959 * Probably the most common case is for the \N to specify a single code point.
11960 * *cp_count will be set to 1, and *code_point_p will be set to that code
11963 * Another possibility is for the input to be an empty \N{}, which for
11964 * backwards compatibility we accept. *cp_count will be set to 0. *node_p
11965 * will be set to a generated NOTHING node.
11967 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
11968 * set to 0. *node_p will be set to a generated REG_ANY node.
11970 * The fourth possibility is that \N resolves to a sequence of more than one
11971 * code points. *cp_count will be set to the number of code points in the
11972 * sequence. *node_p * will be set to a generated node returned by this
11973 * function calling S_reg().
11975 * The final possibility is that it is premature to be calling this function;
11976 * that pass1 needs to be restarted. This can happen when this changes from
11977 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
11978 * latter occurs only when the fourth possibility would otherwise be in
11979 * effect, and is because one of those code points requires the pattern to be
11980 * recompiled as UTF-8. The function returns FALSE, and sets the
11981 * RESTART_PASS1 and NEED_UTF8 flags in *flagp, as appropriate. When this
11982 * happens, the caller needs to desist from continuing parsing, and return
11983 * this information to its caller. This is not set for when there is only one
11984 * code point, as this can be called as part of an ANYOF node, and they can
11985 * store above-Latin1 code points without the pattern having to be in UTF-8.
11987 * For non-single-quoted regexes, the tokenizer has resolved character and
11988 * sequence names inside \N{...} into their Unicode values, normalizing the
11989 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
11990 * hex-represented code points in the sequence. This is done there because
11991 * the names can vary based on what charnames pragma is in scope at the time,
11992 * so we need a way to take a snapshot of what they resolve to at the time of
11993 * the original parse. [perl #56444].
11995 * That parsing is skipped for single-quoted regexes, so we may here get
11996 * '\N{NAME}'. This is a fatal error. These names have to be resolved by the
11997 * parser. But if the single-quoted regex is something like '\N{U+41}', that
11998 * is legal and handled here. The code point is Unicode, and has to be
11999 * translated into the native character set for non-ASCII platforms.
12002 char * endbrace; /* points to '}' following the name */
12003 char *endchar; /* Points to '.' or '}' ending cur char in the input
12005 char* p = RExC_parse; /* Temporary */
12007 GET_RE_DEBUG_FLAGS_DECL;
12009 PERL_ARGS_ASSERT_GROK_BSLASH_N;
12011 GET_RE_DEBUG_FLAGS;
12013 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
12014 assert(! (node_p && cp_count)); /* At most 1 should be set */
12016 if (cp_count) { /* Initialize return for the most common case */
12020 /* The [^\n] meaning of \N ignores spaces and comments under the /x
12021 * modifier. The other meanings do not, so use a temporary until we find
12022 * out which we are being called with */
12023 skip_to_be_ignored_text(pRExC_state, &p,
12024 FALSE /* Don't force to /x */ );
12026 /* Disambiguate between \N meaning a named character versus \N meaning
12027 * [^\n]. The latter is assumed when the {...} following the \N is a legal
12028 * quantifier, or there is no '{' at all */
12029 if (*p != '{' || regcurly(p)) {
12039 *node_p = reg_node(pRExC_state, REG_ANY);
12040 *flagp |= HASWIDTH|SIMPLE;
12042 Set_Node_Length(*node_p, 1); /* MJD */
12046 /* Here, we have decided it should be a named character or sequence */
12048 /* The test above made sure that the next real character is a '{', but
12049 * under the /x modifier, it could be separated by space (or a comment and
12050 * \n) and this is not allowed (for consistency with \x{...} and the
12051 * tokenizer handling of \N{NAME}). */
12052 if (*RExC_parse != '{') {
12053 vFAIL("Missing braces on \\N{}");
12056 RExC_parse++; /* Skip past the '{' */
12058 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
12059 if (! endbrace) { /* no trailing brace */
12060 vFAIL2("Missing right brace on \\%c{}", 'N');
12062 else if (!( endbrace == RExC_parse /* nothing between the {} */
12063 || memBEGINs(RExC_parse, /* U+ (bad hex is checked below
12064 for a better error msg) */
12065 (STRLEN) (RExC_end - RExC_parse),
12068 RExC_parse = endbrace; /* position msg's '<--HERE' */
12069 vFAIL("\\N{NAME} must be resolved by the lexer");
12072 REQUIRE_UNI_RULES(flagp, FALSE); /* Unicode named chars imply Unicode
12075 if (endbrace == RExC_parse) { /* empty: \N{} */
12077 RExC_parse++; /* Position after the "}" */
12078 vFAIL("Zero length \\N{}");
12083 nextchar(pRExC_state);
12088 *node_p = reg_node(pRExC_state,NOTHING);
12092 RExC_parse += 2; /* Skip past the 'U+' */
12094 /* Because toke.c has generated a special construct for us guaranteed not
12095 * to have NULs, we can use a str function */
12096 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12098 /* Code points are separated by dots. If none, there is only one code
12099 * point, and is terminated by the brace */
12101 if (endchar >= endbrace) {
12102 STRLEN length_of_hex;
12103 I32 grok_hex_flags;
12105 /* Here, exactly one code point. If that isn't what is wanted, fail */
12106 if (! code_point_p) {
12111 /* Convert code point from hex */
12112 length_of_hex = (STRLEN)(endchar - RExC_parse);
12113 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
12114 | PERL_SCAN_DISALLOW_PREFIX
12116 /* No errors in the first pass (See [perl
12117 * #122671].) We let the code below find the
12118 * errors when there are multiple chars. */
12120 ? PERL_SCAN_SILENT_ILLDIGIT
12123 /* This routine is the one place where both single- and double-quotish
12124 * \N{U+xxxx} are evaluated. The value is a Unicode code point which
12125 * must be converted to native. */
12126 *code_point_p = UNI_TO_NATIVE(grok_hex(RExC_parse,
12131 /* The tokenizer should have guaranteed validity, but it's possible to
12132 * bypass it by using single quoting, so check. Don't do the check
12133 * here when there are multiple chars; we do it below anyway. */
12134 if (length_of_hex == 0
12135 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
12137 RExC_parse += length_of_hex; /* Includes all the valid */
12138 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
12139 ? UTF8SKIP(RExC_parse)
12141 /* Guard against malformed utf8 */
12142 if (RExC_parse >= endchar) {
12143 RExC_parse = endchar;
12145 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12148 RExC_parse = endbrace + 1;
12151 else { /* Is a multiple character sequence */
12152 SV * substitute_parse;
12154 char *orig_end = RExC_end;
12155 char *save_start = RExC_start;
12158 /* Count the code points, if desired, in the sequence */
12161 while (RExC_parse < endbrace) {
12162 /* Point to the beginning of the next character in the sequence. */
12163 RExC_parse = endchar + 1;
12164 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12169 /* Fail if caller doesn't want to handle a multi-code-point sequence.
12170 * But don't backup up the pointer if the caller wants to know how many
12171 * code points there are (they can then handle things) */
12179 /* What is done here is to convert this to a sub-pattern of the form
12180 * \x{char1}\x{char2}... and then call reg recursively to parse it
12181 * (enclosing in "(?: ... )" ). That way, it retains its atomicness,
12182 * while not having to worry about special handling that some code
12183 * points may have. */
12185 substitute_parse = newSVpvs("?:");
12187 while (RExC_parse < endbrace) {
12189 /* Convert to notation the rest of the code understands */
12190 sv_catpv(substitute_parse, "\\x{");
12191 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
12192 sv_catpv(substitute_parse, "}");
12194 /* Point to the beginning of the next character in the sequence. */
12195 RExC_parse = endchar + 1;
12196 endchar = RExC_parse + strcspn(RExC_parse, ".}");
12199 sv_catpv(substitute_parse, ")");
12201 len = SvCUR(substitute_parse);
12203 /* Don't allow empty number */
12204 if (len < (STRLEN) 8) {
12205 RExC_parse = endbrace;
12206 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12209 RExC_parse = RExC_start = RExC_adjusted_start
12210 = SvPV_nolen(substitute_parse);
12211 RExC_end = RExC_parse + len;
12213 /* The values are Unicode, and therefore not subject to recoding, but
12214 * have to be converted to native on a non-Unicode (meaning non-ASCII)
12217 RExC_recode_x_to_native = 1;
12220 *node_p = reg(pRExC_state, 1, &flags, depth+1);
12222 /* Restore the saved values */
12223 RExC_start = RExC_adjusted_start = save_start;
12224 RExC_parse = endbrace;
12225 RExC_end = orig_end;
12227 RExC_recode_x_to_native = 0;
12229 SvREFCNT_dec_NN(substitute_parse);
12232 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12233 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12236 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#" UVxf,
12239 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12241 nextchar(pRExC_state);
12248 PERL_STATIC_INLINE U8
12249 S_compute_EXACTish(RExC_state_t *pRExC_state)
12253 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
12261 op = get_regex_charset(RExC_flags);
12262 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
12263 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
12264 been, so there is no hole */
12267 return op + EXACTF;
12270 PERL_STATIC_INLINE void
12271 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
12272 regnode *node, I32* flagp, STRLEN len, UV code_point,
12275 /* This knows the details about sizing an EXACTish node, setting flags for
12276 * it (by setting <*flagp>, and potentially populating it with a single
12279 * If <len> (the length in bytes) is non-zero, this function assumes that
12280 * the node has already been populated, and just does the sizing. In this
12281 * case <code_point> should be the final code point that has already been
12282 * placed into the node. This value will be ignored except that under some
12283 * circumstances <*flagp> is set based on it.
12285 * If <len> is zero, the function assumes that the node is to contain only
12286 * the single character given by <code_point> and calculates what <len>
12287 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
12288 * additionally will populate the node's STRING with <code_point> or its
12291 * In both cases <*flagp> is appropriately set
12293 * It knows that under FOLD, the Latin Sharp S and UTF characters above
12294 * 255, must be folded (the former only when the rules indicate it can
12297 * When it does the populating, it looks at the flag 'downgradable'. If
12298 * true with a node that folds, it checks if the single code point
12299 * participates in a fold, and if not downgrades the node to an EXACT.
12300 * This helps the optimizer */
12302 bool len_passed_in = cBOOL(len != 0);
12303 U8 character[UTF8_MAXBYTES_CASE+1];
12305 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
12307 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
12308 * sizing difference, and is extra work that is thrown away */
12309 if (downgradable && ! PASS2) {
12310 downgradable = FALSE;
12313 if (! len_passed_in) {
12315 if (UVCHR_IS_INVARIANT(code_point)) {
12316 if (LOC || ! FOLD) { /* /l defers folding until runtime */
12317 *character = (U8) code_point;
12319 else { /* Here is /i and not /l. (toFOLD() is defined on just
12320 ASCII, which isn't the same thing as INVARIANT on
12321 EBCDIC, but it works there, as the extra invariants
12322 fold to themselves) */
12323 *character = toFOLD((U8) code_point);
12325 /* We can downgrade to an EXACT node if this character
12326 * isn't a folding one. Note that this assumes that
12327 * nothing above Latin1 folds to some other invariant than
12328 * one of these alphabetics; otherwise we would also have
12330 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12331 * || ASCII_FOLD_RESTRICTED))
12333 if (downgradable && PL_fold[code_point] == code_point) {
12339 else if (FOLD && (! LOC
12340 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
12341 { /* Folding, and ok to do so now */
12342 UV folded = _to_uni_fold_flags(
12346 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12347 ? FOLD_FLAGS_NOMIX_ASCII
12350 && folded == code_point /* This quickly rules out many
12351 cases, avoiding the
12352 _invlist_contains_cp() overhead
12354 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
12361 else if (code_point <= MAX_UTF8_TWO_BYTE) {
12363 /* Not folding this cp, and can output it directly */
12364 *character = UTF8_TWO_BYTE_HI(code_point);
12365 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
12369 uvchr_to_utf8( character, code_point);
12370 len = UTF8SKIP(character);
12372 } /* Else pattern isn't UTF8. */
12374 *character = (U8) code_point;
12376 } /* Else is folded non-UTF8 */
12377 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12378 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12379 || UNICODE_DOT_DOT_VERSION > 0)
12380 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
12384 /* We don't fold any non-UTF8 except possibly the Sharp s (see
12385 * comments at join_exact()); */
12386 *character = (U8) code_point;
12389 /* Can turn into an EXACT node if we know the fold at compile time,
12390 * and it folds to itself and doesn't particpate in other folds */
12393 && PL_fold_latin1[code_point] == code_point
12394 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
12395 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
12399 } /* else is Sharp s. May need to fold it */
12400 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
12402 *(character + 1) = 's';
12406 *character = LATIN_SMALL_LETTER_SHARP_S;
12412 RExC_size += STR_SZ(len);
12415 RExC_emit += STR_SZ(len);
12416 STR_LEN(node) = len;
12417 if (! len_passed_in) {
12418 Copy((char *) character, STRING(node), len, char);
12422 *flagp |= HASWIDTH;
12424 /* A single character node is SIMPLE, except for the special-cased SHARP S
12426 if ((len == 1 || (UTF && len == UVCHR_SKIP(code_point)))
12427 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12428 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12429 || UNICODE_DOT_DOT_VERSION > 0)
12430 && ( code_point != LATIN_SMALL_LETTER_SHARP_S
12431 || ! FOLD || ! DEPENDS_SEMANTICS)
12437 /* The OP may not be well defined in PASS1 */
12438 if (PASS2 && OP(node) == EXACTFL) {
12439 RExC_contains_locale = 1;
12444 S_new_regcurly(const char *s, const char *e)
12446 /* This is a temporary function designed to match the most lenient form of
12447 * a {m,n} quantifier we ever envision, with either number omitted, and
12448 * spaces anywhere between/before/after them.
12450 * If this function fails, then the string it matches is very unlikely to
12451 * ever be considered a valid quantifier, so we can allow the '{' that
12452 * begins it to be considered as a literal */
12454 bool has_min = FALSE;
12455 bool has_max = FALSE;
12457 PERL_ARGS_ASSERT_NEW_REGCURLY;
12459 if (s >= e || *s++ != '{')
12462 while (s < e && isSPACE(*s)) {
12465 while (s < e && isDIGIT(*s)) {
12469 while (s < e && isSPACE(*s)) {
12475 while (s < e && isSPACE(*s)) {
12478 while (s < e && isDIGIT(*s)) {
12482 while (s < e && isSPACE(*s)) {
12487 return s < e && *s == '}' && (has_min || has_max);
12490 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
12491 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
12494 S_backref_value(char *p)
12496 const char* endptr;
12498 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
12505 - regatom - the lowest level
12507 Try to identify anything special at the start of the current parse position.
12508 If there is, then handle it as required. This may involve generating a
12509 single regop, such as for an assertion; or it may involve recursing, such as
12510 to handle a () structure.
12512 If the string doesn't start with something special then we gobble up
12513 as much literal text as we can. If we encounter a quantifier, we have to
12514 back off the final literal character, as that quantifier applies to just it
12515 and not to the whole string of literals.
12517 Once we have been able to handle whatever type of thing started the
12518 sequence, we return.
12520 Note: we have to be careful with escapes, as they can be both literal
12521 and special, and in the case of \10 and friends, context determines which.
12523 A summary of the code structure is:
12525 switch (first_byte) {
12526 cases for each special:
12527 handle this special;
12530 switch (2nd byte) {
12531 cases for each unambiguous special:
12532 handle this special;
12534 cases for each ambigous special/literal:
12536 if (special) handle here
12538 default: // unambiguously literal:
12541 default: // is a literal char
12544 create EXACTish node for literal;
12545 while (more input and node isn't full) {
12546 switch (input_byte) {
12547 cases for each special;
12548 make sure parse pointer is set so that the next call to
12549 regatom will see this special first
12550 goto loopdone; // EXACTish node terminated by prev. char
12552 append char to EXACTISH node;
12554 get next input byte;
12558 return the generated node;
12560 Specifically there are two separate switches for handling
12561 escape sequences, with the one for handling literal escapes requiring
12562 a dummy entry for all of the special escapes that are actually handled
12565 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
12567 Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
12568 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
12569 Otherwise does not return NULL.
12573 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12575 regnode *ret = NULL;
12582 GET_RE_DEBUG_FLAGS_DECL;
12584 *flagp = WORST; /* Tentatively. */
12586 DEBUG_PARSE("atom");
12588 PERL_ARGS_ASSERT_REGATOM;
12591 parse_start = RExC_parse;
12592 assert(RExC_parse < RExC_end);
12593 switch ((U8)*RExC_parse) {
12595 RExC_seen_zerolen++;
12596 nextchar(pRExC_state);
12597 if (RExC_flags & RXf_PMf_MULTILINE)
12598 ret = reg_node(pRExC_state, MBOL);
12600 ret = reg_node(pRExC_state, SBOL);
12601 Set_Node_Length(ret, 1); /* MJD */
12604 nextchar(pRExC_state);
12606 RExC_seen_zerolen++;
12607 if (RExC_flags & RXf_PMf_MULTILINE)
12608 ret = reg_node(pRExC_state, MEOL);
12610 ret = reg_node(pRExC_state, SEOL);
12611 Set_Node_Length(ret, 1); /* MJD */
12614 nextchar(pRExC_state);
12615 if (RExC_flags & RXf_PMf_SINGLELINE)
12616 ret = reg_node(pRExC_state, SANY);
12618 ret = reg_node(pRExC_state, REG_ANY);
12619 *flagp |= HASWIDTH|SIMPLE;
12621 Set_Node_Length(ret, 1); /* MJD */
12625 char * const oregcomp_parse = ++RExC_parse;
12626 ret = regclass(pRExC_state, flagp,depth+1,
12627 FALSE, /* means parse the whole char class */
12628 TRUE, /* allow multi-char folds */
12629 FALSE, /* don't silence non-portable warnings. */
12630 (bool) RExC_strict,
12631 TRUE, /* Allow an optimized regnode result */
12635 if (*flagp & (RESTART_PASS1|NEED_UTF8))
12637 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12640 if (*RExC_parse != ']') {
12641 RExC_parse = oregcomp_parse;
12642 vFAIL("Unmatched [");
12644 nextchar(pRExC_state);
12645 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
12649 nextchar(pRExC_state);
12650 ret = reg(pRExC_state, 2, &flags,depth+1);
12652 if (flags & TRYAGAIN) {
12653 if (RExC_parse >= RExC_end) {
12654 /* Make parent create an empty node if needed. */
12655 *flagp |= TRYAGAIN;
12660 if (flags & (RESTART_PASS1|NEED_UTF8)) {
12661 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
12664 FAIL2("panic: reg returned NULL to regatom, flags=%#" UVxf,
12667 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
12671 if (flags & TRYAGAIN) {
12672 *flagp |= TRYAGAIN;
12675 vFAIL("Internal urp");
12676 /* Supposed to be caught earlier. */
12682 vFAIL("Quantifier follows nothing");
12687 This switch handles escape sequences that resolve to some kind
12688 of special regop and not to literal text. Escape sequnces that
12689 resolve to literal text are handled below in the switch marked
12692 Every entry in this switch *must* have a corresponding entry
12693 in the literal escape switch. However, the opposite is not
12694 required, as the default for this switch is to jump to the
12695 literal text handling code.
12698 switch ((U8)*RExC_parse) {
12699 /* Special Escapes */
12701 RExC_seen_zerolen++;
12702 ret = reg_node(pRExC_state, SBOL);
12703 /* SBOL is shared with /^/ so we set the flags so we can tell
12704 * /\A/ from /^/ in split. We check ret because first pass we
12705 * have no regop struct to set the flags on. */
12709 goto finish_meta_pat;
12711 ret = reg_node(pRExC_state, GPOS);
12712 RExC_seen |= REG_GPOS_SEEN;
12714 goto finish_meta_pat;
12716 RExC_seen_zerolen++;
12717 ret = reg_node(pRExC_state, KEEPS);
12719 /* XXX:dmq : disabling in-place substitution seems to
12720 * be necessary here to avoid cases of memory corruption, as
12721 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
12723 RExC_seen |= REG_LOOKBEHIND_SEEN;
12724 goto finish_meta_pat;
12726 ret = reg_node(pRExC_state, SEOL);
12728 RExC_seen_zerolen++; /* Do not optimize RE away */
12729 goto finish_meta_pat;
12731 ret = reg_node(pRExC_state, EOS);
12733 RExC_seen_zerolen++; /* Do not optimize RE away */
12734 goto finish_meta_pat;
12736 vFAIL("\\C no longer supported");
12738 ret = reg_node(pRExC_state, CLUMP);
12739 *flagp |= HASWIDTH;
12740 goto finish_meta_pat;
12746 arg = ANYOF_WORDCHAR;
12754 regex_charset charset = get_regex_charset(RExC_flags);
12756 RExC_seen_zerolen++;
12757 RExC_seen |= REG_LOOKBEHIND_SEEN;
12758 op = BOUND + charset;
12760 if (op == BOUNDL) {
12761 RExC_contains_locale = 1;
12764 ret = reg_node(pRExC_state, op);
12766 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
12767 FLAGS(ret) = TRADITIONAL_BOUND;
12768 if (PASS2 && op > BOUNDA) { /* /aa is same as /a */
12774 char name = *RExC_parse;
12775 char * endbrace = NULL;
12777 if (RExC_parse < RExC_end) {
12778 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
12782 vFAIL2("Missing right brace on \\%c{}", name);
12784 /* XXX Need to decide whether to take spaces or not. Should be
12785 * consistent with \p{}, but that currently is SPACE, which
12786 * means vertical too, which seems wrong
12787 * while (isBLANK(*RExC_parse)) {
12790 if (endbrace == RExC_parse) {
12791 RExC_parse++; /* After the '}' */
12792 vFAIL2("Empty \\%c{}", name);
12794 length = endbrace - RExC_parse;
12795 /*while (isBLANK(*(RExC_parse + length - 1))) {
12798 switch (*RExC_parse) {
12801 && (memNEs(RExC_parse + 1, length - 1, "cb")))
12803 goto bad_bound_type;
12805 FLAGS(ret) = GCB_BOUND;
12808 if (length != 2 || *(RExC_parse + 1) != 'b') {
12809 goto bad_bound_type;
12811 FLAGS(ret) = LB_BOUND;
12814 if (length != 2 || *(RExC_parse + 1) != 'b') {
12815 goto bad_bound_type;
12817 FLAGS(ret) = SB_BOUND;
12820 if (length != 2 || *(RExC_parse + 1) != 'b') {
12821 goto bad_bound_type;
12823 FLAGS(ret) = WB_BOUND;
12827 RExC_parse = endbrace;
12829 "'%" UTF8f "' is an unknown bound type",
12830 UTF8fARG(UTF, length, endbrace - length));
12831 NOT_REACHED; /*NOTREACHED*/
12833 RExC_parse = endbrace;
12834 REQUIRE_UNI_RULES(flagp, NULL);
12836 if (PASS2 && op >= BOUNDA) { /* /aa is same as /a */
12840 /* Don't have to worry about UTF-8, in this message because
12841 * to get here the contents of the \b must be ASCII */
12842 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
12843 "Using /u for '%.*s' instead of /%s",
12845 endbrace - length + 1,
12846 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
12847 ? ASCII_RESTRICT_PAT_MODS
12848 : ASCII_MORE_RESTRICT_PAT_MODS);
12852 if (PASS2 && invert) {
12853 OP(ret) += NBOUND - BOUND;
12855 goto finish_meta_pat;
12863 if (! DEPENDS_SEMANTICS) {
12867 /* \d doesn't have any matches in the upper Latin1 range, hence /d
12868 * is equivalent to /u. Changing to /u saves some branches at
12871 goto join_posix_op_known;
12874 ret = reg_node(pRExC_state, LNBREAK);
12875 *flagp |= HASWIDTH|SIMPLE;
12876 goto finish_meta_pat;
12884 goto join_posix_op_known;
12890 arg = ANYOF_VERTWS;
12892 goto join_posix_op_known;
12902 op = POSIXD + get_regex_charset(RExC_flags);
12903 if (op > POSIXA) { /* /aa is same as /a */
12906 else if (op == POSIXL) {
12907 RExC_contains_locale = 1;
12910 join_posix_op_known:
12913 op += NPOSIXD - POSIXD;
12916 ret = reg_node(pRExC_state, op);
12918 FLAGS(ret) = namedclass_to_classnum(arg);
12921 *flagp |= HASWIDTH|SIMPLE;
12925 if ( UCHARAT(RExC_parse + 1) == '{'
12926 && UNLIKELY(! new_regcurly(RExC_parse + 1, RExC_end)))
12929 vFAIL("Unescaped left brace in regex is illegal here");
12931 nextchar(pRExC_state);
12932 Set_Node_Length(ret, 2); /* MJD */
12938 ret = regclass(pRExC_state, flagp,depth+1,
12939 TRUE, /* means just parse this element */
12940 FALSE, /* don't allow multi-char folds */
12941 FALSE, /* don't silence non-portable warnings. It
12942 would be a bug if these returned
12944 (bool) RExC_strict,
12945 TRUE, /* Allow an optimized regnode result */
12948 if (*flagp & RESTART_PASS1)
12950 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
12951 * multi-char folds are allowed. */
12953 FAIL2("panic: regclass returned NULL to regatom, flags=%#" UVxf,
12958 Set_Node_Offset(ret, parse_start);
12959 Set_Node_Cur_Length(ret, parse_start - 2);
12960 nextchar(pRExC_state);
12963 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
12964 * \N{...} evaluates to a sequence of more than one code points).
12965 * The function call below returns a regnode, which is our result.
12966 * The parameters cause it to fail if the \N{} evaluates to a
12967 * single code point; we handle those like any other literal. The
12968 * reason that the multicharacter case is handled here and not as
12969 * part of the EXACtish code is because of quantifiers. In
12970 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
12971 * this way makes that Just Happen. dmq.
12972 * join_exact() will join this up with adjacent EXACTish nodes
12973 * later on, if appropriate. */
12975 if (grok_bslash_N(pRExC_state,
12976 &ret, /* Want a regnode returned */
12977 NULL, /* Fail if evaluates to a single code
12979 NULL, /* Don't need a count of how many code
12988 if (*flagp & RESTART_PASS1)
12991 /* Here, evaluates to a single code point. Go get that */
12992 RExC_parse = parse_start;
12995 case 'k': /* Handle \k<NAME> and \k'NAME' */
12999 if ( RExC_parse >= RExC_end - 1
13000 || (( ch = RExC_parse[1]) != '<'
13005 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
13006 vFAIL2("Sequence %.2s... not terminated",parse_start);
13009 ret = handle_named_backref(pRExC_state,
13021 case '1': case '2': case '3': case '4':
13022 case '5': case '6': case '7': case '8': case '9':
13027 if (*RExC_parse == 'g') {
13031 if (*RExC_parse == '{') {
13035 if (*RExC_parse == '-') {
13039 if (hasbrace && !isDIGIT(*RExC_parse)) {
13040 if (isrel) RExC_parse--;
13042 goto parse_named_seq;
13045 if (RExC_parse >= RExC_end) {
13046 goto unterminated_g;
13048 num = S_backref_value(RExC_parse);
13050 vFAIL("Reference to invalid group 0");
13051 else if (num == I32_MAX) {
13052 if (isDIGIT(*RExC_parse))
13053 vFAIL("Reference to nonexistent group");
13056 vFAIL("Unterminated \\g... pattern");
13060 num = RExC_npar - num;
13062 vFAIL("Reference to nonexistent or unclosed group");
13066 num = S_backref_value(RExC_parse);
13067 /* bare \NNN might be backref or octal - if it is larger
13068 * than or equal RExC_npar then it is assumed to be an
13069 * octal escape. Note RExC_npar is +1 from the actual
13070 * number of parens. */
13071 /* Note we do NOT check if num == I32_MAX here, as that is
13072 * handled by the RExC_npar check */
13075 /* any numeric escape < 10 is always a backref */
13077 /* any numeric escape < RExC_npar is a backref */
13078 && num >= RExC_npar
13079 /* cannot be an octal escape if it starts with 8 */
13080 && *RExC_parse != '8'
13081 /* cannot be an octal escape it it starts with 9 */
13082 && *RExC_parse != '9'
13085 /* Probably not a backref, instead likely to be an
13086 * octal character escape, e.g. \35 or \777.
13087 * The above logic should make it obvious why using
13088 * octal escapes in patterns is problematic. - Yves */
13089 RExC_parse = parse_start;
13094 /* At this point RExC_parse points at a numeric escape like
13095 * \12 or \88 or something similar, which we should NOT treat
13096 * as an octal escape. It may or may not be a valid backref
13097 * escape. For instance \88888888 is unlikely to be a valid
13099 while (isDIGIT(*RExC_parse))
13102 if (*RExC_parse != '}')
13103 vFAIL("Unterminated \\g{...} pattern");
13107 if (num > (I32)RExC_rx->nparens)
13108 vFAIL("Reference to nonexistent group");
13111 ret = reganode(pRExC_state,
13114 : (ASCII_FOLD_RESTRICTED)
13116 : (AT_LEAST_UNI_SEMANTICS)
13122 *flagp |= HASWIDTH;
13124 /* override incorrect value set in reganode MJD */
13125 Set_Node_Offset(ret, parse_start);
13126 Set_Node_Cur_Length(ret, parse_start-1);
13127 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13128 FALSE /* Don't force to /x */ );
13132 if (RExC_parse >= RExC_end)
13133 FAIL("Trailing \\");
13136 /* Do not generate "unrecognized" warnings here, we fall
13137 back into the quick-grab loop below */
13138 RExC_parse = parse_start;
13140 } /* end of switch on a \foo sequence */
13145 /* '#' comments should have been spaced over before this function was
13147 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13149 if (RExC_flags & RXf_PMf_EXTENDED) {
13150 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13151 if (RExC_parse < RExC_end)
13161 /* Here, we have determined that the next thing is probably a
13162 * literal character. RExC_parse points to the first byte of its
13163 * definition. (It still may be an escape sequence that evaluates
13164 * to a single character) */
13170 #define MAX_NODE_STRING_SIZE 127
13171 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
13173 U8 upper_parse = MAX_NODE_STRING_SIZE;
13174 U8 node_type = compute_EXACTish(pRExC_state);
13175 bool next_is_quantifier;
13176 char * oldp = NULL;
13178 /* We can convert EXACTF nodes to EXACTFU if they contain only
13179 * characters that match identically regardless of the target
13180 * string's UTF8ness. The reason to do this is that EXACTF is not
13181 * trie-able, EXACTFU is.
13183 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13184 * contain only above-Latin1 characters (hence must be in UTF8),
13185 * which don't participate in folds with Latin1-range characters,
13186 * as the latter's folds aren't known until runtime. (We don't
13187 * need to figure this out until pass 2) */
13188 bool maybe_exactfu = PASS2
13189 && (node_type == EXACTF || node_type == EXACTFL);
13191 /* If a folding node contains only code points that don't
13192 * participate in folds, it can be changed into an EXACT node,
13193 * which allows the optimizer more things to look for */
13196 ret = reg_node(pRExC_state, node_type);
13198 /* In pass1, folded, we use a temporary buffer instead of the
13199 * actual node, as the node doesn't exist yet */
13200 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
13206 /* We look for the EXACTFish to EXACT node optimizaton only if
13207 * folding. (And we don't need to figure this out until pass 2).
13208 * XXX It might actually make sense to split the node into portions
13209 * that are exact and ones that aren't, so that we could later use
13210 * the exact ones to find the longest fixed and floating strings.
13211 * One would want to join them back into a larger node. One could
13212 * use a pseudo regnode like 'EXACT_ORIG_FOLD' */
13213 maybe_exact = FOLD && PASS2;
13215 /* XXX The node can hold up to 255 bytes, yet this only goes to
13216 * 127. I (khw) do not know why. Keeping it somewhat less than
13217 * 255 allows us to not have to worry about overflow due to
13218 * converting to utf8 and fold expansion, but that value is
13219 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
13220 * split up by this limit into a single one using the real max of
13221 * 255. Even at 127, this breaks under rare circumstances. If
13222 * folding, we do not want to split a node at a character that is a
13223 * non-final in a multi-char fold, as an input string could just
13224 * happen to want to match across the node boundary. The join
13225 * would solve that problem if the join actually happens. But a
13226 * series of more than two nodes in a row each of 127 would cause
13227 * the first join to succeed to get to 254, but then there wouldn't
13228 * be room for the next one, which could at be one of those split
13229 * multi-char folds. I don't know of any fool-proof solution. One
13230 * could back off to end with only a code point that isn't such a
13231 * non-final, but it is possible for there not to be any in the
13234 assert( ! UTF /* Is at the beginning of a character */
13235 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13236 || UTF8_IS_START(UCHARAT(RExC_parse)));
13238 /* Here, we have a literal character. Find the maximal string of
13239 * them in the input that we can fit into a single EXACTish node.
13240 * We quit at the first non-literal or when the node gets full */
13241 for (p = RExC_parse;
13242 len < upper_parse && p < RExC_end;
13247 /* White space has already been ignored */
13248 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13249 || ! is_PATWS_safe((p), RExC_end, UTF));
13261 /* Literal Escapes Switch
13263 This switch is meant to handle escape sequences that
13264 resolve to a literal character.
13266 Every escape sequence that represents something
13267 else, like an assertion or a char class, is handled
13268 in the switch marked 'Special Escapes' above in this
13269 routine, but also has an entry here as anything that
13270 isn't explicitly mentioned here will be treated as
13271 an unescaped equivalent literal.
13274 switch ((U8)*++p) {
13275 /* These are all the special escapes. */
13276 case 'A': /* Start assertion */
13277 case 'b': case 'B': /* Word-boundary assertion*/
13278 case 'C': /* Single char !DANGEROUS! */
13279 case 'd': case 'D': /* digit class */
13280 case 'g': case 'G': /* generic-backref, pos assertion */
13281 case 'h': case 'H': /* HORIZWS */
13282 case 'k': case 'K': /* named backref, keep marker */
13283 case 'p': case 'P': /* Unicode property */
13284 case 'R': /* LNBREAK */
13285 case 's': case 'S': /* space class */
13286 case 'v': case 'V': /* VERTWS */
13287 case 'w': case 'W': /* word class */
13288 case 'X': /* eXtended Unicode "combining
13289 character sequence" */
13290 case 'z': case 'Z': /* End of line/string assertion */
13294 /* Anything after here is an escape that resolves to a
13295 literal. (Except digits, which may or may not)
13301 case 'N': /* Handle a single-code point named character. */
13302 RExC_parse = p + 1;
13303 if (! grok_bslash_N(pRExC_state,
13304 NULL, /* Fail if evaluates to
13305 anything other than a
13306 single code point */
13307 &ender, /* The returned single code
13309 NULL, /* Don't need a count of
13310 how many code points */
13315 if (*flagp & NEED_UTF8)
13316 FAIL("panic: grok_bslash_N set NEED_UTF8");
13317 if (*flagp & RESTART_PASS1)
13320 /* Here, it wasn't a single code point. Go close
13321 * up this EXACTish node. The switch() prior to
13322 * this switch handles the other cases */
13323 RExC_parse = p = oldp;
13327 RExC_parse = parse_start;
13328 if (ender > 0xff) {
13329 REQUIRE_UTF8(flagp);
13345 ender = ESC_NATIVE;
13355 const char* error_msg;
13357 bool valid = grok_bslash_o(&p,
13361 PASS2, /* out warnings */
13362 (bool) RExC_strict,
13363 TRUE, /* Output warnings
13368 RExC_parse = p; /* going to die anyway; point
13369 to exact spot of failure */
13373 if (ender > 0xff) {
13374 REQUIRE_UTF8(flagp);
13380 UV result = UV_MAX; /* initialize to erroneous
13382 const char* error_msg;
13384 bool valid = grok_bslash_x(&p,
13388 PASS2, /* out warnings */
13389 (bool) RExC_strict,
13390 TRUE, /* Silence warnings
13395 RExC_parse = p; /* going to die anyway; point
13396 to exact spot of failure */
13401 if (ender < 0x100) {
13403 if (RExC_recode_x_to_native) {
13404 ender = LATIN1_TO_NATIVE(ender);
13409 REQUIRE_UTF8(flagp);
13415 ender = grok_bslash_c(*p++, PASS2);
13417 case '8': case '9': /* must be a backreference */
13419 /* we have an escape like \8 which cannot be an octal escape
13420 * so we exit the loop, and let the outer loop handle this
13421 * escape which may or may not be a legitimate backref. */
13423 case '1': case '2': case '3':case '4':
13424 case '5': case '6': case '7':
13425 /* When we parse backslash escapes there is ambiguity
13426 * between backreferences and octal escapes. Any escape
13427 * from \1 - \9 is a backreference, any multi-digit
13428 * escape which does not start with 0 and which when
13429 * evaluated as decimal could refer to an already
13430 * parsed capture buffer is a back reference. Anything
13433 * Note this implies that \118 could be interpreted as
13434 * 118 OR as "\11" . "8" depending on whether there
13435 * were 118 capture buffers defined already in the
13438 /* NOTE, RExC_npar is 1 more than the actual number of
13439 * parens we have seen so far, hence the < RExC_npar below. */
13441 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
13442 { /* Not to be treated as an octal constant, go
13450 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13452 ender = grok_oct(p, &numlen, &flags, NULL);
13453 if (ender > 0xff) {
13454 REQUIRE_UTF8(flagp);
13457 if (PASS2 /* like \08, \178 */
13459 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
13461 reg_warn_non_literal_string(
13463 form_short_octal_warning(p, numlen));
13469 FAIL("Trailing \\");
13472 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
13473 /* Include any left brace following the alpha to emphasize
13474 * that it could be part of an escape at some point
13476 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
13477 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
13479 goto normal_default;
13480 } /* End of switch on '\' */
13483 /* Currently we allow an lbrace at the start of a construct
13484 * without raising a warning. This is because we think we
13485 * will never want such a brace to be meant to be other
13486 * than taken literally. */
13487 if (len || (p > RExC_start && isALPHA_A(*(p - 1)))) {
13489 /* But, we raise a fatal warning otherwise, as the
13490 * deprecation cycle has come and gone. Except that it
13491 * turns out that some heavily-relied on upstream
13492 * software, notably GNU Autoconf, have failed to fix
13493 * their uses. For these, don't make it fatal unless
13494 * we anticipate using the '{' for something else.
13495 * This happens after any alpha, and for a looser {m,n}
13496 * quantifier specification */
13498 || ( p > parse_start + 1
13499 && isALPHA_A(*(p - 1))
13500 && *(p - 2) == '\\')
13501 || new_regcurly(p, RExC_end))
13503 RExC_parse = p + 1;
13504 vFAIL("Unescaped left brace in regex is "
13508 ckWARNregdep(p + 1,
13509 "Unescaped left brace in regex is "
13510 "deprecated here (and will be fatal "
13511 "in Perl 5.30), passed through");
13514 goto normal_default;
13517 if (PASS2 && p > RExC_parse && RExC_strict) {
13518 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
13521 default: /* A literal character */
13523 if (! UTF8_IS_INVARIANT(*p) && UTF) {
13525 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
13526 &numlen, UTF8_ALLOW_DEFAULT);
13532 } /* End of switch on the literal */
13534 /* Here, have looked at the literal character and <ender>
13535 * contains its ordinal, <p> points to the character after it.
13536 * We need to check if the next non-ignored thing is a
13537 * quantifier. Move <p> to after anything that should be
13538 * ignored, which, as a side effect, positions <p> for the next
13539 * loop iteration */
13540 skip_to_be_ignored_text(pRExC_state, &p,
13541 FALSE /* Don't force to /x */ );
13543 /* If the next thing is a quantifier, it applies to this
13544 * character only, which means that this character has to be in
13545 * its own node and can't just be appended to the string in an
13546 * existing node, so if there are already other characters in
13547 * the node, close the node with just them, and set up to do
13548 * this character again next time through, when it will be the
13549 * only thing in its new node */
13551 next_is_quantifier = LIKELY(p < RExC_end)
13552 && UNLIKELY(ISMULT2(p));
13554 if (next_is_quantifier && LIKELY(len)) {
13559 /* Ready to add 'ender' to the node */
13561 if (! FOLD) { /* The simple case, just append the literal */
13563 /* In the sizing pass, we need only the size of the
13564 * character we are appending, hence we can delay getting
13565 * its representation until PASS2. */
13567 if (UTF && ! UVCHR_IS_INVARIANT(ender)) {
13568 const STRLEN unilen = UVCHR_SKIP(ender);
13571 /* We have to subtract 1 just below (and again in
13572 * the corresponding PASS2 code) because the loop
13573 * increments <len> each time, as all but this path
13574 * (and one other) through it add a single byte to
13575 * the EXACTish node. But these paths would change
13576 * len to be the correct final value, so cancel out
13577 * the increment that follows */
13583 } else { /* PASS2 */
13585 if (UTF && ! UVCHR_IS_INVARIANT(ender)) {
13586 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
13587 len += (char *) new_s - s - 1;
13588 s = (char *) new_s;
13591 *(s++) = (char) ender;
13595 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
13597 /* Here are folding under /l, and the code point is
13598 * problematic. First, we know we can't simplify things */
13599 maybe_exact = FALSE;
13600 maybe_exactfu = FALSE;
13602 /* A problematic code point in this context means that its
13603 * fold isn't known until runtime, so we can't fold it now.
13604 * (The non-problematic code points are the above-Latin1
13605 * ones that fold to also all above-Latin1. Their folds
13606 * don't vary no matter what the locale is.) But here we
13607 * have characters whose fold depends on the locale.
13608 * Unlike the non-folding case above, we have to keep track
13609 * of these in the sizing pass, so that we can make sure we
13610 * don't split too-long nodes in the middle of a potential
13611 * multi-char fold. And unlike the regular fold case
13612 * handled in the else clauses below, we don't actually
13613 * fold and don't have special cases to consider. What we
13614 * do for both passes is the PASS2 code for non-folding */
13615 goto not_fold_common;
13617 else /* A regular FOLD code point */
13619 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13620 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13621 || UNICODE_DOT_DOT_VERSION > 0)
13622 /* See comments for join_exact() as to why we fold
13623 * this non-UTF at compile time */
13624 || ( node_type == EXACTFU
13625 && ender == LATIN_SMALL_LETTER_SHARP_S)
13628 /* Here, are folding and are not UTF-8 encoded; therefore
13629 * the character must be in the range 0-255, and is not /l
13630 * (Not /l because we already handled these under /l in
13631 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
13632 if (IS_IN_SOME_FOLD_L1(ender)) {
13633 maybe_exact = FALSE;
13635 /* See if the character's fold differs between /d and
13636 * /u. This includes the multi-char fold SHARP S to
13638 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
13639 RExC_seen_unfolded_sharp_s = 1;
13640 maybe_exactfu = FALSE;
13642 else if (maybe_exactfu
13643 && (PL_fold[ender] != PL_fold_latin1[ender]
13644 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
13645 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
13646 || UNICODE_DOT_DOT_VERSION > 0)
13648 && isALPHA_FOLD_EQ(ender, 's')
13649 && isALPHA_FOLD_EQ(*(s-1), 's'))
13652 maybe_exactfu = FALSE;
13656 /* Even when folding, we store just the input character, as
13657 * we have an array that finds its fold quickly */
13658 *(s++) = (char) ender;
13660 else { /* FOLD, and UTF (or sharp s) */
13661 /* Unlike the non-fold case, we do actually have to
13662 * calculate the results here in pass 1. This is for two
13663 * reasons, the folded length may be longer than the
13664 * unfolded, and we have to calculate how many EXACTish
13665 * nodes it will take; and we may run out of room in a node
13666 * in the middle of a potential multi-char fold, and have
13667 * to back off accordingly. */
13670 if (isASCII_uni(ender)) {
13671 folded = toFOLD(ender);
13672 *(s)++ = (U8) folded;
13677 folded = _to_uni_fold_flags(
13681 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
13682 ? FOLD_FLAGS_NOMIX_ASCII
13686 /* The loop increments <len> each time, as all but this
13687 * path (and one other) through it add a single byte to
13688 * the EXACTish node. But this one has changed len to
13689 * be the correct final value, so subtract one to
13690 * cancel out the increment that follows */
13691 len += foldlen - 1;
13693 /* If this node only contains non-folding code points so
13694 * far, see if this new one is also non-folding */
13696 if (folded != ender) {
13697 maybe_exact = FALSE;
13700 /* Here the fold is the original; we have to check
13701 * further to see if anything folds to it */
13702 if (_invlist_contains_cp(PL_utf8_foldable,
13705 maybe_exact = FALSE;
13712 if (next_is_quantifier) {
13714 /* Here, the next input is a quantifier, and to get here,
13715 * the current character is the only one in the node.
13716 * Also, here <len> doesn't include the final byte for this
13722 } /* End of loop through literal characters */
13724 /* Here we have either exhausted the input or ran out of room in
13725 * the node. (If we encountered a character that can't be in the
13726 * node, transfer is made directly to <loopdone>, and so we
13727 * wouldn't have fallen off the end of the loop.) In the latter
13728 * case, we artificially have to split the node into two, because
13729 * we just don't have enough space to hold everything. This
13730 * creates a problem if the final character participates in a
13731 * multi-character fold in the non-final position, as a match that
13732 * should have occurred won't, due to the way nodes are matched,
13733 * and our artificial boundary. So back off until we find a non-
13734 * problematic character -- one that isn't at the beginning or
13735 * middle of such a fold. (Either it doesn't participate in any
13736 * folds, or appears only in the final position of all the folds it
13737 * does participate in.) A better solution with far fewer false
13738 * positives, and that would fill the nodes more completely, would
13739 * be to actually have available all the multi-character folds to
13740 * test against, and to back-off only far enough to be sure that
13741 * this node isn't ending with a partial one. <upper_parse> is set
13742 * further below (if we need to reparse the node) to include just
13743 * up through that final non-problematic character that this code
13744 * identifies, so when it is set to less than the full node, we can
13745 * skip the rest of this */
13746 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
13748 const STRLEN full_len = len;
13750 assert(len >= MAX_NODE_STRING_SIZE);
13752 /* Here, <s> points to the final byte of the final character.
13753 * Look backwards through the string until find a non-
13754 * problematic character */
13758 /* This has no multi-char folds to non-UTF characters */
13759 if (ASCII_FOLD_RESTRICTED) {
13763 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
13767 if (! PL_NonL1NonFinalFold) {
13768 PL_NonL1NonFinalFold = _new_invlist_C_array(
13769 NonL1_Perl_Non_Final_Folds_invlist);
13772 /* Point to the first byte of the final character */
13773 s = (char *) utf8_hop((U8 *) s, -1);
13775 while (s >= s0) { /* Search backwards until find
13776 non-problematic char */
13777 if (UTF8_IS_INVARIANT(*s)) {
13779 /* There are no ascii characters that participate
13780 * in multi-char folds under /aa. In EBCDIC, the
13781 * non-ascii invariants are all control characters,
13782 * so don't ever participate in any folds. */
13783 if (ASCII_FOLD_RESTRICTED
13784 || ! IS_NON_FINAL_FOLD(*s))
13789 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
13790 if (! IS_NON_FINAL_FOLD(EIGHT_BIT_UTF8_TO_NATIVE(
13796 else if (! _invlist_contains_cp(
13797 PL_NonL1NonFinalFold,
13798 valid_utf8_to_uvchr((U8 *) s, NULL)))
13803 /* Here, the current character is problematic in that
13804 * it does occur in the non-final position of some
13805 * fold, so try the character before it, but have to
13806 * special case the very first byte in the string, so
13807 * we don't read outside the string */
13808 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
13809 } /* End of loop backwards through the string */
13811 /* If there were only problematic characters in the string,
13812 * <s> will point to before s0, in which case the length
13813 * should be 0, otherwise include the length of the
13814 * non-problematic character just found */
13815 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
13818 /* Here, have found the final character, if any, that is
13819 * non-problematic as far as ending the node without splitting
13820 * it across a potential multi-char fold. <len> contains the
13821 * number of bytes in the node up-to and including that
13822 * character, or is 0 if there is no such character, meaning
13823 * the whole node contains only problematic characters. In
13824 * this case, give up and just take the node as-is. We can't
13829 /* If the node ends in an 's' we make sure it stays EXACTF,
13830 * as if it turns into an EXACTFU, it could later get
13831 * joined with another 's' that would then wrongly match
13833 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
13835 maybe_exactfu = FALSE;
13839 /* Here, the node does contain some characters that aren't
13840 * problematic. If one such is the final character in the
13841 * node, we are done */
13842 if (len == full_len) {
13845 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
13847 /* If the final character is problematic, but the
13848 * penultimate is not, back-off that last character to
13849 * later start a new node with it */
13854 /* Here, the final non-problematic character is earlier
13855 * in the input than the penultimate character. What we do
13856 * is reparse from the beginning, going up only as far as
13857 * this final ok one, thus guaranteeing that the node ends
13858 * in an acceptable character. The reason we reparse is
13859 * that we know how far in the character is, but we don't
13860 * know how to correlate its position with the input parse.
13861 * An alternate implementation would be to build that
13862 * correlation as we go along during the original parse,
13863 * but that would entail extra work for every node, whereas
13864 * this code gets executed only when the string is too
13865 * large for the node, and the final two characters are
13866 * problematic, an infrequent occurrence. Yet another
13867 * possible strategy would be to save the tail of the
13868 * string, and the next time regatom is called, initialize
13869 * with that. The problem with this is that unless you
13870 * back off one more character, you won't be guaranteed
13871 * regatom will get called again, unless regbranch,
13872 * regpiece ... are also changed. If you do back off that
13873 * extra character, so that there is input guaranteed to
13874 * force calling regatom, you can't handle the case where
13875 * just the first character in the node is acceptable. I
13876 * (khw) decided to try this method which doesn't have that
13877 * pitfall; if performance issues are found, we can do a
13878 * combination of the current approach plus that one */
13884 } /* End of verifying node ends with an appropriate char */
13886 loopdone: /* Jumped to when encounters something that shouldn't be
13889 /* I (khw) don't know if you can get here with zero length, but the
13890 * old code handled this situation by creating a zero-length EXACT
13891 * node. Might as well be NOTHING instead */
13897 /* If 'maybe_exact' is still set here, means there are no
13898 * code points in the node that participate in folds;
13899 * similarly for 'maybe_exactfu' and code points that match
13900 * differently depending on UTF8ness of the target string
13901 * (for /u), or depending on locale for /l */
13907 else if (maybe_exactfu) {
13913 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
13914 FALSE /* Don't look to see if could
13915 be turned into an EXACT
13916 node, as we have already
13921 RExC_parse = p - 1;
13922 Set_Node_Cur_Length(ret, parse_start);
13925 /* len is STRLEN which is unsigned, need to copy to signed */
13928 vFAIL("Internal disaster");
13931 } /* End of label 'defchar:' */
13933 } /* End of giant switch on input character */
13935 /* Position parse to next real character */
13936 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13937 FALSE /* Don't force to /x */ );
13938 if (PASS2 && *RExC_parse == '{' && OP(ret) != SBOL && ! regcurly(RExC_parse)) {
13939 ckWARNregdep(RExC_parse + 1, "Unescaped left brace in regex is deprecated here (and will be fatal in Perl 5.30), passed through");
13947 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
13949 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
13950 * sets up the bitmap and any flags, removing those code points from the
13951 * inversion list, setting it to NULL should it become completely empty */
13953 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
13954 assert(PL_regkind[OP(node)] == ANYOF);
13956 ANYOF_BITMAP_ZERO(node);
13957 if (*invlist_ptr) {
13959 /* This gets set if we actually need to modify things */
13960 bool change_invlist = FALSE;
13964 /* Start looking through *invlist_ptr */
13965 invlist_iterinit(*invlist_ptr);
13966 while (invlist_iternext(*invlist_ptr, &start, &end)) {
13970 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
13971 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
13974 /* Quit if are above what we should change */
13975 if (start >= NUM_ANYOF_CODE_POINTS) {
13979 change_invlist = TRUE;
13981 /* Set all the bits in the range, up to the max that we are doing */
13982 high = (end < NUM_ANYOF_CODE_POINTS - 1)
13984 : NUM_ANYOF_CODE_POINTS - 1;
13985 for (i = start; i <= (int) high; i++) {
13986 if (! ANYOF_BITMAP_TEST(node, i)) {
13987 ANYOF_BITMAP_SET(node, i);
13991 invlist_iterfinish(*invlist_ptr);
13993 /* Done with loop; remove any code points that are in the bitmap from
13994 * *invlist_ptr; similarly for code points above the bitmap if we have
13995 * a flag to match all of them anyways */
13996 if (change_invlist) {
13997 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
13999 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
14000 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
14003 /* If have completely emptied it, remove it completely */
14004 if (_invlist_len(*invlist_ptr) == 0) {
14005 SvREFCNT_dec_NN(*invlist_ptr);
14006 *invlist_ptr = NULL;
14011 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
14012 Character classes ([:foo:]) can also be negated ([:^foo:]).
14013 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
14014 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
14015 but trigger failures because they are currently unimplemented. */
14017 #define POSIXCC_DONE(c) ((c) == ':')
14018 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
14019 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
14020 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
14022 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
14023 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
14024 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
14026 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
14028 /* 'posix_warnings' and 'warn_text' are names of variables in the following
14030 #define ADD_POSIX_WARNING(p, text) STMT_START { \
14031 if (posix_warnings) { \
14032 if (! RExC_warn_text ) RExC_warn_text = (AV *) sv_2mortal((SV *) newAV()); \
14033 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
14037 REPORT_LOCATION_ARGS(p))); \
14040 #define CLEAR_POSIX_WARNINGS() \
14042 if (posix_warnings && RExC_warn_text) \
14043 av_clear(RExC_warn_text); \
14046 #define CLEAR_POSIX_WARNINGS_AND_RETURN(ret) \
14048 CLEAR_POSIX_WARNINGS(); \
14053 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
14055 const char * const s, /* Where the putative posix class begins.
14056 Normally, this is one past the '['. This
14057 parameter exists so it can be somewhere
14058 besides RExC_parse. */
14059 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
14061 AV ** posix_warnings, /* Where to place any generated warnings, or
14063 const bool check_only /* Don't die if error */
14066 /* This parses what the caller thinks may be one of the three POSIX
14068 * 1) a character class, like [:blank:]
14069 * 2) a collating symbol, like [. .]
14070 * 3) an equivalence class, like [= =]
14071 * In the latter two cases, it croaks if it finds a syntactically legal
14072 * one, as these are not handled by Perl.
14074 * The main purpose is to look for a POSIX character class. It returns:
14075 * a) the class number
14076 * if it is a completely syntactically and semantically legal class.
14077 * 'updated_parse_ptr', if not NULL, is set to point to just after the
14078 * closing ']' of the class
14079 * b) OOB_NAMEDCLASS
14080 * if it appears that one of the three POSIX constructs was meant, but
14081 * its specification was somehow defective. 'updated_parse_ptr', if
14082 * not NULL, is set to point to the character just after the end
14083 * character of the class. See below for handling of warnings.
14084 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
14085 * if it doesn't appear that a POSIX construct was intended.
14086 * 'updated_parse_ptr' is not changed. No warnings nor errors are
14089 * In b) there may be errors or warnings generated. If 'check_only' is
14090 * TRUE, then any errors are discarded. Warnings are returned to the
14091 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
14092 * instead it is NULL, warnings are suppressed. This is done in all
14093 * passes. The reason for this is that the rest of the parsing is heavily
14094 * dependent on whether this routine found a valid posix class or not. If
14095 * it did, the closing ']' is absorbed as part of the class. If no class,
14096 * or an invalid one is found, any ']' will be considered the terminator of
14097 * the outer bracketed character class, leading to very different results.
14098 * In particular, a '(?[ ])' construct will likely have a syntax error if
14099 * the class is parsed other than intended, and this will happen in pass1,
14100 * before the warnings would normally be output. This mechanism allows the
14101 * caller to output those warnings in pass1 just before dieing, giving a
14102 * much better clue as to what is wrong.
14104 * The reason for this function, and its complexity is that a bracketed
14105 * character class can contain just about anything. But it's easy to
14106 * mistype the very specific posix class syntax but yielding a valid
14107 * regular bracketed class, so it silently gets compiled into something
14108 * quite unintended.
14110 * The solution adopted here maintains backward compatibility except that
14111 * it adds a warning if it looks like a posix class was intended but
14112 * improperly specified. The warning is not raised unless what is input
14113 * very closely resembles one of the 14 legal posix classes. To do this,
14114 * it uses fuzzy parsing. It calculates how many single-character edits it
14115 * would take to transform what was input into a legal posix class. Only
14116 * if that number is quite small does it think that the intention was a
14117 * posix class. Obviously these are heuristics, and there will be cases
14118 * where it errs on one side or another, and they can be tweaked as
14119 * experience informs.
14121 * The syntax for a legal posix class is:
14123 * qr/(?xa: \[ : \^? [[:lower:]]{4,6} : \] )/
14125 * What this routine considers syntactically to be an intended posix class
14126 * is this (the comments indicate some restrictions that the pattern
14129 * qr/(?x: \[? # The left bracket, possibly
14131 * \h* # possibly followed by blanks
14132 * (?: \^ \h* )? # possibly a misplaced caret
14133 * [:;]? # The opening class character,
14134 * # possibly omitted. A typo
14135 * # semi-colon can also be used.
14137 * \^? # possibly a correctly placed
14138 * # caret, but not if there was also
14139 * # a misplaced one
14141 * .{3,15} # The class name. If there are
14142 * # deviations from the legal syntax,
14143 * # its edit distance must be close
14144 * # to a real class name in order
14145 * # for it to be considered to be
14146 * # an intended posix class.
14148 * [[:punct:]]? # The closing class character,
14149 * # possibly omitted. If not a colon
14150 * # nor semi colon, the class name
14151 * # must be even closer to a valid
14154 * \]? # The right bracket, possibly
14158 * In the above, \h must be ASCII-only.
14160 * These are heuristics, and can be tweaked as field experience dictates.
14161 * There will be cases when someone didn't intend to specify a posix class
14162 * that this warns as being so. The goal is to minimize these, while
14163 * maximizing the catching of things intended to be a posix class that
14164 * aren't parsed as such.
14168 const char * const e = RExC_end;
14169 unsigned complement = 0; /* If to complement the class */
14170 bool found_problem = FALSE; /* Assume OK until proven otherwise */
14171 bool has_opening_bracket = FALSE;
14172 bool has_opening_colon = FALSE;
14173 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
14175 const char * possible_end = NULL; /* used for a 2nd parse pass */
14176 const char* name_start; /* ptr to class name first char */
14178 /* If the number of single-character typos the input name is away from a
14179 * legal name is no more than this number, it is considered to have meant
14180 * the legal name */
14181 int max_distance = 2;
14183 /* to store the name. The size determines the maximum length before we
14184 * decide that no posix class was intended. Should be at least
14185 * sizeof("alphanumeric") */
14187 STATIC_ASSERT_DECL(C_ARRAY_LENGTH(input_text) >= sizeof "alphanumeric");
14189 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
14191 CLEAR_POSIX_WARNINGS();
14194 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
14197 if (*(p - 1) != '[') {
14198 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
14199 found_problem = TRUE;
14202 has_opening_bracket = TRUE;
14205 /* They could be confused and think you can put spaces between the
14208 found_problem = TRUE;
14212 } while (p < e && isBLANK(*p));
14214 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14217 /* For [. .] and [= =]. These are quite different internally from [: :],
14218 * so they are handled separately. */
14219 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
14220 and 1 for at least one char in it
14223 const char open_char = *p;
14224 const char * temp_ptr = p + 1;
14226 /* These two constructs are not handled by perl, and if we find a
14227 * syntactically valid one, we croak. khw, who wrote this code, finds
14228 * this explanation of them very unclear:
14229 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
14230 * And searching the rest of the internet wasn't very helpful either.
14231 * It looks like just about any byte can be in these constructs,
14232 * depending on the locale. But unless the pattern is being compiled
14233 * under /l, which is very rare, Perl runs under the C or POSIX locale.
14234 * In that case, it looks like [= =] isn't allowed at all, and that
14235 * [. .] could be any single code point, but for longer strings the
14236 * constituent characters would have to be the ASCII alphabetics plus
14237 * the minus-hyphen. Any sensible locale definition would limit itself
14238 * to these. And any portable one definitely should. Trying to parse
14239 * the general case is a nightmare (see [perl #127604]). So, this code
14240 * looks only for interiors of these constructs that match:
14242 * Using \w relaxes the apparent rules a little, without adding much
14243 * danger of mistaking something else for one of these constructs.
14245 * [. .] in some implementations described on the internet is usable to
14246 * escape a character that otherwise is special in bracketed character
14247 * classes. For example [.].] means a literal right bracket instead of
14248 * the ending of the class
14250 * [= =] can legitimately contain a [. .] construct, but we don't
14251 * handle this case, as that [. .] construct will later get parsed
14252 * itself and croak then. And [= =] is checked for even when not under
14253 * /l, as Perl has long done so.
14255 * The code below relies on there being a trailing NUL, so it doesn't
14256 * have to keep checking if the parse ptr < e.
14258 if (temp_ptr[1] == open_char) {
14261 else while ( temp_ptr < e
14262 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
14267 if (*temp_ptr == open_char) {
14269 if (*temp_ptr == ']') {
14271 if (! found_problem && ! check_only) {
14272 RExC_parse = (char *) temp_ptr;
14273 vFAIL3("POSIX syntax [%c %c] is reserved for future "
14274 "extensions", open_char, open_char);
14277 /* Here, the syntax wasn't completely valid, or else the call
14278 * is to check-only */
14279 if (updated_parse_ptr) {
14280 *updated_parse_ptr = (char *) temp_ptr;
14283 CLEAR_POSIX_WARNINGS_AND_RETURN(OOB_NAMEDCLASS);
14287 /* If we find something that started out to look like one of these
14288 * constructs, but isn't, we continue below so that it can be checked
14289 * for being a class name with a typo of '.' or '=' instead of a colon.
14293 /* Here, we think there is a possibility that a [: :] class was meant, and
14294 * we have the first real character. It could be they think the '^' comes
14297 found_problem = TRUE;
14298 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
14303 found_problem = TRUE;
14307 } while (p < e && isBLANK(*p));
14309 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14313 /* But the first character should be a colon, which they could have easily
14314 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
14315 * distinguish from a colon, so treat that as a colon). */
14318 has_opening_colon = TRUE;
14320 else if (*p == ';') {
14321 found_problem = TRUE;
14323 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14324 has_opening_colon = TRUE;
14327 found_problem = TRUE;
14328 ADD_POSIX_WARNING(p, "there must be a starting ':'");
14330 /* Consider an initial punctuation (not one of the recognized ones) to
14331 * be a left terminator */
14332 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
14337 /* They may think that you can put spaces between the components */
14339 found_problem = TRUE;
14343 } while (p < e && isBLANK(*p));
14345 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14350 /* We consider something like [^:^alnum:]] to not have been intended to
14351 * be a posix class, but XXX maybe we should */
14353 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14360 /* Again, they may think that you can put spaces between the components */
14362 found_problem = TRUE;
14366 } while (p < e && isBLANK(*p));
14368 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14373 /* XXX This ']' may be a typo, and something else was meant. But
14374 * treating it as such creates enough complications, that that
14375 * possibility isn't currently considered here. So we assume that the
14376 * ']' is what is intended, and if we've already found an initial '[',
14377 * this leaves this construct looking like [:] or [:^], which almost
14378 * certainly weren't intended to be posix classes */
14379 if (has_opening_bracket) {
14380 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14383 /* But this function can be called when we parse the colon for
14384 * something like qr/[alpha:]]/, so we back up to look for the
14389 found_problem = TRUE;
14390 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14392 else if (*p != ':') {
14394 /* XXX We are currently very restrictive here, so this code doesn't
14395 * consider the possibility that, say, /[alpha.]]/ was intended to
14396 * be a posix class. */
14397 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14400 /* Here we have something like 'foo:]'. There was no initial colon,
14401 * and we back up over 'foo. XXX Unlike the going forward case, we
14402 * don't handle typos of non-word chars in the middle */
14403 has_opening_colon = FALSE;
14406 while (p > RExC_start && isWORDCHAR(*p)) {
14411 /* Here, we have positioned ourselves to where we think the first
14412 * character in the potential class is */
14415 /* Now the interior really starts. There are certain key characters that
14416 * can end the interior, or these could just be typos. To catch both
14417 * cases, we may have to do two passes. In the first pass, we keep on
14418 * going unless we come to a sequence that matches
14419 * qr/ [[:punct:]] [[:blank:]]* \] /xa
14420 * This means it takes a sequence to end the pass, so two typos in a row if
14421 * that wasn't what was intended. If the class is perfectly formed, just
14422 * this one pass is needed. We also stop if there are too many characters
14423 * being accumulated, but this number is deliberately set higher than any
14424 * real class. It is set high enough so that someone who thinks that
14425 * 'alphanumeric' is a correct name would get warned that it wasn't.
14426 * While doing the pass, we keep track of where the key characters were in
14427 * it. If we don't find an end to the class, and one of the key characters
14428 * was found, we redo the pass, but stop when we get to that character.
14429 * Thus the key character was considered a typo in the first pass, but a
14430 * terminator in the second. If two key characters are found, we stop at
14431 * the second one in the first pass. Again this can miss two typos, but
14432 * catches a single one
14434 * In the first pass, 'possible_end' starts as NULL, and then gets set to
14435 * point to the first key character. For the second pass, it starts as -1.
14441 bool has_blank = FALSE;
14442 bool has_upper = FALSE;
14443 bool has_terminating_colon = FALSE;
14444 bool has_terminating_bracket = FALSE;
14445 bool has_semi_colon = FALSE;
14446 unsigned int name_len = 0;
14447 int punct_count = 0;
14451 /* Squeeze out blanks when looking up the class name below */
14452 if (isBLANK(*p) ) {
14454 found_problem = TRUE;
14459 /* The name will end with a punctuation */
14461 const char * peek = p + 1;
14463 /* Treat any non-']' punctuation followed by a ']' (possibly
14464 * with intervening blanks) as trying to terminate the class.
14465 * ']]' is very likely to mean a class was intended (but
14466 * missing the colon), but the warning message that gets
14467 * generated shows the error position better if we exit the
14468 * loop at the bottom (eventually), so skip it here. */
14470 if (peek < e && isBLANK(*peek)) {
14472 found_problem = TRUE;
14475 } while (peek < e && isBLANK(*peek));
14478 if (peek < e && *peek == ']') {
14479 has_terminating_bracket = TRUE;
14481 has_terminating_colon = TRUE;
14483 else if (*p == ';') {
14484 has_semi_colon = TRUE;
14485 has_terminating_colon = TRUE;
14488 found_problem = TRUE;
14495 /* Here we have punctuation we thought didn't end the class.
14496 * Keep track of the position of the key characters that are
14497 * more likely to have been class-enders */
14498 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
14500 /* Allow just one such possible class-ender not actually
14501 * ending the class. */
14502 if (possible_end) {
14508 /* If we have too many punctuation characters, no use in
14510 if (++punct_count > max_distance) {
14514 /* Treat the punctuation as a typo. */
14515 input_text[name_len++] = *p;
14518 else if (isUPPER(*p)) { /* Use lowercase for lookup */
14519 input_text[name_len++] = toLOWER(*p);
14521 found_problem = TRUE;
14523 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
14524 input_text[name_len++] = *p;
14528 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
14532 /* The declaration of 'input_text' is how long we allow a potential
14533 * class name to be, before saying they didn't mean a class name at
14535 if (name_len >= C_ARRAY_LENGTH(input_text)) {
14540 /* We get to here when the possible class name hasn't been properly
14541 * terminated before:
14542 * 1) we ran off the end of the pattern; or
14543 * 2) found two characters, each of which might have been intended to
14544 * be the name's terminator
14545 * 3) found so many punctuation characters in the purported name,
14546 * that the edit distance to a valid one is exceeded
14547 * 4) we decided it was more characters than anyone could have
14548 * intended to be one. */
14550 found_problem = TRUE;
14552 /* In the final two cases, we know that looking up what we've
14553 * accumulated won't lead to a match, even a fuzzy one. */
14554 if ( name_len >= C_ARRAY_LENGTH(input_text)
14555 || punct_count > max_distance)
14557 /* If there was an intermediate key character that could have been
14558 * an intended end, redo the parse, but stop there */
14559 if (possible_end && possible_end != (char *) -1) {
14560 possible_end = (char *) -1; /* Special signal value to say
14561 we've done a first pass */
14566 /* Otherwise, it can't have meant to have been a class */
14567 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14570 /* If we ran off the end, and the final character was a punctuation
14571 * one, back up one, to look at that final one just below. Later, we
14572 * will restore the parse pointer if appropriate */
14573 if (name_len && p == e && isPUNCT(*(p-1))) {
14578 if (p < e && isPUNCT(*p)) {
14580 has_terminating_bracket = TRUE;
14582 /* If this is a 2nd ']', and the first one is just below this
14583 * one, consider that to be the real terminator. This gives a
14584 * uniform and better positioning for the warning message */
14586 && possible_end != (char *) -1
14587 && *possible_end == ']'
14588 && name_len && input_text[name_len - 1] == ']')
14593 /* And this is actually equivalent to having done the 2nd
14594 * pass now, so set it to not try again */
14595 possible_end = (char *) -1;
14600 has_terminating_colon = TRUE;
14602 else if (*p == ';') {
14603 has_semi_colon = TRUE;
14604 has_terminating_colon = TRUE;
14612 /* Here, we have a class name to look up. We can short circuit the
14613 * stuff below for short names that can't possibly be meant to be a
14614 * class name. (We can do this on the first pass, as any second pass
14615 * will yield an even shorter name) */
14616 if (name_len < 3) {
14617 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14620 /* Find which class it is. Initially switch on the length of the name.
14622 switch (name_len) {
14624 if (memEQs(name_start, 4, "word")) {
14625 /* this is not POSIX, this is the Perl \w */
14626 class_number = ANYOF_WORDCHAR;
14630 /* Names all of length 5: alnum alpha ascii blank cntrl digit
14631 * graph lower print punct space upper
14632 * Offset 4 gives the best switch position. */
14633 switch (name_start[4]) {
14635 if (memBEGINs(name_start, 5, "alph")) /* alpha */
14636 class_number = ANYOF_ALPHA;
14639 if (memBEGINs(name_start, 5, "spac")) /* space */
14640 class_number = ANYOF_SPACE;
14643 if (memBEGINs(name_start, 5, "grap")) /* graph */
14644 class_number = ANYOF_GRAPH;
14647 if (memBEGINs(name_start, 5, "asci")) /* ascii */
14648 class_number = ANYOF_ASCII;
14651 if (memBEGINs(name_start, 5, "blan")) /* blank */
14652 class_number = ANYOF_BLANK;
14655 if (memBEGINs(name_start, 5, "cntr")) /* cntrl */
14656 class_number = ANYOF_CNTRL;
14659 if (memBEGINs(name_start, 5, "alnu")) /* alnum */
14660 class_number = ANYOF_ALPHANUMERIC;
14663 if (memBEGINs(name_start, 5, "lowe")) /* lower */
14664 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
14665 else if (memBEGINs(name_start, 5, "uppe")) /* upper */
14666 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
14669 if (memBEGINs(name_start, 5, "digi")) /* digit */
14670 class_number = ANYOF_DIGIT;
14671 else if (memBEGINs(name_start, 5, "prin")) /* print */
14672 class_number = ANYOF_PRINT;
14673 else if (memBEGINs(name_start, 5, "punc")) /* punct */
14674 class_number = ANYOF_PUNCT;
14679 if (memEQs(name_start, 6, "xdigit"))
14680 class_number = ANYOF_XDIGIT;
14684 /* If the name exactly matches a posix class name the class number will
14685 * here be set to it, and the input almost certainly was meant to be a
14686 * posix class, so we can skip further checking. If instead the syntax
14687 * is exactly correct, but the name isn't one of the legal ones, we
14688 * will return that as an error below. But if neither of these apply,
14689 * it could be that no posix class was intended at all, or that one
14690 * was, but there was a typo. We tease these apart by doing fuzzy
14691 * matching on the name */
14692 if (class_number == OOB_NAMEDCLASS && found_problem) {
14693 const UV posix_names[][6] = {
14694 { 'a', 'l', 'n', 'u', 'm' },
14695 { 'a', 'l', 'p', 'h', 'a' },
14696 { 'a', 's', 'c', 'i', 'i' },
14697 { 'b', 'l', 'a', 'n', 'k' },
14698 { 'c', 'n', 't', 'r', 'l' },
14699 { 'd', 'i', 'g', 'i', 't' },
14700 { 'g', 'r', 'a', 'p', 'h' },
14701 { 'l', 'o', 'w', 'e', 'r' },
14702 { 'p', 'r', 'i', 'n', 't' },
14703 { 'p', 'u', 'n', 'c', 't' },
14704 { 's', 'p', 'a', 'c', 'e' },
14705 { 'u', 'p', 'p', 'e', 'r' },
14706 { 'w', 'o', 'r', 'd' },
14707 { 'x', 'd', 'i', 'g', 'i', 't' }
14709 /* The names of the above all have added NULs to make them the same
14710 * size, so we need to also have the real lengths */
14711 const UV posix_name_lengths[] = {
14712 sizeof("alnum") - 1,
14713 sizeof("alpha") - 1,
14714 sizeof("ascii") - 1,
14715 sizeof("blank") - 1,
14716 sizeof("cntrl") - 1,
14717 sizeof("digit") - 1,
14718 sizeof("graph") - 1,
14719 sizeof("lower") - 1,
14720 sizeof("print") - 1,
14721 sizeof("punct") - 1,
14722 sizeof("space") - 1,
14723 sizeof("upper") - 1,
14724 sizeof("word") - 1,
14725 sizeof("xdigit")- 1
14728 int temp_max = max_distance; /* Use a temporary, so if we
14729 reparse, we haven't changed the
14732 /* Use a smaller max edit distance if we are missing one of the
14734 if ( has_opening_bracket + has_opening_colon < 2
14735 || has_terminating_bracket + has_terminating_colon < 2)
14740 /* See if the input name is close to a legal one */
14741 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
14743 /* Short circuit call if the lengths are too far apart to be
14745 if (abs( (int) (name_len - posix_name_lengths[i]))
14751 if (edit_distance(input_text,
14754 posix_name_lengths[i],
14758 { /* If it is close, it probably was intended to be a class */
14759 goto probably_meant_to_be;
14763 /* Here the input name is not close enough to a valid class name
14764 * for us to consider it to be intended to be a posix class. If
14765 * we haven't already done so, and the parse found a character that
14766 * could have been terminators for the name, but which we absorbed
14767 * as typos during the first pass, repeat the parse, signalling it
14768 * to stop at that character */
14769 if (possible_end && possible_end != (char *) -1) {
14770 possible_end = (char *) -1;
14775 /* Here neither pass found a close-enough class name */
14776 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
14779 probably_meant_to_be:
14781 /* Here we think that a posix specification was intended. Update any
14783 if (updated_parse_ptr) {
14784 *updated_parse_ptr = (char *) p;
14787 /* If a posix class name was intended but incorrectly specified, we
14788 * output or return the warnings */
14789 if (found_problem) {
14791 /* We set flags for these issues in the parse loop above instead of
14792 * adding them to the list of warnings, because we can parse it
14793 * twice, and we only want one warning instance */
14795 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
14798 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
14800 if (has_semi_colon) {
14801 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
14803 else if (! has_terminating_colon) {
14804 ADD_POSIX_WARNING(p, "there is no terminating ':'");
14806 if (! has_terminating_bracket) {
14807 ADD_POSIX_WARNING(p, "there is no terminating ']'");
14810 if (posix_warnings && RExC_warn_text && av_top_index(RExC_warn_text) > -1) {
14811 *posix_warnings = RExC_warn_text;
14814 else if (class_number != OOB_NAMEDCLASS) {
14815 /* If it is a known class, return the class. The class number
14816 * #defines are structured so each complement is +1 to the normal
14818 CLEAR_POSIX_WARNINGS_AND_RETURN(class_number + complement);
14820 else if (! check_only) {
14822 /* Here, it is an unrecognized class. This is an error (unless the
14823 * call is to check only, which we've already handled above) */
14824 const char * const complement_string = (complement)
14827 RExC_parse = (char *) p;
14828 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
14830 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
14834 return OOB_NAMEDCLASS;
14836 #undef ADD_POSIX_WARNING
14838 STATIC unsigned int
14839 S_regex_set_precedence(const U8 my_operator) {
14841 /* Returns the precedence in the (?[...]) construct of the input operator,
14842 * specified by its character representation. The precedence follows
14843 * general Perl rules, but it extends this so that ')' and ']' have (low)
14844 * precedence even though they aren't really operators */
14846 switch (my_operator) {
14862 NOT_REACHED; /* NOTREACHED */
14863 return 0; /* Silence compiler warning */
14867 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
14868 I32 *flagp, U32 depth,
14869 char * const oregcomp_parse)
14871 /* Handle the (?[...]) construct to do set operations */
14873 U8 curchar; /* Current character being parsed */
14874 UV start, end; /* End points of code point ranges */
14875 SV* final = NULL; /* The end result inversion list */
14876 SV* result_string; /* 'final' stringified */
14877 AV* stack; /* stack of operators and operands not yet
14879 AV* fence_stack = NULL; /* A stack containing the positions in
14880 'stack' of where the undealt-with left
14881 parens would be if they were actually
14883 /* The 'volatile' is a workaround for an optimiser bug
14884 * in Solaris Studio 12.3. See RT #127455 */
14885 volatile IV fence = 0; /* Position of where most recent undealt-
14886 with left paren in stack is; -1 if none.
14888 STRLEN len; /* Temporary */
14889 regnode* node; /* Temporary, and final regnode returned by
14891 const bool save_fold = FOLD; /* Temporary */
14892 char *save_end, *save_parse; /* Temporaries */
14893 const bool in_locale = LOC; /* we turn off /l during processing */
14894 AV* posix_warnings = NULL;
14896 GET_RE_DEBUG_FLAGS_DECL;
14898 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
14901 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
14904 REQUIRE_UNI_RULES(flagp, NULL); /* The use of this operator implies /u.
14905 This is required so that the compile
14906 time values are valid in all runtime
14909 /* This will return only an ANYOF regnode, or (unlikely) something smaller
14910 * (such as EXACT). Thus we can skip most everything if just sizing. We
14911 * call regclass to handle '[]' so as to not have to reinvent its parsing
14912 * rules here (throwing away the size it computes each time). And, we exit
14913 * upon an unescaped ']' that isn't one ending a regclass. To do both
14914 * these things, we need to realize that something preceded by a backslash
14915 * is escaped, so we have to keep track of backslashes */
14917 UV depth = 0; /* how many nested (?[...]) constructs */
14919 while (RExC_parse < RExC_end) {
14920 SV* current = NULL;
14922 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14923 TRUE /* Force /x */ );
14925 switch (*RExC_parse) {
14927 if (RExC_parse[1] == '[') depth++, RExC_parse++;
14932 /* Skip past this, so the next character gets skipped, after
14935 if (*RExC_parse == 'c') {
14936 /* Skip the \cX notation for control characters */
14937 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
14943 /* See if this is a [:posix:] class. */
14944 bool is_posix_class = (OOB_NAMEDCLASS
14945 < handle_possible_posix(pRExC_state,
14949 TRUE /* checking only */));
14950 /* If it is a posix class, leave the parse pointer at the
14951 * '[' to fool regclass() into thinking it is part of a
14952 * '[[:posix:]]'. */
14953 if (! is_posix_class) {
14957 /* regclass() can only return RESTART_PASS1 and NEED_UTF8
14958 * if multi-char folds are allowed. */
14959 if (!regclass(pRExC_state, flagp,depth+1,
14960 is_posix_class, /* parse the whole char
14961 class only if not a
14963 FALSE, /* don't allow multi-char folds */
14964 TRUE, /* silence non-portable warnings. */
14966 FALSE, /* Require return to be an ANYOF */
14970 FAIL2("panic: regclass returned NULL to handle_sets, "
14971 "flags=%#" UVxf, (UV) *flagp);
14973 /* function call leaves parse pointing to the ']', except
14974 * if we faked it */
14975 if (is_posix_class) {
14979 SvREFCNT_dec(current); /* In case it returned something */
14984 if (depth--) break;
14986 if (*RExC_parse == ')') {
14987 node = reganode(pRExC_state, ANYOF, 0);
14988 RExC_size += ANYOF_SKIP;
14989 nextchar(pRExC_state);
14990 Set_Node_Length(node,
14991 RExC_parse - oregcomp_parse + 1); /* MJD */
14993 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
15001 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
15005 /* We output the messages even if warnings are off, because we'll fail
15006 * the very next thing, and these give a likely diagnosis for that */
15007 if (posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
15008 output_or_return_posix_warnings(pRExC_state, posix_warnings, NULL);
15011 FAIL("Syntax error in (?[...])");
15014 /* Pass 2 only after this. */
15015 Perl_ck_warner_d(aTHX_
15016 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
15017 "The regex_sets feature is experimental" REPORT_LOCATION,
15018 REPORT_LOCATION_ARGS(RExC_parse));
15020 /* Everything in this construct is a metacharacter. Operands begin with
15021 * either a '\' (for an escape sequence), or a '[' for a bracketed
15022 * character class. Any other character should be an operator, or
15023 * parenthesis for grouping. Both types of operands are handled by calling
15024 * regclass() to parse them. It is called with a parameter to indicate to
15025 * return the computed inversion list. The parsing here is implemented via
15026 * a stack. Each entry on the stack is a single character representing one
15027 * of the operators; or else a pointer to an operand inversion list. */
15029 #define IS_OPERATOR(a) SvIOK(a)
15030 #define IS_OPERAND(a) (! IS_OPERATOR(a))
15032 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
15033 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
15034 * with pronouncing it called it Reverse Polish instead, but now that YOU
15035 * know how to pronounce it you can use the correct term, thus giving due
15036 * credit to the person who invented it, and impressing your geek friends.
15037 * Wikipedia says that the pronounciation of "Ł" has been changing so that
15038 * it is now more like an English initial W (as in wonk) than an L.)
15040 * This means that, for example, 'a | b & c' is stored on the stack as
15048 * where the numbers in brackets give the stack [array] element number.
15049 * In this implementation, parentheses are not stored on the stack.
15050 * Instead a '(' creates a "fence" so that the part of the stack below the
15051 * fence is invisible except to the corresponding ')' (this allows us to
15052 * replace testing for parens, by using instead subtraction of the fence
15053 * position). As new operands are processed they are pushed onto the stack
15054 * (except as noted in the next paragraph). New operators of higher
15055 * precedence than the current final one are inserted on the stack before
15056 * the lhs operand (so that when the rhs is pushed next, everything will be
15057 * in the correct positions shown above. When an operator of equal or
15058 * lower precedence is encountered in parsing, all the stacked operations
15059 * of equal or higher precedence are evaluated, leaving the result as the
15060 * top entry on the stack. This makes higher precedence operations
15061 * evaluate before lower precedence ones, and causes operations of equal
15062 * precedence to left associate.
15064 * The only unary operator '!' is immediately pushed onto the stack when
15065 * encountered. When an operand is encountered, if the top of the stack is
15066 * a '!", the complement is immediately performed, and the '!' popped. The
15067 * resulting value is treated as a new operand, and the logic in the
15068 * previous paragraph is executed. Thus in the expression
15070 * the stack looks like
15076 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
15083 * A ')' is treated as an operator with lower precedence than all the
15084 * aforementioned ones, which causes all operations on the stack above the
15085 * corresponding '(' to be evaluated down to a single resultant operand.
15086 * Then the fence for the '(' is removed, and the operand goes through the
15087 * algorithm above, without the fence.
15089 * A separate stack is kept of the fence positions, so that the position of
15090 * the latest so-far unbalanced '(' is at the top of it.
15092 * The ']' ending the construct is treated as the lowest operator of all,
15093 * so that everything gets evaluated down to a single operand, which is the
15096 sv_2mortal((SV *)(stack = newAV()));
15097 sv_2mortal((SV *)(fence_stack = newAV()));
15099 while (RExC_parse < RExC_end) {
15100 I32 top_index; /* Index of top-most element in 'stack' */
15101 SV** top_ptr; /* Pointer to top 'stack' element */
15102 SV* current = NULL; /* To contain the current inversion list
15104 SV* only_to_avoid_leaks;
15106 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
15107 TRUE /* Force /x */ );
15108 if (RExC_parse >= RExC_end) {
15109 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
15112 curchar = UCHARAT(RExC_parse);
15116 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15117 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
15118 DEBUG_U(dump_regex_sets_structures(pRExC_state,
15119 stack, fence, fence_stack));
15122 top_index = av_tindex_skip_len_mg(stack);
15125 SV** stacked_ptr; /* Ptr to something already on 'stack' */
15126 char stacked_operator; /* The topmost operator on the 'stack'. */
15127 SV* lhs; /* Operand to the left of the operator */
15128 SV* rhs; /* Operand to the right of the operator */
15129 SV* fence_ptr; /* Pointer to top element of the fence
15134 if ( RExC_parse < RExC_end - 1
15135 && (UCHARAT(RExC_parse + 1) == '?'))
15137 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
15138 * This happens when we have some thing like
15140 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
15142 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
15144 * Here we would be handling the interpolated
15145 * '$thai_or_lao'. We handle this by a recursive call to
15146 * ourselves which returns the inversion list the
15147 * interpolated expression evaluates to. We use the flags
15148 * from the interpolated pattern. */
15149 U32 save_flags = RExC_flags;
15150 const char * save_parse;
15152 RExC_parse += 2; /* Skip past the '(?' */
15153 save_parse = RExC_parse;
15155 /* Parse any flags for the '(?' */
15156 parse_lparen_question_flags(pRExC_state);
15158 if (RExC_parse == save_parse /* Makes sure there was at
15159 least one flag (or else
15160 this embedding wasn't
15162 || RExC_parse >= RExC_end - 4
15163 || UCHARAT(RExC_parse) != ':'
15164 || UCHARAT(++RExC_parse) != '('
15165 || UCHARAT(++RExC_parse) != '?'
15166 || UCHARAT(++RExC_parse) != '[')
15169 /* In combination with the above, this moves the
15170 * pointer to the point just after the first erroneous
15171 * character (or if there are no flags, to where they
15172 * should have been) */
15173 if (RExC_parse >= RExC_end - 4) {
15174 RExC_parse = RExC_end;
15176 else if (RExC_parse != save_parse) {
15177 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15179 vFAIL("Expecting '(?flags:(?[...'");
15182 /* Recurse, with the meat of the embedded expression */
15184 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
15185 depth+1, oregcomp_parse);
15187 /* Here, 'current' contains the embedded expression's
15188 * inversion list, and RExC_parse points to the trailing
15189 * ']'; the next character should be the ')' */
15191 assert(UCHARAT(RExC_parse) == ')');
15193 /* Then the ')' matching the original '(' handled by this
15194 * case: statement */
15196 assert(UCHARAT(RExC_parse) == ')');
15199 RExC_flags = save_flags;
15200 goto handle_operand;
15203 /* A regular '('. Look behind for illegal syntax */
15204 if (top_index - fence >= 0) {
15205 /* If the top entry on the stack is an operator, it had
15206 * better be a '!', otherwise the entry below the top
15207 * operand should be an operator */
15208 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
15209 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
15210 || ( IS_OPERAND(*top_ptr)
15211 && ( top_index - fence < 1
15212 || ! (stacked_ptr = av_fetch(stack,
15215 || ! IS_OPERATOR(*stacked_ptr))))
15218 vFAIL("Unexpected '(' with no preceding operator");
15222 /* Stack the position of this undealt-with left paren */
15223 av_push(fence_stack, newSViv(fence));
15224 fence = top_index + 1;
15228 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15229 * multi-char folds are allowed. */
15230 if (!regclass(pRExC_state, flagp,depth+1,
15231 TRUE, /* means parse just the next thing */
15232 FALSE, /* don't allow multi-char folds */
15233 FALSE, /* don't silence non-portable warnings. */
15235 FALSE, /* Require return to be an ANYOF */
15239 FAIL2("panic: regclass returned NULL to handle_sets, "
15240 "flags=%#" UVxf, (UV) *flagp);
15243 /* regclass() will return with parsing just the \ sequence,
15244 * leaving the parse pointer at the next thing to parse */
15246 goto handle_operand;
15248 case '[': /* Is a bracketed character class */
15250 /* See if this is a [:posix:] class. */
15251 bool is_posix_class = (OOB_NAMEDCLASS
15252 < handle_possible_posix(pRExC_state,
15256 TRUE /* checking only */));
15257 /* If it is a posix class, leave the parse pointer at the '['
15258 * to fool regclass() into thinking it is part of a
15259 * '[[:posix:]]'. */
15260 if (! is_posix_class) {
15264 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
15265 * multi-char folds are allowed. */
15266 if (!regclass(pRExC_state, flagp,depth+1,
15267 is_posix_class, /* parse the whole char
15268 class only if not a
15270 FALSE, /* don't allow multi-char folds */
15271 TRUE, /* silence non-portable warnings. */
15273 FALSE, /* Require return to be an ANYOF */
15278 FAIL2("panic: regclass returned NULL to handle_sets, "
15279 "flags=%#" UVxf, (UV) *flagp);
15282 /* function call leaves parse pointing to the ']', except if we
15284 if (is_posix_class) {
15288 goto handle_operand;
15292 if (top_index >= 1) {
15293 goto join_operators;
15296 /* Only a single operand on the stack: are done */
15300 if (av_tindex_skip_len_mg(fence_stack) < 0) {
15302 vFAIL("Unexpected ')'");
15305 /* If nothing after the fence, is missing an operand */
15306 if (top_index - fence < 0) {
15310 /* If at least two things on the stack, treat this as an
15312 if (top_index - fence >= 1) {
15313 goto join_operators;
15316 /* Here only a single thing on the fenced stack, and there is a
15317 * fence. Get rid of it */
15318 fence_ptr = av_pop(fence_stack);
15320 fence = SvIV(fence_ptr) - 1;
15321 SvREFCNT_dec_NN(fence_ptr);
15328 /* Having gotten rid of the fence, we pop the operand at the
15329 * stack top and process it as a newly encountered operand */
15330 current = av_pop(stack);
15331 if (IS_OPERAND(current)) {
15332 goto handle_operand;
15344 /* These binary operators should have a left operand already
15346 if ( top_index - fence < 0
15347 || top_index - fence == 1
15348 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
15349 || ! IS_OPERAND(*top_ptr))
15351 goto unexpected_binary;
15354 /* If only the one operand is on the part of the stack visible
15355 * to us, we just place this operator in the proper position */
15356 if (top_index - fence < 2) {
15358 /* Place the operator before the operand */
15360 SV* lhs = av_pop(stack);
15361 av_push(stack, newSVuv(curchar));
15362 av_push(stack, lhs);
15366 /* But if there is something else on the stack, we need to
15367 * process it before this new operator if and only if the
15368 * stacked operation has equal or higher precedence than the
15373 /* The operator on the stack is supposed to be below both its
15375 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
15376 || IS_OPERAND(*stacked_ptr))
15378 /* But if not, it's legal and indicates we are completely
15379 * done if and only if we're currently processing a ']',
15380 * which should be the final thing in the expression */
15381 if (curchar == ']') {
15387 vFAIL2("Unexpected binary operator '%c' with no "
15388 "preceding operand", curchar);
15390 stacked_operator = (char) SvUV(*stacked_ptr);
15392 if (regex_set_precedence(curchar)
15393 > regex_set_precedence(stacked_operator))
15395 /* Here, the new operator has higher precedence than the
15396 * stacked one. This means we need to add the new one to
15397 * the stack to await its rhs operand (and maybe more
15398 * stuff). We put it before the lhs operand, leaving
15399 * untouched the stacked operator and everything below it
15401 lhs = av_pop(stack);
15402 assert(IS_OPERAND(lhs));
15404 av_push(stack, newSVuv(curchar));
15405 av_push(stack, lhs);
15409 /* Here, the new operator has equal or lower precedence than
15410 * what's already there. This means the operation already
15411 * there should be performed now, before the new one. */
15413 rhs = av_pop(stack);
15414 if (! IS_OPERAND(rhs)) {
15416 /* This can happen when a ! is not followed by an operand,
15417 * like in /(?[\t &!])/ */
15421 lhs = av_pop(stack);
15423 if (! IS_OPERAND(lhs)) {
15425 /* This can happen when there is an empty (), like in
15426 * /(?[[0]+()+])/ */
15430 switch (stacked_operator) {
15432 _invlist_intersection(lhs, rhs, &rhs);
15437 _invlist_union(lhs, rhs, &rhs);
15441 _invlist_subtract(lhs, rhs, &rhs);
15444 case '^': /* The union minus the intersection */
15449 _invlist_union(lhs, rhs, &u);
15450 _invlist_intersection(lhs, rhs, &i);
15451 _invlist_subtract(u, i, &rhs);
15452 SvREFCNT_dec_NN(i);
15453 SvREFCNT_dec_NN(u);
15459 /* Here, the higher precedence operation has been done, and the
15460 * result is in 'rhs'. We overwrite the stacked operator with
15461 * the result. Then we redo this code to either push the new
15462 * operator onto the stack or perform any higher precedence
15463 * stacked operation */
15464 only_to_avoid_leaks = av_pop(stack);
15465 SvREFCNT_dec(only_to_avoid_leaks);
15466 av_push(stack, rhs);
15469 case '!': /* Highest priority, right associative */
15471 /* If what's already at the top of the stack is another '!",
15472 * they just cancel each other out */
15473 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
15474 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
15476 only_to_avoid_leaks = av_pop(stack);
15477 SvREFCNT_dec(only_to_avoid_leaks);
15479 else { /* Otherwise, since it's right associative, just push
15481 av_push(stack, newSVuv(curchar));
15486 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15487 vFAIL("Unexpected character");
15491 /* Here 'current' is the operand. If something is already on the
15492 * stack, we have to check if it is a !. But first, the code above
15493 * may have altered the stack in the time since we earlier set
15496 top_index = av_tindex_skip_len_mg(stack);
15497 if (top_index - fence >= 0) {
15498 /* If the top entry on the stack is an operator, it had better
15499 * be a '!', otherwise the entry below the top operand should
15500 * be an operator */
15501 top_ptr = av_fetch(stack, top_index, FALSE);
15503 if (IS_OPERATOR(*top_ptr)) {
15505 /* The only permissible operator at the top of the stack is
15506 * '!', which is applied immediately to this operand. */
15507 curchar = (char) SvUV(*top_ptr);
15508 if (curchar != '!') {
15509 SvREFCNT_dec(current);
15510 vFAIL2("Unexpected binary operator '%c' with no "
15511 "preceding operand", curchar);
15514 _invlist_invert(current);
15516 only_to_avoid_leaks = av_pop(stack);
15517 SvREFCNT_dec(only_to_avoid_leaks);
15519 /* And we redo with the inverted operand. This allows
15520 * handling multiple ! in a row */
15521 goto handle_operand;
15523 /* Single operand is ok only for the non-binary ')'
15525 else if ((top_index - fence == 0 && curchar != ')')
15526 || (top_index - fence > 0
15527 && (! (stacked_ptr = av_fetch(stack,
15530 || IS_OPERAND(*stacked_ptr))))
15532 SvREFCNT_dec(current);
15533 vFAIL("Operand with no preceding operator");
15537 /* Here there was nothing on the stack or the top element was
15538 * another operand. Just add this new one */
15539 av_push(stack, current);
15541 } /* End of switch on next parse token */
15543 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
15544 } /* End of loop parsing through the construct */
15547 if (av_tindex_skip_len_mg(fence_stack) >= 0) {
15548 vFAIL("Unmatched (");
15551 if (av_tindex_skip_len_mg(stack) < 0 /* Was empty */
15552 || ((final = av_pop(stack)) == NULL)
15553 || ! IS_OPERAND(final)
15554 || SvTYPE(final) != SVt_INVLIST
15555 || av_tindex_skip_len_mg(stack) >= 0) /* More left on stack */
15558 SvREFCNT_dec(final);
15559 vFAIL("Incomplete expression within '(?[ ])'");
15562 /* Here, 'final' is the resultant inversion list from evaluating the
15563 * expression. Return it if so requested */
15564 if (return_invlist) {
15565 *return_invlist = final;
15569 /* Otherwise generate a resultant node, based on 'final'. regclass() is
15570 * expecting a string of ranges and individual code points */
15571 invlist_iterinit(final);
15572 result_string = newSVpvs("");
15573 while (invlist_iternext(final, &start, &end)) {
15574 if (start == end) {
15575 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
15578 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%" UVXf "}",
15583 /* About to generate an ANYOF (or similar) node from the inversion list we
15584 * have calculated */
15585 save_parse = RExC_parse;
15586 RExC_parse = SvPV(result_string, len);
15587 save_end = RExC_end;
15588 RExC_end = RExC_parse + len;
15590 /* We turn off folding around the call, as the class we have constructed
15591 * already has all folding taken into consideration, and we don't want
15592 * regclass() to add to that */
15593 RExC_flags &= ~RXf_PMf_FOLD;
15594 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if multi-char
15595 * folds are allowed. */
15596 node = regclass(pRExC_state, flagp,depth+1,
15597 FALSE, /* means parse the whole char class */
15598 FALSE, /* don't allow multi-char folds */
15599 TRUE, /* silence non-portable warnings. The above may very
15600 well have generated non-portable code points, but
15601 they're valid on this machine */
15602 FALSE, /* similarly, no need for strict */
15603 FALSE, /* Require return to be an ANYOF */
15608 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#" UVxf,
15611 /* Fix up the node type if we are in locale. (We have pretended we are
15612 * under /u for the purposes of regclass(), as this construct will only
15613 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
15614 * as to cause any warnings about bad locales to be output in regexec.c),
15615 * and add the flag that indicates to check if not in a UTF-8 locale. The
15616 * reason we above forbid optimization into something other than an ANYOF
15617 * node is simply to minimize the number of code changes in regexec.c.
15618 * Otherwise we would have to create new EXACTish node types and deal with
15619 * them. This decision could be revisited should this construct become
15622 * (One might think we could look at the resulting ANYOF node and suppress
15623 * the flag if everything is above 255, as those would be UTF-8 only,
15624 * but this isn't true, as the components that led to that result could
15625 * have been locale-affected, and just happen to cancel each other out
15626 * under UTF-8 locales.) */
15628 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
15630 assert(OP(node) == ANYOF);
15634 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
15638 RExC_flags |= RXf_PMf_FOLD;
15641 RExC_parse = save_parse + 1;
15642 RExC_end = save_end;
15643 SvREFCNT_dec_NN(final);
15644 SvREFCNT_dec_NN(result_string);
15646 nextchar(pRExC_state);
15647 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
15651 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15654 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
15655 AV * stack, const IV fence, AV * fence_stack)
15656 { /* Dumps the stacks in handle_regex_sets() */
15658 const SSize_t stack_top = av_tindex_skip_len_mg(stack);
15659 const SSize_t fence_stack_top = av_tindex_skip_len_mg(fence_stack);
15662 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
15664 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
15666 if (stack_top < 0) {
15667 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
15670 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
15671 for (i = stack_top; i >= 0; i--) {
15672 SV ** element_ptr = av_fetch(stack, i, FALSE);
15673 if (! element_ptr) {
15676 if (IS_OPERATOR(*element_ptr)) {
15677 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
15678 (int) i, (int) SvIV(*element_ptr));
15681 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
15682 sv_dump(*element_ptr);
15687 if (fence_stack_top < 0) {
15688 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
15691 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
15692 for (i = fence_stack_top; i >= 0; i--) {
15693 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
15694 if (! element_ptr) {
15697 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
15698 (int) i, (int) SvIV(*element_ptr));
15709 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
15711 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
15712 * innocent-looking character class, like /[ks]/i won't have to go out to
15713 * disk to find the possible matches.
15715 * This should be called only for a Latin1-range code points, cp, which is
15716 * known to be involved in a simple fold with other code points above
15717 * Latin1. It would give false results if /aa has been specified.
15718 * Multi-char folds are outside the scope of this, and must be handled
15721 * XXX It would be better to generate these via regen, in case a new
15722 * version of the Unicode standard adds new mappings, though that is not
15723 * really likely, and may be caught by the default: case of the switch
15726 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
15728 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
15734 add_cp_to_invlist(*invlist, KELVIN_SIGN);
15738 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
15741 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
15742 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
15744 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
15745 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
15746 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
15748 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
15749 *invlist = add_cp_to_invlist(*invlist,
15750 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
15753 #ifdef LATIN_CAPITAL_LETTER_SHARP_S /* not defined in early Unicode releases */
15755 case LATIN_SMALL_LETTER_SHARP_S:
15756 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
15761 #if UNICODE_MAJOR_VERSION < 3 \
15762 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0)
15764 /* In 3.0 and earlier, U+0130 folded simply to 'i'; and in 3.0.1 so did
15769 add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
15770 # if UNICODE_DOT_DOT_VERSION == 1
15771 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_DOTLESS_I);
15777 /* Use deprecated warning to increase the chances of this being
15780 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
15787 S_output_or_return_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings, AV** return_posix_warnings)
15789 /* If the final parameter is NULL, output the elements of the array given
15790 * by '*posix_warnings' as REGEXP warnings. Otherwise, the elements are
15791 * pushed onto it, (creating if necessary) */
15794 const bool first_is_fatal = ! return_posix_warnings
15795 && ckDEAD(packWARN(WARN_REGEXP));
15797 PERL_ARGS_ASSERT_OUTPUT_OR_RETURN_POSIX_WARNINGS;
15799 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
15800 if (return_posix_warnings) {
15801 if (! *return_posix_warnings) { /* mortalize to not leak if
15802 warnings are fatal */
15803 *return_posix_warnings = (AV *) sv_2mortal((SV *) newAV());
15805 av_push(*return_posix_warnings, msg);
15808 if (first_is_fatal) { /* Avoid leaking this */
15809 av_undef(posix_warnings); /* This isn't necessary if the
15810 array is mortal, but is a
15812 (void) sv_2mortal(msg);
15814 SAVEFREESV(RExC_rx_sv);
15817 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
15818 SvREFCNT_dec_NN(msg);
15824 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
15826 /* This adds the string scalar <multi_string> to the array
15827 * <multi_char_matches>. <multi_string> is known to have exactly
15828 * <cp_count> code points in it. This is used when constructing a
15829 * bracketed character class and we find something that needs to match more
15830 * than a single character.
15832 * <multi_char_matches> is actually an array of arrays. Each top-level
15833 * element is an array that contains all the strings known so far that are
15834 * the same length. And that length (in number of code points) is the same
15835 * as the index of the top-level array. Hence, the [2] element is an
15836 * array, each element thereof is a string containing TWO code points;
15837 * while element [3] is for strings of THREE characters, and so on. Since
15838 * this is for multi-char strings there can never be a [0] nor [1] element.
15840 * When we rewrite the character class below, we will do so such that the
15841 * longest strings are written first, so that it prefers the longest
15842 * matching strings first. This is done even if it turns out that any
15843 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
15844 * Christiansen has agreed that this is ok. This makes the test for the
15845 * ligature 'ffi' come before the test for 'ff', for example */
15848 AV** this_array_ptr;
15850 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
15852 if (! multi_char_matches) {
15853 multi_char_matches = newAV();
15856 if (av_exists(multi_char_matches, cp_count)) {
15857 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
15858 this_array = *this_array_ptr;
15861 this_array = newAV();
15862 av_store(multi_char_matches, cp_count,
15865 av_push(this_array, multi_string);
15867 return multi_char_matches;
15870 /* The names of properties whose definitions are not known at compile time are
15871 * stored in this SV, after a constant heading. So if the length has been
15872 * changed since initialization, then there is a run-time definition. */
15873 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
15874 (SvCUR(listsv) != initial_listsv_len)
15876 /* There is a restricted set of white space characters that are legal when
15877 * ignoring white space in a bracketed character class. This generates the
15878 * code to skip them.
15880 * There is a line below that uses the same white space criteria but is outside
15881 * this macro. Both here and there must use the same definition */
15882 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
15885 while (isBLANK_A(UCHARAT(p))) \
15893 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
15894 const bool stop_at_1, /* Just parse the next thing, don't
15895 look for a full character class */
15896 bool allow_multi_folds,
15897 const bool silence_non_portable, /* Don't output warnings
15901 bool optimizable, /* ? Allow a non-ANYOF return
15903 SV** ret_invlist, /* Return an inversion list, not a node */
15904 AV** return_posix_warnings
15907 /* parse a bracketed class specification. Most of these will produce an
15908 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
15909 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
15910 * under /i with multi-character folds: it will be rewritten following the
15911 * paradigm of this example, where the <multi-fold>s are characters which
15912 * fold to multiple character sequences:
15913 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
15914 * gets effectively rewritten as:
15915 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
15916 * reg() gets called (recursively) on the rewritten version, and this
15917 * function will return what it constructs. (Actually the <multi-fold>s
15918 * aren't physically removed from the [abcdefghi], it's just that they are
15919 * ignored in the recursion by means of a flag:
15920 * <RExC_in_multi_char_class>.)
15922 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
15923 * characters, with the corresponding bit set if that character is in the
15924 * list. For characters above this, a range list or swash is used. There
15925 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
15926 * determinable at compile time
15928 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs
15929 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded
15930 * to UTF-8. This can only happen if ret_invlist is non-NULL.
15933 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
15935 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
15938 int namedclass = OOB_NAMEDCLASS;
15939 char *rangebegin = NULL;
15940 bool need_class = 0;
15942 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
15943 than just initialized. */
15944 SV* properties = NULL; /* Code points that match \p{} \P{} */
15945 SV* posixes = NULL; /* Code points that match classes like [:word:],
15946 extended beyond the Latin1 range. These have to
15947 be kept separate from other code points for much
15948 of this function because their handling is
15949 different under /i, and for most classes under
15951 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
15952 separate for a while from the non-complemented
15953 versions because of complications with /d
15955 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
15956 treated more simply than the general case,
15957 leading to less compilation and execution
15959 UV element_count = 0; /* Number of distinct elements in the class.
15960 Optimizations may be possible if this is tiny */
15961 AV * multi_char_matches = NULL; /* Code points that fold to more than one
15962 character; used under /i */
15964 char * stop_ptr = RExC_end; /* where to stop parsing */
15966 /* ignore unescaped whitespace? */
15967 const bool skip_white = cBOOL( ret_invlist
15968 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
15970 /* Unicode properties are stored in a swash; this holds the current one
15971 * being parsed. If this swash is the only above-latin1 component of the
15972 * character class, an optimization is to pass it directly on to the
15973 * execution engine. Otherwise, it is set to NULL to indicate that there
15974 * are other things in the class that have to be dealt with at execution
15976 SV* swash = NULL; /* Code points that match \p{} \P{} */
15978 /* Set if a component of this character class is user-defined; just passed
15979 * on to the engine */
15980 bool has_user_defined_property = FALSE;
15982 /* inversion list of code points this node matches only when the target
15983 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
15985 SV* has_upper_latin1_only_utf8_matches = NULL;
15987 /* Inversion list of code points this node matches regardless of things
15988 * like locale, folding, utf8ness of the target string */
15989 SV* cp_list = NULL;
15991 /* Like cp_list, but code points on this list need to be checked for things
15992 * that fold to/from them under /i */
15993 SV* cp_foldable_list = NULL;
15995 /* Like cp_list, but code points on this list are valid only when the
15996 * runtime locale is UTF-8 */
15997 SV* only_utf8_locale_list = NULL;
15999 /* In a range, if one of the endpoints is non-character-set portable,
16000 * meaning that it hard-codes a code point that may mean a different
16001 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
16002 * mnemonic '\t' which each mean the same character no matter which
16003 * character set the platform is on. */
16004 unsigned int non_portable_endpoint = 0;
16006 /* Is the range unicode? which means on a platform that isn't 1-1 native
16007 * to Unicode (i.e. non-ASCII), each code point in it should be considered
16008 * to be a Unicode value. */
16009 bool unicode_range = FALSE;
16010 bool invert = FALSE; /* Is this class to be complemented */
16012 bool warn_super = ALWAYS_WARN_SUPER;
16014 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
16015 case we need to change the emitted regop to an EXACT. */
16016 const char * orig_parse = RExC_parse;
16017 const SSize_t orig_size = RExC_size;
16018 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
16020 /* This variable is used to mark where the end in the input is of something
16021 * that looks like a POSIX construct but isn't. During the parse, when
16022 * something looks like it could be such a construct is encountered, it is
16023 * checked for being one, but not if we've already checked this area of the
16024 * input. Only after this position is reached do we check again */
16025 char *not_posix_region_end = RExC_parse - 1;
16027 AV* posix_warnings = NULL;
16028 const bool do_posix_warnings = return_posix_warnings
16029 || (PASS2 && ckWARN(WARN_REGEXP));
16031 GET_RE_DEBUG_FLAGS_DECL;
16033 PERL_ARGS_ASSERT_REGCLASS;
16035 PERL_UNUSED_ARG(depth);
16038 DEBUG_PARSE("clas");
16040 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
16041 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
16042 && UNICODE_DOT_DOT_VERSION == 0)
16043 allow_multi_folds = FALSE;
16046 /* Assume we are going to generate an ANYOF node. */
16047 ret = reganode(pRExC_state,
16054 RExC_size += ANYOF_SKIP;
16055 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
16058 ANYOF_FLAGS(ret) = 0;
16060 RExC_emit += ANYOF_SKIP;
16061 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
16062 initial_listsv_len = SvCUR(listsv);
16063 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
16066 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16068 assert(RExC_parse <= RExC_end);
16070 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
16073 allow_multi_folds = FALSE;
16075 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16078 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
16079 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
16080 int maybe_class = handle_possible_posix(pRExC_state,
16082 ¬_posix_region_end,
16084 TRUE /* checking only */);
16085 if (PASS2 && maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
16086 SAVEFREESV(RExC_rx_sv);
16087 ckWARN4reg(not_posix_region_end,
16088 "POSIX syntax [%c %c] belongs inside character classes%s",
16089 *RExC_parse, *RExC_parse,
16090 (maybe_class == OOB_NAMEDCLASS)
16091 ? ((POSIXCC_NOTYET(*RExC_parse))
16092 ? " (but this one isn't implemented)"
16093 : " (but this one isn't fully valid)")
16096 (void)ReREFCNT_inc(RExC_rx_sv);
16100 /* If the caller wants us to just parse a single element, accomplish this
16101 * by faking the loop ending condition */
16102 if (stop_at_1 && RExC_end > RExC_parse) {
16103 stop_ptr = RExC_parse + 1;
16106 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
16107 if (UCHARAT(RExC_parse) == ']')
16108 goto charclassloop;
16112 if ( posix_warnings
16113 && av_tindex_skip_len_mg(posix_warnings) >= 0
16114 && RExC_parse > not_posix_region_end)
16116 /* Warnings about posix class issues are considered tentative until
16117 * we are far enough along in the parse that we can no longer
16118 * change our mind, at which point we either output them or add
16119 * them, if it has so specified, to what gets returned to the
16120 * caller. This is done each time through the loop so that a later
16121 * class won't zap them before they have been dealt with. */
16122 output_or_return_posix_warnings(pRExC_state, posix_warnings,
16123 return_posix_warnings);
16126 if (RExC_parse >= stop_ptr) {
16130 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16132 if (UCHARAT(RExC_parse) == ']') {
16138 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
16139 save_value = value;
16140 save_prevvalue = prevvalue;
16143 rangebegin = RExC_parse;
16145 non_portable_endpoint = 0;
16147 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
16148 value = utf8n_to_uvchr((U8*)RExC_parse,
16149 RExC_end - RExC_parse,
16150 &numlen, UTF8_ALLOW_DEFAULT);
16151 RExC_parse += numlen;
16154 value = UCHARAT(RExC_parse++);
16156 if (value == '[') {
16157 char * posix_class_end;
16158 namedclass = handle_possible_posix(pRExC_state,
16161 do_posix_warnings ? &posix_warnings : NULL,
16162 FALSE /* die if error */);
16163 if (namedclass > OOB_NAMEDCLASS) {
16165 /* If there was an earlier attempt to parse this particular
16166 * posix class, and it failed, it was a false alarm, as this
16167 * successful one proves */
16168 if ( posix_warnings
16169 && av_tindex_skip_len_mg(posix_warnings) >= 0
16170 && not_posix_region_end >= RExC_parse
16171 && not_posix_region_end <= posix_class_end)
16173 av_undef(posix_warnings);
16176 RExC_parse = posix_class_end;
16178 else if (namedclass == OOB_NAMEDCLASS) {
16179 not_posix_region_end = posix_class_end;
16182 namedclass = OOB_NAMEDCLASS;
16185 else if ( RExC_parse - 1 > not_posix_region_end
16186 && MAYBE_POSIXCC(value))
16188 (void) handle_possible_posix(
16190 RExC_parse - 1, /* -1 because parse has already been
16192 ¬_posix_region_end,
16193 do_posix_warnings ? &posix_warnings : NULL,
16194 TRUE /* checking only */);
16196 else if (value == '\\') {
16197 /* Is a backslash; get the code point of the char after it */
16199 if (RExC_parse >= RExC_end) {
16200 vFAIL("Unmatched [");
16203 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
16204 value = utf8n_to_uvchr((U8*)RExC_parse,
16205 RExC_end - RExC_parse,
16206 &numlen, UTF8_ALLOW_DEFAULT);
16207 RExC_parse += numlen;
16210 value = UCHARAT(RExC_parse++);
16212 /* Some compilers cannot handle switching on 64-bit integer
16213 * values, therefore value cannot be an UV. Yes, this will
16214 * be a problem later if we want switch on Unicode.
16215 * A similar issue a little bit later when switching on
16216 * namedclass. --jhi */
16218 /* If the \ is escaping white space when white space is being
16219 * skipped, it means that that white space is wanted literally, and
16220 * is already in 'value'. Otherwise, need to translate the escape
16221 * into what it signifies. */
16222 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
16224 case 'w': namedclass = ANYOF_WORDCHAR; break;
16225 case 'W': namedclass = ANYOF_NWORDCHAR; break;
16226 case 's': namedclass = ANYOF_SPACE; break;
16227 case 'S': namedclass = ANYOF_NSPACE; break;
16228 case 'd': namedclass = ANYOF_DIGIT; break;
16229 case 'D': namedclass = ANYOF_NDIGIT; break;
16230 case 'v': namedclass = ANYOF_VERTWS; break;
16231 case 'V': namedclass = ANYOF_NVERTWS; break;
16232 case 'h': namedclass = ANYOF_HORIZWS; break;
16233 case 'H': namedclass = ANYOF_NHORIZWS; break;
16234 case 'N': /* Handle \N{NAME} in class */
16236 const char * const backslash_N_beg = RExC_parse - 2;
16239 if (! grok_bslash_N(pRExC_state,
16240 NULL, /* No regnode */
16241 &value, /* Yes single value */
16242 &cp_count, /* Multiple code pt count */
16248 if (*flagp & NEED_UTF8)
16249 FAIL("panic: grok_bslash_N set NEED_UTF8");
16250 if (*flagp & RESTART_PASS1)
16253 if (cp_count < 0) {
16254 vFAIL("\\N in a character class must be a named character: \\N{...}");
16256 else if (cp_count == 0) {
16258 ckWARNreg(RExC_parse,
16259 "Ignoring zero length \\N{} in character class");
16262 else { /* cp_count > 1 */
16263 if (! RExC_in_multi_char_class) {
16264 if (invert || range || *RExC_parse == '-') {
16267 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
16270 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
16272 break; /* <value> contains the first code
16273 point. Drop out of the switch to
16277 SV * multi_char_N = newSVpvn(backslash_N_beg,
16278 RExC_parse - backslash_N_beg);
16280 = add_multi_match(multi_char_matches,
16285 } /* End of cp_count != 1 */
16287 /* This element should not be processed further in this
16290 value = save_value;
16291 prevvalue = save_prevvalue;
16292 continue; /* Back to top of loop to get next char */
16295 /* Here, is a single code point, and <value> contains it */
16296 unicode_range = TRUE; /* \N{} are Unicode */
16304 /* We will handle any undefined properties ourselves */
16305 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
16306 /* And we actually would prefer to get
16307 * the straight inversion list of the
16308 * swash, since we will be accessing it
16309 * anyway, to save a little time */
16310 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
16312 if (RExC_parse >= RExC_end)
16313 vFAIL2("Empty \\%c", (U8)value);
16314 if (*RExC_parse == '{') {
16315 const U8 c = (U8)value;
16316 e = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
16319 vFAIL2("Missing right brace on \\%c{}", c);
16323 while (isSPACE(*RExC_parse)) {
16327 if (UCHARAT(RExC_parse) == '^') {
16329 /* toggle. (The rhs xor gets the single bit that
16330 * differs between P and p; the other xor inverts just
16332 value ^= 'P' ^ 'p';
16335 while (isSPACE(*RExC_parse)) {
16340 if (e == RExC_parse)
16341 vFAIL2("Empty \\%c{}", c);
16343 n = e - RExC_parse;
16344 while (isSPACE(*(RExC_parse + n - 1)))
16346 } /* The \p isn't immediately followed by a '{' */
16347 else if (! isALPHA(*RExC_parse)) {
16348 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16349 vFAIL2("Character following \\%c must be '{' or a "
16350 "single-character Unicode property name",
16360 char* base_name; /* name after any packages are stripped */
16361 char* lookup_name = NULL;
16362 const char * const colon_colon = "::";
16364 /* Try to get the definition of the property into
16365 * <invlist>. If /i is in effect, the effective property
16366 * will have its name be <__NAME_i>. The design is
16367 * discussed in commit
16368 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
16369 name = savepv(Perl_form(aTHX_ "%.*s", (int)n, RExC_parse));
16372 lookup_name = savepv(Perl_form(aTHX_ "__%s_i", name));
16374 /* The function call just below that uses this can fail
16375 * to return, leaking memory if we don't do this */
16376 SAVEFREEPV(lookup_name);
16379 /* Look up the property name, and get its swash and
16380 * inversion list, if the property is found */
16381 SvREFCNT_dec(swash); /* Free any left-overs */
16382 swash = _core_swash_init("utf8",
16389 NULL, /* No inversion list */
16392 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
16393 HV* curpkg = (IN_PERL_COMPILETIME)
16395 : CopSTASH(PL_curcop);
16399 if (swash) { /* Got a swash but no inversion list.
16400 Something is likely wrong that will
16401 be sorted-out later */
16402 SvREFCNT_dec_NN(swash);
16406 /* Here didn't find it. It could be a an error (like a
16407 * typo) in specifying a Unicode property, or it could
16408 * be a user-defined property that will be available at
16409 * run-time. The names of these must begin with 'In'
16410 * or 'Is' (after any packages are stripped off). So
16411 * if not one of those, or if we accept only
16412 * compile-time properties, is an error; otherwise add
16413 * it to the list for run-time look up. */
16414 if ((base_name = rninstr(name, name + n,
16415 colon_colon, colon_colon + 2)))
16416 { /* Has ::. We know this must be a user-defined
16419 final_n -= base_name - name;
16428 || base_name[0] != 'I'
16429 || (base_name[1] != 's' && base_name[1] != 'n')
16432 const char * const msg
16434 ? "Illegal user-defined property name"
16435 : "Can't find Unicode property definition";
16436 RExC_parse = e + 1;
16438 /* diag_listed_as: Can't find Unicode property definition "%s" */
16439 vFAIL3utf8f("%s \"%" UTF8f "\"",
16440 msg, UTF8fARG(UTF, n, name));
16443 /* If the property name doesn't already have a package
16444 * name, add the current one to it so that it can be
16445 * referred to outside it. [perl #121777] */
16446 if (! has_pkg && curpkg) {
16447 char* pkgname = HvNAME(curpkg);
16448 if (memNEs(pkgname, HvNAMELEN(curpkg), "main")) {
16449 char* full_name = Perl_form(aTHX_
16453 n = strlen(full_name);
16454 name = savepvn(full_name, n);
16458 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s%" UTF8f "%s\n",
16459 (value == 'p' ? '+' : '!'),
16460 (FOLD) ? "__" : "",
16461 UTF8fARG(UTF, n, name),
16462 (FOLD) ? "_i" : "");
16463 has_user_defined_property = TRUE;
16464 optimizable = FALSE; /* Will have to leave this an
16467 /* We don't know yet what this matches, so have to flag
16469 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
16473 /* Here, did get the swash and its inversion list. If
16474 * the swash is from a user-defined property, then this
16475 * whole character class should be regarded as such */
16476 if (swash_init_flags
16477 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
16479 has_user_defined_property = TRUE;
16482 /* We warn on matching an above-Unicode code point
16483 * if the match would return true, except don't
16484 * warn for \p{All}, which has exactly one element
16486 (_invlist_contains_cp(invlist, 0x110000)
16487 && (! (_invlist_len(invlist) == 1
16488 && *invlist_array(invlist) == 0)))
16494 /* Invert if asking for the complement */
16495 if (value == 'P') {
16496 _invlist_union_complement_2nd(properties,
16500 /* The swash can't be used as-is, because we've
16501 * inverted things; delay removing it to here after
16502 * have copied its invlist above */
16503 SvREFCNT_dec_NN(swash);
16507 _invlist_union(properties, invlist, &properties);
16511 RExC_parse = e + 1;
16512 namedclass = ANYOF_UNIPROP; /* no official name, but it's
16515 /* \p means they want Unicode semantics */
16516 REQUIRE_UNI_RULES(flagp, NULL);
16519 case 'n': value = '\n'; break;
16520 case 'r': value = '\r'; break;
16521 case 't': value = '\t'; break;
16522 case 'f': value = '\f'; break;
16523 case 'b': value = '\b'; break;
16524 case 'e': value = ESC_NATIVE; break;
16525 case 'a': value = '\a'; break;
16527 RExC_parse--; /* function expects to be pointed at the 'o' */
16529 const char* error_msg;
16530 bool valid = grok_bslash_o(&RExC_parse,
16534 PASS2, /* warnings only in
16537 silence_non_portable,
16543 non_portable_endpoint++;
16546 RExC_parse--; /* function expects to be pointed at the 'x' */
16548 const char* error_msg;
16549 bool valid = grok_bslash_x(&RExC_parse,
16553 PASS2, /* Output warnings */
16555 silence_non_portable,
16561 non_portable_endpoint++;
16564 value = grok_bslash_c(*RExC_parse++, PASS2);
16565 non_portable_endpoint++;
16567 case '0': case '1': case '2': case '3': case '4':
16568 case '5': case '6': case '7':
16570 /* Take 1-3 octal digits */
16571 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
16572 numlen = (strict) ? 4 : 3;
16573 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
16574 RExC_parse += numlen;
16577 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16578 vFAIL("Need exactly 3 octal digits");
16580 else if (! SIZE_ONLY /* like \08, \178 */
16582 && RExC_parse < RExC_end
16583 && isDIGIT(*RExC_parse)
16584 && ckWARN(WARN_REGEXP))
16586 SAVEFREESV(RExC_rx_sv);
16587 reg_warn_non_literal_string(
16589 form_short_octal_warning(RExC_parse, numlen));
16590 (void)ReREFCNT_inc(RExC_rx_sv);
16593 non_portable_endpoint++;
16597 /* Allow \_ to not give an error */
16598 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
16600 vFAIL2("Unrecognized escape \\%c in character class",
16604 SAVEFREESV(RExC_rx_sv);
16605 ckWARN2reg(RExC_parse,
16606 "Unrecognized escape \\%c in character class passed through",
16608 (void)ReREFCNT_inc(RExC_rx_sv);
16612 } /* End of switch on char following backslash */
16613 } /* end of handling backslash escape sequences */
16615 /* Here, we have the current token in 'value' */
16617 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
16620 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
16621 * literal, as is the character that began the false range, i.e.
16622 * the 'a' in the examples */
16625 const int w = (RExC_parse >= rangebegin)
16626 ? RExC_parse - rangebegin
16630 "False [] range \"%" UTF8f "\"",
16631 UTF8fARG(UTF, w, rangebegin));
16634 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
16635 ckWARN2reg(RExC_parse,
16636 "False [] range \"%" UTF8f "\"",
16637 UTF8fARG(UTF, w, rangebegin));
16638 (void)ReREFCNT_inc(RExC_rx_sv);
16639 cp_list = add_cp_to_invlist(cp_list, '-');
16640 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
16645 range = 0; /* this was not a true range */
16646 element_count += 2; /* So counts for three values */
16649 classnum = namedclass_to_classnum(namedclass);
16651 if (LOC && namedclass < ANYOF_POSIXL_MAX
16652 #ifndef HAS_ISASCII
16653 && classnum != _CC_ASCII
16656 /* What the Posix classes (like \w, [:space:]) match in locale
16657 * isn't knowable under locale until actual match time. Room
16658 * must be reserved (one time per outer bracketed class) to
16659 * store such classes. The space will contain a bit for each
16660 * named class that is to be matched against. This isn't
16661 * needed for \p{} and pseudo-classes, as they are not affected
16662 * by locale, and hence are dealt with separately */
16663 if (! need_class) {
16666 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16669 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
16671 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
16672 ANYOF_POSIXL_ZERO(ret);
16674 /* We can't change this into some other type of node
16675 * (unless this is the only element, in which case there
16676 * are nodes that mean exactly this) as has runtime
16678 optimizable = FALSE;
16681 /* Coverity thinks it is possible for this to be negative; both
16682 * jhi and khw think it's not, but be safer */
16683 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16684 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
16686 /* See if it already matches the complement of this POSIX
16688 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
16689 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
16693 posixl_matches_all = TRUE;
16694 break; /* No need to continue. Since it matches both
16695 e.g., \w and \W, it matches everything, and the
16696 bracketed class can be optimized into qr/./s */
16699 /* Add this class to those that should be checked at runtime */
16700 ANYOF_POSIXL_SET(ret, namedclass);
16702 /* The above-Latin1 characters are not subject to locale rules.
16703 * Just add them, in the second pass, to the
16704 * unconditionally-matched list */
16706 SV* scratch_list = NULL;
16708 /* Get the list of the above-Latin1 code points this
16710 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
16711 PL_XPosix_ptrs[classnum],
16713 /* Odd numbers are complements, like
16714 * NDIGIT, NASCII, ... */
16715 namedclass % 2 != 0,
16717 /* Checking if 'cp_list' is NULL first saves an extra
16718 * clone. Its reference count will be decremented at the
16719 * next union, etc, or if this is the only instance, at the
16720 * end of the routine */
16722 cp_list = scratch_list;
16725 _invlist_union(cp_list, scratch_list, &cp_list);
16726 SvREFCNT_dec_NN(scratch_list);
16728 continue; /* Go get next character */
16731 else if (! SIZE_ONLY) {
16733 /* Here, not in pass1 (in that pass we skip calculating the
16734 * contents of this class), and is not /l, or is a POSIX class
16735 * for which /l doesn't matter (or is a Unicode property, which
16736 * is skipped here). */
16737 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
16738 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
16740 /* Here, should be \h, \H, \v, or \V. None of /d, /i
16741 * nor /l make a difference in what these match,
16742 * therefore we just add what they match to cp_list. */
16743 if (classnum != _CC_VERTSPACE) {
16744 assert( namedclass == ANYOF_HORIZWS
16745 || namedclass == ANYOF_NHORIZWS);
16747 /* It turns out that \h is just a synonym for
16749 classnum = _CC_BLANK;
16752 _invlist_union_maybe_complement_2nd(
16754 PL_XPosix_ptrs[classnum],
16755 namedclass % 2 != 0, /* Complement if odd
16756 (NHORIZWS, NVERTWS)
16761 else if ( UNI_SEMANTICS
16762 || classnum == _CC_ASCII
16763 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
16764 || classnum == _CC_XDIGIT)))
16766 /* We usually have to worry about /d and /a affecting what
16767 * POSIX classes match, with special code needed for /d
16768 * because we won't know until runtime what all matches.
16769 * But there is no extra work needed under /u, and
16770 * [:ascii:] is unaffected by /a and /d; and :digit: and
16771 * :xdigit: don't have runtime differences under /d. So we
16772 * can special case these, and avoid some extra work below,
16773 * and at runtime. */
16774 _invlist_union_maybe_complement_2nd(
16776 PL_XPosix_ptrs[classnum],
16777 namedclass % 2 != 0,
16780 else { /* Garden variety class. If is NUPPER, NALPHA, ...
16781 complement and use nposixes */
16782 SV** posixes_ptr = namedclass % 2 == 0
16785 _invlist_union_maybe_complement_2nd(
16787 PL_XPosix_ptrs[classnum],
16788 namedclass % 2 != 0,
16792 } /* end of namedclass \blah */
16794 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16796 /* If 'range' is set, 'value' is the ending of a range--check its
16797 * validity. (If value isn't a single code point in the case of a
16798 * range, we should have figured that out above in the code that
16799 * catches false ranges). Later, we will handle each individual code
16800 * point in the range. If 'range' isn't set, this could be the
16801 * beginning of a range, so check for that by looking ahead to see if
16802 * the next real character to be processed is the range indicator--the
16807 /* For unicode ranges, we have to test that the Unicode as opposed
16808 * to the native values are not decreasing. (Above 255, there is
16809 * no difference between native and Unicode) */
16810 if (unicode_range && prevvalue < 255 && value < 255) {
16811 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
16812 goto backwards_range;
16817 if (prevvalue > value) /* b-a */ {
16822 w = RExC_parse - rangebegin;
16824 "Invalid [] range \"%" UTF8f "\"",
16825 UTF8fARG(UTF, w, rangebegin));
16826 NOT_REACHED; /* NOTREACHED */
16830 prevvalue = value; /* save the beginning of the potential range */
16831 if (! stop_at_1 /* Can't be a range if parsing just one thing */
16832 && *RExC_parse == '-')
16834 char* next_char_ptr = RExC_parse + 1;
16836 /* Get the next real char after the '-' */
16837 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
16839 /* If the '-' is at the end of the class (just before the ']',
16840 * it is a literal minus; otherwise it is a range */
16841 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
16842 RExC_parse = next_char_ptr;
16844 /* a bad range like \w-, [:word:]- ? */
16845 if (namedclass > OOB_NAMEDCLASS) {
16846 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
16847 const int w = RExC_parse >= rangebegin
16848 ? RExC_parse - rangebegin
16851 vFAIL4("False [] range \"%*.*s\"",
16856 "False [] range \"%*.*s\"",
16861 cp_list = add_cp_to_invlist(cp_list, '-');
16865 range = 1; /* yeah, it's a range! */
16866 continue; /* but do it the next time */
16871 if (namedclass > OOB_NAMEDCLASS) {
16875 /* Here, we have a single value this time through the loop, and
16876 * <prevvalue> is the beginning of the range, if any; or <value> if
16879 /* non-Latin1 code point implies unicode semantics. Must be set in
16880 * pass1 so is there for the whole of pass 2 */
16882 REQUIRE_UNI_RULES(flagp, NULL);
16885 /* Ready to process either the single value, or the completed range.
16886 * For single-valued non-inverted ranges, we consider the possibility
16887 * of multi-char folds. (We made a conscious decision to not do this
16888 * for the other cases because it can often lead to non-intuitive
16889 * results. For example, you have the peculiar case that:
16890 * "s s" =~ /^[^\xDF]+$/i => Y
16891 * "ss" =~ /^[^\xDF]+$/i => N
16893 * See [perl #89750] */
16894 if (FOLD && allow_multi_folds && value == prevvalue) {
16895 if (value == LATIN_SMALL_LETTER_SHARP_S
16896 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
16899 /* Here <value> is indeed a multi-char fold. Get what it is */
16901 U8 foldbuf[UTF8_MAXBYTES_CASE];
16904 UV folded = _to_uni_fold_flags(
16908 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
16909 ? FOLD_FLAGS_NOMIX_ASCII
16913 /* Here, <folded> should be the first character of the
16914 * multi-char fold of <value>, with <foldbuf> containing the
16915 * whole thing. But, if this fold is not allowed (because of
16916 * the flags), <fold> will be the same as <value>, and should
16917 * be processed like any other character, so skip the special
16919 if (folded != value) {
16921 /* Skip if we are recursed, currently parsing the class
16922 * again. Otherwise add this character to the list of
16923 * multi-char folds. */
16924 if (! RExC_in_multi_char_class) {
16925 STRLEN cp_count = utf8_length(foldbuf,
16926 foldbuf + foldlen);
16927 SV* multi_fold = sv_2mortal(newSVpvs(""));
16929 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
16932 = add_multi_match(multi_char_matches,
16938 /* This element should not be processed further in this
16941 value = save_value;
16942 prevvalue = save_prevvalue;
16948 if (strict && PASS2 && ckWARN(WARN_REGEXP)) {
16951 /* If the range starts above 255, everything is portable and
16952 * likely to be so for any forseeable character set, so don't
16954 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
16955 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
16957 else if (prevvalue != value) {
16959 /* Under strict, ranges that stop and/or end in an ASCII
16960 * printable should have each end point be a portable value
16961 * for it (preferably like 'A', but we don't warn if it is
16962 * a (portable) Unicode name or code point), and the range
16963 * must be be all digits or all letters of the same case.
16964 * Otherwise, the range is non-portable and unclear as to
16965 * what it contains */
16966 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
16967 && ( non_portable_endpoint
16968 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
16969 || (isLOWER_A(prevvalue) && isLOWER_A(value))
16970 || (isUPPER_A(prevvalue) && isUPPER_A(value))
16972 vWARN(RExC_parse, "Ranges of ASCII printables should"
16973 " be some subset of \"0-9\","
16974 " \"A-Z\", or \"a-z\"");
16976 else if (prevvalue >= 0x660) { /* ARABIC_INDIC_DIGIT_ZERO */
16977 SSize_t index_start;
16978 SSize_t index_final;
16980 /* But the nature of Unicode and languages mean we
16981 * can't do the same checks for above-ASCII ranges,
16982 * except in the case of digit ones. These should
16983 * contain only digits from the same group of 10. The
16984 * ASCII case is handled just above. 0x660 is the
16985 * first digit character beyond ASCII. Hence here, the
16986 * range could be a range of digits. First some
16987 * unlikely special cases. Grandfather in that a range
16988 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
16989 * if its starting value is one of the 10 digits prior
16990 * to it. This is because it is an alternate way of
16991 * writing 19D1, and some people may expect it to be in
16992 * that group. But it is bad, because it won't give
16993 * the expected results. In Unicode 5.2 it was
16994 * considered to be in that group (of 11, hence), but
16995 * this was fixed in the next version */
16997 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
16998 goto warn_bad_digit_range;
17000 else if (UNLIKELY( prevvalue >= 0x1D7CE
17001 && value <= 0x1D7FF))
17003 /* This is the only other case currently in Unicode
17004 * where the algorithm below fails. The code
17005 * points just above are the end points of a single
17006 * range containing only decimal digits. It is 5
17007 * different series of 0-9. All other ranges of
17008 * digits currently in Unicode are just a single
17009 * series. (And mktables will notify us if a later
17010 * Unicode version breaks this.)
17012 * If the range being checked is at most 9 long,
17013 * and the digit values represented are in
17014 * numerical order, they are from the same series.
17016 if ( value - prevvalue > 9
17017 || ((( value - 0x1D7CE) % 10)
17018 <= (prevvalue - 0x1D7CE) % 10))
17020 goto warn_bad_digit_range;
17025 /* For all other ranges of digits in Unicode, the
17026 * algorithm is just to check if both end points
17027 * are in the same series, which is the same range.
17029 index_start = _invlist_search(
17030 PL_XPosix_ptrs[_CC_DIGIT],
17033 /* Warn if the range starts and ends with a digit,
17034 * and they are not in the same group of 10. */
17035 if ( index_start >= 0
17036 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
17038 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
17039 value)) != index_start
17040 && index_final >= 0
17041 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
17043 warn_bad_digit_range:
17044 vWARN(RExC_parse, "Ranges of digits should be"
17045 " from the same group of"
17052 if ((! range || prevvalue == value) && non_portable_endpoint) {
17053 if (isPRINT_A(value)) {
17056 if (isBACKSLASHED_PUNCT(value)) {
17057 literal[d++] = '\\';
17059 literal[d++] = (char) value;
17060 literal[d++] = '\0';
17063 "\"%.*s\" is more clearly written simply as \"%s\"",
17064 (int) (RExC_parse - rangebegin),
17069 else if isMNEMONIC_CNTRL(value) {
17071 "\"%.*s\" is more clearly written simply as \"%s\"",
17072 (int) (RExC_parse - rangebegin),
17074 cntrl_to_mnemonic((U8) value)
17080 /* Deal with this element of the class */
17084 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17087 /* On non-ASCII platforms, for ranges that span all of 0..255, and
17088 * ones that don't require special handling, we can just add the
17089 * range like we do for ASCII platforms */
17090 if ((UNLIKELY(prevvalue == 0) && value >= 255)
17091 || ! (prevvalue < 256
17093 || (! non_portable_endpoint
17094 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
17095 || (isUPPER_A(prevvalue)
17096 && isUPPER_A(value)))))))
17098 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17102 /* Here, requires special handling. This can be because it is
17103 * a range whose code points are considered to be Unicode, and
17104 * so must be individually translated into native, or because
17105 * its a subrange of 'A-Z' or 'a-z' which each aren't
17106 * contiguous in EBCDIC, but we have defined them to include
17107 * only the "expected" upper or lower case ASCII alphabetics.
17108 * Subranges above 255 are the same in native and Unicode, so
17109 * can be added as a range */
17110 U8 start = NATIVE_TO_LATIN1(prevvalue);
17112 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
17113 for (j = start; j <= end; j++) {
17114 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
17117 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17124 range = 0; /* this range (if it was one) is done now */
17125 } /* End of loop through all the text within the brackets */
17128 if ( posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
17129 output_or_return_posix_warnings(pRExC_state, posix_warnings,
17130 return_posix_warnings);
17133 /* If anything in the class expands to more than one character, we have to
17134 * deal with them by building up a substitute parse string, and recursively
17135 * calling reg() on it, instead of proceeding */
17136 if (multi_char_matches) {
17137 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
17140 char *save_end = RExC_end;
17141 char *save_parse = RExC_parse;
17142 char *save_start = RExC_start;
17143 STRLEN prefix_end = 0; /* We copy the character class after a
17144 prefix supplied here. This is the size
17145 + 1 of that prefix */
17146 bool first_time = TRUE; /* First multi-char occurrence doesn't get
17151 assert(RExC_precomp_adj == 0); /* Only one level of recursion allowed */
17153 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
17154 because too confusing */
17156 sv_catpv(substitute_parse, "(?:");
17160 /* Look at the longest folds first */
17161 for (cp_count = av_tindex_skip_len_mg(multi_char_matches);
17166 if (av_exists(multi_char_matches, cp_count)) {
17167 AV** this_array_ptr;
17170 this_array_ptr = (AV**) av_fetch(multi_char_matches,
17172 while ((this_sequence = av_pop(*this_array_ptr)) !=
17175 if (! first_time) {
17176 sv_catpv(substitute_parse, "|");
17178 first_time = FALSE;
17180 sv_catpv(substitute_parse, SvPVX(this_sequence));
17185 /* If the character class contains anything else besides these
17186 * multi-character folds, have to include it in recursive parsing */
17187 if (element_count) {
17188 sv_catpv(substitute_parse, "|[");
17189 prefix_end = SvCUR(substitute_parse);
17190 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
17192 /* Put in a closing ']' only if not going off the end, as otherwise
17193 * we are adding something that really isn't there */
17194 if (RExC_parse < RExC_end) {
17195 sv_catpv(substitute_parse, "]");
17199 sv_catpv(substitute_parse, ")");
17202 /* This is a way to get the parse to skip forward a whole named
17203 * sequence instead of matching the 2nd character when it fails the
17205 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
17209 /* Set up the data structure so that any errors will be properly
17210 * reported. See the comments at the definition of
17211 * REPORT_LOCATION_ARGS for details */
17212 RExC_precomp_adj = orig_parse - RExC_precomp;
17213 RExC_start = RExC_parse = SvPV(substitute_parse, len);
17214 RExC_adjusted_start = RExC_start + prefix_end;
17215 RExC_end = RExC_parse + len;
17216 RExC_in_multi_char_class = 1;
17217 RExC_emit = (regnode *)orig_emit;
17219 ret = reg(pRExC_state, 1, ®_flags, depth+1);
17221 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PASS1|NEED_UTF8);
17223 /* And restore so can parse the rest of the pattern */
17224 RExC_parse = save_parse;
17225 RExC_start = RExC_adjusted_start = save_start;
17226 RExC_precomp_adj = 0;
17227 RExC_end = save_end;
17228 RExC_in_multi_char_class = 0;
17229 SvREFCNT_dec_NN(multi_char_matches);
17233 /* Here, we've gone through the entire class and dealt with multi-char
17234 * folds. We are now in a position that we can do some checks to see if we
17235 * can optimize this ANYOF node into a simpler one, even in Pass 1.
17236 * Currently we only do two checks:
17237 * 1) is in the unlikely event that the user has specified both, eg. \w and
17238 * \W under /l, then the class matches everything. (This optimization
17239 * is done only to make the optimizer code run later work.)
17240 * 2) if the character class contains only a single element (including a
17241 * single range), we see if there is an equivalent node for it.
17242 * Other checks are possible */
17244 && ! ret_invlist /* Can't optimize if returning the constructed
17246 && (UNLIKELY(posixl_matches_all) || element_count == 1))
17251 if (UNLIKELY(posixl_matches_all)) {
17254 else if (namedclass > OOB_NAMEDCLASS) { /* this is a single named
17255 class, like \w or [:digit:]
17258 /* All named classes are mapped into POSIXish nodes, with its FLAG
17259 * argument giving which class it is */
17260 switch ((I32)namedclass) {
17261 case ANYOF_UNIPROP:
17264 /* These don't depend on the charset modifiers. They always
17265 * match under /u rules */
17266 case ANYOF_NHORIZWS:
17267 case ANYOF_HORIZWS:
17268 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
17271 case ANYOF_NVERTWS:
17276 /* The actual POSIXish node for all the rest depends on the
17277 * charset modifier. The ones in the first set depend only on
17278 * ASCII or, if available on this platform, also locale */
17282 op = (LOC) ? POSIXL : POSIXA;
17288 /* The following don't have any matches in the upper Latin1
17289 * range, hence /d is equivalent to /u for them. Making it /u
17290 * saves some branches at runtime */
17294 case ANYOF_NXDIGIT:
17295 if (! DEPENDS_SEMANTICS) {
17296 goto treat_as_default;
17302 /* The following change to CASED under /i */
17308 namedclass = ANYOF_CASED + (namedclass % 2);
17312 /* The rest have more possibilities depending on the charset.
17313 * We take advantage of the enum ordering of the charset
17314 * modifiers to get the exact node type, */
17317 op = POSIXD + get_regex_charset(RExC_flags);
17318 if (op > POSIXA) { /* /aa is same as /a */
17323 /* The odd numbered ones are the complements of the
17324 * next-lower even number one */
17325 if (namedclass % 2 == 1) {
17329 arg = namedclass_to_classnum(namedclass);
17333 else if (value == prevvalue) {
17335 /* Here, the class consists of just a single code point */
17338 if (! LOC && value == '\n') {
17339 op = REG_ANY; /* Optimize [^\n] */
17340 *flagp |= HASWIDTH|SIMPLE;
17344 else if (value < 256 || UTF) {
17346 /* Optimize a single value into an EXACTish node, but not if it
17347 * would require converting the pattern to UTF-8. */
17348 op = compute_EXACTish(pRExC_state);
17350 } /* Otherwise is a range */
17351 else if (! LOC) { /* locale could vary these */
17352 if (prevvalue == '0') {
17353 if (value == '9') {
17358 else if (! FOLD || ASCII_FOLD_RESTRICTED) {
17359 /* We can optimize A-Z or a-z, but not if they could match
17360 * something like the KELVIN SIGN under /i. */
17361 if (prevvalue == 'A') {
17364 && ! non_portable_endpoint
17367 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
17371 else if (prevvalue == 'a') {
17374 && ! non_portable_endpoint
17377 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
17384 /* Here, we have changed <op> away from its initial value iff we found
17385 * an optimization */
17388 /* Throw away this ANYOF regnode, and emit the calculated one,
17389 * which should correspond to the beginning, not current, state of
17391 const char * cur_parse = RExC_parse;
17392 RExC_parse = (char *)orig_parse;
17396 /* To get locale nodes to not use the full ANYOF size would
17397 * require moving the code above that writes the portions
17398 * of it that aren't in other nodes to after this point.
17399 * e.g. ANYOF_POSIXL_SET */
17400 RExC_size = orig_size;
17404 RExC_emit = (regnode *)orig_emit;
17405 if (PL_regkind[op] == POSIXD) {
17406 if (op == POSIXL) {
17407 RExC_contains_locale = 1;
17410 op += NPOSIXD - POSIXD;
17415 ret = reg_node(pRExC_state, op);
17417 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
17421 *flagp |= HASWIDTH|SIMPLE;
17423 else if (PL_regkind[op] == EXACT) {
17424 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
17425 TRUE /* downgradable to EXACT */
17429 RExC_parse = (char *) cur_parse;
17431 SvREFCNT_dec(posixes);
17432 SvREFCNT_dec(nposixes);
17433 SvREFCNT_dec(simple_posixes);
17434 SvREFCNT_dec(cp_list);
17435 SvREFCNT_dec(cp_foldable_list);
17442 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
17444 /* If folding, we calculate all characters that could fold to or from the
17445 * ones already on the list */
17446 if (cp_foldable_list) {
17448 UV start, end; /* End points of code point ranges */
17450 SV* fold_intersection = NULL;
17453 /* Our calculated list will be for Unicode rules. For locale
17454 * matching, we have to keep a separate list that is consulted at
17455 * runtime only when the locale indicates Unicode rules. For
17456 * non-locale, we just use the general list */
17458 use_list = &only_utf8_locale_list;
17461 use_list = &cp_list;
17464 /* Only the characters in this class that participate in folds need
17465 * be checked. Get the intersection of this class and all the
17466 * possible characters that are foldable. This can quickly narrow
17467 * down a large class */
17468 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
17469 &fold_intersection);
17471 /* The folds for all the Latin1 characters are hard-coded into this
17472 * program, but we have to go out to disk to get the others. */
17473 if (invlist_highest(cp_foldable_list) >= 256) {
17475 /* This is a hash that for a particular fold gives all
17476 * characters that are involved in it */
17477 if (! PL_utf8_foldclosures) {
17478 _load_PL_utf8_foldclosures();
17482 /* Now look at the foldable characters in this class individually */
17483 invlist_iterinit(fold_intersection);
17484 while (invlist_iternext(fold_intersection, &start, &end)) {
17487 /* Look at every character in the range */
17488 for (j = start; j <= end; j++) {
17489 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17495 if (IS_IN_SOME_FOLD_L1(j)) {
17497 /* ASCII is always matched; non-ASCII is matched
17498 * only under Unicode rules (which could happen
17499 * under /l if the locale is a UTF-8 one */
17500 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
17501 *use_list = add_cp_to_invlist(*use_list,
17502 PL_fold_latin1[j]);
17505 has_upper_latin1_only_utf8_matches
17506 = add_cp_to_invlist(
17507 has_upper_latin1_only_utf8_matches,
17508 PL_fold_latin1[j]);
17512 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
17513 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
17515 add_above_Latin1_folds(pRExC_state,
17522 /* Here is an above Latin1 character. We don't have the
17523 * rules hard-coded for it. First, get its fold. This is
17524 * the simple fold, as the multi-character folds have been
17525 * handled earlier and separated out */
17526 _to_uni_fold_flags(j, foldbuf, &foldlen,
17527 (ASCII_FOLD_RESTRICTED)
17528 ? FOLD_FLAGS_NOMIX_ASCII
17531 /* Single character fold of above Latin1. Add everything in
17532 * its fold closure to the list that this node should match.
17533 * The fold closures data structure is a hash with the keys
17534 * being the UTF-8 of every character that is folded to, like
17535 * 'k', and the values each an array of all code points that
17536 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
17537 * Multi-character folds are not included */
17538 if ((listp = hv_fetch(PL_utf8_foldclosures,
17539 (char *) foldbuf, foldlen, FALSE)))
17541 AV* list = (AV*) *listp;
17543 for (k = 0; k <= av_tindex_skip_len_mg(list); k++) {
17544 SV** c_p = av_fetch(list, k, FALSE);
17550 /* /aa doesn't allow folds between ASCII and non- */
17551 if ((ASCII_FOLD_RESTRICTED
17552 && (isASCII(c) != isASCII(j))))
17557 /* Folds under /l which cross the 255/256 boundary
17558 * are added to a separate list. (These are valid
17559 * only when the locale is UTF-8.) */
17560 if (c < 256 && LOC) {
17561 *use_list = add_cp_to_invlist(*use_list, c);
17565 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
17567 cp_list = add_cp_to_invlist(cp_list, c);
17570 /* Similarly folds involving non-ascii Latin1
17571 * characters under /d are added to their list */
17572 has_upper_latin1_only_utf8_matches
17573 = add_cp_to_invlist(
17574 has_upper_latin1_only_utf8_matches,
17581 SvREFCNT_dec_NN(fold_intersection);
17584 /* Now that we have finished adding all the folds, there is no reason
17585 * to keep the foldable list separate */
17586 _invlist_union(cp_list, cp_foldable_list, &cp_list);
17587 SvREFCNT_dec_NN(cp_foldable_list);
17590 /* And combine the result (if any) with any inversion lists from posix
17591 * classes. The lists are kept separate up to now because we don't want to
17592 * fold the classes (folding of those is automatically handled by the swash
17593 * fetching code) */
17594 if (simple_posixes) { /* These are the classes known to be unaffected by
17597 _invlist_union(cp_list, simple_posixes, &cp_list);
17598 SvREFCNT_dec_NN(simple_posixes);
17601 cp_list = simple_posixes;
17604 if (posixes || nposixes) {
17606 /* We have to adjust /a and /aa */
17607 if (AT_LEAST_ASCII_RESTRICTED) {
17609 /* Under /a and /aa, nothing above ASCII matches these */
17611 _invlist_intersection(posixes,
17612 PL_XPosix_ptrs[_CC_ASCII],
17616 /* Under /a and /aa, everything above ASCII matches these
17619 _invlist_union_complement_2nd(nposixes,
17620 PL_XPosix_ptrs[_CC_ASCII],
17625 if (! DEPENDS_SEMANTICS) {
17627 /* For everything but /d, we can just add the current 'posixes' and
17628 * 'nposixes' to the main list */
17631 _invlist_union(cp_list, posixes, &cp_list);
17632 SvREFCNT_dec_NN(posixes);
17640 _invlist_union(cp_list, nposixes, &cp_list);
17641 SvREFCNT_dec_NN(nposixes);
17644 cp_list = nposixes;
17649 /* Under /d, things like \w match upper Latin1 characters only if
17650 * the target string is in UTF-8. But things like \W match all the
17651 * upper Latin1 characters if the target string is not in UTF-8.
17653 * Handle the case where there something like \W separately */
17655 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1);
17657 /* A complemented posix class matches all upper Latin1
17658 * characters if not in UTF-8. And it matches just certain
17659 * ones when in UTF-8. That means those certain ones are
17660 * matched regardless, so can just be added to the
17661 * unconditional list */
17663 _invlist_union(cp_list, nposixes, &cp_list);
17664 SvREFCNT_dec_NN(nposixes);
17668 cp_list = nposixes;
17671 /* Likewise for 'posixes' */
17672 _invlist_union(posixes, cp_list, &cp_list);
17674 /* Likewise for anything else in the range that matched only
17676 if (has_upper_latin1_only_utf8_matches) {
17677 _invlist_union(cp_list,
17678 has_upper_latin1_only_utf8_matches,
17680 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17681 has_upper_latin1_only_utf8_matches = NULL;
17684 /* If we don't match all the upper Latin1 characters regardless
17685 * of UTF-8ness, we have to set a flag to match the rest when
17687 _invlist_subtract(only_non_utf8_list, cp_list,
17688 &only_non_utf8_list);
17689 if (_invlist_len(only_non_utf8_list) != 0) {
17690 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17694 /* Here there were no complemented posix classes. That means
17695 * the upper Latin1 characters in 'posixes' match only when the
17696 * target string is in UTF-8. So we have to add them to the
17697 * list of those types of code points, while adding the
17698 * remainder to the unconditional list.
17700 * First calculate what they are */
17701 SV* nonascii_but_latin1_properties = NULL;
17702 _invlist_intersection(posixes, PL_UpperLatin1,
17703 &nonascii_but_latin1_properties);
17705 /* And add them to the final list of such characters. */
17706 _invlist_union(has_upper_latin1_only_utf8_matches,
17707 nonascii_but_latin1_properties,
17708 &has_upper_latin1_only_utf8_matches);
17710 /* Remove them from what now becomes the unconditional list */
17711 _invlist_subtract(posixes, nonascii_but_latin1_properties,
17714 /* And add those unconditional ones to the final list */
17716 _invlist_union(cp_list, posixes, &cp_list);
17717 SvREFCNT_dec_NN(posixes);
17724 SvREFCNT_dec(nonascii_but_latin1_properties);
17726 /* Get rid of any characters that we now know are matched
17727 * unconditionally from the conditional list, which may make
17728 * that list empty */
17729 _invlist_subtract(has_upper_latin1_only_utf8_matches,
17731 &has_upper_latin1_only_utf8_matches);
17732 if (_invlist_len(has_upper_latin1_only_utf8_matches) == 0) {
17733 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
17734 has_upper_latin1_only_utf8_matches = NULL;
17740 /* And combine the result (if any) with any inversion list from properties.
17741 * The lists are kept separate up to now so that we can distinguish the two
17742 * in regards to matching above-Unicode. A run-time warning is generated
17743 * if a Unicode property is matched against a non-Unicode code point. But,
17744 * we allow user-defined properties to match anything, without any warning,
17745 * and we also suppress the warning if there is a portion of the character
17746 * class that isn't a Unicode property, and which matches above Unicode, \W
17747 * or [\x{110000}] for example.
17748 * (Note that in this case, unlike the Posix one above, there is no
17749 * <has_upper_latin1_only_utf8_matches>, because having a Unicode property
17750 * forces Unicode semantics */
17754 /* If it matters to the final outcome, see if a non-property
17755 * component of the class matches above Unicode. If so, the
17756 * warning gets suppressed. This is true even if just a single
17757 * such code point is specified, as, though not strictly correct if
17758 * another such code point is matched against, the fact that they
17759 * are using above-Unicode code points indicates they should know
17760 * the issues involved */
17762 warn_super = ! (invert
17763 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
17766 _invlist_union(properties, cp_list, &cp_list);
17767 SvREFCNT_dec_NN(properties);
17770 cp_list = properties;
17775 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
17777 /* Because an ANYOF node is the only one that warns, this node
17778 * can't be optimized into something else */
17779 optimizable = FALSE;
17783 /* Here, we have calculated what code points should be in the character
17786 * Now we can see about various optimizations. Fold calculation (which we
17787 * did above) needs to take place before inversion. Otherwise /[^k]/i
17788 * would invert to include K, which under /i would match k, which it
17789 * shouldn't. Therefore we can't invert folded locale now, as it won't be
17790 * folded until runtime */
17792 /* If we didn't do folding, it's because some information isn't available
17793 * until runtime; set the run-time fold flag for these. (We don't have to
17794 * worry about properties folding, as that is taken care of by the swash
17795 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
17796 * locales, or the class matches at least one 0-255 range code point */
17799 /* Some things on the list might be unconditionally included because of
17800 * other components. Remove them, and clean up the list if it goes to
17802 if (only_utf8_locale_list && cp_list) {
17803 _invlist_subtract(only_utf8_locale_list, cp_list,
17804 &only_utf8_locale_list);
17806 if (_invlist_len(only_utf8_locale_list) == 0) {
17807 SvREFCNT_dec_NN(only_utf8_locale_list);
17808 only_utf8_locale_list = NULL;
17811 if (only_utf8_locale_list) {
17814 |ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
17816 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
17818 invlist_iterinit(cp_list);
17819 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
17820 ANYOF_FLAGS(ret) |= ANYOFL_FOLD;
17822 invlist_iterfinish(cp_list);
17825 else if ( DEPENDS_SEMANTICS
17826 && ( has_upper_latin1_only_utf8_matches
17827 || (ANYOF_FLAGS(ret) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
17830 optimizable = FALSE;
17834 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
17835 * at compile time. Besides not inverting folded locale now, we can't
17836 * invert if there are things such as \w, which aren't known until runtime
17840 && OP(ret) != ANYOFD
17841 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
17842 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
17844 _invlist_invert(cp_list);
17846 /* Any swash can't be used as-is, because we've inverted things */
17848 SvREFCNT_dec_NN(swash);
17852 /* Clear the invert flag since have just done it here */
17859 *ret_invlist = cp_list;
17860 SvREFCNT_dec(swash);
17862 /* Discard the generated node */
17864 RExC_size = orig_size;
17867 RExC_emit = orig_emit;
17872 /* Some character classes are equivalent to other nodes. Such nodes take
17873 * up less room and generally fewer operations to execute than ANYOF nodes.
17874 * Above, we checked for and optimized into some such equivalents for
17875 * certain common classes that are easy to test. Getting to this point in
17876 * the code means that the class didn't get optimized there. Since this
17877 * code is only executed in Pass 2, it is too late to save space--it has
17878 * been allocated in Pass 1, and currently isn't given back. But turning
17879 * things into an EXACTish node can allow the optimizer to join it to any
17880 * adjacent such nodes. And if the class is equivalent to things like /./,
17881 * expensive run-time swashes can be avoided. Now that we have more
17882 * complete information, we can find things necessarily missed by the
17883 * earlier code. Another possible "optimization" that isn't done is that
17884 * something like [Ee] could be changed into an EXACTFU. khw tried this
17885 * and found that the ANYOF is faster, including for code points not in the
17886 * bitmap. This still might make sense to do, provided it got joined with
17887 * an adjacent node(s) to create a longer EXACTFU one. This could be
17888 * accomplished by creating a pseudo ANYOF_EXACTFU node type that the join
17889 * routine would know is joinable. If that didn't happen, the node type
17890 * could then be made a straight ANYOF */
17892 if (optimizable && cp_list && ! invert) {
17894 U8 op = END; /* The optimzation node-type */
17895 int posix_class = -1; /* Illegal value */
17896 const char * cur_parse= RExC_parse;
17898 invlist_iterinit(cp_list);
17899 if (! invlist_iternext(cp_list, &start, &end)) {
17901 /* Here, the list is empty. This happens, for example, when a
17902 * Unicode property that doesn't match anything is the only element
17903 * in the character class (perluniprops.pod notes such properties).
17906 *flagp |= HASWIDTH|SIMPLE;
17908 else if (start == end) { /* The range is a single code point */
17909 if (! invlist_iternext(cp_list, &start, &end)
17911 /* Don't do this optimization if it would require changing
17912 * the pattern to UTF-8 */
17913 && (start < 256 || UTF))
17915 /* Here, the list contains a single code point. Can optimize
17916 * into an EXACTish node */
17927 /* A locale node under folding with one code point can be
17928 * an EXACTFL, as its fold won't be calculated until
17934 /* Here, we are generally folding, but there is only one
17935 * code point to match. If we have to, we use an EXACT
17936 * node, but it would be better for joining with adjacent
17937 * nodes in the optimization pass if we used the same
17938 * EXACTFish node that any such are likely to be. We can
17939 * do this iff the code point doesn't participate in any
17940 * folds. For example, an EXACTF of a colon is the same as
17941 * an EXACT one, since nothing folds to or from a colon. */
17943 if (IS_IN_SOME_FOLD_L1(value)) {
17948 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
17953 /* If we haven't found the node type, above, it means we
17954 * can use the prevailing one */
17956 op = compute_EXACTish(pRExC_state);
17960 } /* End of first range contains just a single code point */
17961 else if (start == 0) {
17962 if (end == UV_MAX) {
17964 *flagp |= HASWIDTH|SIMPLE;
17967 else if (end == '\n' - 1
17968 && invlist_iternext(cp_list, &start, &end)
17969 && start == '\n' + 1 && end == UV_MAX)
17972 *flagp |= HASWIDTH|SIMPLE;
17976 invlist_iterfinish(cp_list);
17979 const UV cp_list_len = _invlist_len(cp_list);
17980 const UV* cp_list_array = invlist_array(cp_list);
17982 /* Here, didn't find an optimization. See if this matches any of
17983 * the POSIX classes. These run slightly faster for above-Unicode
17984 * code points, so don't bother with POSIXA ones nor the 2 that
17985 * have no above-Unicode matches. We can avoid these checks unless
17986 * the ANYOF matches at least as high as the lowest POSIX one
17987 * (which was manually found to be \v. The actual code point may
17988 * increase in later Unicode releases, if a higher code point is
17989 * assigned to be \v, but this code will never break. It would
17990 * just mean we could execute the checks for posix optimizations
17991 * unnecessarily) */
17993 if (cp_list_array[cp_list_len-1] > 0x2029) {
17994 for (posix_class = 0;
17995 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
17999 if (posix_class == _CC_ASCII || posix_class == _CC_CNTRL) {
18002 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
18004 /* Check if matches normal or inverted */
18005 if (_invlistEQ(cp_list,
18006 PL_XPosix_ptrs[posix_class],
18009 op = (try_inverted)
18012 *flagp |= HASWIDTH|SIMPLE;
18022 RExC_parse = (char *)orig_parse;
18023 RExC_emit = (regnode *)orig_emit;
18025 if (regarglen[op]) {
18026 ret = reganode(pRExC_state, op, 0);
18028 ret = reg_node(pRExC_state, op);
18031 RExC_parse = (char *)cur_parse;
18033 if (PL_regkind[op] == EXACT) {
18034 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
18035 TRUE /* downgradable to EXACT */
18038 else if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
18039 FLAGS(ret) = posix_class;
18042 SvREFCNT_dec_NN(cp_list);
18047 /* Here, <cp_list> contains all the code points we can determine at
18048 * compile time that match under all conditions. Go through it, and
18049 * for things that belong in the bitmap, put them there, and delete from
18050 * <cp_list>. While we are at it, see if everything above 255 is in the
18051 * list, and if so, set a flag to speed up execution */
18053 populate_ANYOF_from_invlist(ret, &cp_list);
18056 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
18059 /* Here, the bitmap has been populated with all the Latin1 code points that
18060 * always match. Can now add to the overall list those that match only
18061 * when the target string is UTF-8 (<has_upper_latin1_only_utf8_matches>).
18063 if (has_upper_latin1_only_utf8_matches) {
18065 _invlist_union(cp_list,
18066 has_upper_latin1_only_utf8_matches,
18068 SvREFCNT_dec_NN(has_upper_latin1_only_utf8_matches);
18071 cp_list = has_upper_latin1_only_utf8_matches;
18073 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
18076 /* If there is a swash and more than one element, we can't use the swash in
18077 * the optimization below. */
18078 if (swash && element_count > 1) {
18079 SvREFCNT_dec_NN(swash);
18083 /* Note that the optimization of using 'swash' if it is the only thing in
18084 * the class doesn't have us change swash at all, so it can include things
18085 * that are also in the bitmap; otherwise we have purposely deleted that
18086 * duplicate information */
18087 set_ANYOF_arg(pRExC_state, ret, cp_list,
18088 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
18090 only_utf8_locale_list,
18091 swash, has_user_defined_property);
18093 *flagp |= HASWIDTH|SIMPLE;
18095 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
18096 RExC_contains_locale = 1;
18102 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
18105 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
18106 regnode* const node,
18108 SV* const runtime_defns,
18109 SV* const only_utf8_locale_list,
18111 const bool has_user_defined_property)
18113 /* Sets the arg field of an ANYOF-type node 'node', using information about
18114 * the node passed-in. If there is nothing outside the node's bitmap, the
18115 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
18116 * the count returned by add_data(), having allocated and stored an array,
18117 * av, that that count references, as follows:
18118 * av[0] stores the character class description in its textual form.
18119 * This is used later (regexec.c:Perl_regclass_swash()) to
18120 * initialize the appropriate swash, and is also useful for dumping
18121 * the regnode. This is set to &PL_sv_undef if the textual
18122 * description is not needed at run-time (as happens if the other
18123 * elements completely define the class)
18124 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
18125 * computed from av[0]. But if no further computation need be done,
18126 * the swash is stored here now (and av[0] is &PL_sv_undef).
18127 * av[2] stores the inversion list of code points that match only if the
18128 * current locale is UTF-8
18129 * av[3] stores the cp_list inversion list for use in addition or instead
18130 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
18131 * (Otherwise everything needed is already in av[0] and av[1])
18132 * av[4] is set if any component of the class is from a user-defined
18133 * property; used only if av[3] exists */
18137 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
18139 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
18140 assert(! (ANYOF_FLAGS(node)
18141 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
18142 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
18145 AV * const av = newAV();
18148 av_store(av, 0, (runtime_defns)
18149 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
18152 av_store(av, 1, swash);
18153 SvREFCNT_dec_NN(cp_list);
18156 av_store(av, 1, &PL_sv_undef);
18158 av_store(av, 3, cp_list);
18159 av_store(av, 4, newSVuv(has_user_defined_property));
18163 if (only_utf8_locale_list) {
18164 av_store(av, 2, only_utf8_locale_list);
18167 av_store(av, 2, &PL_sv_undef);
18170 rv = newRV_noinc(MUTABLE_SV(av));
18171 n = add_data(pRExC_state, STR_WITH_LEN("s"));
18172 RExC_rxi->data->data[n] = (void*)rv;
18177 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
18179 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
18180 const regnode* node,
18183 SV** only_utf8_locale_ptr,
18184 SV** output_invlist)
18187 /* For internal core use only.
18188 * Returns the swash for the input 'node' in the regex 'prog'.
18189 * If <doinit> is 'true', will attempt to create the swash if not already
18191 * If <listsvp> is non-null, will return the printable contents of the
18192 * swash. This can be used to get debugging information even before the
18193 * swash exists, by calling this function with 'doinit' set to false, in
18194 * which case the components that will be used to eventually create the
18195 * swash are returned (in a printable form).
18196 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
18197 * store an inversion list of code points that should match only if the
18198 * execution-time locale is a UTF-8 one.
18199 * If <output_invlist> is not NULL, it is where this routine is to store an
18200 * inversion list of the code points that would be instead returned in
18201 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
18202 * when this parameter is used, is just the non-code point data that
18203 * will go into creating the swash. This currently should be just
18204 * user-defined properties whose definitions were not known at compile
18205 * time. Using this parameter allows for easier manipulation of the
18206 * swash's data by the caller. It is illegal to call this function with
18207 * this parameter set, but not <listsvp>
18209 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
18210 * that, in spite of this function's name, the swash it returns may include
18211 * the bitmap data as well */
18214 SV *si = NULL; /* Input swash initialization string */
18215 SV* invlist = NULL;
18217 RXi_GET_DECL(prog,progi);
18218 const struct reg_data * const data = prog ? progi->data : NULL;
18220 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
18221 assert(! output_invlist || listsvp);
18223 if (data && data->count) {
18224 const U32 n = ARG(node);
18226 if (data->what[n] == 's') {
18227 SV * const rv = MUTABLE_SV(data->data[n]);
18228 AV * const av = MUTABLE_AV(SvRV(rv));
18229 SV **const ary = AvARRAY(av);
18230 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
18232 si = *ary; /* ary[0] = the string to initialize the swash with */
18234 if (av_tindex_skip_len_mg(av) >= 2) {
18235 if (only_utf8_locale_ptr
18237 && ary[2] != &PL_sv_undef)
18239 *only_utf8_locale_ptr = ary[2];
18242 assert(only_utf8_locale_ptr);
18243 *only_utf8_locale_ptr = NULL;
18246 /* Elements 3 and 4 are either both present or both absent. [3]
18247 * is any inversion list generated at compile time; [4]
18248 * indicates if that inversion list has any user-defined
18249 * properties in it. */
18250 if (av_tindex_skip_len_mg(av) >= 3) {
18252 if (SvUV(ary[4])) {
18253 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
18261 /* Element [1] is reserved for the set-up swash. If already there,
18262 * return it; if not, create it and store it there */
18263 if (ary[1] && SvROK(ary[1])) {
18266 else if (doinit && ((si && si != &PL_sv_undef)
18267 || (invlist && invlist != &PL_sv_undef))) {
18269 sw = _core_swash_init("utf8", /* the utf8 package */
18273 0, /* not from tr/// */
18275 &swash_init_flags);
18276 (void)av_store(av, 1, sw);
18281 /* If requested, return a printable version of what this swash matches */
18283 SV* matches_string = NULL;
18285 /* The swash should be used, if possible, to get the data, as it
18286 * contains the resolved data. But this function can be called at
18287 * compile-time, before everything gets resolved, in which case we
18288 * return the currently best available information, which is the string
18289 * that will eventually be used to do that resolving, 'si' */
18290 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
18291 && (si && si != &PL_sv_undef))
18293 /* Here, we only have 'si' (and possibly some passed-in data in
18294 * 'invlist', which is handled below) If the caller only wants
18295 * 'si', use that. */
18296 if (! output_invlist) {
18297 matches_string = newSVsv(si);
18300 /* But if the caller wants an inversion list of the node, we
18301 * need to parse 'si' and place as much as possible in the
18302 * desired output inversion list, making 'matches_string' only
18303 * contain the currently unresolvable things */
18304 const char *si_string = SvPVX(si);
18305 STRLEN remaining = SvCUR(si);
18309 /* Ignore everything before the first new-line */
18310 while (*si_string != '\n' && remaining > 0) {
18314 assert(remaining > 0);
18319 while (remaining > 0) {
18321 /* The data consists of just strings defining user-defined
18322 * property names, but in prior incarnations, and perhaps
18323 * somehow from pluggable regex engines, it could still
18324 * hold hex code point definitions. Each component of a
18325 * range would be separated by a tab, and each range by a
18326 * new-line. If these are found, instead add them to the
18327 * inversion list */
18328 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
18329 |PERL_SCAN_SILENT_NON_PORTABLE;
18330 STRLEN len = remaining;
18331 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
18333 /* If the hex decode routine found something, it should go
18334 * up to the next \n */
18335 if ( *(si_string + len) == '\n') {
18336 if (count) { /* 2nd code point on line */
18337 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
18340 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
18343 goto prepare_for_next_iteration;
18346 /* If the hex decode was instead for the lower range limit,
18347 * save it, and go parse the upper range limit */
18348 if (*(si_string + len) == '\t') {
18349 assert(count == 0);
18353 prepare_for_next_iteration:
18354 si_string += len + 1;
18355 remaining -= len + 1;
18359 /* Here, didn't find a legal hex number. Just add it from
18360 * here to the next \n */
18363 while (*(si_string + len) != '\n' && remaining > 0) {
18367 if (*(si_string + len) == '\n') {
18371 if (matches_string) {
18372 sv_catpvn(matches_string, si_string, len - 1);
18375 matches_string = newSVpvn(si_string, len - 1);
18378 sv_catpvs(matches_string, " ");
18379 } /* end of loop through the text */
18381 assert(matches_string);
18382 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
18383 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
18385 } /* end of has an 'si' but no swash */
18388 /* If we have a swash in place, its equivalent inversion list was above
18389 * placed into 'invlist'. If not, this variable may contain a stored
18390 * inversion list which is information beyond what is in 'si' */
18393 /* Again, if the caller doesn't want the output inversion list, put
18394 * everything in 'matches-string' */
18395 if (! output_invlist) {
18396 if ( ! matches_string) {
18397 matches_string = newSVpvs("\n");
18399 sv_catsv(matches_string, invlist_contents(invlist,
18400 TRUE /* traditional style */
18403 else if (! *output_invlist) {
18404 *output_invlist = invlist_clone(invlist);
18407 _invlist_union(*output_invlist, invlist, output_invlist);
18411 *listsvp = matches_string;
18416 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
18418 /* reg_skipcomment()
18420 Absorbs an /x style # comment from the input stream,
18421 returning a pointer to the first character beyond the comment, or if the
18422 comment terminates the pattern without anything following it, this returns
18423 one past the final character of the pattern (in other words, RExC_end) and
18424 sets the REG_RUN_ON_COMMENT_SEEN flag.
18426 Note it's the callers responsibility to ensure that we are
18427 actually in /x mode
18431 PERL_STATIC_INLINE char*
18432 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
18434 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
18438 while (p < RExC_end) {
18439 if (*(++p) == '\n') {
18444 /* we ran off the end of the pattern without ending the comment, so we have
18445 * to add an \n when wrapping */
18446 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
18451 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
18453 const bool force_to_xmod
18456 /* If the text at the current parse position '*p' is a '(?#...)' comment,
18457 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
18458 * is /x whitespace, advance '*p' so that on exit it points to the first
18459 * byte past all such white space and comments */
18461 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
18463 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
18465 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
18468 if (RExC_end - (*p) >= 3
18470 && *(*p + 1) == '?'
18471 && *(*p + 2) == '#')
18473 while (*(*p) != ')') {
18474 if ((*p) == RExC_end)
18475 FAIL("Sequence (?#... not terminated");
18483 const char * save_p = *p;
18484 while ((*p) < RExC_end) {
18486 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
18489 else if (*(*p) == '#') {
18490 (*p) = reg_skipcomment(pRExC_state, (*p));
18496 if (*p != save_p) {
18509 Advances the parse position by one byte, unless that byte is the beginning
18510 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
18511 those two cases, the parse position is advanced beyond all such comments and
18514 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
18518 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
18520 PERL_ARGS_ASSERT_NEXTCHAR;
18522 if (RExC_parse < RExC_end) {
18524 || UTF8_IS_INVARIANT(*RExC_parse)
18525 || UTF8_IS_START(*RExC_parse));
18527 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
18529 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
18530 FALSE /* Don't force /x */ );
18535 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
18537 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
18538 * space. In pass1, it aligns and increments RExC_size; in pass2,
18541 regnode * const ret = RExC_emit;
18542 GET_RE_DEBUG_FLAGS_DECL;
18544 PERL_ARGS_ASSERT_REGNODE_GUTS;
18546 assert(extra_size >= regarglen[op]);
18549 SIZE_ALIGN(RExC_size);
18550 RExC_size += 1 + extra_size;
18553 if (RExC_emit >= RExC_emit_bound)
18554 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
18555 op, (void*)RExC_emit, (void*)RExC_emit_bound);
18557 NODE_ALIGN_FILL(ret);
18558 #ifndef RE_TRACK_PATTERN_OFFSETS
18559 PERL_UNUSED_ARG(name);
18561 if (RExC_offsets) { /* MJD */
18563 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
18566 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
18567 ? "Overwriting end of array!\n" : "OK",
18568 (UV)(RExC_emit - RExC_emit_start),
18569 (UV)(RExC_parse - RExC_start),
18570 (UV)RExC_offsets[0]));
18571 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
18578 - reg_node - emit a node
18580 STATIC regnode * /* Location. */
18581 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
18583 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
18585 PERL_ARGS_ASSERT_REG_NODE;
18587 assert(regarglen[op] == 0);
18590 regnode *ptr = ret;
18591 FILL_ADVANCE_NODE(ptr, op);
18598 - reganode - emit a node with an argument
18600 STATIC regnode * /* Location. */
18601 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
18603 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
18605 PERL_ARGS_ASSERT_REGANODE;
18607 assert(regarglen[op] == 1);
18610 regnode *ptr = ret;
18611 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
18618 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
18620 /* emit a node with U32 and I32 arguments */
18622 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
18624 PERL_ARGS_ASSERT_REG2LANODE;
18626 assert(regarglen[op] == 2);
18629 regnode *ptr = ret;
18630 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
18637 - reginsert - insert an operator in front of already-emitted operand
18639 * Means relocating the operand.
18641 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
18642 * set up NEXT_OFF() of the inserted node if needed. Something like this:
18644 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
18646 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
18648 * ALSO NOTE - operand->flags will be set to 0 as well.
18651 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *operand, U32 depth)
18656 const int offset = regarglen[(U8)op];
18657 const int size = NODE_STEP_REGNODE + offset;
18658 GET_RE_DEBUG_FLAGS_DECL;
18660 PERL_ARGS_ASSERT_REGINSERT;
18661 PERL_UNUSED_CONTEXT;
18662 PERL_UNUSED_ARG(depth);
18663 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
18664 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
18669 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
18670 studying. If this is wrong then we need to adjust RExC_recurse
18671 below like we do with RExC_open_parens/RExC_close_parens. */
18675 if (RExC_open_parens) {
18677 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
18678 /* remember that RExC_npar is rex->nparens + 1,
18679 * iow it is 1 more than the number of parens seen in
18680 * the pattern so far. */
18681 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
18682 /* note, RExC_open_parens[0] is the start of the
18683 * regex, it can't move. RExC_close_parens[0] is the end
18684 * of the regex, it *can* move. */
18685 if ( paren && RExC_open_parens[paren] >= operand ) {
18686 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
18687 RExC_open_parens[paren] += size;
18689 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
18691 if ( RExC_close_parens[paren] >= operand ) {
18692 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
18693 RExC_close_parens[paren] += size;
18695 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
18700 RExC_end_op += size;
18702 while (src > operand) {
18703 StructCopy(--src, --dst, regnode);
18704 #ifdef RE_TRACK_PATTERN_OFFSETS
18705 if (RExC_offsets) { /* MJD 20010112 */
18707 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
18711 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
18712 ? "Overwriting end of array!\n" : "OK",
18713 (UV)(src - RExC_emit_start),
18714 (UV)(dst - RExC_emit_start),
18715 (UV)RExC_offsets[0]));
18716 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
18717 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
18722 place = operand; /* Op node, where operand used to be. */
18723 #ifdef RE_TRACK_PATTERN_OFFSETS
18724 if (RExC_offsets) { /* MJD */
18726 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
18730 (UV)(place - RExC_emit_start) > RExC_offsets[0]
18731 ? "Overwriting end of array!\n" : "OK",
18732 (UV)(place - RExC_emit_start),
18733 (UV)(RExC_parse - RExC_start),
18734 (UV)RExC_offsets[0]));
18735 Set_Node_Offset(place, RExC_parse);
18736 Set_Node_Length(place, 1);
18739 src = NEXTOPER(place);
18741 FILL_ADVANCE_NODE(place, op);
18742 Zero(src, offset, regnode);
18746 - regtail - set the next-pointer at the end of a node chain of p to val.
18747 - SEE ALSO: regtail_study
18750 S_regtail(pTHX_ RExC_state_t * pRExC_state,
18751 const regnode * const p,
18752 const regnode * const val,
18756 GET_RE_DEBUG_FLAGS_DECL;
18758 PERL_ARGS_ASSERT_REGTAIL;
18760 PERL_UNUSED_ARG(depth);
18766 /* Find last node. */
18767 scan = (regnode *) p;
18769 regnode * const temp = regnext(scan);
18771 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
18772 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18773 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
18774 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
18775 (temp == NULL ? "->" : ""),
18776 (temp == NULL ? PL_reg_name[OP(val)] : "")
18784 if (reg_off_by_arg[OP(scan)]) {
18785 ARG_SET(scan, val - scan);
18788 NEXT_OFF(scan) = val - scan;
18794 - regtail_study - set the next-pointer at the end of a node chain of p to val.
18795 - Look for optimizable sequences at the same time.
18796 - currently only looks for EXACT chains.
18798 This is experimental code. The idea is to use this routine to perform
18799 in place optimizations on branches and groups as they are constructed,
18800 with the long term intention of removing optimization from study_chunk so
18801 that it is purely analytical.
18803 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
18804 to control which is which.
18807 /* TODO: All four parms should be const */
18810 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
18811 const regnode *val,U32 depth)
18815 #ifdef EXPERIMENTAL_INPLACESCAN
18818 GET_RE_DEBUG_FLAGS_DECL;
18820 PERL_ARGS_ASSERT_REGTAIL_STUDY;
18826 /* Find last node. */
18830 regnode * const temp = regnext(scan);
18831 #ifdef EXPERIMENTAL_INPLACESCAN
18832 if (PL_regkind[OP(scan)] == EXACT) {
18833 bool unfolded_multi_char; /* Unexamined in this routine */
18834 if (join_exact(pRExC_state, scan, &min,
18835 &unfolded_multi_char, 1, val, depth+1))
18840 switch (OP(scan)) {
18844 case EXACTFA_NO_TRIE:
18850 if( exact == PSEUDO )
18852 else if ( exact != OP(scan) )
18861 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
18862 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
18863 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
18864 SvPV_nolen_const(RExC_mysv),
18865 REG_NODE_NUM(scan),
18866 PL_reg_name[exact]);
18873 DEBUG_PARSE_MSG("");
18874 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
18875 Perl_re_printf( aTHX_
18876 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
18877 SvPV_nolen_const(RExC_mysv),
18878 (IV)REG_NODE_NUM(val),
18882 if (reg_off_by_arg[OP(scan)]) {
18883 ARG_SET(scan, val - scan);
18886 NEXT_OFF(scan) = val - scan;
18894 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
18899 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
18904 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18906 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
18907 if (flags & (1<<bit)) {
18908 if (!set++ && lead)
18909 Perl_re_printf( aTHX_ "%s",lead);
18910 Perl_re_printf( aTHX_ "%s ",PL_reg_intflags_name[bit]);
18915 Perl_re_printf( aTHX_ "\n");
18917 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18922 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
18928 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
18930 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
18931 if (flags & (1<<bit)) {
18932 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
18935 if (!set++ && lead)
18936 Perl_re_printf( aTHX_ "%s",lead);
18937 Perl_re_printf( aTHX_ "%s ",PL_reg_extflags_name[bit]);
18940 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
18941 if (!set++ && lead) {
18942 Perl_re_printf( aTHX_ "%s",lead);
18945 case REGEX_UNICODE_CHARSET:
18946 Perl_re_printf( aTHX_ "UNICODE");
18948 case REGEX_LOCALE_CHARSET:
18949 Perl_re_printf( aTHX_ "LOCALE");
18951 case REGEX_ASCII_RESTRICTED_CHARSET:
18952 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
18954 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
18955 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
18958 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
18964 Perl_re_printf( aTHX_ "\n");
18966 Perl_re_printf( aTHX_ "%s[none-set]\n",lead);
18972 Perl_regdump(pTHX_ const regexp *r)
18976 SV * const sv = sv_newmortal();
18977 SV *dsv= sv_newmortal();
18978 RXi_GET_DECL(r,ri);
18979 GET_RE_DEBUG_FLAGS_DECL;
18981 PERL_ARGS_ASSERT_REGDUMP;
18983 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
18985 /* Header fields of interest. */
18986 for (i = 0; i < 2; i++) {
18987 if (r->substrs->data[i].substr) {
18988 RE_PV_QUOTED_DECL(s, 0, dsv,
18989 SvPVX_const(r->substrs->data[i].substr),
18990 RE_SV_DUMPLEN(r->substrs->data[i].substr),
18991 PL_dump_re_max_len);
18992 Perl_re_printf( aTHX_
18993 "%s %s%s at %" IVdf "..%" UVuf " ",
18994 i ? "floating" : "anchored",
18996 RE_SV_TAIL(r->substrs->data[i].substr),
18997 (IV)r->substrs->data[i].min_offset,
18998 (UV)r->substrs->data[i].max_offset);
19000 else if (r->substrs->data[i].utf8_substr) {
19001 RE_PV_QUOTED_DECL(s, 1, dsv,
19002 SvPVX_const(r->substrs->data[i].utf8_substr),
19003 RE_SV_DUMPLEN(r->substrs->data[i].utf8_substr),
19005 Perl_re_printf( aTHX_
19006 "%s utf8 %s%s at %" IVdf "..%" UVuf " ",
19007 i ? "floating" : "anchored",
19009 RE_SV_TAIL(r->substrs->data[i].utf8_substr),
19010 (IV)r->substrs->data[i].min_offset,
19011 (UV)r->substrs->data[i].max_offset);
19015 if (r->check_substr || r->check_utf8)
19016 Perl_re_printf( aTHX_
19018 ( r->check_substr == r->substrs->data[1].substr
19019 && r->check_utf8 == r->substrs->data[1].utf8_substr
19020 ? "(checking floating" : "(checking anchored"));
19021 if (r->intflags & PREGf_NOSCAN)
19022 Perl_re_printf( aTHX_ " noscan");
19023 if (r->extflags & RXf_CHECK_ALL)
19024 Perl_re_printf( aTHX_ " isall");
19025 if (r->check_substr || r->check_utf8)
19026 Perl_re_printf( aTHX_ ") ");
19028 if (ri->regstclass) {
19029 regprop(r, sv, ri->regstclass, NULL, NULL);
19030 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
19032 if (r->intflags & PREGf_ANCH) {
19033 Perl_re_printf( aTHX_ "anchored");
19034 if (r->intflags & PREGf_ANCH_MBOL)
19035 Perl_re_printf( aTHX_ "(MBOL)");
19036 if (r->intflags & PREGf_ANCH_SBOL)
19037 Perl_re_printf( aTHX_ "(SBOL)");
19038 if (r->intflags & PREGf_ANCH_GPOS)
19039 Perl_re_printf( aTHX_ "(GPOS)");
19040 Perl_re_printf( aTHX_ " ");
19042 if (r->intflags & PREGf_GPOS_SEEN)
19043 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
19044 if (r->intflags & PREGf_SKIP)
19045 Perl_re_printf( aTHX_ "plus ");
19046 if (r->intflags & PREGf_IMPLICIT)
19047 Perl_re_printf( aTHX_ "implicit ");
19048 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
19049 if (r->extflags & RXf_EVAL_SEEN)
19050 Perl_re_printf( aTHX_ "with eval ");
19051 Perl_re_printf( aTHX_ "\n");
19053 regdump_extflags("r->extflags: ",r->extflags);
19054 regdump_intflags("r->intflags: ",r->intflags);
19057 PERL_ARGS_ASSERT_REGDUMP;
19058 PERL_UNUSED_CONTEXT;
19059 PERL_UNUSED_ARG(r);
19060 #endif /* DEBUGGING */
19063 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
19066 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
19067 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
19068 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
19069 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
19070 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
19071 || _CC_VERTSPACE != 15
19072 # error Need to adjust order of anyofs[]
19074 static const char * const anyofs[] = {
19111 - regprop - printable representation of opcode, with run time support
19115 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
19119 RXi_GET_DECL(prog,progi);
19120 GET_RE_DEBUG_FLAGS_DECL;
19122 PERL_ARGS_ASSERT_REGPROP;
19126 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
19127 /* It would be nice to FAIL() here, but this may be called from
19128 regexec.c, and it would be hard to supply pRExC_state. */
19129 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
19130 (int)OP(o), (int)REGNODE_MAX);
19131 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
19133 k = PL_regkind[OP(o)];
19136 sv_catpvs(sv, " ");
19137 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
19138 * is a crude hack but it may be the best for now since
19139 * we have no flag "this EXACTish node was UTF-8"
19141 pv_pretty(sv, STRING(o), STR_LEN(o), PL_dump_re_max_len,
19142 PL_colors[0], PL_colors[1],
19143 PERL_PV_ESCAPE_UNI_DETECT |
19144 PERL_PV_ESCAPE_NONASCII |
19145 PERL_PV_PRETTY_ELLIPSES |
19146 PERL_PV_PRETTY_LTGT |
19147 PERL_PV_PRETTY_NOCLEAR
19149 } else if (k == TRIE) {
19150 /* print the details of the trie in dumpuntil instead, as
19151 * progi->data isn't available here */
19152 const char op = OP(o);
19153 const U32 n = ARG(o);
19154 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
19155 (reg_ac_data *)progi->data->data[n] :
19157 const reg_trie_data * const trie
19158 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
19160 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
19161 DEBUG_TRIE_COMPILE_r({
19163 sv_catpvs(sv, "(JUMP)");
19164 Perl_sv_catpvf(aTHX_ sv,
19165 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
19166 (UV)trie->startstate,
19167 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
19168 (UV)trie->wordcount,
19171 (UV)TRIE_CHARCOUNT(trie),
19172 (UV)trie->uniquecharcount
19175 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
19176 sv_catpvs(sv, "[");
19177 (void) put_charclass_bitmap_innards(sv,
19178 ((IS_ANYOF_TRIE(op))
19180 : TRIE_BITMAP(trie)),
19186 sv_catpvs(sv, "]");
19188 } else if (k == CURLY) {
19189 U32 lo = ARG1(o), hi = ARG2(o);
19190 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
19191 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
19192 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
19193 if (hi == REG_INFTY)
19194 sv_catpvs(sv, "INFTY");
19196 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
19197 sv_catpvs(sv, "}");
19199 else if (k == WHILEM && o->flags) /* Ordinal/of */
19200 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
19201 else if (k == REF || k == OPEN || k == CLOSE
19202 || k == GROUPP || OP(o)==ACCEPT)
19204 AV *name_list= NULL;
19205 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
19206 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
19207 if ( RXp_PAREN_NAMES(prog) ) {
19208 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19209 } else if ( pRExC_state ) {
19210 name_list= RExC_paren_name_list;
19213 if ( k != REF || (OP(o) < NREF)) {
19214 SV **name= av_fetch(name_list, parno, 0 );
19216 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19219 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
19220 I32 *nums=(I32*)SvPVX(sv_dat);
19221 SV **name= av_fetch(name_list, nums[0], 0 );
19224 for ( n=0; n<SvIVX(sv_dat); n++ ) {
19225 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
19226 (n ? "," : ""), (IV)nums[n]);
19228 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19232 if ( k == REF && reginfo) {
19233 U32 n = ARG(o); /* which paren pair */
19234 I32 ln = prog->offs[n].start;
19235 if (prog->lastparen < n || ln == -1)
19236 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
19237 else if (ln == prog->offs[n].end)
19238 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
19240 const char *s = reginfo->strbeg + ln;
19241 Perl_sv_catpvf(aTHX_ sv, ": ");
19242 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
19243 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
19246 } else if (k == GOSUB) {
19247 AV *name_list= NULL;
19248 if ( RXp_PAREN_NAMES(prog) ) {
19249 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
19250 } else if ( pRExC_state ) {
19251 name_list= RExC_paren_name_list;
19254 /* Paren and offset */
19255 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
19256 (int)((o + (int)ARG2L(o)) - progi->program) );
19258 SV **name= av_fetch(name_list, ARG(o), 0 );
19260 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
19263 else if (k == LOGICAL)
19264 /* 2: embedded, otherwise 1 */
19265 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
19266 else if (k == ANYOF) {
19267 const U8 flags = ANYOF_FLAGS(o);
19268 bool do_sep = FALSE; /* Do we need to separate various components of
19270 /* Set if there is still an unresolved user-defined property */
19271 SV *unresolved = NULL;
19273 /* Things that are ignored except when the runtime locale is UTF-8 */
19274 SV *only_utf8_locale_invlist = NULL;
19276 /* Code points that don't fit in the bitmap */
19277 SV *nonbitmap_invlist = NULL;
19279 /* And things that aren't in the bitmap, but are small enough to be */
19280 SV* bitmap_range_not_in_bitmap = NULL;
19282 const bool inverted = flags & ANYOF_INVERT;
19284 if (OP(o) == ANYOFL) {
19285 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
19286 sv_catpvs(sv, "{utf8-locale-reqd}");
19288 if (flags & ANYOFL_FOLD) {
19289 sv_catpvs(sv, "{i}");
19293 /* If there is stuff outside the bitmap, get it */
19294 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
19295 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
19297 &only_utf8_locale_invlist,
19298 &nonbitmap_invlist);
19299 /* The non-bitmap data may contain stuff that could fit in the
19300 * bitmap. This could come from a user-defined property being
19301 * finally resolved when this call was done; or much more likely
19302 * because there are matches that require UTF-8 to be valid, and so
19303 * aren't in the bitmap. This is teased apart later */
19304 _invlist_intersection(nonbitmap_invlist,
19306 &bitmap_range_not_in_bitmap);
19307 /* Leave just the things that don't fit into the bitmap */
19308 _invlist_subtract(nonbitmap_invlist,
19310 &nonbitmap_invlist);
19313 /* Obey this flag to add all above-the-bitmap code points */
19314 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
19315 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
19316 NUM_ANYOF_CODE_POINTS,
19320 /* Ready to start outputting. First, the initial left bracket */
19321 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
19323 /* Then all the things that could fit in the bitmap */
19324 do_sep = put_charclass_bitmap_innards(sv,
19326 bitmap_range_not_in_bitmap,
19327 only_utf8_locale_invlist,
19330 /* Can't try inverting for a
19331 * better display if there are
19332 * things that haven't been
19334 unresolved != NULL);
19335 SvREFCNT_dec(bitmap_range_not_in_bitmap);
19337 /* If there are user-defined properties which haven't been defined yet,
19338 * output them. If the result is not to be inverted, it is clearest to
19339 * output them in a separate [] from the bitmap range stuff. If the
19340 * result is to be complemented, we have to show everything in one [],
19341 * as the inversion applies to the whole thing. Use {braces} to
19342 * separate them from anything in the bitmap and anything above the
19346 if (! do_sep) { /* If didn't output anything in the bitmap */
19347 sv_catpvs(sv, "^");
19349 sv_catpvs(sv, "{");
19352 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19354 sv_catsv(sv, unresolved);
19356 sv_catpvs(sv, "}");
19358 do_sep = ! inverted;
19361 /* And, finally, add the above-the-bitmap stuff */
19362 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
19365 /* See if truncation size is overridden */
19366 const STRLEN dump_len = (PL_dump_re_max_len > 256)
19367 ? PL_dump_re_max_len
19370 /* This is output in a separate [] */
19372 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
19375 /* And, for easy of understanding, it is shown in the
19376 * uncomplemented form if possible. The one exception being if
19377 * there are unresolved items, where the inversion has to be
19378 * delayed until runtime */
19379 if (inverted && ! unresolved) {
19380 _invlist_invert(nonbitmap_invlist);
19381 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
19384 contents = invlist_contents(nonbitmap_invlist,
19385 FALSE /* output suitable for catsv */
19388 /* If the output is shorter than the permissible maximum, just do it. */
19389 if (SvCUR(contents) <= dump_len) {
19390 sv_catsv(sv, contents);
19393 const char * contents_string = SvPVX(contents);
19394 STRLEN i = dump_len;
19396 /* Otherwise, start at the permissible max and work back to the
19397 * first break possibility */
19398 while (i > 0 && contents_string[i] != ' ') {
19401 if (i == 0) { /* Fail-safe. Use the max if we couldn't
19402 find a legal break */
19406 sv_catpvn(sv, contents_string, i);
19407 sv_catpvs(sv, "...");
19410 SvREFCNT_dec_NN(contents);
19411 SvREFCNT_dec_NN(nonbitmap_invlist);
19414 /* And finally the matching, closing ']' */
19415 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
19417 SvREFCNT_dec(unresolved);
19419 else if (k == POSIXD || k == NPOSIXD) {
19420 U8 index = FLAGS(o) * 2;
19421 if (index < C_ARRAY_LENGTH(anyofs)) {
19422 if (*anyofs[index] != '[') {
19425 sv_catpv(sv, anyofs[index]);
19426 if (*anyofs[index] != '[') {
19431 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
19434 else if (k == BOUND || k == NBOUND) {
19435 /* Must be synced with order of 'bound_type' in regcomp.h */
19436 const char * const bounds[] = {
19437 "", /* Traditional */
19443 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
19444 sv_catpv(sv, bounds[FLAGS(o)]);
19446 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
19447 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
19448 else if (OP(o) == SBOL)
19449 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
19451 /* add on the verb argument if there is one */
19452 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
19454 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
19455 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
19457 sv_catpvs(sv, ":NULL");
19460 PERL_UNUSED_CONTEXT;
19461 PERL_UNUSED_ARG(sv);
19462 PERL_UNUSED_ARG(o);
19463 PERL_UNUSED_ARG(prog);
19464 PERL_UNUSED_ARG(reginfo);
19465 PERL_UNUSED_ARG(pRExC_state);
19466 #endif /* DEBUGGING */
19472 Perl_re_intuit_string(pTHX_ REGEXP * const r)
19473 { /* Assume that RE_INTUIT is set */
19474 struct regexp *const prog = ReANY(r);
19475 GET_RE_DEBUG_FLAGS_DECL;
19477 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
19478 PERL_UNUSED_CONTEXT;
19482 const char * const s = SvPV_nolen_const(RX_UTF8(r)
19483 ? prog->check_utf8 : prog->check_substr);
19485 if (!PL_colorset) reginitcolors();
19486 Perl_re_printf( aTHX_
19487 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
19489 RX_UTF8(r) ? "utf8 " : "",
19490 PL_colors[5],PL_colors[0],
19493 (strlen(s) > PL_dump_re_max_len ? "..." : ""));
19496 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
19497 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
19503 handles refcounting and freeing the perl core regexp structure. When
19504 it is necessary to actually free the structure the first thing it
19505 does is call the 'free' method of the regexp_engine associated to
19506 the regexp, allowing the handling of the void *pprivate; member
19507 first. (This routine is not overridable by extensions, which is why
19508 the extensions free is called first.)
19510 See regdupe and regdupe_internal if you change anything here.
19512 #ifndef PERL_IN_XSUB_RE
19514 Perl_pregfree(pTHX_ REGEXP *r)
19520 Perl_pregfree2(pTHX_ REGEXP *rx)
19522 struct regexp *const r = ReANY(rx);
19523 GET_RE_DEBUG_FLAGS_DECL;
19525 PERL_ARGS_ASSERT_PREGFREE2;
19527 if (r->mother_re) {
19528 ReREFCNT_dec(r->mother_re);
19530 CALLREGFREE_PVT(rx); /* free the private data */
19531 SvREFCNT_dec(RXp_PAREN_NAMES(r));
19535 for (i = 0; i < 2; i++) {
19536 SvREFCNT_dec(r->substrs->data[i].substr);
19537 SvREFCNT_dec(r->substrs->data[i].utf8_substr);
19539 Safefree(r->substrs);
19541 RX_MATCH_COPY_FREE(rx);
19542 #ifdef PERL_ANY_COW
19543 SvREFCNT_dec(r->saved_copy);
19546 SvREFCNT_dec(r->qr_anoncv);
19547 if (r->recurse_locinput)
19548 Safefree(r->recurse_locinput);
19554 Copy ssv to dsv, both of which should of type SVt_REGEXP or SVt_PVLV,
19555 except that dsv will be created if NULL.
19557 This function is used in two main ways. First to implement
19558 $r = qr/....; $s = $$r;
19560 Secondly, it is used as a hacky workaround to the structural issue of
19562 being stored in the regexp structure which is in turn stored in
19563 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
19564 could be PL_curpm in multiple contexts, and could require multiple
19565 result sets being associated with the pattern simultaneously, such
19566 as when doing a recursive match with (??{$qr})
19568 The solution is to make a lightweight copy of the regexp structure
19569 when a qr// is returned from the code executed by (??{$qr}) this
19570 lightweight copy doesn't actually own any of its data except for
19571 the starp/end and the actual regexp structure itself.
19577 Perl_reg_temp_copy(pTHX_ REGEXP *dsv, REGEXP *ssv)
19579 struct regexp *drx;
19580 struct regexp *const srx = ReANY(ssv);
19581 const bool islv = dsv && SvTYPE(dsv) == SVt_PVLV;
19583 PERL_ARGS_ASSERT_REG_TEMP_COPY;
19586 dsv = (REGEXP*) newSV_type(SVt_REGEXP);
19588 SvOK_off((SV *)dsv);
19590 /* For PVLVs, the head (sv_any) points to an XPVLV, while
19591 * the LV's xpvlenu_rx will point to a regexp body, which
19592 * we allocate here */
19593 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
19594 assert(!SvPVX(dsv));
19595 ((XPV*)SvANY(dsv))->xpv_len_u.xpvlenu_rx = temp->sv_any;
19596 temp->sv_any = NULL;
19597 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
19598 SvREFCNT_dec_NN(temp);
19599 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
19600 ing below will not set it. */
19601 SvCUR_set(dsv, SvCUR(ssv));
19604 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
19605 sv_force_normal(sv) is called. */
19609 SvFLAGS(dsv) |= SvFLAGS(ssv) & (SVf_POK|SVp_POK|SVf_UTF8);
19610 SvPV_set(dsv, RX_WRAPPED(ssv));
19611 /* We share the same string buffer as the original regexp, on which we
19612 hold a reference count, incremented when mother_re is set below.
19613 The string pointer is copied here, being part of the regexp struct.
19615 memcpy(&(drx->xpv_cur), &(srx->xpv_cur),
19616 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
19620 const I32 npar = srx->nparens+1;
19621 Newx(drx->offs, npar, regexp_paren_pair);
19622 Copy(srx->offs, drx->offs, npar, regexp_paren_pair);
19624 if (srx->substrs) {
19626 Newx(drx->substrs, 1, struct reg_substr_data);
19627 StructCopy(srx->substrs, drx->substrs, struct reg_substr_data);
19629 for (i = 0; i < 2; i++) {
19630 SvREFCNT_inc_void(drx->substrs->data[i].substr);
19631 SvREFCNT_inc_void(drx->substrs->data[i].utf8_substr);
19634 /* check_substr and check_utf8, if non-NULL, point to either their
19635 anchored or float namesakes, and don't hold a second reference. */
19637 RX_MATCH_COPIED_off(dsv);
19638 #ifdef PERL_ANY_COW
19639 drx->saved_copy = NULL;
19641 drx->mother_re = ReREFCNT_inc(srx->mother_re ? srx->mother_re : ssv);
19642 SvREFCNT_inc_void(drx->qr_anoncv);
19643 if (srx->recurse_locinput)
19644 Newx(drx->recurse_locinput,srx->nparens + 1,char *);
19651 /* regfree_internal()
19653 Free the private data in a regexp. This is overloadable by
19654 extensions. Perl takes care of the regexp structure in pregfree(),
19655 this covers the *pprivate pointer which technically perl doesn't
19656 know about, however of course we have to handle the
19657 regexp_internal structure when no extension is in use.
19659 Note this is called before freeing anything in the regexp
19664 Perl_regfree_internal(pTHX_ REGEXP * const rx)
19666 struct regexp *const r = ReANY(rx);
19667 RXi_GET_DECL(r,ri);
19668 GET_RE_DEBUG_FLAGS_DECL;
19670 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
19676 SV *dsv= sv_newmortal();
19677 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
19678 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), PL_dump_re_max_len);
19679 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
19680 PL_colors[4],PL_colors[5],s);
19683 #ifdef RE_TRACK_PATTERN_OFFSETS
19685 Safefree(ri->u.offsets); /* 20010421 MJD */
19687 if (ri->code_blocks)
19688 S_free_codeblocks(aTHX_ ri->code_blocks);
19691 int n = ri->data->count;
19694 /* If you add a ->what type here, update the comment in regcomp.h */
19695 switch (ri->data->what[n]) {
19701 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
19704 Safefree(ri->data->data[n]);
19710 { /* Aho Corasick add-on structure for a trie node.
19711 Used in stclass optimization only */
19713 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
19714 #ifdef USE_ITHREADS
19718 refcount = --aho->refcount;
19721 PerlMemShared_free(aho->states);
19722 PerlMemShared_free(aho->fail);
19723 /* do this last!!!! */
19724 PerlMemShared_free(ri->data->data[n]);
19725 /* we should only ever get called once, so
19726 * assert as much, and also guard the free
19727 * which /might/ happen twice. At the least
19728 * it will make code anlyzers happy and it
19729 * doesn't cost much. - Yves */
19730 assert(ri->regstclass);
19731 if (ri->regstclass) {
19732 PerlMemShared_free(ri->regstclass);
19733 ri->regstclass = 0;
19740 /* trie structure. */
19742 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
19743 #ifdef USE_ITHREADS
19747 refcount = --trie->refcount;
19750 PerlMemShared_free(trie->charmap);
19751 PerlMemShared_free(trie->states);
19752 PerlMemShared_free(trie->trans);
19754 PerlMemShared_free(trie->bitmap);
19756 PerlMemShared_free(trie->jump);
19757 PerlMemShared_free(trie->wordinfo);
19758 /* do this last!!!! */
19759 PerlMemShared_free(ri->data->data[n]);
19764 Perl_croak(aTHX_ "panic: regfree data code '%c'",
19765 ri->data->what[n]);
19768 Safefree(ri->data->what);
19769 Safefree(ri->data);
19775 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
19776 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
19777 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
19780 re_dup_guts - duplicate a regexp.
19782 This routine is expected to clone a given regexp structure. It is only
19783 compiled under USE_ITHREADS.
19785 After all of the core data stored in struct regexp is duplicated
19786 the regexp_engine.dupe method is used to copy any private data
19787 stored in the *pprivate pointer. This allows extensions to handle
19788 any duplication it needs to do.
19790 See pregfree() and regfree_internal() if you change anything here.
19792 #if defined(USE_ITHREADS)
19793 #ifndef PERL_IN_XSUB_RE
19795 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
19799 const struct regexp *r = ReANY(sstr);
19800 struct regexp *ret = ReANY(dstr);
19802 PERL_ARGS_ASSERT_RE_DUP_GUTS;
19804 npar = r->nparens+1;
19805 Newx(ret->offs, npar, regexp_paren_pair);
19806 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
19808 if (ret->substrs) {
19809 /* Do it this way to avoid reading from *r after the StructCopy().
19810 That way, if any of the sv_dup_inc()s dislodge *r from the L1
19811 cache, it doesn't matter. */
19813 const bool anchored = r->check_substr
19814 ? r->check_substr == r->substrs->data[0].substr
19815 : r->check_utf8 == r->substrs->data[0].utf8_substr;
19816 Newx(ret->substrs, 1, struct reg_substr_data);
19817 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
19819 for (i = 0; i < 2; i++) {
19820 ret->substrs->data[i].substr =
19821 sv_dup_inc(ret->substrs->data[i].substr, param);
19822 ret->substrs->data[i].utf8_substr =
19823 sv_dup_inc(ret->substrs->data[i].utf8_substr, param);
19826 /* check_substr and check_utf8, if non-NULL, point to either their
19827 anchored or float namesakes, and don't hold a second reference. */
19829 if (ret->check_substr) {
19831 assert(r->check_utf8 == r->substrs->data[0].utf8_substr);
19833 ret->check_substr = ret->substrs->data[0].substr;
19834 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
19836 assert(r->check_substr == r->substrs->data[1].substr);
19837 assert(r->check_utf8 == r->substrs->data[1].utf8_substr);
19839 ret->check_substr = ret->substrs->data[1].substr;
19840 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
19842 } else if (ret->check_utf8) {
19844 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
19846 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
19851 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
19852 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
19853 if (r->recurse_locinput)
19854 Newx(ret->recurse_locinput,r->nparens + 1,char *);
19857 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
19859 if (RX_MATCH_COPIED(dstr))
19860 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
19862 ret->subbeg = NULL;
19863 #ifdef PERL_ANY_COW
19864 ret->saved_copy = NULL;
19867 /* Whether mother_re be set or no, we need to copy the string. We
19868 cannot refrain from copying it when the storage points directly to
19869 our mother regexp, because that's
19870 1: a buffer in a different thread
19871 2: something we no longer hold a reference on
19872 so we need to copy it locally. */
19873 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED_const(sstr), SvCUR(sstr)+1);
19874 ret->mother_re = NULL;
19876 #endif /* PERL_IN_XSUB_RE */
19881 This is the internal complement to regdupe() which is used to copy
19882 the structure pointed to by the *pprivate pointer in the regexp.
19883 This is the core version of the extension overridable cloning hook.
19884 The regexp structure being duplicated will be copied by perl prior
19885 to this and will be provided as the regexp *r argument, however
19886 with the /old/ structures pprivate pointer value. Thus this routine
19887 may override any copying normally done by perl.
19889 It returns a pointer to the new regexp_internal structure.
19893 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
19896 struct regexp *const r = ReANY(rx);
19897 regexp_internal *reti;
19899 RXi_GET_DECL(r,ri);
19901 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
19905 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
19906 char, regexp_internal);
19907 Copy(ri->program, reti->program, len+1, regnode);
19910 if (ri->code_blocks) {
19912 Newx(reti->code_blocks, 1, struct reg_code_blocks);
19913 Newx(reti->code_blocks->cb, ri->code_blocks->count,
19914 struct reg_code_block);
19915 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
19916 ri->code_blocks->count, struct reg_code_block);
19917 for (n = 0; n < ri->code_blocks->count; n++)
19918 reti->code_blocks->cb[n].src_regex = (REGEXP*)
19919 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
19920 reti->code_blocks->count = ri->code_blocks->count;
19921 reti->code_blocks->refcnt = 1;
19924 reti->code_blocks = NULL;
19926 reti->regstclass = NULL;
19929 struct reg_data *d;
19930 const int count = ri->data->count;
19933 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
19934 char, struct reg_data);
19935 Newx(d->what, count, U8);
19938 for (i = 0; i < count; i++) {
19939 d->what[i] = ri->data->what[i];
19940 switch (d->what[i]) {
19941 /* see also regcomp.h and regfree_internal() */
19942 case 'a': /* actually an AV, but the dup function is identical.
19943 values seem to be "plain sv's" generally. */
19944 case 'r': /* a compiled regex (but still just another SV) */
19945 case 's': /* an RV (currently only used for an RV to an AV by the ANYOF code)
19946 this use case should go away, the code could have used
19947 'a' instead - see S_set_ANYOF_arg() for array contents. */
19948 case 'S': /* actually an SV, but the dup function is identical. */
19949 case 'u': /* actually an HV, but the dup function is identical.
19950 values are "plain sv's" */
19951 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
19954 /* Synthetic Start Class - "Fake" charclass we generate to optimize
19955 * patterns which could start with several different things. Pre-TRIE
19956 * this was more important than it is now, however this still helps
19957 * in some places, for instance /x?a+/ might produce a SSC equivalent
19958 * to [xa]. This is used by Perl_re_intuit_start() and S_find_byclass()
19961 /* This is cheating. */
19962 Newx(d->data[i], 1, regnode_ssc);
19963 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
19964 reti->regstclass = (regnode*)d->data[i];
19967 /* AHO-CORASICK fail table */
19968 /* Trie stclasses are readonly and can thus be shared
19969 * without duplication. We free the stclass in pregfree
19970 * when the corresponding reg_ac_data struct is freed.
19972 reti->regstclass= ri->regstclass;
19975 /* TRIE transition table */
19977 ((reg_trie_data*)ri->data->data[i])->refcount++;
19980 case 'l': /* (?{...}) or (??{ ... }) code (cb->block) */
19981 case 'L': /* same when RExC_pm_flags & PMf_HAS_CV and code
19982 is not from another regexp */
19983 d->data[i] = ri->data->data[i];
19986 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
19987 ri->data->what[i]);
19996 reti->name_list_idx = ri->name_list_idx;
19998 #ifdef RE_TRACK_PATTERN_OFFSETS
19999 if (ri->u.offsets) {
20000 Newx(reti->u.offsets, 2*len+1, U32);
20001 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
20004 SetProgLen(reti,len);
20007 return (void*)reti;
20010 #endif /* USE_ITHREADS */
20012 #ifndef PERL_IN_XSUB_RE
20015 - regnext - dig the "next" pointer out of a node
20018 Perl_regnext(pTHX_ regnode *p)
20025 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
20026 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
20027 (int)OP(p), (int)REGNODE_MAX);
20030 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
20039 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
20042 STRLEN l1 = strlen(pat1);
20043 STRLEN l2 = strlen(pat2);
20046 const char *message;
20048 PERL_ARGS_ASSERT_RE_CROAK2;
20054 Copy(pat1, buf, l1 , char);
20055 Copy(pat2, buf + l1, l2 , char);
20056 buf[l1 + l2] = '\n';
20057 buf[l1 + l2 + 1] = '\0';
20058 va_start(args, pat2);
20059 msv = vmess(buf, &args);
20061 message = SvPV_const(msv,l1);
20064 Copy(message, buf, l1 , char);
20065 /* l1-1 to avoid \n */
20066 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, l1-1, buf));
20069 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
20071 #ifndef PERL_IN_XSUB_RE
20073 Perl_save_re_context(pTHX)
20078 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
20081 const REGEXP * const rx = PM_GETRE(PL_curpm);
20083 nparens = RX_NPARENS(rx);
20086 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
20087 * that PL_curpm will be null, but that utf8.pm and the modules it
20088 * loads will only use $1..$3.
20089 * The t/porting/re_context.t test file checks this assumption.
20094 for (i = 1; i <= nparens; i++) {
20095 char digits[TYPE_CHARS(long)];
20096 const STRLEN len = my_snprintf(digits, sizeof(digits),
20098 GV *const *const gvp
20099 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
20102 GV * const gv = *gvp;
20103 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
20113 S_put_code_point(pTHX_ SV *sv, UV c)
20115 PERL_ARGS_ASSERT_PUT_CODE_POINT;
20118 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
20120 else if (isPRINT(c)) {
20121 const char string = (char) c;
20123 /* We use {phrase} as metanotation in the class, so also escape literal
20125 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
20126 sv_catpvs(sv, "\\");
20127 sv_catpvn(sv, &string, 1);
20129 else if (isMNEMONIC_CNTRL(c)) {
20130 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
20133 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
20137 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
20140 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
20142 /* Appends to 'sv' a displayable version of the range of code points from
20143 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
20144 * that have them, when they occur at the beginning or end of the range.
20145 * It uses hex to output the remaining code points, unless 'allow_literals'
20146 * is true, in which case the printable ASCII ones are output as-is (though
20147 * some of these will be escaped by put_code_point()).
20149 * NOTE: This is designed only for printing ranges of code points that fit
20150 * inside an ANYOF bitmap. Higher code points are simply suppressed
20153 const unsigned int min_range_count = 3;
20155 assert(start <= end);
20157 PERL_ARGS_ASSERT_PUT_RANGE;
20159 while (start <= end) {
20161 const char * format;
20163 if (end - start < min_range_count) {
20165 /* Output chars individually when they occur in short ranges */
20166 for (; start <= end; start++) {
20167 put_code_point(sv, start);
20172 /* If permitted by the input options, and there is a possibility that
20173 * this range contains a printable literal, look to see if there is
20175 if (allow_literals && start <= MAX_PRINT_A) {
20177 /* If the character at the beginning of the range isn't an ASCII
20178 * printable, effectively split the range into two parts:
20179 * 1) the portion before the first such printable,
20181 * and output them separately. */
20182 if (! isPRINT_A(start)) {
20183 UV temp_end = start + 1;
20185 /* There is no point looking beyond the final possible
20186 * printable, in MAX_PRINT_A */
20187 UV max = MIN(end, MAX_PRINT_A);
20189 while (temp_end <= max && ! isPRINT_A(temp_end)) {
20193 /* Here, temp_end points to one beyond the first printable if
20194 * found, or to one beyond 'max' if not. If none found, make
20195 * sure that we use the entire range */
20196 if (temp_end > MAX_PRINT_A) {
20197 temp_end = end + 1;
20200 /* Output the first part of the split range: the part that
20201 * doesn't have printables, with the parameter set to not look
20202 * for literals (otherwise we would infinitely recurse) */
20203 put_range(sv, start, temp_end - 1, FALSE);
20205 /* The 2nd part of the range (if any) starts here. */
20208 /* We do a continue, instead of dropping down, because even if
20209 * the 2nd part is non-empty, it could be so short that we want
20210 * to output it as individual characters, as tested for at the
20211 * top of this loop. */
20215 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
20216 * output a sub-range of just the digits or letters, then process
20217 * the remaining portion as usual. */
20218 if (isALPHANUMERIC_A(start)) {
20219 UV mask = (isDIGIT_A(start))
20224 UV temp_end = start + 1;
20226 /* Find the end of the sub-range that includes just the
20227 * characters in the same class as the first character in it */
20228 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
20233 /* For short ranges, don't duplicate the code above to output
20234 * them; just call recursively */
20235 if (temp_end - start < min_range_count) {
20236 put_range(sv, start, temp_end, FALSE);
20238 else { /* Output as a range */
20239 put_code_point(sv, start);
20240 sv_catpvs(sv, "-");
20241 put_code_point(sv, temp_end);
20243 start = temp_end + 1;
20247 /* We output any other printables as individual characters */
20248 if (isPUNCT_A(start) || isSPACE_A(start)) {
20249 while (start <= end && (isPUNCT_A(start)
20250 || isSPACE_A(start)))
20252 put_code_point(sv, start);
20257 } /* End of looking for literals */
20259 /* Here is not to output as a literal. Some control characters have
20260 * mnemonic names. Split off any of those at the beginning and end of
20261 * the range to print mnemonically. It isn't possible for many of
20262 * these to be in a row, so this won't overwhelm with output */
20264 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
20266 while (isMNEMONIC_CNTRL(start) && start <= end) {
20267 put_code_point(sv, start);
20271 /* If this didn't take care of the whole range ... */
20272 if (start <= end) {
20274 /* Look backwards from the end to find the final non-mnemonic
20277 while (isMNEMONIC_CNTRL(temp_end)) {
20281 /* And separately output the interior range that doesn't start
20282 * or end with mnemonics */
20283 put_range(sv, start, temp_end, FALSE);
20285 /* Then output the mnemonic trailing controls */
20286 start = temp_end + 1;
20287 while (start <= end) {
20288 put_code_point(sv, start);
20295 /* As a final resort, output the range or subrange as hex. */
20297 this_end = (end < NUM_ANYOF_CODE_POINTS)
20299 : NUM_ANYOF_CODE_POINTS - 1;
20300 #if NUM_ANYOF_CODE_POINTS > 256
20301 format = (this_end < 256)
20302 ? "\\x%02" UVXf "-\\x%02" UVXf
20303 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
20305 format = "\\x%02" UVXf "-\\x%02" UVXf;
20307 GCC_DIAG_IGNORE(-Wformat-nonliteral);
20308 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
20315 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
20317 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
20321 bool allow_literals = TRUE;
20323 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
20325 /* Generally, it is more readable if printable characters are output as
20326 * literals, but if a range (nearly) spans all of them, it's best to output
20327 * it as a single range. This code will use a single range if all but 2
20328 * ASCII printables are in it */
20329 invlist_iterinit(invlist);
20330 while (invlist_iternext(invlist, &start, &end)) {
20332 /* If the range starts beyond the final printable, it doesn't have any
20334 if (start > MAX_PRINT_A) {
20338 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
20339 * all but two, the range must start and end no later than 2 from
20341 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
20342 if (end > MAX_PRINT_A) {
20348 if (end - start >= MAX_PRINT_A - ' ' - 2) {
20349 allow_literals = FALSE;
20354 invlist_iterfinish(invlist);
20356 /* Here we have figured things out. Output each range */
20357 invlist_iterinit(invlist);
20358 while (invlist_iternext(invlist, &start, &end)) {
20359 if (start >= NUM_ANYOF_CODE_POINTS) {
20362 put_range(sv, start, end, allow_literals);
20364 invlist_iterfinish(invlist);
20370 S_put_charclass_bitmap_innards_common(pTHX_
20371 SV* invlist, /* The bitmap */
20372 SV* posixes, /* Under /l, things like [:word:], \S */
20373 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
20374 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
20375 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
20376 const bool invert /* Is the result to be inverted? */
20379 /* Create and return an SV containing a displayable version of the bitmap
20380 * and associated information determined by the input parameters. If the
20381 * output would have been only the inversion indicator '^', NULL is instead
20386 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
20389 output = newSVpvs("^");
20392 output = newSVpvs("");
20395 /* First, the code points in the bitmap that are unconditionally there */
20396 put_charclass_bitmap_innards_invlist(output, invlist);
20398 /* Traditionally, these have been placed after the main code points */
20400 sv_catsv(output, posixes);
20403 if (only_utf8 && _invlist_len(only_utf8)) {
20404 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
20405 put_charclass_bitmap_innards_invlist(output, only_utf8);
20408 if (not_utf8 && _invlist_len(not_utf8)) {
20409 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
20410 put_charclass_bitmap_innards_invlist(output, not_utf8);
20413 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
20414 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
20415 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
20417 /* This is the only list in this routine that can legally contain code
20418 * points outside the bitmap range. The call just above to
20419 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
20420 * output them here. There's about a half-dozen possible, and none in
20421 * contiguous ranges longer than 2 */
20422 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20424 SV* above_bitmap = NULL;
20426 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
20428 invlist_iterinit(above_bitmap);
20429 while (invlist_iternext(above_bitmap, &start, &end)) {
20432 for (i = start; i <= end; i++) {
20433 put_code_point(output, i);
20436 invlist_iterfinish(above_bitmap);
20437 SvREFCNT_dec_NN(above_bitmap);
20441 if (invert && SvCUR(output) == 1) {
20449 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
20451 SV *nonbitmap_invlist,
20452 SV *only_utf8_locale_invlist,
20453 const regnode * const node,
20454 const bool force_as_is_display)
20456 /* Appends to 'sv' a displayable version of the innards of the bracketed
20457 * character class defined by the other arguments:
20458 * 'bitmap' points to the bitmap.
20459 * 'nonbitmap_invlist' is an inversion list of the code points that are in
20460 * the bitmap range, but for some reason aren't in the bitmap; NULL if
20461 * none. The reasons for this could be that they require some
20462 * condition such as the target string being or not being in UTF-8
20463 * (under /d), or because they came from a user-defined property that
20464 * was not resolved at the time of the regex compilation (under /u)
20465 * 'only_utf8_locale_invlist' is an inversion list of the code points that
20466 * are valid only if the runtime locale is a UTF-8 one; NULL if none
20467 * 'node' is the regex pattern node. It is needed only when the above two
20468 * parameters are not null, and is passed so that this routine can
20469 * tease apart the various reasons for them.
20470 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
20471 * to invert things to see if that leads to a cleaner display. If
20472 * FALSE, this routine is free to use its judgment about doing this.
20474 * It returns TRUE if there was actually something output. (It may be that
20475 * the bitmap, etc is empty.)
20477 * When called for outputting the bitmap of a non-ANYOF node, just pass the
20478 * bitmap, with the succeeding parameters set to NULL, and the final one to
20482 /* In general, it tries to display the 'cleanest' representation of the
20483 * innards, choosing whether to display them inverted or not, regardless of
20484 * whether the class itself is to be inverted. However, there are some
20485 * cases where it can't try inverting, as what actually matches isn't known
20486 * until runtime, and hence the inversion isn't either. */
20487 bool inverting_allowed = ! force_as_is_display;
20490 STRLEN orig_sv_cur = SvCUR(sv);
20492 SV* invlist; /* Inversion list we accumulate of code points that
20493 are unconditionally matched */
20494 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
20496 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
20498 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
20499 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
20502 SV* as_is_display; /* The output string when we take the inputs
20504 SV* inverted_display; /* The output string when we invert the inputs */
20506 U8 flags = (node) ? ANYOF_FLAGS(node) : 0;
20508 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
20510 /* We are biased in favor of displaying things without them being inverted,
20511 * as that is generally easier to understand */
20512 const int bias = 5;
20514 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
20516 /* Start off with whatever code points are passed in. (We clone, so we
20517 * don't change the caller's list) */
20518 if (nonbitmap_invlist) {
20519 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
20520 invlist = invlist_clone(nonbitmap_invlist);
20522 else { /* Worst case size is every other code point is matched */
20523 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
20527 if (OP(node) == ANYOFD) {
20529 /* This flag indicates that the code points below 0x100 in the
20530 * nonbitmap list are precisely the ones that match only when the
20531 * target is UTF-8 (they should all be non-ASCII). */
20532 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
20534 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
20535 _invlist_subtract(invlist, only_utf8, &invlist);
20538 /* And this flag for matching all non-ASCII 0xFF and below */
20539 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
20541 not_utf8 = invlist_clone(PL_UpperLatin1);
20544 else if (OP(node) == ANYOFL) {
20546 /* If either of these flags are set, what matches isn't
20547 * determinable except during execution, so don't know enough here
20549 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
20550 inverting_allowed = FALSE;
20553 /* What the posix classes match also varies at runtime, so these
20554 * will be output symbolically. */
20555 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
20558 posixes = newSVpvs("");
20559 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
20560 if (ANYOF_POSIXL_TEST(node,i)) {
20561 sv_catpv(posixes, anyofs[i]);
20568 /* Accumulate the bit map into the unconditional match list */
20569 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
20570 if (BITMAP_TEST(bitmap, i)) {
20572 for (; i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i); i++) {
20575 invlist = _add_range_to_invlist(invlist, start, i-1);
20579 /* Make sure that the conditional match lists don't have anything in them
20580 * that match unconditionally; otherwise the output is quite confusing.
20581 * This could happen if the code that populates these misses some
20584 _invlist_subtract(only_utf8, invlist, &only_utf8);
20587 _invlist_subtract(not_utf8, invlist, ¬_utf8);
20590 if (only_utf8_locale_invlist) {
20592 /* Since this list is passed in, we have to make a copy before
20594 only_utf8_locale = invlist_clone(only_utf8_locale_invlist);
20596 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
20598 /* And, it can get really weird for us to try outputting an inverted
20599 * form of this list when it has things above the bitmap, so don't even
20601 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
20602 inverting_allowed = FALSE;
20606 /* Calculate what the output would be if we take the input as-is */
20607 as_is_display = put_charclass_bitmap_innards_common(invlist,
20614 /* If have to take the output as-is, just do that */
20615 if (! inverting_allowed) {
20616 if (as_is_display) {
20617 sv_catsv(sv, as_is_display);
20618 SvREFCNT_dec_NN(as_is_display);
20621 else { /* But otherwise, create the output again on the inverted input, and
20622 use whichever version is shorter */
20624 int inverted_bias, as_is_bias;
20626 /* We will apply our bias to whichever of the the results doesn't have
20636 inverted_bias = bias;
20639 /* Now invert each of the lists that contribute to the output,
20640 * excluding from the result things outside the possible range */
20642 /* For the unconditional inversion list, we have to add in all the
20643 * conditional code points, so that when inverted, they will be gone
20645 _invlist_union(only_utf8, invlist, &invlist);
20646 _invlist_union(not_utf8, invlist, &invlist);
20647 _invlist_union(only_utf8_locale, invlist, &invlist);
20648 _invlist_invert(invlist);
20649 _invlist_intersection(invlist, PL_InBitmap, &invlist);
20652 _invlist_invert(only_utf8);
20653 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
20655 else if (not_utf8) {
20657 /* If a code point matches iff the target string is not in UTF-8,
20658 * then complementing the result has it not match iff not in UTF-8,
20659 * which is the same thing as matching iff it is UTF-8. */
20660 only_utf8 = not_utf8;
20664 if (only_utf8_locale) {
20665 _invlist_invert(only_utf8_locale);
20666 _invlist_intersection(only_utf8_locale,
20668 &only_utf8_locale);
20671 inverted_display = put_charclass_bitmap_innards_common(
20676 only_utf8_locale, invert);
20678 /* Use the shortest representation, taking into account our bias
20679 * against showing it inverted */
20680 if ( inverted_display
20681 && ( ! as_is_display
20682 || ( SvCUR(inverted_display) + inverted_bias
20683 < SvCUR(as_is_display) + as_is_bias)))
20685 sv_catsv(sv, inverted_display);
20687 else if (as_is_display) {
20688 sv_catsv(sv, as_is_display);
20691 SvREFCNT_dec(as_is_display);
20692 SvREFCNT_dec(inverted_display);
20695 SvREFCNT_dec_NN(invlist);
20696 SvREFCNT_dec(only_utf8);
20697 SvREFCNT_dec(not_utf8);
20698 SvREFCNT_dec(posixes);
20699 SvREFCNT_dec(only_utf8_locale);
20701 return SvCUR(sv) > orig_sv_cur;
20704 #define CLEAR_OPTSTART \
20705 if (optstart) STMT_START { \
20706 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
20707 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
20711 #define DUMPUNTIL(b,e) \
20713 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
20715 STATIC const regnode *
20716 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
20717 const regnode *last, const regnode *plast,
20718 SV* sv, I32 indent, U32 depth)
20720 U8 op = PSEUDO; /* Arbitrary non-END op. */
20721 const regnode *next;
20722 const regnode *optstart= NULL;
20724 RXi_GET_DECL(r,ri);
20725 GET_RE_DEBUG_FLAGS_DECL;
20727 PERL_ARGS_ASSERT_DUMPUNTIL;
20729 #ifdef DEBUG_DUMPUNTIL
20730 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n",indent,node-start,
20731 last ? last-start : 0,plast ? plast-start : 0);
20734 if (plast && plast < last)
20737 while (PL_regkind[op] != END && (!last || node < last)) {
20739 /* While that wasn't END last time... */
20742 if (op == CLOSE || op == WHILEM)
20744 next = regnext((regnode *)node);
20747 if (OP(node) == OPTIMIZED) {
20748 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
20755 regprop(r, sv, node, NULL, NULL);
20756 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
20757 (int)(2*indent + 1), "", SvPVX_const(sv));
20759 if (OP(node) != OPTIMIZED) {
20760 if (next == NULL) /* Next ptr. */
20761 Perl_re_printf( aTHX_ " (0)");
20762 else if (PL_regkind[(U8)op] == BRANCH
20763 && PL_regkind[OP(next)] != BRANCH )
20764 Perl_re_printf( aTHX_ " (FAIL)");
20766 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
20767 Perl_re_printf( aTHX_ "\n");
20771 if (PL_regkind[(U8)op] == BRANCHJ) {
20774 const regnode *nnode = (OP(next) == LONGJMP
20775 ? regnext((regnode *)next)
20777 if (last && nnode > last)
20779 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
20782 else if (PL_regkind[(U8)op] == BRANCH) {
20784 DUMPUNTIL(NEXTOPER(node), next);
20786 else if ( PL_regkind[(U8)op] == TRIE ) {
20787 const regnode *this_trie = node;
20788 const char op = OP(node);
20789 const U32 n = ARG(node);
20790 const reg_ac_data * const ac = op>=AHOCORASICK ?
20791 (reg_ac_data *)ri->data->data[n] :
20793 const reg_trie_data * const trie =
20794 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
20796 AV *const trie_words
20797 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
20799 const regnode *nextbranch= NULL;
20802 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
20803 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
20805 Perl_re_indentf( aTHX_ "%s ",
20808 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
20809 SvCUR(*elem_ptr), PL_dump_re_max_len,
20810 PL_colors[0], PL_colors[1],
20812 ? PERL_PV_ESCAPE_UNI
20814 | PERL_PV_PRETTY_ELLIPSES
20815 | PERL_PV_PRETTY_LTGT
20820 U16 dist= trie->jump[word_idx+1];
20821 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
20822 (UV)((dist ? this_trie + dist : next) - start));
20825 nextbranch= this_trie + trie->jump[0];
20826 DUMPUNTIL(this_trie + dist, nextbranch);
20828 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
20829 nextbranch= regnext((regnode *)nextbranch);
20831 Perl_re_printf( aTHX_ "\n");
20834 if (last && next > last)
20839 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
20840 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
20841 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
20843 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
20845 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
20847 else if ( op == PLUS || op == STAR) {
20848 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
20850 else if (PL_regkind[(U8)op] == ANYOF) {
20851 /* arglen 1 + class block */
20852 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
20853 ? ANYOF_POSIXL_SKIP
20855 node = NEXTOPER(node);
20857 else if (PL_regkind[(U8)op] == EXACT) {
20858 /* Literal string, where present. */
20859 node += NODE_SZ_STR(node) - 1;
20860 node = NEXTOPER(node);
20863 node = NEXTOPER(node);
20864 node += regarglen[(U8)op];
20866 if (op == CURLYX || op == OPEN)
20870 #ifdef DEBUG_DUMPUNTIL
20871 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
20876 #endif /* DEBUGGING */
20879 * ex: set ts=8 sts=4 sw=4 et: