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
105 #define MIN(a,b) ((a) < (b) ? (a) : (b))
108 /* this is a chain of data about sub patterns we are processing that
109 need to be handled separately/specially in study_chunk. Its so
110 we can simulate recursion without losing state. */
112 typedef struct scan_frame {
113 regnode *last_regnode; /* last node to process in this frame */
114 regnode *next_regnode; /* next node to process when last is reached */
115 U32 prev_recursed_depth;
116 I32 stopparen; /* what stopparen do we use */
117 U32 is_top_frame; /* what flags do we use? */
119 struct scan_frame *this_prev_frame; /* this previous frame */
120 struct scan_frame *prev_frame; /* previous frame */
121 struct scan_frame *next_frame; /* next frame */
124 /* Certain characters are output as a sequence with the first being a
126 #define isBACKSLASHED_PUNCT(c) \
127 ((c) == '-' || (c) == ']' || (c) == '\\' || (c) == '^')
130 struct RExC_state_t {
131 U32 flags; /* RXf_* are we folding, multilining? */
132 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
133 char *precomp; /* uncompiled string. */
134 REGEXP *rx_sv; /* The SV that is the regexp. */
135 regexp *rx; /* perl core regexp structure */
136 regexp_internal *rxi; /* internal data for regexp object
138 char *start; /* Start of input for compile */
139 char *end; /* End of input for compile */
140 char *parse; /* Input-scan pointer. */
141 SSize_t whilem_seen; /* number of WHILEM in this expr */
142 regnode *emit_start; /* Start of emitted-code area */
143 regnode *emit_bound; /* First regnode outside of the
145 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
146 implies compiling, so don't emit */
147 regnode_ssc emit_dummy; /* placeholder for emit to point to;
148 large enough for the largest
149 non-EXACTish node, so can use it as
151 I32 naughty; /* How bad is this pattern? */
152 I32 sawback; /* Did we see \1, ...? */
154 SSize_t size; /* Code size. */
155 I32 npar; /* Capture buffer count, (OPEN) plus
156 one. ("par" 0 is the whole
158 I32 nestroot; /* root parens we are in - used by
162 regnode **open_parens; /* pointers to open parens */
163 regnode **close_parens; /* pointers to close parens */
164 regnode *opend; /* END node in program */
165 I32 utf8; /* whether the pattern is utf8 or not */
166 I32 orig_utf8; /* whether the pattern was originally in utf8 */
167 /* XXX use this for future optimisation of case
168 * where pattern must be upgraded to utf8. */
169 I32 uni_semantics; /* If a d charset modifier should use unicode
170 rules, even if the pattern is not in
172 HV *paren_names; /* Paren names */
174 regnode **recurse; /* Recurse regops */
175 I32 recurse_count; /* Number of recurse regops */
176 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
178 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
182 I32 override_recoding;
184 I32 recode_x_to_native;
186 I32 in_multi_char_class;
187 struct reg_code_block *code_blocks; /* positions of literal (?{})
189 int num_code_blocks; /* size of code_blocks[] */
190 int code_index; /* next code_blocks[] slot */
191 SSize_t maxlen; /* mininum possible number of chars in string to match */
192 scan_frame *frame_head;
193 scan_frame *frame_last;
196 #ifdef ADD_TO_REGEXEC
197 char *starttry; /* -Dr: where regtry was called. */
198 #define RExC_starttry (pRExC_state->starttry)
200 SV *runtime_code_qr; /* qr with the runtime code blocks */
202 const char *lastparse;
204 AV *paren_name_list; /* idx -> name */
205 U32 study_chunk_recursed_count;
208 #define RExC_lastparse (pRExC_state->lastparse)
209 #define RExC_lastnum (pRExC_state->lastnum)
210 #define RExC_paren_name_list (pRExC_state->paren_name_list)
211 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
212 #define RExC_mysv (pRExC_state->mysv1)
213 #define RExC_mysv1 (pRExC_state->mysv1)
214 #define RExC_mysv2 (pRExC_state->mysv2)
217 bool seen_unfolded_sharp_s;
220 #define RExC_flags (pRExC_state->flags)
221 #define RExC_pm_flags (pRExC_state->pm_flags)
222 #define RExC_precomp (pRExC_state->precomp)
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_opend (pRExC_state->opend)
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)
271 #define RExC_contains_i (pRExC_state->contains_i)
272 #define RExC_override_recoding (pRExC_state->override_recoding)
274 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
276 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
277 #define RExC_frame_head (pRExC_state->frame_head)
278 #define RExC_frame_last (pRExC_state->frame_last)
279 #define RExC_frame_count (pRExC_state->frame_count)
280 #define RExC_strict (pRExC_state->strict)
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:
394 - offset or min_offset
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.
408 A pointer to the minimum number of characters of the pattern that the
409 string was found inside. This is important as in the case of positive
410 lookahead or positive lookbehind we can have multiple patterns
415 The minimum length of the pattern overall is 3, the minimum length
416 of the lookahead part is 3, but the minimum length of the part that
417 will actually match is 1. So 'FOO's minimum length is 3, but the
418 minimum length for the F is 1. This is important as the minimum length
419 is used to determine offsets in front of and behind the string being
420 looked for. Since strings can be composites this is the length of the
421 pattern at the time it was committed with a scan_commit. Note that
422 the length is calculated by study_chunk, so that the minimum lengths
423 are not known until the full pattern has been compiled, thus the
424 pointer to the value.
428 In the case of lookbehind the string being searched for can be
429 offset past the start point of the final matching string.
430 If this value was just blithely removed from the min_offset it would
431 invalidate some of the calculations for how many chars must match
432 before or after (as they are derived from min_offset and minlen and
433 the length of the string being searched for).
434 When the final pattern is compiled and the data is moved from the
435 scan_data_t structure into the regexp structure the information
436 about lookbehind is factored in, with the information that would
437 have been lost precalculated in the end_shift field for the
440 The fields pos_min and pos_delta are used to store the minimum offset
441 and the delta to the maximum offset at the current point in the pattern.
445 typedef struct scan_data_t {
446 /*I32 len_min; unused */
447 /*I32 len_delta; unused */
451 SSize_t last_end; /* min value, <0 unless valid. */
452 SSize_t last_start_min;
453 SSize_t last_start_max;
454 SV **longest; /* Either &l_fixed, or &l_float. */
455 SV *longest_fixed; /* longest fixed string found in pattern */
456 SSize_t offset_fixed; /* offset where it starts */
457 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
458 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
459 SV *longest_float; /* longest floating string found in pattern */
460 SSize_t offset_float_min; /* earliest point in string it can appear */
461 SSize_t offset_float_max; /* latest point in string it can appear */
462 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
463 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
466 SSize_t *last_closep;
467 regnode_ssc *start_class;
471 * Forward declarations for pregcomp()'s friends.
474 static const scan_data_t zero_scan_data =
475 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
477 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
478 #define SF_BEFORE_SEOL 0x0001
479 #define SF_BEFORE_MEOL 0x0002
480 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
481 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
483 #define SF_FIX_SHIFT_EOL (+2)
484 #define SF_FL_SHIFT_EOL (+4)
486 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
487 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
489 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
490 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
491 #define SF_IS_INF 0x0040
492 #define SF_HAS_PAR 0x0080
493 #define SF_IN_PAR 0x0100
494 #define SF_HAS_EVAL 0x0200
495 #define SCF_DO_SUBSTR 0x0400
496 #define SCF_DO_STCLASS_AND 0x0800
497 #define SCF_DO_STCLASS_OR 0x1000
498 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
499 #define SCF_WHILEM_VISITED_POS 0x2000
501 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
502 #define SCF_SEEN_ACCEPT 0x8000
503 #define SCF_TRIE_DOING_RESTUDY 0x10000
504 #define SCF_IN_DEFINE 0x20000
509 #define UTF cBOOL(RExC_utf8)
511 /* The enums for all these are ordered so things work out correctly */
512 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
513 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
514 == REGEX_DEPENDS_CHARSET)
515 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
516 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
517 >= REGEX_UNICODE_CHARSET)
518 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
519 == REGEX_ASCII_RESTRICTED_CHARSET)
520 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
521 >= REGEX_ASCII_RESTRICTED_CHARSET)
522 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
523 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
525 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
527 /* For programs that want to be strictly Unicode compatible by dying if any
528 * attempt is made to match a non-Unicode code point against a Unicode
530 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
532 #define OOB_NAMEDCLASS -1
534 /* There is no code point that is out-of-bounds, so this is problematic. But
535 * its only current use is to initialize a variable that is always set before
537 #define OOB_UNICODE 0xDEADBEEF
539 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
540 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
543 /* length of regex to show in messages that don't mark a position within */
544 #define RegexLengthToShowInErrorMessages 127
547 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
548 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
549 * op/pragma/warn/regcomp.
551 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
552 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
554 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
555 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
557 #define REPORT_LOCATION_ARGS(offset) \
558 UTF8fARG(UTF, offset, RExC_precomp), \
559 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
561 /* Used to point after bad bytes for an error message, but avoid skipping
562 * past a nul byte. */
563 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
566 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
567 * arg. Show regex, up to a maximum length. If it's too long, chop and add
570 #define _FAIL(code) STMT_START { \
571 const char *ellipses = ""; \
572 IV len = RExC_end - RExC_precomp; \
575 SAVEFREESV(RExC_rx_sv); \
576 if (len > RegexLengthToShowInErrorMessages) { \
577 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
578 len = RegexLengthToShowInErrorMessages - 10; \
584 #define FAIL(msg) _FAIL( \
585 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
586 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
588 #define FAIL2(msg,arg) _FAIL( \
589 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
590 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
593 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
595 #define Simple_vFAIL(m) STMT_START { \
597 (RExC_parse > RExC_end ? RExC_end : RExC_parse) - RExC_precomp; \
598 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
599 m, REPORT_LOCATION_ARGS(offset)); \
603 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
605 #define vFAIL(m) STMT_START { \
607 SAVEFREESV(RExC_rx_sv); \
612 * Like Simple_vFAIL(), but accepts two arguments.
614 #define Simple_vFAIL2(m,a1) STMT_START { \
615 const IV offset = RExC_parse - RExC_precomp; \
616 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
617 REPORT_LOCATION_ARGS(offset)); \
621 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
623 #define vFAIL2(m,a1) STMT_START { \
625 SAVEFREESV(RExC_rx_sv); \
626 Simple_vFAIL2(m, a1); \
631 * Like Simple_vFAIL(), but accepts three arguments.
633 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
634 const IV offset = RExC_parse - RExC_precomp; \
635 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
636 REPORT_LOCATION_ARGS(offset)); \
640 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
642 #define vFAIL3(m,a1,a2) STMT_START { \
644 SAVEFREESV(RExC_rx_sv); \
645 Simple_vFAIL3(m, a1, a2); \
649 * Like Simple_vFAIL(), but accepts four arguments.
651 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
652 const IV offset = RExC_parse - RExC_precomp; \
653 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
654 REPORT_LOCATION_ARGS(offset)); \
657 #define vFAIL4(m,a1,a2,a3) STMT_START { \
659 SAVEFREESV(RExC_rx_sv); \
660 Simple_vFAIL4(m, a1, a2, a3); \
663 /* A specialized version of vFAIL2 that works with UTF8f */
664 #define vFAIL2utf8f(m, a1) STMT_START { \
665 const IV offset = RExC_parse - RExC_precomp; \
667 SAVEFREESV(RExC_rx_sv); \
668 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
669 REPORT_LOCATION_ARGS(offset)); \
672 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
673 const IV offset = RExC_parse - RExC_precomp; \
675 SAVEFREESV(RExC_rx_sv); \
676 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
677 REPORT_LOCATION_ARGS(offset)); \
680 /* These have asserts in them because of [perl #122671] Many warnings in
681 * regcomp.c can occur twice. If they get output in pass1 and later in that
682 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
683 * would get output again. So they should be output in pass2, and these
684 * asserts make sure new warnings follow that paradigm. */
686 /* m is not necessarily a "literal string", in this macro */
687 #define reg_warn_non_literal_string(loc, m) STMT_START { \
688 const IV offset = loc - RExC_precomp; \
689 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
690 m, REPORT_LOCATION_ARGS(offset)); \
693 #define ckWARNreg(loc,m) STMT_START { \
694 const IV offset = loc - RExC_precomp; \
695 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
696 REPORT_LOCATION_ARGS(offset)); \
699 #define vWARN(loc, m) STMT_START { \
700 const IV offset = loc - RExC_precomp; \
701 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
702 REPORT_LOCATION_ARGS(offset)); \
705 #define vWARN_dep(loc, m) STMT_START { \
706 const IV offset = loc - RExC_precomp; \
707 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
708 REPORT_LOCATION_ARGS(offset)); \
711 #define ckWARNdep(loc,m) STMT_START { \
712 const IV offset = loc - RExC_precomp; \
713 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
715 REPORT_LOCATION_ARGS(offset)); \
718 #define ckWARNregdep(loc,m) STMT_START { \
719 const IV offset = loc - RExC_precomp; \
720 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
722 REPORT_LOCATION_ARGS(offset)); \
725 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
726 const IV offset = loc - RExC_precomp; \
727 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
729 a1, REPORT_LOCATION_ARGS(offset)); \
732 #define ckWARN2reg(loc, m, a1) STMT_START { \
733 const IV offset = loc - RExC_precomp; \
734 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
735 a1, REPORT_LOCATION_ARGS(offset)); \
738 #define vWARN3(loc, m, a1, a2) STMT_START { \
739 const IV offset = loc - RExC_precomp; \
740 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
741 a1, a2, REPORT_LOCATION_ARGS(offset)); \
744 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
745 const IV offset = loc - RExC_precomp; \
746 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
747 a1, a2, REPORT_LOCATION_ARGS(offset)); \
750 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
751 const IV offset = loc - RExC_precomp; \
752 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
753 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
756 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
757 const IV offset = loc - RExC_precomp; \
758 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
759 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
762 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
763 const IV offset = loc - RExC_precomp; \
764 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
765 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
768 /* Macros for recording node offsets. 20001227 mjd@plover.com
769 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
770 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
771 * Element 0 holds the number n.
772 * Position is 1 indexed.
774 #ifndef RE_TRACK_PATTERN_OFFSETS
775 #define Set_Node_Offset_To_R(node,byte)
776 #define Set_Node_Offset(node,byte)
777 #define Set_Cur_Node_Offset
778 #define Set_Node_Length_To_R(node,len)
779 #define Set_Node_Length(node,len)
780 #define Set_Node_Cur_Length(node,start)
781 #define Node_Offset(n)
782 #define Node_Length(n)
783 #define Set_Node_Offset_Length(node,offset,len)
784 #define ProgLen(ri) ri->u.proglen
785 #define SetProgLen(ri,x) ri->u.proglen = x
787 #define ProgLen(ri) ri->u.offsets[0]
788 #define SetProgLen(ri,x) ri->u.offsets[0] = x
789 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
791 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
792 __LINE__, (int)(node), (int)(byte))); \
794 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
797 RExC_offsets[2*(node)-1] = (byte); \
802 #define Set_Node_Offset(node,byte) \
803 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
804 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
806 #define Set_Node_Length_To_R(node,len) STMT_START { \
808 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
809 __LINE__, (int)(node), (int)(len))); \
811 Perl_croak(aTHX_ "value of node is %d in Length macro", \
814 RExC_offsets[2*(node)] = (len); \
819 #define Set_Node_Length(node,len) \
820 Set_Node_Length_To_R((node)-RExC_emit_start, len)
821 #define Set_Node_Cur_Length(node, start) \
822 Set_Node_Length(node, RExC_parse - start)
824 /* Get offsets and lengths */
825 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
826 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
828 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
829 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
830 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
834 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
835 #define EXPERIMENTAL_INPLACESCAN
836 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
838 #define DEBUG_RExC_seen() \
839 DEBUG_OPTIMISE_MORE_r({ \
840 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
842 if (RExC_seen & REG_ZERO_LEN_SEEN) \
843 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
845 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
846 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
848 if (RExC_seen & REG_GPOS_SEEN) \
849 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
851 if (RExC_seen & REG_RECURSE_SEEN) \
852 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
854 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
855 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
857 if (RExC_seen & REG_VERBARG_SEEN) \
858 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
860 if (RExC_seen & REG_CUTGROUP_SEEN) \
861 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
863 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
864 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
866 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
867 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
869 if (RExC_seen & REG_GOSTART_SEEN) \
870 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
872 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
873 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
875 PerlIO_printf(Perl_debug_log,"\n"); \
878 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
879 if ((flags) & flag) PerlIO_printf(Perl_debug_log, "%s ", #flag)
881 #define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
883 PerlIO_printf(Perl_debug_log, "%s", open_str); \
884 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
885 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
886 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
887 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
888 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
889 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
890 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
891 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
892 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
893 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
894 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
895 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
896 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
897 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
898 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
899 PerlIO_printf(Perl_debug_log, "%s", close_str); \
903 #define DEBUG_STUDYDATA(str,data,depth) \
904 DEBUG_OPTIMISE_MORE_r(if(data){ \
905 PerlIO_printf(Perl_debug_log, \
906 "%*s" str "Pos:%"IVdf"/%"IVdf \
908 (int)(depth)*2, "", \
909 (IV)((data)->pos_min), \
910 (IV)((data)->pos_delta), \
911 (UV)((data)->flags) \
913 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
914 PerlIO_printf(Perl_debug_log, \
915 " Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
916 (IV)((data)->whilem_c), \
917 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
918 is_inf ? "INF " : "" \
920 if ((data)->last_found) \
921 PerlIO_printf(Perl_debug_log, \
922 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
923 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
924 SvPVX_const((data)->last_found), \
925 (IV)((data)->last_end), \
926 (IV)((data)->last_start_min), \
927 (IV)((data)->last_start_max), \
928 ((data)->longest && \
929 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
930 SvPVX_const((data)->longest_fixed), \
931 (IV)((data)->offset_fixed), \
932 ((data)->longest && \
933 (data)->longest==&((data)->longest_float)) ? "*" : "", \
934 SvPVX_const((data)->longest_float), \
935 (IV)((data)->offset_float_min), \
936 (IV)((data)->offset_float_max) \
938 PerlIO_printf(Perl_debug_log,"\n"); \
941 /* is c a control character for which we have a mnemonic? */
942 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
945 S_cntrl_to_mnemonic(const U8 c)
947 /* Returns the mnemonic string that represents character 'c', if one
948 * exists; NULL otherwise. The only ones that exist for the purposes of
949 * this routine are a few control characters */
952 case '\a': return "\\a";
953 case '\b': return "\\b";
954 case ESC_NATIVE: return "\\e";
955 case '\f': return "\\f";
956 case '\n': return "\\n";
957 case '\r': return "\\r";
958 case '\t': return "\\t";
964 /* Mark that we cannot extend a found fixed substring at this point.
965 Update the longest found anchored substring and the longest found
966 floating substrings if needed. */
969 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
970 SSize_t *minlenp, int is_inf)
972 const STRLEN l = CHR_SVLEN(data->last_found);
973 const STRLEN old_l = CHR_SVLEN(*data->longest);
974 GET_RE_DEBUG_FLAGS_DECL;
976 PERL_ARGS_ASSERT_SCAN_COMMIT;
978 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
979 SvSetMagicSV(*data->longest, data->last_found);
980 if (*data->longest == data->longest_fixed) {
981 data->offset_fixed = l ? data->last_start_min : data->pos_min;
982 if (data->flags & SF_BEFORE_EOL)
984 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
986 data->flags &= ~SF_FIX_BEFORE_EOL;
987 data->minlen_fixed=minlenp;
988 data->lookbehind_fixed=0;
990 else { /* *data->longest == data->longest_float */
991 data->offset_float_min = l ? data->last_start_min : data->pos_min;
992 data->offset_float_max = (l
993 ? data->last_start_max
994 : (data->pos_delta > SSize_t_MAX - data->pos_min
996 : data->pos_min + data->pos_delta));
998 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
999 data->offset_float_max = SSize_t_MAX;
1000 if (data->flags & SF_BEFORE_EOL)
1002 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
1004 data->flags &= ~SF_FL_BEFORE_EOL;
1005 data->minlen_float=minlenp;
1006 data->lookbehind_float=0;
1009 SvCUR_set(data->last_found, 0);
1011 SV * const sv = data->last_found;
1012 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1013 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1018 data->last_end = -1;
1019 data->flags &= ~SF_BEFORE_EOL;
1020 DEBUG_STUDYDATA("commit: ",data,0);
1023 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1024 * list that describes which code points it matches */
1027 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1029 /* Set the SSC 'ssc' to match an empty string or any code point */
1031 PERL_ARGS_ASSERT_SSC_ANYTHING;
1033 assert(is_ANYOF_SYNTHETIC(ssc));
1035 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
1036 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
1037 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1041 S_ssc_is_anything(const regnode_ssc *ssc)
1043 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1044 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1045 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1046 * in any way, so there's no point in using it */
1051 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1053 assert(is_ANYOF_SYNTHETIC(ssc));
1055 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1059 /* See if the list consists solely of the range 0 - Infinity */
1060 invlist_iterinit(ssc->invlist);
1061 ret = invlist_iternext(ssc->invlist, &start, &end)
1065 invlist_iterfinish(ssc->invlist);
1071 /* If e.g., both \w and \W are set, matches everything */
1072 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1074 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1075 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1085 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1087 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1088 * string, any code point, or any posix class under locale */
1090 PERL_ARGS_ASSERT_SSC_INIT;
1092 Zero(ssc, 1, regnode_ssc);
1093 set_ANYOF_SYNTHETIC(ssc);
1094 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1097 /* If any portion of the regex is to operate under locale rules that aren't
1098 * fully known at compile time, initialization includes it. The reason
1099 * this isn't done for all regexes is that the optimizer was written under
1100 * the assumption that locale was all-or-nothing. Given the complexity and
1101 * lack of documentation in the optimizer, and that there are inadequate
1102 * test cases for locale, many parts of it may not work properly, it is
1103 * safest to avoid locale unless necessary. */
1104 if (RExC_contains_locale) {
1105 ANYOF_POSIXL_SETALL(ssc);
1108 ANYOF_POSIXL_ZERO(ssc);
1113 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1114 const regnode_ssc *ssc)
1116 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1117 * to the list of code points matched, and locale posix classes; hence does
1118 * not check its flags) */
1123 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1125 assert(is_ANYOF_SYNTHETIC(ssc));
1127 invlist_iterinit(ssc->invlist);
1128 ret = invlist_iternext(ssc->invlist, &start, &end)
1132 invlist_iterfinish(ssc->invlist);
1138 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1146 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1147 const regnode_charclass* const node)
1149 /* Returns a mortal inversion list defining which code points are matched
1150 * by 'node', which is of type ANYOF. Handles complementing the result if
1151 * appropriate. If some code points aren't knowable at this time, the
1152 * returned list must, and will, contain every code point that is a
1155 SV* invlist = sv_2mortal(_new_invlist(0));
1156 SV* only_utf8_locale_invlist = NULL;
1158 const U32 n = ARG(node);
1159 bool new_node_has_latin1 = FALSE;
1161 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1163 /* Look at the data structure created by S_set_ANYOF_arg() */
1164 if (n != ANYOF_ONLY_HAS_BITMAP) {
1165 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1166 AV * const av = MUTABLE_AV(SvRV(rv));
1167 SV **const ary = AvARRAY(av);
1168 assert(RExC_rxi->data->what[n] == 's');
1170 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1171 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1173 else if (ary[0] && ary[0] != &PL_sv_undef) {
1175 /* Here, no compile-time swash, and there are things that won't be
1176 * known until runtime -- we have to assume it could be anything */
1177 return _add_range_to_invlist(invlist, 0, UV_MAX);
1179 else if (ary[3] && ary[3] != &PL_sv_undef) {
1181 /* Here no compile-time swash, and no run-time only data. Use the
1182 * node's inversion list */
1183 invlist = sv_2mortal(invlist_clone(ary[3]));
1186 /* Get the code points valid only under UTF-8 locales */
1187 if ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
1188 && ary[2] && ary[2] != &PL_sv_undef)
1190 only_utf8_locale_invlist = ary[2];
1194 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1195 * code points, and an inversion list for the others, but if there are code
1196 * points that should match only conditionally on the target string being
1197 * UTF-8, those are placed in the inversion list, and not the bitmap.
1198 * Since there are circumstances under which they could match, they are
1199 * included in the SSC. But if the ANYOF node is to be inverted, we have
1200 * to exclude them here, so that when we invert below, the end result
1201 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1202 * have to do this here before we add the unconditionally matched code
1204 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1205 _invlist_intersection_complement_2nd(invlist,
1210 /* Add in the points from the bit map */
1211 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1212 if (ANYOF_BITMAP_TEST(node, i)) {
1213 invlist = add_cp_to_invlist(invlist, i);
1214 new_node_has_latin1 = TRUE;
1218 /* If this can match all upper Latin1 code points, have to add them
1220 if (OP(node) == ANYOFD
1221 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1223 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1226 /* Similarly for these */
1227 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1228 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1231 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1232 _invlist_invert(invlist);
1234 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1236 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1237 * locale. We can skip this if there are no 0-255 at all. */
1238 _invlist_union(invlist, PL_Latin1, &invlist);
1241 /* Similarly add the UTF-8 locale possible matches. These have to be
1242 * deferred until after the non-UTF-8 locale ones are taken care of just
1243 * above, or it leads to wrong results under ANYOF_INVERT */
1244 if (only_utf8_locale_invlist) {
1245 _invlist_union_maybe_complement_2nd(invlist,
1246 only_utf8_locale_invlist,
1247 ANYOF_FLAGS(node) & ANYOF_INVERT,
1254 /* These two functions currently do the exact same thing */
1255 #define ssc_init_zero ssc_init
1257 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1258 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1260 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1261 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1262 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1265 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1266 const regnode_charclass *and_with)
1268 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1269 * another SSC or a regular ANYOF class. Can create false positives. */
1274 PERL_ARGS_ASSERT_SSC_AND;
1276 assert(is_ANYOF_SYNTHETIC(ssc));
1278 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1279 * the code point inversion list and just the relevant flags */
1280 if (is_ANYOF_SYNTHETIC(and_with)) {
1281 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1282 anded_flags = ANYOF_FLAGS(and_with);
1284 /* XXX This is a kludge around what appears to be deficiencies in the
1285 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1286 * there are paths through the optimizer where it doesn't get weeded
1287 * out when it should. And if we don't make some extra provision for
1288 * it like the code just below, it doesn't get added when it should.
1289 * This solution is to add it only when AND'ing, which is here, and
1290 * only when what is being AND'ed is the pristine, original node
1291 * matching anything. Thus it is like adding it to ssc_anything() but
1292 * only when the result is to be AND'ed. Probably the same solution
1293 * could be adopted for the same problem we have with /l matching,
1294 * which is solved differently in S_ssc_init(), and that would lead to
1295 * fewer false positives than that solution has. But if this solution
1296 * creates bugs, the consequences are only that a warning isn't raised
1297 * that should be; while the consequences for having /l bugs is
1298 * incorrect matches */
1299 if (ssc_is_anything((regnode_ssc *)and_with)) {
1300 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1304 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1305 if (OP(and_with) == ANYOFD) {
1306 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1309 anded_flags = ANYOF_FLAGS(and_with)
1310 &( ANYOF_COMMON_FLAGS
1311 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER);
1315 ANYOF_FLAGS(ssc) &= anded_flags;
1317 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1318 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1319 * 'and_with' may be inverted. When not inverted, we have the situation of
1321 * (C1 | P1) & (C2 | P2)
1322 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1323 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1324 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1325 * <= ((C1 & C2) | P1 | P2)
1326 * Alternatively, the last few steps could be:
1327 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1328 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1329 * <= (C1 | C2 | (P1 & P2))
1330 * We favor the second approach if either P1 or P2 is non-empty. This is
1331 * because these components are a barrier to doing optimizations, as what
1332 * they match cannot be known until the moment of matching as they are
1333 * dependent on the current locale, 'AND"ing them likely will reduce or
1335 * But we can do better if we know that C1,P1 are in their initial state (a
1336 * frequent occurrence), each matching everything:
1337 * (<everything>) & (C2 | P2) = C2 | P2
1338 * Similarly, if C2,P2 are in their initial state (again a frequent
1339 * occurrence), the result is a no-op
1340 * (C1 | P1) & (<everything>) = C1 | P1
1343 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1344 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1345 * <= (C1 & ~C2) | (P1 & ~P2)
1348 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1349 && ! is_ANYOF_SYNTHETIC(and_with))
1353 ssc_intersection(ssc,
1355 FALSE /* Has already been inverted */
1358 /* If either P1 or P2 is empty, the intersection will be also; can skip
1360 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1361 ANYOF_POSIXL_ZERO(ssc);
1363 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1365 /* Note that the Posix class component P from 'and_with' actually
1367 * P = Pa | Pb | ... | Pn
1368 * where each component is one posix class, such as in [\w\s].
1370 * ~P = ~(Pa | Pb | ... | Pn)
1371 * = ~Pa & ~Pb & ... & ~Pn
1372 * <= ~Pa | ~Pb | ... | ~Pn
1373 * The last is something we can easily calculate, but unfortunately
1374 * is likely to have many false positives. We could do better
1375 * in some (but certainly not all) instances if two classes in
1376 * P have known relationships. For example
1377 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1379 * :lower: & :print: = :lower:
1380 * And similarly for classes that must be disjoint. For example,
1381 * since \s and \w can have no elements in common based on rules in
1382 * the POSIX standard,
1383 * \w & ^\S = nothing
1384 * Unfortunately, some vendor locales do not meet the Posix
1385 * standard, in particular almost everything by Microsoft.
1386 * The loop below just changes e.g., \w into \W and vice versa */
1388 regnode_charclass_posixl temp;
1389 int add = 1; /* To calculate the index of the complement */
1391 ANYOF_POSIXL_ZERO(&temp);
1392 for (i = 0; i < ANYOF_MAX; i++) {
1394 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1395 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1397 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1398 ANYOF_POSIXL_SET(&temp, i + add);
1400 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1402 ANYOF_POSIXL_AND(&temp, ssc);
1404 } /* else ssc already has no posixes */
1405 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1406 in its initial state */
1407 else if (! is_ANYOF_SYNTHETIC(and_with)
1408 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1410 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1411 * copy it over 'ssc' */
1412 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1413 if (is_ANYOF_SYNTHETIC(and_with)) {
1414 StructCopy(and_with, ssc, regnode_ssc);
1417 ssc->invlist = anded_cp_list;
1418 ANYOF_POSIXL_ZERO(ssc);
1419 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1420 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1424 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1425 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1427 /* One or the other of P1, P2 is non-empty. */
1428 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1429 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1431 ssc_union(ssc, anded_cp_list, FALSE);
1433 else { /* P1 = P2 = empty */
1434 ssc_intersection(ssc, anded_cp_list, FALSE);
1440 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1441 const regnode_charclass *or_with)
1443 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1444 * another SSC or a regular ANYOF class. Can create false positives if
1445 * 'or_with' is to be inverted. */
1450 PERL_ARGS_ASSERT_SSC_OR;
1452 assert(is_ANYOF_SYNTHETIC(ssc));
1454 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1455 * the code point inversion list and just the relevant flags */
1456 if (is_ANYOF_SYNTHETIC(or_with)) {
1457 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1458 ored_flags = ANYOF_FLAGS(or_with);
1461 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1462 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1463 if (OP(or_with) != ANYOFD) {
1465 |= ANYOF_FLAGS(or_with)
1466 & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1470 ANYOF_FLAGS(ssc) |= ored_flags;
1472 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1473 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1474 * 'or_with' may be inverted. When not inverted, we have the simple
1475 * situation of computing:
1476 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1477 * If P1|P2 yields a situation with both a class and its complement are
1478 * set, like having both \w and \W, this matches all code points, and we
1479 * can delete these from the P component of the ssc going forward. XXX We
1480 * might be able to delete all the P components, but I (khw) am not certain
1481 * about this, and it is better to be safe.
1484 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1485 * <= (C1 | P1) | ~C2
1486 * <= (C1 | ~C2) | P1
1487 * (which results in actually simpler code than the non-inverted case)
1490 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1491 && ! is_ANYOF_SYNTHETIC(or_with))
1493 /* We ignore P2, leaving P1 going forward */
1494 } /* else Not inverted */
1495 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1496 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1497 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1499 for (i = 0; i < ANYOF_MAX; i += 2) {
1500 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1502 ssc_match_all_cp(ssc);
1503 ANYOF_POSIXL_CLEAR(ssc, i);
1504 ANYOF_POSIXL_CLEAR(ssc, i+1);
1512 FALSE /* Already has been inverted */
1516 PERL_STATIC_INLINE void
1517 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1519 PERL_ARGS_ASSERT_SSC_UNION;
1521 assert(is_ANYOF_SYNTHETIC(ssc));
1523 _invlist_union_maybe_complement_2nd(ssc->invlist,
1529 PERL_STATIC_INLINE void
1530 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1532 const bool invert2nd)
1534 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1536 assert(is_ANYOF_SYNTHETIC(ssc));
1538 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1544 PERL_STATIC_INLINE void
1545 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1547 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1549 assert(is_ANYOF_SYNTHETIC(ssc));
1551 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1554 PERL_STATIC_INLINE void
1555 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1557 /* AND just the single code point 'cp' into the SSC 'ssc' */
1559 SV* cp_list = _new_invlist(2);
1561 PERL_ARGS_ASSERT_SSC_CP_AND;
1563 assert(is_ANYOF_SYNTHETIC(ssc));
1565 cp_list = add_cp_to_invlist(cp_list, cp);
1566 ssc_intersection(ssc, cp_list,
1567 FALSE /* Not inverted */
1569 SvREFCNT_dec_NN(cp_list);
1572 PERL_STATIC_INLINE void
1573 S_ssc_clear_locale(regnode_ssc *ssc)
1575 /* Set the SSC 'ssc' to not match any locale things */
1576 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1578 assert(is_ANYOF_SYNTHETIC(ssc));
1580 ANYOF_POSIXL_ZERO(ssc);
1581 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1584 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1587 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1589 /* The synthetic start class is used to hopefully quickly winnow down
1590 * places where a pattern could start a match in the target string. If it
1591 * doesn't really narrow things down that much, there isn't much point to
1592 * having the overhead of using it. This function uses some very crude
1593 * heuristics to decide if to use the ssc or not.
1595 * It returns TRUE if 'ssc' rules out more than half what it considers to
1596 * be the "likely" possible matches, but of course it doesn't know what the
1597 * actual things being matched are going to be; these are only guesses
1599 * For /l matches, it assumes that the only likely matches are going to be
1600 * in the 0-255 range, uniformly distributed, so half of that is 127
1601 * For /a and /d matches, it assumes that the likely matches will be just
1602 * the ASCII range, so half of that is 63
1603 * For /u and there isn't anything matching above the Latin1 range, it
1604 * assumes that that is the only range likely to be matched, and uses
1605 * half that as the cut-off: 127. If anything matches above Latin1,
1606 * it assumes that all of Unicode could match (uniformly), except for
1607 * non-Unicode code points and things in the General Category "Other"
1608 * (unassigned, private use, surrogates, controls and formats). This
1609 * is a much large number. */
1611 const U32 max_match = (LOC)
1615 : (invlist_highest(ssc->invlist) < 256)
1617 : ((NON_OTHER_COUNT + 1) / 2) - 1;
1618 U32 count = 0; /* Running total of number of code points matched by
1620 UV start, end; /* Start and end points of current range in inversion
1623 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1625 invlist_iterinit(ssc->invlist);
1626 while (invlist_iternext(ssc->invlist, &start, &end)) {
1628 /* /u is the only thing that we expect to match above 255; so if not /u
1629 * and even if there are matches above 255, ignore them. This catches
1630 * things like \d under /d which does match the digits above 255, but
1631 * since the pattern is /d, it is not likely to be expecting them */
1632 if (! UNI_SEMANTICS) {
1636 end = MIN(end, 255);
1638 count += end - start + 1;
1639 if (count > max_match) {
1640 invlist_iterfinish(ssc->invlist);
1650 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1652 /* The inversion list in the SSC is marked mortal; now we need a more
1653 * permanent copy, which is stored the same way that is done in a regular
1654 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1657 SV* invlist = invlist_clone(ssc->invlist);
1659 PERL_ARGS_ASSERT_SSC_FINALIZE;
1661 assert(is_ANYOF_SYNTHETIC(ssc));
1663 /* The code in this file assumes that all but these flags aren't relevant
1664 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1665 * by the time we reach here */
1666 assert(! (ANYOF_FLAGS(ssc)
1667 & ~( ANYOF_COMMON_FLAGS
1668 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)));
1670 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1672 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1673 NULL, NULL, NULL, FALSE);
1675 /* Make sure is clone-safe */
1676 ssc->invlist = NULL;
1678 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1679 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1682 if (RExC_contains_locale) {
1686 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1689 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1690 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1691 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1692 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1693 ? (TRIE_LIST_CUR( idx ) - 1) \
1699 dump_trie(trie,widecharmap,revcharmap)
1700 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1701 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1703 These routines dump out a trie in a somewhat readable format.
1704 The _interim_ variants are used for debugging the interim
1705 tables that are used to generate the final compressed
1706 representation which is what dump_trie expects.
1708 Part of the reason for their existence is to provide a form
1709 of documentation as to how the different representations function.
1714 Dumps the final compressed table form of the trie to Perl_debug_log.
1715 Used for debugging make_trie().
1719 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1720 AV *revcharmap, U32 depth)
1723 SV *sv=sv_newmortal();
1724 int colwidth= widecharmap ? 6 : 4;
1726 GET_RE_DEBUG_FLAGS_DECL;
1728 PERL_ARGS_ASSERT_DUMP_TRIE;
1730 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1731 (int)depth * 2 + 2,"",
1732 "Match","Base","Ofs" );
1734 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1735 SV ** const tmp = av_fetch( revcharmap, state, 0);
1737 PerlIO_printf( Perl_debug_log, "%*s",
1739 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1740 PL_colors[0], PL_colors[1],
1741 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1742 PERL_PV_ESCAPE_FIRSTCHAR
1747 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1748 (int)depth * 2 + 2,"");
1750 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1751 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1752 PerlIO_printf( Perl_debug_log, "\n");
1754 for( state = 1 ; state < trie->statecount ; state++ ) {
1755 const U32 base = trie->states[ state ].trans.base;
1757 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1758 (int)depth * 2 + 2,"", (UV)state);
1760 if ( trie->states[ state ].wordnum ) {
1761 PerlIO_printf( Perl_debug_log, " W%4X",
1762 trie->states[ state ].wordnum );
1764 PerlIO_printf( Perl_debug_log, "%6s", "" );
1767 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1772 while( ( base + ofs < trie->uniquecharcount ) ||
1773 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1774 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1778 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1780 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1781 if ( ( base + ofs >= trie->uniquecharcount )
1782 && ( base + ofs - trie->uniquecharcount
1784 && trie->trans[ base + ofs
1785 - trie->uniquecharcount ].check == state )
1787 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1789 (UV)trie->trans[ base + ofs
1790 - trie->uniquecharcount ].next );
1792 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1796 PerlIO_printf( Perl_debug_log, "]");
1799 PerlIO_printf( Perl_debug_log, "\n" );
1801 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1803 for (word=1; word <= trie->wordcount; word++) {
1804 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1805 (int)word, (int)(trie->wordinfo[word].prev),
1806 (int)(trie->wordinfo[word].len));
1808 PerlIO_printf(Perl_debug_log, "\n" );
1811 Dumps a fully constructed but uncompressed trie in list form.
1812 List tries normally only are used for construction when the number of
1813 possible chars (trie->uniquecharcount) is very high.
1814 Used for debugging make_trie().
1817 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1818 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1822 SV *sv=sv_newmortal();
1823 int colwidth= widecharmap ? 6 : 4;
1824 GET_RE_DEBUG_FLAGS_DECL;
1826 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1828 /* print out the table precompression. */
1829 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1830 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1831 "------:-----+-----------------\n" );
1833 for( state=1 ; state < next_alloc ; state ++ ) {
1836 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1837 (int)depth * 2 + 2,"", (UV)state );
1838 if ( ! trie->states[ state ].wordnum ) {
1839 PerlIO_printf( Perl_debug_log, "%5s| ","");
1841 PerlIO_printf( Perl_debug_log, "W%4x| ",
1842 trie->states[ state ].wordnum
1845 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1846 SV ** const tmp = av_fetch( revcharmap,
1847 TRIE_LIST_ITEM(state,charid).forid, 0);
1849 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1851 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1853 PL_colors[0], PL_colors[1],
1854 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1855 | PERL_PV_ESCAPE_FIRSTCHAR
1857 TRIE_LIST_ITEM(state,charid).forid,
1858 (UV)TRIE_LIST_ITEM(state,charid).newstate
1861 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1862 (int)((depth * 2) + 14), "");
1865 PerlIO_printf( Perl_debug_log, "\n");
1870 Dumps a fully constructed but uncompressed trie in table form.
1871 This is the normal DFA style state transition table, with a few
1872 twists to facilitate compression later.
1873 Used for debugging make_trie().
1876 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1877 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1882 SV *sv=sv_newmortal();
1883 int colwidth= widecharmap ? 6 : 4;
1884 GET_RE_DEBUG_FLAGS_DECL;
1886 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1889 print out the table precompression so that we can do a visual check
1890 that they are identical.
1893 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1895 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1896 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1898 PerlIO_printf( Perl_debug_log, "%*s",
1900 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1901 PL_colors[0], PL_colors[1],
1902 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1903 PERL_PV_ESCAPE_FIRSTCHAR
1909 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1911 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1912 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1915 PerlIO_printf( Perl_debug_log, "\n" );
1917 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1919 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1920 (int)depth * 2 + 2,"",
1921 (UV)TRIE_NODENUM( state ) );
1923 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1924 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1926 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1928 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1930 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1931 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1932 (UV)trie->trans[ state ].check );
1934 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1935 (UV)trie->trans[ state ].check,
1936 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1944 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1945 startbranch: the first branch in the whole branch sequence
1946 first : start branch of sequence of branch-exact nodes.
1947 May be the same as startbranch
1948 last : Thing following the last branch.
1949 May be the same as tail.
1950 tail : item following the branch sequence
1951 count : words in the sequence
1952 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
1953 depth : indent depth
1955 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1957 A trie is an N'ary tree where the branches are determined by digital
1958 decomposition of the key. IE, at the root node you look up the 1st character and
1959 follow that branch repeat until you find the end of the branches. Nodes can be
1960 marked as "accepting" meaning they represent a complete word. Eg:
1964 would convert into the following structure. Numbers represent states, letters
1965 following numbers represent valid transitions on the letter from that state, if
1966 the number is in square brackets it represents an accepting state, otherwise it
1967 will be in parenthesis.
1969 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1973 (1) +-i->(6)-+-s->[7]
1975 +-s->(3)-+-h->(4)-+-e->[5]
1977 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1979 This shows that when matching against the string 'hers' we will begin at state 1
1980 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1981 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1982 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1983 single traverse. We store a mapping from accepting to state to which word was
1984 matched, and then when we have multiple possibilities we try to complete the
1985 rest of the regex in the order in which they occurred in the alternation.
1987 The only prior NFA like behaviour that would be changed by the TRIE support is
1988 the silent ignoring of duplicate alternations which are of the form:
1990 / (DUPE|DUPE) X? (?{ ... }) Y /x
1992 Thus EVAL blocks following a trie may be called a different number of times with
1993 and without the optimisation. With the optimisations dupes will be silently
1994 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1995 the following demonstrates:
1997 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1999 which prints out 'word' three times, but
2001 'words'=~/(word|word|word)(?{ print $1 })S/
2003 which doesnt print it out at all. This is due to other optimisations kicking in.
2005 Example of what happens on a structural level:
2007 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2009 1: CURLYM[1] {1,32767}(18)
2020 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2021 and should turn into:
2023 1: CURLYM[1] {1,32767}(18)
2025 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2033 Cases where tail != last would be like /(?foo|bar)baz/:
2043 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2044 and would end up looking like:
2047 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2054 d = uvchr_to_utf8_flags(d, uv, 0);
2056 is the recommended Unicode-aware way of saying
2061 #define TRIE_STORE_REVCHAR(val) \
2064 SV *zlopp = newSV(UTF8_MAXBYTES); \
2065 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2066 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2067 SvCUR_set(zlopp, kapow - flrbbbbb); \
2070 av_push(revcharmap, zlopp); \
2072 char ooooff = (char)val; \
2073 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2077 /* This gets the next character from the input, folding it if not already
2079 #define TRIE_READ_CHAR STMT_START { \
2082 /* if it is UTF then it is either already folded, or does not need \
2084 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2086 else if (folder == PL_fold_latin1) { \
2087 /* This folder implies Unicode rules, which in the range expressible \
2088 * by not UTF is the lower case, with the two exceptions, one of \
2089 * which should have been taken care of before calling this */ \
2090 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2091 uvc = toLOWER_L1(*uc); \
2092 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2095 /* raw data, will be folded later if needed */ \
2103 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2104 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2105 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2106 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2108 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2109 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2110 TRIE_LIST_CUR( state )++; \
2113 #define TRIE_LIST_NEW(state) STMT_START { \
2114 Newxz( trie->states[ state ].trans.list, \
2115 4, reg_trie_trans_le ); \
2116 TRIE_LIST_CUR( state ) = 1; \
2117 TRIE_LIST_LEN( state ) = 4; \
2120 #define TRIE_HANDLE_WORD(state) STMT_START { \
2121 U16 dupe= trie->states[ state ].wordnum; \
2122 regnode * const noper_next = regnext( noper ); \
2125 /* store the word for dumping */ \
2127 if (OP(noper) != NOTHING) \
2128 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2130 tmp = newSVpvn_utf8( "", 0, UTF ); \
2131 av_push( trie_words, tmp ); \
2135 trie->wordinfo[curword].prev = 0; \
2136 trie->wordinfo[curword].len = wordlen; \
2137 trie->wordinfo[curword].accept = state; \
2139 if ( noper_next < tail ) { \
2141 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2143 trie->jump[curword] = (U16)(noper_next - convert); \
2145 jumper = noper_next; \
2147 nextbranch= regnext(cur); \
2151 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2152 /* chain, so that when the bits of chain are later */\
2153 /* linked together, the dups appear in the chain */\
2154 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2155 trie->wordinfo[dupe].prev = curword; \
2157 /* we haven't inserted this word yet. */ \
2158 trie->states[ state ].wordnum = curword; \
2163 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2164 ( ( base + charid >= ucharcount \
2165 && base + charid < ubound \
2166 && state == trie->trans[ base - ucharcount + charid ].check \
2167 && trie->trans[ base - ucharcount + charid ].next ) \
2168 ? trie->trans[ base - ucharcount + charid ].next \
2169 : ( state==1 ? special : 0 ) \
2173 #define MADE_JUMP_TRIE 2
2174 #define MADE_EXACT_TRIE 4
2177 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2178 regnode *first, regnode *last, regnode *tail,
2179 U32 word_count, U32 flags, U32 depth)
2181 /* first pass, loop through and scan words */
2182 reg_trie_data *trie;
2183 HV *widecharmap = NULL;
2184 AV *revcharmap = newAV();
2190 regnode *jumper = NULL;
2191 regnode *nextbranch = NULL;
2192 regnode *convert = NULL;
2193 U32 *prev_states; /* temp array mapping each state to previous one */
2194 /* we just use folder as a flag in utf8 */
2195 const U8 * folder = NULL;
2198 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2199 AV *trie_words = NULL;
2200 /* along with revcharmap, this only used during construction but both are
2201 * useful during debugging so we store them in the struct when debugging.
2204 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2205 STRLEN trie_charcount=0;
2207 SV *re_trie_maxbuff;
2208 GET_RE_DEBUG_FLAGS_DECL;
2210 PERL_ARGS_ASSERT_MAKE_TRIE;
2212 PERL_UNUSED_ARG(depth);
2216 case EXACT: case EXACTL: break;
2220 case EXACTFLU8: folder = PL_fold_latin1; break;
2221 case EXACTF: folder = PL_fold; break;
2222 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2225 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2227 trie->startstate = 1;
2228 trie->wordcount = word_count;
2229 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2230 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2231 if (flags == EXACT || flags == EXACTL)
2232 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2233 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2234 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2237 trie_words = newAV();
2240 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2241 assert(re_trie_maxbuff);
2242 if (!SvIOK(re_trie_maxbuff)) {
2243 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2245 DEBUG_TRIE_COMPILE_r({
2246 PerlIO_printf( Perl_debug_log,
2247 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2248 (int)depth * 2 + 2, "",
2249 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2250 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2253 /* Find the node we are going to overwrite */
2254 if ( first == startbranch && OP( last ) != BRANCH ) {
2255 /* whole branch chain */
2258 /* branch sub-chain */
2259 convert = NEXTOPER( first );
2262 /* -- First loop and Setup --
2264 We first traverse the branches and scan each word to determine if it
2265 contains widechars, and how many unique chars there are, this is
2266 important as we have to build a table with at least as many columns as we
2269 We use an array of integers to represent the character codes 0..255
2270 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2271 the native representation of the character value as the key and IV's for
2274 *TODO* If we keep track of how many times each character is used we can
2275 remap the columns so that the table compression later on is more
2276 efficient in terms of memory by ensuring the most common value is in the
2277 middle and the least common are on the outside. IMO this would be better
2278 than a most to least common mapping as theres a decent chance the most
2279 common letter will share a node with the least common, meaning the node
2280 will not be compressible. With a middle is most common approach the worst
2281 case is when we have the least common nodes twice.
2285 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2286 regnode *noper = NEXTOPER( cur );
2287 const U8 *uc = (U8*)STRING( noper );
2288 const U8 *e = uc + STR_LEN( noper );
2290 U32 wordlen = 0; /* required init */
2291 STRLEN minchars = 0;
2292 STRLEN maxchars = 0;
2293 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2296 if (OP(noper) == NOTHING) {
2297 regnode *noper_next= regnext(noper);
2298 if (noper_next != tail && OP(noper_next) == flags) {
2300 uc= (U8*)STRING(noper);
2301 e= uc + STR_LEN(noper);
2302 trie->minlen= STR_LEN(noper);
2309 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2310 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2311 regardless of encoding */
2312 if (OP( noper ) == EXACTFU_SS) {
2313 /* false positives are ok, so just set this */
2314 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2317 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2319 TRIE_CHARCOUNT(trie)++;
2322 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2323 * is in effect. Under /i, this character can match itself, or
2324 * anything that folds to it. If not under /i, it can match just
2325 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2326 * all fold to k, and all are single characters. But some folds
2327 * expand to more than one character, so for example LATIN SMALL
2328 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2329 * the string beginning at 'uc' is 'ffi', it could be matched by
2330 * three characters, or just by the one ligature character. (It
2331 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2332 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2333 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2334 * match.) The trie needs to know the minimum and maximum number
2335 * of characters that could match so that it can use size alone to
2336 * quickly reject many match attempts. The max is simple: it is
2337 * the number of folded characters in this branch (since a fold is
2338 * never shorter than what folds to it. */
2342 /* And the min is equal to the max if not under /i (indicated by
2343 * 'folder' being NULL), or there are no multi-character folds. If
2344 * there is a multi-character fold, the min is incremented just
2345 * once, for the character that folds to the sequence. Each
2346 * character in the sequence needs to be added to the list below of
2347 * characters in the trie, but we count only the first towards the
2348 * min number of characters needed. This is done through the
2349 * variable 'foldlen', which is returned by the macros that look
2350 * for these sequences as the number of bytes the sequence
2351 * occupies. Each time through the loop, we decrement 'foldlen' by
2352 * how many bytes the current char occupies. Only when it reaches
2353 * 0 do we increment 'minchars' or look for another multi-character
2355 if (folder == NULL) {
2358 else if (foldlen > 0) {
2359 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2364 /* See if *uc is the beginning of a multi-character fold. If
2365 * so, we decrement the length remaining to look at, to account
2366 * for the current character this iteration. (We can use 'uc'
2367 * instead of the fold returned by TRIE_READ_CHAR because for
2368 * non-UTF, the latin1_safe macro is smart enough to account
2369 * for all the unfolded characters, and because for UTF, the
2370 * string will already have been folded earlier in the
2371 * compilation process */
2373 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2374 foldlen -= UTF8SKIP(uc);
2377 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2382 /* The current character (and any potential folds) should be added
2383 * to the possible matching characters for this position in this
2387 U8 folded= folder[ (U8) uvc ];
2388 if ( !trie->charmap[ folded ] ) {
2389 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2390 TRIE_STORE_REVCHAR( folded );
2393 if ( !trie->charmap[ uvc ] ) {
2394 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2395 TRIE_STORE_REVCHAR( uvc );
2398 /* store the codepoint in the bitmap, and its folded
2400 TRIE_BITMAP_SET(trie, uvc);
2402 /* store the folded codepoint */
2403 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2406 /* store first byte of utf8 representation of
2407 variant codepoints */
2408 if (! UVCHR_IS_INVARIANT(uvc)) {
2409 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2412 set_bit = 0; /* We've done our bit :-) */
2416 /* XXX We could come up with the list of code points that fold
2417 * to this using PL_utf8_foldclosures, except not for
2418 * multi-char folds, as there may be multiple combinations
2419 * there that could work, which needs to wait until runtime to
2420 * resolve (The comment about LIGATURE FFI above is such an
2425 widecharmap = newHV();
2427 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2430 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2432 if ( !SvTRUE( *svpp ) ) {
2433 sv_setiv( *svpp, ++trie->uniquecharcount );
2434 TRIE_STORE_REVCHAR(uvc);
2437 } /* end loop through characters in this branch of the trie */
2439 /* We take the min and max for this branch and combine to find the min
2440 * and max for all branches processed so far */
2441 if( cur == first ) {
2442 trie->minlen = minchars;
2443 trie->maxlen = maxchars;
2444 } else if (minchars < trie->minlen) {
2445 trie->minlen = minchars;
2446 } else if (maxchars > trie->maxlen) {
2447 trie->maxlen = maxchars;
2449 } /* end first pass */
2450 DEBUG_TRIE_COMPILE_r(
2451 PerlIO_printf( Perl_debug_log,
2452 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2453 (int)depth * 2 + 2,"",
2454 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2455 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2456 (int)trie->minlen, (int)trie->maxlen )
2460 We now know what we are dealing with in terms of unique chars and
2461 string sizes so we can calculate how much memory a naive
2462 representation using a flat table will take. If it's over a reasonable
2463 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2464 conservative but potentially much slower representation using an array
2467 At the end we convert both representations into the same compressed
2468 form that will be used in regexec.c for matching with. The latter
2469 is a form that cannot be used to construct with but has memory
2470 properties similar to the list form and access properties similar
2471 to the table form making it both suitable for fast searches and
2472 small enough that its feasable to store for the duration of a program.
2474 See the comment in the code where the compressed table is produced
2475 inplace from the flat tabe representation for an explanation of how
2476 the compression works.
2481 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2484 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2485 > SvIV(re_trie_maxbuff) )
2488 Second Pass -- Array Of Lists Representation
2490 Each state will be represented by a list of charid:state records
2491 (reg_trie_trans_le) the first such element holds the CUR and LEN
2492 points of the allocated array. (See defines above).
2494 We build the initial structure using the lists, and then convert
2495 it into the compressed table form which allows faster lookups
2496 (but cant be modified once converted).
2499 STRLEN transcount = 1;
2501 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2502 "%*sCompiling trie using list compiler\n",
2503 (int)depth * 2 + 2, ""));
2505 trie->states = (reg_trie_state *)
2506 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2507 sizeof(reg_trie_state) );
2511 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2513 regnode *noper = NEXTOPER( cur );
2514 U8 *uc = (U8*)STRING( noper );
2515 const U8 *e = uc + STR_LEN( noper );
2516 U32 state = 1; /* required init */
2517 U16 charid = 0; /* sanity init */
2518 U32 wordlen = 0; /* required init */
2520 if (OP(noper) == NOTHING) {
2521 regnode *noper_next= regnext(noper);
2522 if (noper_next != tail && OP(noper_next) == flags) {
2524 uc= (U8*)STRING(noper);
2525 e= uc + STR_LEN(noper);
2529 if (OP(noper) != NOTHING) {
2530 for ( ; uc < e ; uc += len ) {
2535 charid = trie->charmap[ uvc ];
2537 SV** const svpp = hv_fetch( widecharmap,
2544 charid=(U16)SvIV( *svpp );
2547 /* charid is now 0 if we dont know the char read, or
2548 * nonzero if we do */
2555 if ( !trie->states[ state ].trans.list ) {
2556 TRIE_LIST_NEW( state );
2559 check <= TRIE_LIST_USED( state );
2562 if ( TRIE_LIST_ITEM( state, check ).forid
2565 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2570 newstate = next_alloc++;
2571 prev_states[newstate] = state;
2572 TRIE_LIST_PUSH( state, charid, newstate );
2577 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2581 TRIE_HANDLE_WORD(state);
2583 } /* end second pass */
2585 /* next alloc is the NEXT state to be allocated */
2586 trie->statecount = next_alloc;
2587 trie->states = (reg_trie_state *)
2588 PerlMemShared_realloc( trie->states,
2590 * sizeof(reg_trie_state) );
2592 /* and now dump it out before we compress it */
2593 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2594 revcharmap, next_alloc,
2598 trie->trans = (reg_trie_trans *)
2599 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2606 for( state=1 ; state < next_alloc ; state ++ ) {
2610 DEBUG_TRIE_COMPILE_MORE_r(
2611 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2615 if (trie->states[state].trans.list) {
2616 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2620 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2621 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2622 if ( forid < minid ) {
2624 } else if ( forid > maxid ) {
2628 if ( transcount < tp + maxid - minid + 1) {
2630 trie->trans = (reg_trie_trans *)
2631 PerlMemShared_realloc( trie->trans,
2633 * sizeof(reg_trie_trans) );
2634 Zero( trie->trans + (transcount / 2),
2638 base = trie->uniquecharcount + tp - minid;
2639 if ( maxid == minid ) {
2641 for ( ; zp < tp ; zp++ ) {
2642 if ( ! trie->trans[ zp ].next ) {
2643 base = trie->uniquecharcount + zp - minid;
2644 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2646 trie->trans[ zp ].check = state;
2652 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2654 trie->trans[ tp ].check = state;
2659 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2660 const U32 tid = base
2661 - trie->uniquecharcount
2662 + TRIE_LIST_ITEM( state, idx ).forid;
2663 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2665 trie->trans[ tid ].check = state;
2667 tp += ( maxid - minid + 1 );
2669 Safefree(trie->states[ state ].trans.list);
2672 DEBUG_TRIE_COMPILE_MORE_r(
2673 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2676 trie->states[ state ].trans.base=base;
2678 trie->lasttrans = tp + 1;
2682 Second Pass -- Flat Table Representation.
2684 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2685 each. We know that we will need Charcount+1 trans at most to store
2686 the data (one row per char at worst case) So we preallocate both
2687 structures assuming worst case.
2689 We then construct the trie using only the .next slots of the entry
2692 We use the .check field of the first entry of the node temporarily
2693 to make compression both faster and easier by keeping track of how
2694 many non zero fields are in the node.
2696 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2699 There are two terms at use here: state as a TRIE_NODEIDX() which is
2700 a number representing the first entry of the node, and state as a
2701 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2702 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2703 if there are 2 entrys per node. eg:
2711 The table is internally in the right hand, idx form. However as we
2712 also have to deal with the states array which is indexed by nodenum
2713 we have to use TRIE_NODENUM() to convert.
2716 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2717 "%*sCompiling trie using table compiler\n",
2718 (int)depth * 2 + 2, ""));
2720 trie->trans = (reg_trie_trans *)
2721 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2722 * trie->uniquecharcount + 1,
2723 sizeof(reg_trie_trans) );
2724 trie->states = (reg_trie_state *)
2725 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2726 sizeof(reg_trie_state) );
2727 next_alloc = trie->uniquecharcount + 1;
2730 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2732 regnode *noper = NEXTOPER( cur );
2733 const U8 *uc = (U8*)STRING( noper );
2734 const U8 *e = uc + STR_LEN( noper );
2736 U32 state = 1; /* required init */
2738 U16 charid = 0; /* sanity init */
2739 U32 accept_state = 0; /* sanity init */
2741 U32 wordlen = 0; /* required init */
2743 if (OP(noper) == NOTHING) {
2744 regnode *noper_next= regnext(noper);
2745 if (noper_next != tail && OP(noper_next) == flags) {
2747 uc= (U8*)STRING(noper);
2748 e= uc + STR_LEN(noper);
2752 if ( OP(noper) != NOTHING ) {
2753 for ( ; uc < e ; uc += len ) {
2758 charid = trie->charmap[ uvc ];
2760 SV* const * const svpp = hv_fetch( widecharmap,
2764 charid = svpp ? (U16)SvIV(*svpp) : 0;
2768 if ( !trie->trans[ state + charid ].next ) {
2769 trie->trans[ state + charid ].next = next_alloc;
2770 trie->trans[ state ].check++;
2771 prev_states[TRIE_NODENUM(next_alloc)]
2772 = TRIE_NODENUM(state);
2773 next_alloc += trie->uniquecharcount;
2775 state = trie->trans[ state + charid ].next;
2777 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2779 /* charid is now 0 if we dont know the char read, or
2780 * nonzero if we do */
2783 accept_state = TRIE_NODENUM( state );
2784 TRIE_HANDLE_WORD(accept_state);
2786 } /* end second pass */
2788 /* and now dump it out before we compress it */
2789 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2791 next_alloc, depth+1));
2795 * Inplace compress the table.*
2797 For sparse data sets the table constructed by the trie algorithm will
2798 be mostly 0/FAIL transitions or to put it another way mostly empty.
2799 (Note that leaf nodes will not contain any transitions.)
2801 This algorithm compresses the tables by eliminating most such
2802 transitions, at the cost of a modest bit of extra work during lookup:
2804 - Each states[] entry contains a .base field which indicates the
2805 index in the state[] array wheres its transition data is stored.
2807 - If .base is 0 there are no valid transitions from that node.
2809 - If .base is nonzero then charid is added to it to find an entry in
2812 -If trans[states[state].base+charid].check!=state then the
2813 transition is taken to be a 0/Fail transition. Thus if there are fail
2814 transitions at the front of the node then the .base offset will point
2815 somewhere inside the previous nodes data (or maybe even into a node
2816 even earlier), but the .check field determines if the transition is
2820 The following process inplace converts the table to the compressed
2821 table: We first do not compress the root node 1,and mark all its
2822 .check pointers as 1 and set its .base pointer as 1 as well. This
2823 allows us to do a DFA construction from the compressed table later,
2824 and ensures that any .base pointers we calculate later are greater
2827 - We set 'pos' to indicate the first entry of the second node.
2829 - We then iterate over the columns of the node, finding the first and
2830 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2831 and set the .check pointers accordingly, and advance pos
2832 appropriately and repreat for the next node. Note that when we copy
2833 the next pointers we have to convert them from the original
2834 NODEIDX form to NODENUM form as the former is not valid post
2837 - If a node has no transitions used we mark its base as 0 and do not
2838 advance the pos pointer.
2840 - If a node only has one transition we use a second pointer into the
2841 structure to fill in allocated fail transitions from other states.
2842 This pointer is independent of the main pointer and scans forward
2843 looking for null transitions that are allocated to a state. When it
2844 finds one it writes the single transition into the "hole". If the
2845 pointer doesnt find one the single transition is appended as normal.
2847 - Once compressed we can Renew/realloc the structures to release the
2850 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2851 specifically Fig 3.47 and the associated pseudocode.
2855 const U32 laststate = TRIE_NODENUM( next_alloc );
2858 trie->statecount = laststate;
2860 for ( state = 1 ; state < laststate ; state++ ) {
2862 const U32 stateidx = TRIE_NODEIDX( state );
2863 const U32 o_used = trie->trans[ stateidx ].check;
2864 U32 used = trie->trans[ stateidx ].check;
2865 trie->trans[ stateidx ].check = 0;
2868 used && charid < trie->uniquecharcount;
2871 if ( flag || trie->trans[ stateidx + charid ].next ) {
2872 if ( trie->trans[ stateidx + charid ].next ) {
2874 for ( ; zp < pos ; zp++ ) {
2875 if ( ! trie->trans[ zp ].next ) {
2879 trie->states[ state ].trans.base
2881 + trie->uniquecharcount
2883 trie->trans[ zp ].next
2884 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2886 trie->trans[ zp ].check = state;
2887 if ( ++zp > pos ) pos = zp;
2894 trie->states[ state ].trans.base
2895 = pos + trie->uniquecharcount - charid ;
2897 trie->trans[ pos ].next
2898 = SAFE_TRIE_NODENUM(
2899 trie->trans[ stateidx + charid ].next );
2900 trie->trans[ pos ].check = state;
2905 trie->lasttrans = pos + 1;
2906 trie->states = (reg_trie_state *)
2907 PerlMemShared_realloc( trie->states, laststate
2908 * sizeof(reg_trie_state) );
2909 DEBUG_TRIE_COMPILE_MORE_r(
2910 PerlIO_printf( Perl_debug_log,
2911 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2912 (int)depth * 2 + 2,"",
2913 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2917 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2920 } /* end table compress */
2922 DEBUG_TRIE_COMPILE_MORE_r(
2923 PerlIO_printf(Perl_debug_log,
2924 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2925 (int)depth * 2 + 2, "",
2926 (UV)trie->statecount,
2927 (UV)trie->lasttrans)
2929 /* resize the trans array to remove unused space */
2930 trie->trans = (reg_trie_trans *)
2931 PerlMemShared_realloc( trie->trans, trie->lasttrans
2932 * sizeof(reg_trie_trans) );
2934 { /* Modify the program and insert the new TRIE node */
2935 U8 nodetype =(U8)(flags & 0xFF);
2939 regnode *optimize = NULL;
2940 #ifdef RE_TRACK_PATTERN_OFFSETS
2943 U32 mjd_nodelen = 0;
2944 #endif /* RE_TRACK_PATTERN_OFFSETS */
2945 #endif /* DEBUGGING */
2947 This means we convert either the first branch or the first Exact,
2948 depending on whether the thing following (in 'last') is a branch
2949 or not and whther first is the startbranch (ie is it a sub part of
2950 the alternation or is it the whole thing.)
2951 Assuming its a sub part we convert the EXACT otherwise we convert
2952 the whole branch sequence, including the first.
2954 /* Find the node we are going to overwrite */
2955 if ( first != startbranch || OP( last ) == BRANCH ) {
2956 /* branch sub-chain */
2957 NEXT_OFF( first ) = (U16)(last - first);
2958 #ifdef RE_TRACK_PATTERN_OFFSETS
2960 mjd_offset= Node_Offset((convert));
2961 mjd_nodelen= Node_Length((convert));
2964 /* whole branch chain */
2966 #ifdef RE_TRACK_PATTERN_OFFSETS
2969 const regnode *nop = NEXTOPER( convert );
2970 mjd_offset= Node_Offset((nop));
2971 mjd_nodelen= Node_Length((nop));
2975 PerlIO_printf(Perl_debug_log,
2976 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2977 (int)depth * 2 + 2, "",
2978 (UV)mjd_offset, (UV)mjd_nodelen)
2981 /* But first we check to see if there is a common prefix we can
2982 split out as an EXACT and put in front of the TRIE node. */
2983 trie->startstate= 1;
2984 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2986 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2990 const U32 base = trie->states[ state ].trans.base;
2992 if ( trie->states[state].wordnum )
2995 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2996 if ( ( base + ofs >= trie->uniquecharcount ) &&
2997 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2998 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3000 if ( ++count > 1 ) {
3001 SV **tmp = av_fetch( revcharmap, ofs, 0);
3002 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
3003 if ( state == 1 ) break;
3005 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3007 PerlIO_printf(Perl_debug_log,
3008 "%*sNew Start State=%"UVuf" Class: [",
3009 (int)depth * 2 + 2, "",
3012 SV ** const tmp = av_fetch( revcharmap, idx, 0);
3013 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3015 TRIE_BITMAP_SET(trie,*ch);
3017 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3019 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3023 TRIE_BITMAP_SET(trie,*ch);
3025 TRIE_BITMAP_SET(trie,folder[ *ch ]);
3026 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
3032 SV **tmp = av_fetch( revcharmap, idx, 0);
3034 char *ch = SvPV( *tmp, len );
3036 SV *sv=sv_newmortal();
3037 PerlIO_printf( Perl_debug_log,
3038 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
3039 (int)depth * 2 + 2, "",
3041 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3042 PL_colors[0], PL_colors[1],
3043 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3044 PERL_PV_ESCAPE_FIRSTCHAR
3049 OP( convert ) = nodetype;
3050 str=STRING(convert);
3053 STR_LEN(convert) += len;
3059 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
3064 trie->prefixlen = (state-1);
3066 regnode *n = convert+NODE_SZ_STR(convert);
3067 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3068 trie->startstate = state;
3069 trie->minlen -= (state - 1);
3070 trie->maxlen -= (state - 1);
3072 /* At least the UNICOS C compiler choked on this
3073 * being argument to DEBUG_r(), so let's just have
3076 #ifdef PERL_EXT_RE_BUILD
3082 regnode *fix = convert;
3083 U32 word = trie->wordcount;
3085 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3086 while( ++fix < n ) {
3087 Set_Node_Offset_Length(fix, 0, 0);
3090 SV ** const tmp = av_fetch( trie_words, word, 0 );
3092 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3093 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3095 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3103 NEXT_OFF(convert) = (U16)(tail - convert);
3104 DEBUG_r(optimize= n);
3110 if ( trie->maxlen ) {
3111 NEXT_OFF( convert ) = (U16)(tail - convert);
3112 ARG_SET( convert, data_slot );
3113 /* Store the offset to the first unabsorbed branch in
3114 jump[0], which is otherwise unused by the jump logic.
3115 We use this when dumping a trie and during optimisation. */
3117 trie->jump[0] = (U16)(nextbranch - convert);
3119 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3120 * and there is a bitmap
3121 * and the first "jump target" node we found leaves enough room
3122 * then convert the TRIE node into a TRIEC node, with the bitmap
3123 * embedded inline in the opcode - this is hypothetically faster.
3125 if ( !trie->states[trie->startstate].wordnum
3127 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3129 OP( convert ) = TRIEC;
3130 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3131 PerlMemShared_free(trie->bitmap);
3134 OP( convert ) = TRIE;
3136 /* store the type in the flags */
3137 convert->flags = nodetype;
3141 + regarglen[ OP( convert ) ];
3143 /* XXX We really should free up the resource in trie now,
3144 as we won't use them - (which resources?) dmq */
3146 /* needed for dumping*/
3147 DEBUG_r(if (optimize) {
3148 regnode *opt = convert;
3150 while ( ++opt < optimize) {
3151 Set_Node_Offset_Length(opt,0,0);
3154 Try to clean up some of the debris left after the
3157 while( optimize < jumper ) {
3158 mjd_nodelen += Node_Length((optimize));
3159 OP( optimize ) = OPTIMIZED;
3160 Set_Node_Offset_Length(optimize,0,0);
3163 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3165 } /* end node insert */
3167 /* Finish populating the prev field of the wordinfo array. Walk back
3168 * from each accept state until we find another accept state, and if
3169 * so, point the first word's .prev field at the second word. If the
3170 * second already has a .prev field set, stop now. This will be the
3171 * case either if we've already processed that word's accept state,
3172 * or that state had multiple words, and the overspill words were
3173 * already linked up earlier.
3180 for (word=1; word <= trie->wordcount; word++) {
3182 if (trie->wordinfo[word].prev)
3184 state = trie->wordinfo[word].accept;
3186 state = prev_states[state];
3189 prev = trie->states[state].wordnum;
3193 trie->wordinfo[word].prev = prev;
3195 Safefree(prev_states);
3199 /* and now dump out the compressed format */
3200 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3202 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3204 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3205 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3207 SvREFCNT_dec_NN(revcharmap);
3211 : trie->startstate>1
3217 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3219 /* The Trie is constructed and compressed now so we can build a fail array if
3222 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3224 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3228 We find the fail state for each state in the trie, this state is the longest
3229 proper suffix of the current state's 'word' that is also a proper prefix of
3230 another word in our trie. State 1 represents the word '' and is thus the
3231 default fail state. This allows the DFA not to have to restart after its
3232 tried and failed a word at a given point, it simply continues as though it
3233 had been matching the other word in the first place.
3235 'abcdgu'=~/abcdefg|cdgu/
3236 When we get to 'd' we are still matching the first word, we would encounter
3237 'g' which would fail, which would bring us to the state representing 'd' in
3238 the second word where we would try 'g' and succeed, proceeding to match
3241 /* add a fail transition */
3242 const U32 trie_offset = ARG(source);
3243 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3245 const U32 ucharcount = trie->uniquecharcount;
3246 const U32 numstates = trie->statecount;
3247 const U32 ubound = trie->lasttrans + ucharcount;
3251 U32 base = trie->states[ 1 ].trans.base;
3254 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3256 GET_RE_DEBUG_FLAGS_DECL;
3258 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3259 PERL_UNUSED_CONTEXT;
3261 PERL_UNUSED_ARG(depth);
3264 if ( OP(source) == TRIE ) {
3265 struct regnode_1 *op = (struct regnode_1 *)
3266 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3267 StructCopy(source,op,struct regnode_1);
3268 stclass = (regnode *)op;
3270 struct regnode_charclass *op = (struct regnode_charclass *)
3271 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3272 StructCopy(source,op,struct regnode_charclass);
3273 stclass = (regnode *)op;
3275 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3277 ARG_SET( stclass, data_slot );
3278 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3279 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3280 aho->trie=trie_offset;
3281 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3282 Copy( trie->states, aho->states, numstates, reg_trie_state );
3283 Newxz( q, numstates, U32);
3284 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3287 /* initialize fail[0..1] to be 1 so that we always have
3288 a valid final fail state */
3289 fail[ 0 ] = fail[ 1 ] = 1;
3291 for ( charid = 0; charid < ucharcount ; charid++ ) {
3292 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3294 q[ q_write ] = newstate;
3295 /* set to point at the root */
3296 fail[ q[ q_write++ ] ]=1;
3299 while ( q_read < q_write) {
3300 const U32 cur = q[ q_read++ % numstates ];
3301 base = trie->states[ cur ].trans.base;
3303 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3304 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3306 U32 fail_state = cur;
3309 fail_state = fail[ fail_state ];
3310 fail_base = aho->states[ fail_state ].trans.base;
3311 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3313 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3314 fail[ ch_state ] = fail_state;
3315 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3317 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3319 q[ q_write++ % numstates] = ch_state;
3323 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3324 when we fail in state 1, this allows us to use the
3325 charclass scan to find a valid start char. This is based on the principle
3326 that theres a good chance the string being searched contains lots of stuff
3327 that cant be a start char.
3329 fail[ 0 ] = fail[ 1 ] = 0;
3330 DEBUG_TRIE_COMPILE_r({
3331 PerlIO_printf(Perl_debug_log,
3332 "%*sStclass Failtable (%"UVuf" states): 0",
3333 (int)(depth * 2), "", (UV)numstates
3335 for( q_read=1; q_read<numstates; q_read++ ) {
3336 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3338 PerlIO_printf(Perl_debug_log, "\n");
3341 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3346 #define DEBUG_PEEP(str,scan,depth) \
3347 DEBUG_OPTIMISE_r({if (scan){ \
3348 regnode *Next = regnext(scan); \
3349 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state); \
3350 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)", \
3351 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3352 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3353 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3354 PerlIO_printf(Perl_debug_log, "\n"); \
3357 /* The below joins as many adjacent EXACTish nodes as possible into a single
3358 * one. The regop may be changed if the node(s) contain certain sequences that
3359 * require special handling. The joining is only done if:
3360 * 1) there is room in the current conglomerated node to entirely contain the
3362 * 2) they are the exact same node type
3364 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3365 * these get optimized out
3367 * If a node is to match under /i (folded), the number of characters it matches
3368 * can be different than its character length if it contains a multi-character
3369 * fold. *min_subtract is set to the total delta number of characters of the
3372 * And *unfolded_multi_char is set to indicate whether or not the node contains
3373 * an unfolded multi-char fold. This happens when whether the fold is valid or
3374 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3375 * SMALL LETTER SHARP S, as only if the target string being matched against
3376 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3377 * folding rules depend on the locale in force at runtime. (Multi-char folds
3378 * whose components are all above the Latin1 range are not run-time locale
3379 * dependent, and have already been folded by the time this function is
3382 * This is as good a place as any to discuss the design of handling these
3383 * multi-character fold sequences. It's been wrong in Perl for a very long
3384 * time. There are three code points in Unicode whose multi-character folds
3385 * were long ago discovered to mess things up. The previous designs for
3386 * dealing with these involved assigning a special node for them. This
3387 * approach doesn't always work, as evidenced by this example:
3388 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3389 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3390 * would match just the \xDF, it won't be able to handle the case where a
3391 * successful match would have to cross the node's boundary. The new approach
3392 * that hopefully generally solves the problem generates an EXACTFU_SS node
3393 * that is "sss" in this case.
3395 * It turns out that there are problems with all multi-character folds, and not
3396 * just these three. Now the code is general, for all such cases. The
3397 * approach taken is:
3398 * 1) This routine examines each EXACTFish node that could contain multi-
3399 * character folded sequences. Since a single character can fold into
3400 * such a sequence, the minimum match length for this node is less than
3401 * the number of characters in the node. This routine returns in
3402 * *min_subtract how many characters to subtract from the the actual
3403 * length of the string to get a real minimum match length; it is 0 if
3404 * there are no multi-char foldeds. This delta is used by the caller to
3405 * adjust the min length of the match, and the delta between min and max,
3406 * so that the optimizer doesn't reject these possibilities based on size
3408 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3409 * is used for an EXACTFU node that contains at least one "ss" sequence in
3410 * it. For non-UTF-8 patterns and strings, this is the only case where
3411 * there is a possible fold length change. That means that a regular
3412 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3413 * with length changes, and so can be processed faster. regexec.c takes
3414 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3415 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3416 * known until runtime). This saves effort in regex matching. However,
3417 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3418 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3419 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3420 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3421 * possibilities for the non-UTF8 patterns are quite simple, except for
3422 * the sharp s. All the ones that don't involve a UTF-8 target string are
3423 * members of a fold-pair, and arrays are set up for all of them so that
3424 * the other member of the pair can be found quickly. Code elsewhere in
3425 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3426 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3427 * described in the next item.
3428 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3429 * validity of the fold won't be known until runtime, and so must remain
3430 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3431 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3432 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3433 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3434 * The reason this is a problem is that the optimizer part of regexec.c
3435 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3436 * that a character in the pattern corresponds to at most a single
3437 * character in the target string. (And I do mean character, and not byte
3438 * here, unlike other parts of the documentation that have never been
3439 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3440 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3441 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3442 * nodes, violate the assumption, and they are the only instances where it
3443 * is violated. I'm reluctant to try to change the assumption, as the
3444 * code involved is impenetrable to me (khw), so instead the code here
3445 * punts. This routine examines EXACTFL nodes, and (when the pattern
3446 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3447 * boolean indicating whether or not the node contains such a fold. When
3448 * it is true, the caller sets a flag that later causes the optimizer in
3449 * this file to not set values for the floating and fixed string lengths,
3450 * and thus avoids the optimizer code in regexec.c that makes the invalid
3451 * assumption. Thus, there is no optimization based on string lengths for
3452 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3453 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3454 * assumption is wrong only in these cases is that all other non-UTF-8
3455 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3456 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3457 * EXACTF nodes because we don't know at compile time if it actually
3458 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3459 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3460 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3461 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3462 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3463 * string would require the pattern to be forced into UTF-8, the overhead
3464 * of which we want to avoid. Similarly the unfolded multi-char folds in
3465 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3468 * Similarly, the code that generates tries doesn't currently handle
3469 * not-already-folded multi-char folds, and it looks like a pain to change
3470 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3471 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3472 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3473 * using /iaa matching will be doing so almost entirely with ASCII
3474 * strings, so this should rarely be encountered in practice */
3476 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3477 if (PL_regkind[OP(scan)] == EXACT) \
3478 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3481 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3482 UV *min_subtract, bool *unfolded_multi_char,
3483 U32 flags,regnode *val, U32 depth)
3485 /* Merge several consecutive EXACTish nodes into one. */
3486 regnode *n = regnext(scan);
3488 regnode *next = scan + NODE_SZ_STR(scan);
3492 regnode *stop = scan;
3493 GET_RE_DEBUG_FLAGS_DECL;
3495 PERL_UNUSED_ARG(depth);
3498 PERL_ARGS_ASSERT_JOIN_EXACT;
3499 #ifndef EXPERIMENTAL_INPLACESCAN
3500 PERL_UNUSED_ARG(flags);
3501 PERL_UNUSED_ARG(val);
3503 DEBUG_PEEP("join",scan,depth);
3505 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3506 * EXACT ones that are mergeable to the current one. */
3508 && (PL_regkind[OP(n)] == NOTHING
3509 || (stringok && OP(n) == OP(scan)))
3511 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3514 if (OP(n) == TAIL || n > next)
3516 if (PL_regkind[OP(n)] == NOTHING) {
3517 DEBUG_PEEP("skip:",n,depth);
3518 NEXT_OFF(scan) += NEXT_OFF(n);
3519 next = n + NODE_STEP_REGNODE;
3526 else if (stringok) {
3527 const unsigned int oldl = STR_LEN(scan);
3528 regnode * const nnext = regnext(n);
3530 /* XXX I (khw) kind of doubt that this works on platforms (should
3531 * Perl ever run on one) where U8_MAX is above 255 because of lots
3532 * of other assumptions */
3533 /* Don't join if the sum can't fit into a single node */
3534 if (oldl + STR_LEN(n) > U8_MAX)
3537 DEBUG_PEEP("merg",n,depth);
3540 NEXT_OFF(scan) += NEXT_OFF(n);
3541 STR_LEN(scan) += STR_LEN(n);
3542 next = n + NODE_SZ_STR(n);
3543 /* Now we can overwrite *n : */
3544 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3552 #ifdef EXPERIMENTAL_INPLACESCAN
3553 if (flags && !NEXT_OFF(n)) {
3554 DEBUG_PEEP("atch", val, depth);
3555 if (reg_off_by_arg[OP(n)]) {
3556 ARG_SET(n, val - n);
3559 NEXT_OFF(n) = val - n;
3567 *unfolded_multi_char = FALSE;
3569 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3570 * can now analyze for sequences of problematic code points. (Prior to
3571 * this final joining, sequences could have been split over boundaries, and
3572 * hence missed). The sequences only happen in folding, hence for any
3573 * non-EXACT EXACTish node */
3574 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3575 U8* s0 = (U8*) STRING(scan);
3577 U8* s_end = s0 + STR_LEN(scan);
3579 int total_count_delta = 0; /* Total delta number of characters that
3580 multi-char folds expand to */
3582 /* One pass is made over the node's string looking for all the
3583 * possibilities. To avoid some tests in the loop, there are two main
3584 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3589 if (OP(scan) == EXACTFL) {
3592 /* An EXACTFL node would already have been changed to another
3593 * node type unless there is at least one character in it that
3594 * is problematic; likely a character whose fold definition
3595 * won't be known until runtime, and so has yet to be folded.
3596 * For all but the UTF-8 locale, folds are 1-1 in length, but
3597 * to handle the UTF-8 case, we need to create a temporary
3598 * folded copy using UTF-8 locale rules in order to analyze it.
3599 * This is because our macros that look to see if a sequence is
3600 * a multi-char fold assume everything is folded (otherwise the
3601 * tests in those macros would be too complicated and slow).
3602 * Note that here, the non-problematic folds will have already
3603 * been done, so we can just copy such characters. We actually
3604 * don't completely fold the EXACTFL string. We skip the
3605 * unfolded multi-char folds, as that would just create work
3606 * below to figure out the size they already are */
3608 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3611 STRLEN s_len = UTF8SKIP(s);
3612 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3613 Copy(s, d, s_len, U8);
3616 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3617 *unfolded_multi_char = TRUE;
3618 Copy(s, d, s_len, U8);
3621 else if (isASCII(*s)) {
3622 *(d++) = toFOLD(*s);
3626 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3632 /* Point the remainder of the routine to look at our temporary
3636 } /* End of creating folded copy of EXACTFL string */
3638 /* Examine the string for a multi-character fold sequence. UTF-8
3639 * patterns have all characters pre-folded by the time this code is
3641 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3642 length sequence we are looking for is 2 */
3644 int count = 0; /* How many characters in a multi-char fold */
3645 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3646 if (! len) { /* Not a multi-char fold: get next char */
3651 /* Nodes with 'ss' require special handling, except for
3652 * EXACTFA-ish for which there is no multi-char fold to this */
3653 if (len == 2 && *s == 's' && *(s+1) == 's'
3654 && OP(scan) != EXACTFA
3655 && OP(scan) != EXACTFA_NO_TRIE)
3658 if (OP(scan) != EXACTFL) {
3659 OP(scan) = EXACTFU_SS;
3663 else { /* Here is a generic multi-char fold. */
3664 U8* multi_end = s + len;
3666 /* Count how many characters are in it. In the case of
3667 * /aa, no folds which contain ASCII code points are
3668 * allowed, so check for those, and skip if found. */
3669 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3670 count = utf8_length(s, multi_end);
3674 while (s < multi_end) {
3677 goto next_iteration;
3687 /* The delta is how long the sequence is minus 1 (1 is how long
3688 * the character that folds to the sequence is) */
3689 total_count_delta += count - 1;
3693 /* We created a temporary folded copy of the string in EXACTFL
3694 * nodes. Therefore we need to be sure it doesn't go below zero,
3695 * as the real string could be shorter */
3696 if (OP(scan) == EXACTFL) {
3697 int total_chars = utf8_length((U8*) STRING(scan),
3698 (U8*) STRING(scan) + STR_LEN(scan));
3699 if (total_count_delta > total_chars) {
3700 total_count_delta = total_chars;
3704 *min_subtract += total_count_delta;
3707 else if (OP(scan) == EXACTFA) {
3709 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3710 * fold to the ASCII range (and there are no existing ones in the
3711 * upper latin1 range). But, as outlined in the comments preceding
3712 * this function, we need to flag any occurrences of the sharp s.
3713 * This character forbids trie formation (because of added
3715 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
3716 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
3717 || UNICODE_DOT_DOT_VERSION > 0)
3719 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3720 OP(scan) = EXACTFA_NO_TRIE;
3721 *unfolded_multi_char = TRUE;
3729 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3730 * folds that are all Latin1. As explained in the comments
3731 * preceding this function, we look also for the sharp s in EXACTF
3732 * and EXACTFL nodes; it can be in the final position. Otherwise
3733 * we can stop looking 1 byte earlier because have to find at least
3734 * two characters for a multi-fold */
3735 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3740 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3741 if (! len) { /* Not a multi-char fold. */
3742 if (*s == LATIN_SMALL_LETTER_SHARP_S
3743 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3745 *unfolded_multi_char = TRUE;
3752 && isALPHA_FOLD_EQ(*s, 's')
3753 && isALPHA_FOLD_EQ(*(s+1), 's'))
3756 /* EXACTF nodes need to know that the minimum length
3757 * changed so that a sharp s in the string can match this
3758 * ss in the pattern, but they remain EXACTF nodes, as they
3759 * won't match this unless the target string is is UTF-8,
3760 * which we don't know until runtime. EXACTFL nodes can't
3761 * transform into EXACTFU nodes */
3762 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
3763 OP(scan) = EXACTFU_SS;
3767 *min_subtract += len - 1;
3775 /* Allow dumping but overwriting the collection of skipped
3776 * ops and/or strings with fake optimized ops */
3777 n = scan + NODE_SZ_STR(scan);
3785 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3789 /* REx optimizer. Converts nodes into quicker variants "in place".
3790 Finds fixed substrings. */
3792 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3793 to the position after last scanned or to NULL. */
3795 #define INIT_AND_WITHP \
3796 assert(!and_withp); \
3797 Newx(and_withp,1, regnode_ssc); \
3798 SAVEFREEPV(and_withp)
3802 S_unwind_scan_frames(pTHX_ const void *p)
3804 scan_frame *f= (scan_frame *)p;
3806 scan_frame *n= f->next_frame;
3814 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3815 SSize_t *minlenp, SSize_t *deltap,
3820 regnode_ssc *and_withp,
3821 U32 flags, U32 depth)
3822 /* scanp: Start here (read-write). */
3823 /* deltap: Write maxlen-minlen here. */
3824 /* last: Stop before this one. */
3825 /* data: string data about the pattern */
3826 /* stopparen: treat close N as END */
3827 /* recursed: which subroutines have we recursed into */
3828 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3830 /* There must be at least this number of characters to match */
3833 regnode *scan = *scanp, *next;
3835 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3836 int is_inf_internal = 0; /* The studied chunk is infinite */
3837 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3838 scan_data_t data_fake;
3839 SV *re_trie_maxbuff = NULL;
3840 regnode *first_non_open = scan;
3841 SSize_t stopmin = SSize_t_MAX;
3842 scan_frame *frame = NULL;
3843 GET_RE_DEBUG_FLAGS_DECL;
3845 PERL_ARGS_ASSERT_STUDY_CHUNK;
3849 while (first_non_open && OP(first_non_open) == OPEN)
3850 first_non_open=regnext(first_non_open);
3856 RExC_study_chunk_recursed_count++;
3858 DEBUG_OPTIMISE_MORE_r(
3860 PerlIO_printf(Perl_debug_log,
3861 "%*sstudy_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
3862 (int)(depth*2), "", (long)stopparen,
3863 (unsigned long)RExC_study_chunk_recursed_count,
3864 (unsigned long)depth, (unsigned long)recursed_depth,
3867 if (recursed_depth) {
3870 for ( j = 0 ; j < recursed_depth ; j++ ) {
3871 for ( i = 0 ; i < (U32)RExC_npar ; i++ ) {
3873 PAREN_TEST(RExC_study_chunk_recursed +
3874 ( j * RExC_study_chunk_recursed_bytes), i )
3877 !PAREN_TEST(RExC_study_chunk_recursed +
3878 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
3881 PerlIO_printf(Perl_debug_log," %d",(int)i);
3885 if ( j + 1 < recursed_depth ) {
3886 PerlIO_printf(Perl_debug_log, ",");
3890 PerlIO_printf(Perl_debug_log,"\n");
3893 while ( scan && OP(scan) != END && scan < last ){
3894 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3895 node length to get a real minimum (because
3896 the folded version may be shorter) */
3897 bool unfolded_multi_char = FALSE;
3898 /* Peephole optimizer: */
3899 DEBUG_STUDYDATA("Peep:", data, depth);
3900 DEBUG_PEEP("Peep", scan, depth);
3903 /* The reason we do this here we need to deal with things like /(?:f)(?:o)(?:o)/
3904 * which cant be dealt with by the normal EXACT parsing code, as each (?:..) is handled
3905 * by a different invocation of reg() -- Yves
3907 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
3909 /* Follow the next-chain of the current node and optimize
3910 away all the NOTHINGs from it. */
3911 if (OP(scan) != CURLYX) {
3912 const int max = (reg_off_by_arg[OP(scan)]
3914 /* I32 may be smaller than U16 on CRAYs! */
3915 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3916 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3920 /* Skip NOTHING and LONGJMP. */
3921 while ((n = regnext(n))
3922 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3923 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3924 && off + noff < max)
3926 if (reg_off_by_arg[OP(scan)])
3929 NEXT_OFF(scan) = off;
3932 /* The principal pseudo-switch. Cannot be a switch, since we
3933 look into several different things. */
3934 if ( OP(scan) == DEFINEP ) {
3936 SSize_t deltanext = 0;
3937 SSize_t fake_last_close = 0;
3938 I32 f = SCF_IN_DEFINE;
3940 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3941 scan = regnext(scan);
3942 assert( OP(scan) == IFTHEN );
3943 DEBUG_PEEP("expect IFTHEN", scan, depth);
3945 data_fake.last_closep= &fake_last_close;
3947 next = regnext(scan);
3948 scan = NEXTOPER(NEXTOPER(scan));
3949 DEBUG_PEEP("scan", scan, depth);
3950 DEBUG_PEEP("next", next, depth);
3952 /* we suppose the run is continuous, last=next...
3953 * NOTE we dont use the return here! */
3954 (void)study_chunk(pRExC_state, &scan, &minlen,
3955 &deltanext, next, &data_fake, stopparen,
3956 recursed_depth, NULL, f, depth+1);
3961 OP(scan) == BRANCH ||
3962 OP(scan) == BRANCHJ ||
3965 next = regnext(scan);
3968 /* The op(next)==code check below is to see if we
3969 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
3970 * IFTHEN is special as it might not appear in&nb
https://perl5.git.perl.org / perl5.git / blob