5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
84 EXTERN_C const struct regexp_engine my_reg_engine;
89 #include "dquote_static.c"
90 #include "charclass_invlists.h"
91 #include "inline_invlist.c"
92 #include "unicode_constants.h"
94 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
95 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
96 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
97 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
100 #define STATIC static
104 struct RExC_state_t {
105 U32 flags; /* RXf_* are we folding, multilining? */
106 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
107 char *precomp; /* uncompiled string. */
108 REGEXP *rx_sv; /* The SV that is the regexp. */
109 regexp *rx; /* perl core regexp structure */
110 regexp_internal *rxi; /* internal data for regexp object
112 char *start; /* Start of input for compile */
113 char *end; /* End of input for compile */
114 char *parse; /* Input-scan pointer. */
115 SSize_t whilem_seen; /* number of WHILEM in this expr */
116 regnode *emit_start; /* Start of emitted-code area */
117 regnode *emit_bound; /* First regnode outside of the
119 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
120 implies compiling, so don't emit */
121 regnode_ssc emit_dummy; /* placeholder for emit to point to;
122 large enough for the largest
123 non-EXACTish node, so can use it as
125 I32 naughty; /* How bad is this pattern? */
126 I32 sawback; /* Did we see \1, ...? */
128 SSize_t size; /* Code size. */
129 I32 npar; /* Capture buffer count, (OPEN) plus
130 one. ("par" 0 is the whole
132 I32 nestroot; /* root parens we are in - used by
136 regnode **open_parens; /* pointers to open parens */
137 regnode **close_parens; /* pointers to close parens */
138 regnode *opend; /* END node in program */
139 I32 utf8; /* whether the pattern is utf8 or not */
140 I32 orig_utf8; /* whether the pattern was originally in utf8 */
141 /* XXX use this for future optimisation of case
142 * where pattern must be upgraded to utf8. */
143 I32 uni_semantics; /* If a d charset modifier should use unicode
144 rules, even if the pattern is not in
146 HV *paren_names; /* Paren names */
148 regnode **recurse; /* Recurse regops */
149 I32 recurse_count; /* Number of recurse regops */
150 U8 *study_chunk_recursed; /* bitmap of which parens we have moved
152 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
156 I32 override_recoding;
157 I32 in_multi_char_class;
158 struct reg_code_block *code_blocks; /* positions of literal (?{})
160 int num_code_blocks; /* size of code_blocks[] */
161 int code_index; /* next code_blocks[] slot */
162 SSize_t maxlen; /* mininum possible number of chars in string to match */
163 #ifdef ADD_TO_REGEXEC
164 char *starttry; /* -Dr: where regtry was called. */
165 #define RExC_starttry (pRExC_state->starttry)
167 SV *runtime_code_qr; /* qr with the runtime code blocks */
169 const char *lastparse;
171 AV *paren_name_list; /* idx -> name */
172 #define RExC_lastparse (pRExC_state->lastparse)
173 #define RExC_lastnum (pRExC_state->lastnum)
174 #define RExC_paren_name_list (pRExC_state->paren_name_list)
178 #define RExC_flags (pRExC_state->flags)
179 #define RExC_pm_flags (pRExC_state->pm_flags)
180 #define RExC_precomp (pRExC_state->precomp)
181 #define RExC_rx_sv (pRExC_state->rx_sv)
182 #define RExC_rx (pRExC_state->rx)
183 #define RExC_rxi (pRExC_state->rxi)
184 #define RExC_start (pRExC_state->start)
185 #define RExC_end (pRExC_state->end)
186 #define RExC_parse (pRExC_state->parse)
187 #define RExC_whilem_seen (pRExC_state->whilem_seen)
188 #ifdef RE_TRACK_PATTERN_OFFSETS
189 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
192 #define RExC_emit (pRExC_state->emit)
193 #define RExC_emit_dummy (pRExC_state->emit_dummy)
194 #define RExC_emit_start (pRExC_state->emit_start)
195 #define RExC_emit_bound (pRExC_state->emit_bound)
196 #define RExC_naughty (pRExC_state->naughty)
197 #define RExC_sawback (pRExC_state->sawback)
198 #define RExC_seen (pRExC_state->seen)
199 #define RExC_size (pRExC_state->size)
200 #define RExC_maxlen (pRExC_state->maxlen)
201 #define RExC_npar (pRExC_state->npar)
202 #define RExC_nestroot (pRExC_state->nestroot)
203 #define RExC_extralen (pRExC_state->extralen)
204 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
205 #define RExC_utf8 (pRExC_state->utf8)
206 #define RExC_uni_semantics (pRExC_state->uni_semantics)
207 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
208 #define RExC_open_parens (pRExC_state->open_parens)
209 #define RExC_close_parens (pRExC_state->close_parens)
210 #define RExC_opend (pRExC_state->opend)
211 #define RExC_paren_names (pRExC_state->paren_names)
212 #define RExC_recurse (pRExC_state->recurse)
213 #define RExC_recurse_count (pRExC_state->recurse_count)
214 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
215 #define RExC_study_chunk_recursed_bytes \
216 (pRExC_state->study_chunk_recursed_bytes)
217 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
218 #define RExC_contains_locale (pRExC_state->contains_locale)
219 #define RExC_contains_i (pRExC_state->contains_i)
220 #define RExC_override_recoding (pRExC_state->override_recoding)
221 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
224 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
225 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
226 ((*s) == '{' && regcurly(s, FALSE)))
229 * Flags to be passed up and down.
231 #define WORST 0 /* Worst case. */
232 #define HASWIDTH 0x01 /* Known to match non-null strings. */
234 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
235 * character. (There needs to be a case: in the switch statement in regexec.c
236 * for any node marked SIMPLE.) Note that this is not the same thing as
239 #define SPSTART 0x04 /* Starts with * or + */
240 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
241 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
242 #define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
244 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
246 /* whether trie related optimizations are enabled */
247 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
248 #define TRIE_STUDY_OPT
249 #define FULL_TRIE_STUDY
255 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
256 #define PBITVAL(paren) (1 << ((paren) & 7))
257 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
258 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
259 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
261 #define REQUIRE_UTF8 STMT_START { \
263 *flagp = RESTART_UTF8; \
268 /* This converts the named class defined in regcomp.h to its equivalent class
269 * number defined in handy.h. */
270 #define namedclass_to_classnum(class) ((int) ((class) / 2))
271 #define classnum_to_namedclass(classnum) ((classnum) * 2)
273 #define _invlist_union_complement_2nd(a, b, output) \
274 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
275 #define _invlist_intersection_complement_2nd(a, b, output) \
276 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
278 /* About scan_data_t.
280 During optimisation we recurse through the regexp program performing
281 various inplace (keyhole style) optimisations. In addition study_chunk
282 and scan_commit populate this data structure with information about
283 what strings MUST appear in the pattern. We look for the longest
284 string that must appear at a fixed location, and we look for the
285 longest string that may appear at a floating location. So for instance
290 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
291 strings (because they follow a .* construct). study_chunk will identify
292 both FOO and BAR as being the longest fixed and floating strings respectively.
294 The strings can be composites, for instance
298 will result in a composite fixed substring 'foo'.
300 For each string some basic information is maintained:
302 - offset or min_offset
303 This is the position the string must appear at, or not before.
304 It also implicitly (when combined with minlenp) tells us how many
305 characters must match before the string we are searching for.
306 Likewise when combined with minlenp and the length of the string it
307 tells us how many characters must appear after the string we have
311 Only used for floating strings. This is the rightmost point that
312 the string can appear at. If set to SSize_t_MAX it indicates that the
313 string can occur infinitely far to the right.
316 A pointer to the minimum number of characters of the pattern that the
317 string was found inside. This is important as in the case of positive
318 lookahead or positive lookbehind we can have multiple patterns
323 The minimum length of the pattern overall is 3, the minimum length
324 of the lookahead part is 3, but the minimum length of the part that
325 will actually match is 1. So 'FOO's minimum length is 3, but the
326 minimum length for the F is 1. This is important as the minimum length
327 is used to determine offsets in front of and behind the string being
328 looked for. Since strings can be composites this is the length of the
329 pattern at the time it was committed with a scan_commit. Note that
330 the length is calculated by study_chunk, so that the minimum lengths
331 are not known until the full pattern has been compiled, thus the
332 pointer to the value.
336 In the case of lookbehind the string being searched for can be
337 offset past the start point of the final matching string.
338 If this value was just blithely removed from the min_offset it would
339 invalidate some of the calculations for how many chars must match
340 before or after (as they are derived from min_offset and minlen and
341 the length of the string being searched for).
342 When the final pattern is compiled and the data is moved from the
343 scan_data_t structure into the regexp structure the information
344 about lookbehind is factored in, with the information that would
345 have been lost precalculated in the end_shift field for the
348 The fields pos_min and pos_delta are used to store the minimum offset
349 and the delta to the maximum offset at the current point in the pattern.
353 typedef struct scan_data_t {
354 /*I32 len_min; unused */
355 /*I32 len_delta; unused */
359 SSize_t last_end; /* min value, <0 unless valid. */
360 SSize_t last_start_min;
361 SSize_t last_start_max;
362 SV **longest; /* Either &l_fixed, or &l_float. */
363 SV *longest_fixed; /* longest fixed string found in pattern */
364 SSize_t offset_fixed; /* offset where it starts */
365 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
366 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
367 SV *longest_float; /* longest floating string found in pattern */
368 SSize_t offset_float_min; /* earliest point in string it can appear */
369 SSize_t offset_float_max; /* latest point in string it can appear */
370 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
371 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
374 SSize_t *last_closep;
375 regnode_ssc *start_class;
378 /* The below is perhaps overboard, but this allows us to save a test at the
379 * expense of a mask. This is because on both EBCDIC and ASCII machines, 'A'
380 * and 'a' differ by a single bit; the same with the upper and lower case of
381 * all other ASCII-range alphabetics. On ASCII platforms, they are 32 apart;
382 * on EBCDIC, they are 64. This uses an exclusive 'or' to find that bit and
383 * then inverts it to form a mask, with just a single 0, in the bit position
384 * where the upper- and lowercase differ. XXX There are about 40 other
385 * instances in the Perl core where this micro-optimization could be used.
386 * Should decide if maintenance cost is worse, before changing those
388 * Returns a boolean as to whether or not 'v' is either a lowercase or
389 * uppercase instance of 'c', where 'c' is in [A-Za-z]. If 'c' is a
390 * compile-time constant, the generated code is better than some optimizing
391 * compilers figure out, amounting to a mask and test. The results are
392 * meaningless if 'c' is not one of [A-Za-z] */
393 #define isARG2_lower_or_UPPER_ARG1(c, v) \
394 (((v) & ~('A' ^ 'a')) == ((c) & ~('A' ^ 'a')))
397 * Forward declarations for pregcomp()'s friends.
400 static const scan_data_t zero_scan_data =
401 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
403 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
404 #define SF_BEFORE_SEOL 0x0001
405 #define SF_BEFORE_MEOL 0x0002
406 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
407 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
409 #define SF_FIX_SHIFT_EOL (+2)
410 #define SF_FL_SHIFT_EOL (+4)
412 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
413 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
415 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
416 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
417 #define SF_IS_INF 0x0040
418 #define SF_HAS_PAR 0x0080
419 #define SF_IN_PAR 0x0100
420 #define SF_HAS_EVAL 0x0200
421 #define SCF_DO_SUBSTR 0x0400
422 #define SCF_DO_STCLASS_AND 0x0800
423 #define SCF_DO_STCLASS_OR 0x1000
424 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
425 #define SCF_WHILEM_VISITED_POS 0x2000
427 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
428 #define SCF_SEEN_ACCEPT 0x8000
429 #define SCF_TRIE_DOING_RESTUDY 0x10000
431 #define UTF cBOOL(RExC_utf8)
433 /* The enums for all these are ordered so things work out correctly */
434 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
435 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
436 == REGEX_DEPENDS_CHARSET)
437 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
438 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
439 >= REGEX_UNICODE_CHARSET)
440 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
441 == REGEX_ASCII_RESTRICTED_CHARSET)
442 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
443 >= REGEX_ASCII_RESTRICTED_CHARSET)
444 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
445 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
447 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
449 /* For programs that want to be strictly Unicode compatible by dying if any
450 * attempt is made to match a non-Unicode code point against a Unicode
452 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
454 #define OOB_NAMEDCLASS -1
456 /* There is no code point that is out-of-bounds, so this is problematic. But
457 * its only current use is to initialize a variable that is always set before
459 #define OOB_UNICODE 0xDEADBEEF
461 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
462 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
465 /* length of regex to show in messages that don't mark a position within */
466 #define RegexLengthToShowInErrorMessages 127
469 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
470 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
471 * op/pragma/warn/regcomp.
473 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
474 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
476 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
477 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
479 #define REPORT_LOCATION_ARGS(offset) \
480 UTF8fARG(UTF, offset, RExC_precomp), \
481 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
484 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
485 * arg. Show regex, up to a maximum length. If it's too long, chop and add
488 #define _FAIL(code) STMT_START { \
489 const char *ellipses = ""; \
490 IV len = RExC_end - RExC_precomp; \
493 SAVEFREESV(RExC_rx_sv); \
494 if (len > RegexLengthToShowInErrorMessages) { \
495 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
496 len = RegexLengthToShowInErrorMessages - 10; \
502 #define FAIL(msg) _FAIL( \
503 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
504 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
506 #define FAIL2(msg,arg) _FAIL( \
507 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
508 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
511 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
513 #define Simple_vFAIL(m) STMT_START { \
514 const IV offset = RExC_parse - RExC_precomp; \
515 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
516 m, REPORT_LOCATION_ARGS(offset)); \
520 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
522 #define vFAIL(m) STMT_START { \
524 SAVEFREESV(RExC_rx_sv); \
529 * Like Simple_vFAIL(), but accepts two arguments.
531 #define Simple_vFAIL2(m,a1) STMT_START { \
532 const IV offset = RExC_parse - RExC_precomp; \
533 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
534 REPORT_LOCATION_ARGS(offset)); \
538 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
540 #define vFAIL2(m,a1) STMT_START { \
542 SAVEFREESV(RExC_rx_sv); \
543 Simple_vFAIL2(m, a1); \
548 * Like Simple_vFAIL(), but accepts three arguments.
550 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
551 const IV offset = RExC_parse - RExC_precomp; \
552 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
553 REPORT_LOCATION_ARGS(offset)); \
557 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
559 #define vFAIL3(m,a1,a2) STMT_START { \
561 SAVEFREESV(RExC_rx_sv); \
562 Simple_vFAIL3(m, a1, a2); \
566 * Like Simple_vFAIL(), but accepts four arguments.
568 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
569 const IV offset = RExC_parse - RExC_precomp; \
570 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
571 REPORT_LOCATION_ARGS(offset)); \
574 #define vFAIL4(m,a1,a2,a3) STMT_START { \
576 SAVEFREESV(RExC_rx_sv); \
577 Simple_vFAIL4(m, a1, a2, a3); \
580 /* A specialized version of vFAIL2 that works with UTF8f */
581 #define vFAIL2utf8f(m, a1) STMT_START { \
582 const IV offset = RExC_parse - RExC_precomp; \
584 SAVEFREESV(RExC_rx_sv); \
585 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
586 REPORT_LOCATION_ARGS(offset)); \
590 /* m is not necessarily a "literal string", in this macro */
591 #define reg_warn_non_literal_string(loc, m) STMT_START { \
592 const IV offset = loc - RExC_precomp; \
593 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
594 m, REPORT_LOCATION_ARGS(offset)); \
597 #define ckWARNreg(loc,m) STMT_START { \
598 const IV offset = loc - RExC_precomp; \
599 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
600 REPORT_LOCATION_ARGS(offset)); \
603 #define vWARN_dep(loc, m) STMT_START { \
604 const IV offset = loc - RExC_precomp; \
605 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
606 REPORT_LOCATION_ARGS(offset)); \
609 #define ckWARNdep(loc,m) STMT_START { \
610 const IV offset = loc - RExC_precomp; \
611 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
613 REPORT_LOCATION_ARGS(offset)); \
616 #define ckWARNregdep(loc,m) STMT_START { \
617 const IV offset = loc - RExC_precomp; \
618 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
620 REPORT_LOCATION_ARGS(offset)); \
623 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
624 const IV offset = loc - RExC_precomp; \
625 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
627 a1, REPORT_LOCATION_ARGS(offset)); \
630 #define ckWARN2reg(loc, m, a1) STMT_START { \
631 const IV offset = loc - RExC_precomp; \
632 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
633 a1, REPORT_LOCATION_ARGS(offset)); \
636 #define vWARN3(loc, m, a1, a2) STMT_START { \
637 const IV offset = loc - RExC_precomp; \
638 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
639 a1, a2, REPORT_LOCATION_ARGS(offset)); \
642 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
643 const IV offset = loc - RExC_precomp; \
644 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
645 a1, a2, REPORT_LOCATION_ARGS(offset)); \
648 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
649 const IV offset = loc - RExC_precomp; \
650 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
651 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
654 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
655 const IV offset = loc - RExC_precomp; \
656 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
657 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
660 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
661 const IV offset = loc - RExC_precomp; \
662 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
663 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
667 /* Allow for side effects in s */
668 #define REGC(c,s) STMT_START { \
669 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
672 /* Macros for recording node offsets. 20001227 mjd@plover.com
673 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
674 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
675 * Element 0 holds the number n.
676 * Position is 1 indexed.
678 #ifndef RE_TRACK_PATTERN_OFFSETS
679 #define Set_Node_Offset_To_R(node,byte)
680 #define Set_Node_Offset(node,byte)
681 #define Set_Cur_Node_Offset
682 #define Set_Node_Length_To_R(node,len)
683 #define Set_Node_Length(node,len)
684 #define Set_Node_Cur_Length(node,start)
685 #define Node_Offset(n)
686 #define Node_Length(n)
687 #define Set_Node_Offset_Length(node,offset,len)
688 #define ProgLen(ri) ri->u.proglen
689 #define SetProgLen(ri,x) ri->u.proglen = x
691 #define ProgLen(ri) ri->u.offsets[0]
692 #define SetProgLen(ri,x) ri->u.offsets[0] = x
693 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
695 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
696 __LINE__, (int)(node), (int)(byte))); \
698 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
701 RExC_offsets[2*(node)-1] = (byte); \
706 #define Set_Node_Offset(node,byte) \
707 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
708 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
710 #define Set_Node_Length_To_R(node,len) STMT_START { \
712 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
713 __LINE__, (int)(node), (int)(len))); \
715 Perl_croak(aTHX_ "value of node is %d in Length macro", \
718 RExC_offsets[2*(node)] = (len); \
723 #define Set_Node_Length(node,len) \
724 Set_Node_Length_To_R((node)-RExC_emit_start, len)
725 #define Set_Node_Cur_Length(node, start) \
726 Set_Node_Length(node, RExC_parse - start)
728 /* Get offsets and lengths */
729 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
730 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
732 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
733 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
734 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
738 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
739 #define EXPERIMENTAL_INPLACESCAN
740 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
742 #define DEBUG_RExC_seen() \
743 DEBUG_OPTIMISE_MORE_r({ \
744 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
746 if (RExC_seen & REG_ZERO_LEN_SEEN) \
747 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
749 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
750 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
752 if (RExC_seen & REG_GPOS_SEEN) \
753 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
755 if (RExC_seen & REG_CANY_SEEN) \
756 PerlIO_printf(Perl_debug_log,"REG_CANY_SEEN "); \
758 if (RExC_seen & REG_RECURSE_SEEN) \
759 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
761 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
762 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
764 if (RExC_seen & REG_VERBARG_SEEN) \
765 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
767 if (RExC_seen & REG_CUTGROUP_SEEN) \
768 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
770 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
771 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
773 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
774 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
776 if (RExC_seen & REG_GOSTART_SEEN) \
777 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
779 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
780 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
782 PerlIO_printf(Perl_debug_log,"\n"); \
785 #define DEBUG_STUDYDATA(str,data,depth) \
786 DEBUG_OPTIMISE_MORE_r(if(data){ \
787 PerlIO_printf(Perl_debug_log, \
788 "%*s" str "Pos:%"IVdf"/%"IVdf \
789 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
790 (int)(depth)*2, "", \
791 (IV)((data)->pos_min), \
792 (IV)((data)->pos_delta), \
793 (UV)((data)->flags), \
794 (IV)((data)->whilem_c), \
795 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
796 is_inf ? "INF " : "" \
798 if ((data)->last_found) \
799 PerlIO_printf(Perl_debug_log, \
800 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
801 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
802 SvPVX_const((data)->last_found), \
803 (IV)((data)->last_end), \
804 (IV)((data)->last_start_min), \
805 (IV)((data)->last_start_max), \
806 ((data)->longest && \
807 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
808 SvPVX_const((data)->longest_fixed), \
809 (IV)((data)->offset_fixed), \
810 ((data)->longest && \
811 (data)->longest==&((data)->longest_float)) ? "*" : "", \
812 SvPVX_const((data)->longest_float), \
813 (IV)((data)->offset_float_min), \
814 (IV)((data)->offset_float_max) \
816 PerlIO_printf(Perl_debug_log,"\n"); \
819 /* Mark that we cannot extend a found fixed substring at this point.
820 Update the longest found anchored substring and the longest found
821 floating substrings if needed. */
824 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
825 SSize_t *minlenp, int is_inf)
827 const STRLEN l = CHR_SVLEN(data->last_found);
828 const STRLEN old_l = CHR_SVLEN(*data->longest);
829 GET_RE_DEBUG_FLAGS_DECL;
831 PERL_ARGS_ASSERT_SCAN_COMMIT;
833 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
834 SvSetMagicSV(*data->longest, data->last_found);
835 if (*data->longest == data->longest_fixed) {
836 data->offset_fixed = l ? data->last_start_min : data->pos_min;
837 if (data->flags & SF_BEFORE_EOL)
839 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
841 data->flags &= ~SF_FIX_BEFORE_EOL;
842 data->minlen_fixed=minlenp;
843 data->lookbehind_fixed=0;
845 else { /* *data->longest == data->longest_float */
846 data->offset_float_min = l ? data->last_start_min : data->pos_min;
847 data->offset_float_max = (l
848 ? data->last_start_max
849 : (data->pos_delta == SSize_t_MAX
851 : data->pos_min + data->pos_delta));
853 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
854 data->offset_float_max = SSize_t_MAX;
855 if (data->flags & SF_BEFORE_EOL)
857 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
859 data->flags &= ~SF_FL_BEFORE_EOL;
860 data->minlen_float=minlenp;
861 data->lookbehind_float=0;
864 SvCUR_set(data->last_found, 0);
866 SV * const sv = data->last_found;
867 if (SvUTF8(sv) && SvMAGICAL(sv)) {
868 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
874 data->flags &= ~SF_BEFORE_EOL;
875 DEBUG_STUDYDATA("commit: ",data,0);
878 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
879 * list that describes which code points it matches */
882 S_ssc_anything(pTHX_ regnode_ssc *ssc)
884 /* Set the SSC 'ssc' to match an empty string or any code point */
886 PERL_ARGS_ASSERT_SSC_ANYTHING;
888 assert(is_ANYOF_SYNTHETIC(ssc));
890 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
891 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
892 ANYOF_FLAGS(ssc) |= ANYOF_EMPTY_STRING; /* Plus match empty string */
896 S_ssc_is_anything(pTHX_ const regnode_ssc *ssc)
898 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
899 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
900 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
901 * in any way, so there's no point in using it */
906 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
908 assert(is_ANYOF_SYNTHETIC(ssc));
910 if (! (ANYOF_FLAGS(ssc) & ANYOF_EMPTY_STRING)) {
914 /* See if the list consists solely of the range 0 - Infinity */
915 invlist_iterinit(ssc->invlist);
916 ret = invlist_iternext(ssc->invlist, &start, &end)
920 invlist_iterfinish(ssc->invlist);
926 /* If e.g., both \w and \W are set, matches everything */
927 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
929 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
930 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
940 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
942 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
943 * string, any code point, or any posix class under locale */
945 PERL_ARGS_ASSERT_SSC_INIT;
947 Zero(ssc, 1, regnode_ssc);
948 set_ANYOF_SYNTHETIC(ssc);
949 ARG_SET(ssc, ANYOF_NONBITMAP_EMPTY);
952 /* If any portion of the regex is to operate under locale rules,
953 * initialization includes it. The reason this isn't done for all regexes
954 * is that the optimizer was written under the assumption that locale was
955 * all-or-nothing. Given the complexity and lack of documentation in the
956 * optimizer, and that there are inadequate test cases for locale, many
957 * parts of it may not work properly, it is safest to avoid locale unless
959 if (RExC_contains_locale) {
960 ANYOF_POSIXL_SETALL(ssc);
963 ANYOF_POSIXL_ZERO(ssc);
968 S_ssc_is_cp_posixl_init(pTHX_ const RExC_state_t *pRExC_state,
969 const regnode_ssc *ssc)
971 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
972 * to the list of code points matched, and locale posix classes; hence does
973 * not check its flags) */
978 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
980 assert(is_ANYOF_SYNTHETIC(ssc));
982 invlist_iterinit(ssc->invlist);
983 ret = invlist_iternext(ssc->invlist, &start, &end)
987 invlist_iterfinish(ssc->invlist);
993 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1001 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1002 const regnode_charclass* const node)
1004 /* Returns a mortal inversion list defining which code points are matched
1005 * by 'node', which is of type ANYOF. Handles complementing the result if
1006 * appropriate. If some code points aren't knowable at this time, the
1007 * returned list must, and will, contain every code point that is a
1010 SV* invlist = sv_2mortal(_new_invlist(0));
1011 SV* only_utf8_locale_invlist = NULL;
1013 const U32 n = ARG(node);
1014 bool new_node_has_latin1 = FALSE;
1016 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1018 /* Look at the data structure created by S_set_ANYOF_arg() */
1019 if (n != ANYOF_NONBITMAP_EMPTY) {
1020 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1021 AV * const av = MUTABLE_AV(SvRV(rv));
1022 SV **const ary = AvARRAY(av);
1023 assert(RExC_rxi->data->what[n] == 's');
1025 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1026 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1028 else if (ary[0] && ary[0] != &PL_sv_undef) {
1030 /* Here, no compile-time swash, and there are things that won't be
1031 * known until runtime -- we have to assume it could be anything */
1032 return _add_range_to_invlist(invlist, 0, UV_MAX);
1034 else if (ary[3] && ary[3] != &PL_sv_undef) {
1036 /* Here no compile-time swash, and no run-time only data. Use the
1037 * node's inversion list */
1038 invlist = sv_2mortal(invlist_clone(ary[3]));
1041 /* Get the code points valid only under UTF-8 locales */
1042 if ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
1043 && ary[2] && ary[2] != &PL_sv_undef)
1045 only_utf8_locale_invlist = ary[2];
1049 /* An ANYOF node contains a bitmap for the first 256 code points, and an
1050 * inversion list for the others, but if there are code points that should
1051 * match only conditionally on the target string being UTF-8, those are
1052 * placed in the inversion list, and not the bitmap. Since there are
1053 * circumstances under which they could match, they are included in the
1054 * SSC. But if the ANYOF node is to be inverted, we have to exclude them
1055 * here, so that when we invert below, the end result actually does include
1056 * them. (Think about "\xe0" =~ /[^\xc0]/di;). We have to do this here
1057 * before we add the unconditionally matched code points */
1058 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1059 _invlist_intersection_complement_2nd(invlist,
1064 /* Add in the points from the bit map */
1065 for (i = 0; i < 256; i++) {
1066 if (ANYOF_BITMAP_TEST(node, i)) {
1067 invlist = add_cp_to_invlist(invlist, i);
1068 new_node_has_latin1 = TRUE;
1072 /* If this can match all upper Latin1 code points, have to add them
1074 if (ANYOF_FLAGS(node) & ANYOF_NON_UTF8_NON_ASCII_ALL) {
1075 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1078 /* Similarly for these */
1079 if (ANYOF_FLAGS(node) & ANYOF_ABOVE_LATIN1_ALL) {
1080 invlist = _add_range_to_invlist(invlist, 256, UV_MAX);
1083 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1084 _invlist_invert(invlist);
1086 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1088 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1089 * locale. We can skip this if there are no 0-255 at all. */
1090 _invlist_union(invlist, PL_Latin1, &invlist);
1093 /* Similarly add the UTF-8 locale possible matches. These have to be
1094 * deferred until after the non-UTF-8 locale ones are taken care of just
1095 * above, or it leads to wrong results under ANYOF_INVERT */
1096 if (only_utf8_locale_invlist) {
1097 _invlist_union_maybe_complement_2nd(invlist,
1098 only_utf8_locale_invlist,
1099 ANYOF_FLAGS(node) & ANYOF_INVERT,
1106 /* These two functions currently do the exact same thing */
1107 #define ssc_init_zero ssc_init
1109 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1110 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1112 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1113 * should not be inverted. 'and_with->flags & ANYOF_POSIXL' should be 0 if
1114 * 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1117 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1118 const regnode_charclass *and_with)
1120 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1121 * another SSC or a regular ANYOF class. Can create false positives. */
1126 PERL_ARGS_ASSERT_SSC_AND;
1128 assert(is_ANYOF_SYNTHETIC(ssc));
1130 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1131 * the code point inversion list and just the relevant flags */
1132 if (is_ANYOF_SYNTHETIC(and_with)) {
1133 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1134 anded_flags = ANYOF_FLAGS(and_with);
1136 /* XXX This is a kludge around what appears to be deficiencies in the
1137 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1138 * there are paths through the optimizer where it doesn't get weeded
1139 * out when it should. And if we don't make some extra provision for
1140 * it like the code just below, it doesn't get added when it should.
1141 * This solution is to add it only when AND'ing, which is here, and
1142 * only when what is being AND'ed is the pristine, original node
1143 * matching anything. Thus it is like adding it to ssc_anything() but
1144 * only when the result is to be AND'ed. Probably the same solution
1145 * could be adopted for the same problem we have with /l matching,
1146 * which is solved differently in S_ssc_init(), and that would lead to
1147 * fewer false positives than that solution has. But if this solution
1148 * creates bugs, the consequences are only that a warning isn't raised
1149 * that should be; while the consequences for having /l bugs is
1150 * incorrect matches */
1151 if (ssc_is_anything((regnode_ssc *)and_with)) {
1152 anded_flags |= ANYOF_WARN_SUPER;
1156 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1157 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1160 ANYOF_FLAGS(ssc) &= anded_flags;
1162 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1163 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1164 * 'and_with' may be inverted. When not inverted, we have the situation of
1166 * (C1 | P1) & (C2 | P2)
1167 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1168 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1169 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1170 * <= ((C1 & C2) | P1 | P2)
1171 * Alternatively, the last few steps could be:
1172 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1173 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1174 * <= (C1 | C2 | (P1 & P2))
1175 * We favor the second approach if either P1 or P2 is non-empty. This is
1176 * because these components are a barrier to doing optimizations, as what
1177 * they match cannot be known until the moment of matching as they are
1178 * dependent on the current locale, 'AND"ing them likely will reduce or
1180 * But we can do better if we know that C1,P1 are in their initial state (a
1181 * frequent occurrence), each matching everything:
1182 * (<everything>) & (C2 | P2) = C2 | P2
1183 * Similarly, if C2,P2 are in their initial state (again a frequent
1184 * occurrence), the result is a no-op
1185 * (C1 | P1) & (<everything>) = C1 | P1
1188 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1189 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1190 * <= (C1 & ~C2) | (P1 & ~P2)
1193 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1194 && ! is_ANYOF_SYNTHETIC(and_with))
1198 ssc_intersection(ssc,
1200 FALSE /* Has already been inverted */
1203 /* If either P1 or P2 is empty, the intersection will be also; can skip
1205 if (! (ANYOF_FLAGS(and_with) & ANYOF_POSIXL)) {
1206 ANYOF_POSIXL_ZERO(ssc);
1208 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1210 /* Note that the Posix class component P from 'and_with' actually
1212 * P = Pa | Pb | ... | Pn
1213 * where each component is one posix class, such as in [\w\s].
1215 * ~P = ~(Pa | Pb | ... | Pn)
1216 * = ~Pa & ~Pb & ... & ~Pn
1217 * <= ~Pa | ~Pb | ... | ~Pn
1218 * The last is something we can easily calculate, but unfortunately
1219 * is likely to have many false positives. We could do better
1220 * in some (but certainly not all) instances if two classes in
1221 * P have known relationships. For example
1222 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1224 * :lower: & :print: = :lower:
1225 * And similarly for classes that must be disjoint. For example,
1226 * since \s and \w can have no elements in common based on rules in
1227 * the POSIX standard,
1228 * \w & ^\S = nothing
1229 * Unfortunately, some vendor locales do not meet the Posix
1230 * standard, in particular almost everything by Microsoft.
1231 * The loop below just changes e.g., \w into \W and vice versa */
1233 regnode_charclass_posixl temp;
1234 int add = 1; /* To calculate the index of the complement */
1236 ANYOF_POSIXL_ZERO(&temp);
1237 for (i = 0; i < ANYOF_MAX; i++) {
1239 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1240 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1242 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1243 ANYOF_POSIXL_SET(&temp, i + add);
1245 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1247 ANYOF_POSIXL_AND(&temp, ssc);
1249 } /* else ssc already has no posixes */
1250 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1251 in its initial state */
1252 else if (! is_ANYOF_SYNTHETIC(and_with)
1253 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1255 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1256 * copy it over 'ssc' */
1257 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1258 if (is_ANYOF_SYNTHETIC(and_with)) {
1259 StructCopy(and_with, ssc, regnode_ssc);
1262 ssc->invlist = anded_cp_list;
1263 ANYOF_POSIXL_ZERO(ssc);
1264 if (ANYOF_FLAGS(and_with) & ANYOF_POSIXL) {
1265 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1269 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1270 || (ANYOF_FLAGS(and_with) & ANYOF_POSIXL))
1272 /* One or the other of P1, P2 is non-empty. */
1273 if (ANYOF_FLAGS(and_with) & ANYOF_POSIXL) {
1274 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1276 ssc_union(ssc, anded_cp_list, FALSE);
1278 else { /* P1 = P2 = empty */
1279 ssc_intersection(ssc, anded_cp_list, FALSE);
1285 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1286 const regnode_charclass *or_with)
1288 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1289 * another SSC or a regular ANYOF class. Can create false positives if
1290 * 'or_with' is to be inverted. */
1295 PERL_ARGS_ASSERT_SSC_OR;
1297 assert(is_ANYOF_SYNTHETIC(ssc));
1299 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1300 * the code point inversion list and just the relevant flags */
1301 if (is_ANYOF_SYNTHETIC(or_with)) {
1302 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1303 ored_flags = ANYOF_FLAGS(or_with);
1306 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1307 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1310 ANYOF_FLAGS(ssc) |= ored_flags;
1312 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1313 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1314 * 'or_with' may be inverted. When not inverted, we have the simple
1315 * situation of computing:
1316 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1317 * If P1|P2 yields a situation with both a class and its complement are
1318 * set, like having both \w and \W, this matches all code points, and we
1319 * can delete these from the P component of the ssc going forward. XXX We
1320 * might be able to delete all the P components, but I (khw) am not certain
1321 * about this, and it is better to be safe.
1324 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1325 * <= (C1 | P1) | ~C2
1326 * <= (C1 | ~C2) | P1
1327 * (which results in actually simpler code than the non-inverted case)
1330 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1331 && ! is_ANYOF_SYNTHETIC(or_with))
1333 /* We ignore P2, leaving P1 going forward */
1334 } /* else Not inverted */
1335 else if (ANYOF_FLAGS(or_with) & ANYOF_POSIXL) {
1336 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1337 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1339 for (i = 0; i < ANYOF_MAX; i += 2) {
1340 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1342 ssc_match_all_cp(ssc);
1343 ANYOF_POSIXL_CLEAR(ssc, i);
1344 ANYOF_POSIXL_CLEAR(ssc, i+1);
1352 FALSE /* Already has been inverted */
1356 PERL_STATIC_INLINE void
1357 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1359 PERL_ARGS_ASSERT_SSC_UNION;
1361 assert(is_ANYOF_SYNTHETIC(ssc));
1363 _invlist_union_maybe_complement_2nd(ssc->invlist,
1369 PERL_STATIC_INLINE void
1370 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1372 const bool invert2nd)
1374 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1376 assert(is_ANYOF_SYNTHETIC(ssc));
1378 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1384 PERL_STATIC_INLINE void
1385 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1387 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1389 assert(is_ANYOF_SYNTHETIC(ssc));
1391 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1394 PERL_STATIC_INLINE void
1395 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1397 /* AND just the single code point 'cp' into the SSC 'ssc' */
1399 SV* cp_list = _new_invlist(2);
1401 PERL_ARGS_ASSERT_SSC_CP_AND;
1403 assert(is_ANYOF_SYNTHETIC(ssc));
1405 cp_list = add_cp_to_invlist(cp_list, cp);
1406 ssc_intersection(ssc, cp_list,
1407 FALSE /* Not inverted */
1409 SvREFCNT_dec_NN(cp_list);
1412 PERL_STATIC_INLINE void
1413 S_ssc_clear_locale(pTHX_ regnode_ssc *ssc)
1415 /* Set the SSC 'ssc' to not match any locale things */
1417 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1419 assert(is_ANYOF_SYNTHETIC(ssc));
1421 ANYOF_POSIXL_ZERO(ssc);
1422 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1426 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1428 /* The inversion list in the SSC is marked mortal; now we need a more
1429 * permanent copy, which is stored the same way that is done in a regular
1430 * ANYOF node, with the first 256 code points in a bit map */
1432 SV* invlist = invlist_clone(ssc->invlist);
1434 PERL_ARGS_ASSERT_SSC_FINALIZE;
1436 assert(is_ANYOF_SYNTHETIC(ssc));
1438 /* The code in this file assumes that all but these flags aren't relevant
1439 * to the SSC, except ANYOF_EMPTY_STRING, which should be cleared by the
1440 * time we reach here */
1441 assert(! (ANYOF_FLAGS(ssc) & ~ANYOF_COMMON_FLAGS));
1443 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1445 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1446 NULL, NULL, NULL, FALSE);
1448 /* Make sure is clone-safe */
1449 ssc->invlist = NULL;
1451 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1452 ANYOF_FLAGS(ssc) |= ANYOF_POSIXL;
1455 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1458 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1459 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1460 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1461 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1462 ? (TRIE_LIST_CUR( idx ) - 1) \
1468 dump_trie(trie,widecharmap,revcharmap)
1469 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1470 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1472 These routines dump out a trie in a somewhat readable format.
1473 The _interim_ variants are used for debugging the interim
1474 tables that are used to generate the final compressed
1475 representation which is what dump_trie expects.
1477 Part of the reason for their existence is to provide a form
1478 of documentation as to how the different representations function.
1483 Dumps the final compressed table form of the trie to Perl_debug_log.
1484 Used for debugging make_trie().
1488 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1489 AV *revcharmap, U32 depth)
1492 SV *sv=sv_newmortal();
1493 int colwidth= widecharmap ? 6 : 4;
1495 GET_RE_DEBUG_FLAGS_DECL;
1497 PERL_ARGS_ASSERT_DUMP_TRIE;
1499 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1500 (int)depth * 2 + 2,"",
1501 "Match","Base","Ofs" );
1503 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1504 SV ** const tmp = av_fetch( revcharmap, state, 0);
1506 PerlIO_printf( Perl_debug_log, "%*s",
1508 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1509 PL_colors[0], PL_colors[1],
1510 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1511 PERL_PV_ESCAPE_FIRSTCHAR
1516 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1517 (int)depth * 2 + 2,"");
1519 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1520 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1521 PerlIO_printf( Perl_debug_log, "\n");
1523 for( state = 1 ; state < trie->statecount ; state++ ) {
1524 const U32 base = trie->states[ state ].trans.base;
1526 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1527 (int)depth * 2 + 2,"", (UV)state);
1529 if ( trie->states[ state ].wordnum ) {
1530 PerlIO_printf( Perl_debug_log, " W%4X",
1531 trie->states[ state ].wordnum );
1533 PerlIO_printf( Perl_debug_log, "%6s", "" );
1536 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1541 while( ( base + ofs < trie->uniquecharcount ) ||
1542 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1543 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1547 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1549 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1550 if ( ( base + ofs >= trie->uniquecharcount )
1551 && ( base + ofs - trie->uniquecharcount
1553 && trie->trans[ base + ofs
1554 - trie->uniquecharcount ].check == state )
1556 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1558 (UV)trie->trans[ base + ofs
1559 - trie->uniquecharcount ].next );
1561 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1565 PerlIO_printf( Perl_debug_log, "]");
1568 PerlIO_printf( Perl_debug_log, "\n" );
1570 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1572 for (word=1; word <= trie->wordcount; word++) {
1573 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1574 (int)word, (int)(trie->wordinfo[word].prev),
1575 (int)(trie->wordinfo[word].len));
1577 PerlIO_printf(Perl_debug_log, "\n" );
1580 Dumps a fully constructed but uncompressed trie in list form.
1581 List tries normally only are used for construction when the number of
1582 possible chars (trie->uniquecharcount) is very high.
1583 Used for debugging make_trie().
1586 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1587 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1591 SV *sv=sv_newmortal();
1592 int colwidth= widecharmap ? 6 : 4;
1593 GET_RE_DEBUG_FLAGS_DECL;
1595 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1597 /* print out the table precompression. */
1598 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1599 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1600 "------:-----+-----------------\n" );
1602 for( state=1 ; state < next_alloc ; state ++ ) {
1605 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1606 (int)depth * 2 + 2,"", (UV)state );
1607 if ( ! trie->states[ state ].wordnum ) {
1608 PerlIO_printf( Perl_debug_log, "%5s| ","");
1610 PerlIO_printf( Perl_debug_log, "W%4x| ",
1611 trie->states[ state ].wordnum
1614 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1615 SV ** const tmp = av_fetch( revcharmap,
1616 TRIE_LIST_ITEM(state,charid).forid, 0);
1618 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1620 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1622 PL_colors[0], PL_colors[1],
1623 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1624 | PERL_PV_ESCAPE_FIRSTCHAR
1626 TRIE_LIST_ITEM(state,charid).forid,
1627 (UV)TRIE_LIST_ITEM(state,charid).newstate
1630 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1631 (int)((depth * 2) + 14), "");
1634 PerlIO_printf( Perl_debug_log, "\n");
1639 Dumps a fully constructed but uncompressed trie in table form.
1640 This is the normal DFA style state transition table, with a few
1641 twists to facilitate compression later.
1642 Used for debugging make_trie().
1645 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1646 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1651 SV *sv=sv_newmortal();
1652 int colwidth= widecharmap ? 6 : 4;
1653 GET_RE_DEBUG_FLAGS_DECL;
1655 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1658 print out the table precompression so that we can do a visual check
1659 that they are identical.
1662 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1664 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1665 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1667 PerlIO_printf( Perl_debug_log, "%*s",
1669 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1670 PL_colors[0], PL_colors[1],
1671 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1672 PERL_PV_ESCAPE_FIRSTCHAR
1678 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1680 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1681 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1684 PerlIO_printf( Perl_debug_log, "\n" );
1686 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1688 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1689 (int)depth * 2 + 2,"",
1690 (UV)TRIE_NODENUM( state ) );
1692 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1693 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1695 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1697 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1699 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1700 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1701 (UV)trie->trans[ state ].check );
1703 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1704 (UV)trie->trans[ state ].check,
1705 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1713 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1714 startbranch: the first branch in the whole branch sequence
1715 first : start branch of sequence of branch-exact nodes.
1716 May be the same as startbranch
1717 last : Thing following the last branch.
1718 May be the same as tail.
1719 tail : item following the branch sequence
1720 count : words in the sequence
1721 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1722 depth : indent depth
1724 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1726 A trie is an N'ary tree where the branches are determined by digital
1727 decomposition of the key. IE, at the root node you look up the 1st character and
1728 follow that branch repeat until you find the end of the branches. Nodes can be
1729 marked as "accepting" meaning they represent a complete word. Eg:
1733 would convert into the following structure. Numbers represent states, letters
1734 following numbers represent valid transitions on the letter from that state, if
1735 the number is in square brackets it represents an accepting state, otherwise it
1736 will be in parenthesis.
1738 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1742 (1) +-i->(6)-+-s->[7]
1744 +-s->(3)-+-h->(4)-+-e->[5]
1746 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1748 This shows that when matching against the string 'hers' we will begin at state 1
1749 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1750 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1751 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1752 single traverse. We store a mapping from accepting to state to which word was
1753 matched, and then when we have multiple possibilities we try to complete the
1754 rest of the regex in the order in which they occured in the alternation.
1756 The only prior NFA like behaviour that would be changed by the TRIE support is
1757 the silent ignoring of duplicate alternations which are of the form:
1759 / (DUPE|DUPE) X? (?{ ... }) Y /x
1761 Thus EVAL blocks following a trie may be called a different number of times with
1762 and without the optimisation. With the optimisations dupes will be silently
1763 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1764 the following demonstrates:
1766 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1768 which prints out 'word' three times, but
1770 'words'=~/(word|word|word)(?{ print $1 })S/
1772 which doesnt print it out at all. This is due to other optimisations kicking in.
1774 Example of what happens on a structural level:
1776 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1778 1: CURLYM[1] {1,32767}(18)
1789 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1790 and should turn into:
1792 1: CURLYM[1] {1,32767}(18)
1794 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1802 Cases where tail != last would be like /(?foo|bar)baz/:
1812 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1813 and would end up looking like:
1816 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1823 d = uvchr_to_utf8_flags(d, uv, 0);
1825 is the recommended Unicode-aware way of saying
1830 #define TRIE_STORE_REVCHAR(val) \
1833 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1834 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1835 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
1836 SvCUR_set(zlopp, kapow - flrbbbbb); \
1839 av_push(revcharmap, zlopp); \
1841 char ooooff = (char)val; \
1842 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1846 /* This gets the next character from the input, folding it if not already
1848 #define TRIE_READ_CHAR STMT_START { \
1851 /* if it is UTF then it is either already folded, or does not need \
1853 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
1855 else if (folder == PL_fold_latin1) { \
1856 /* This folder implies Unicode rules, which in the range expressible \
1857 * by not UTF is the lower case, with the two exceptions, one of \
1858 * which should have been taken care of before calling this */ \
1859 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
1860 uvc = toLOWER_L1(*uc); \
1861 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
1864 /* raw data, will be folded later if needed */ \
1872 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1873 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1874 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1875 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1877 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1878 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1879 TRIE_LIST_CUR( state )++; \
1882 #define TRIE_LIST_NEW(state) STMT_START { \
1883 Newxz( trie->states[ state ].trans.list, \
1884 4, reg_trie_trans_le ); \
1885 TRIE_LIST_CUR( state ) = 1; \
1886 TRIE_LIST_LEN( state ) = 4; \
1889 #define TRIE_HANDLE_WORD(state) STMT_START { \
1890 U16 dupe= trie->states[ state ].wordnum; \
1891 regnode * const noper_next = regnext( noper ); \
1894 /* store the word for dumping */ \
1896 if (OP(noper) != NOTHING) \
1897 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1899 tmp = newSVpvn_utf8( "", 0, UTF ); \
1900 av_push( trie_words, tmp ); \
1904 trie->wordinfo[curword].prev = 0; \
1905 trie->wordinfo[curword].len = wordlen; \
1906 trie->wordinfo[curword].accept = state; \
1908 if ( noper_next < tail ) { \
1910 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
1912 trie->jump[curword] = (U16)(noper_next - convert); \
1914 jumper = noper_next; \
1916 nextbranch= regnext(cur); \
1920 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1921 /* chain, so that when the bits of chain are later */\
1922 /* linked together, the dups appear in the chain */\
1923 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1924 trie->wordinfo[dupe].prev = curword; \
1926 /* we haven't inserted this word yet. */ \
1927 trie->states[ state ].wordnum = curword; \
1932 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1933 ( ( base + charid >= ucharcount \
1934 && base + charid < ubound \
1935 && state == trie->trans[ base - ucharcount + charid ].check \
1936 && trie->trans[ base - ucharcount + charid ].next ) \
1937 ? trie->trans[ base - ucharcount + charid ].next \
1938 : ( state==1 ? special : 0 ) \
1942 #define MADE_JUMP_TRIE 2
1943 #define MADE_EXACT_TRIE 4
1946 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
1947 regnode *first, regnode *last, regnode *tail,
1948 U32 word_count, U32 flags, U32 depth)
1951 /* first pass, loop through and scan words */
1952 reg_trie_data *trie;
1953 HV *widecharmap = NULL;
1954 AV *revcharmap = newAV();
1960 regnode *jumper = NULL;
1961 regnode *nextbranch = NULL;
1962 regnode *convert = NULL;
1963 U32 *prev_states; /* temp array mapping each state to previous one */
1964 /* we just use folder as a flag in utf8 */
1965 const U8 * folder = NULL;
1968 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
1969 AV *trie_words = NULL;
1970 /* along with revcharmap, this only used during construction but both are
1971 * useful during debugging so we store them in the struct when debugging.
1974 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
1975 STRLEN trie_charcount=0;
1977 SV *re_trie_maxbuff;
1978 GET_RE_DEBUG_FLAGS_DECL;
1980 PERL_ARGS_ASSERT_MAKE_TRIE;
1982 PERL_UNUSED_ARG(depth);
1989 case EXACTFU: folder = PL_fold_latin1; break;
1990 case EXACTF: folder = PL_fold; break;
1991 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1994 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1996 trie->startstate = 1;
1997 trie->wordcount = word_count;
1998 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1999 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2001 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2002 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2003 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2006 trie_words = newAV();
2009 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2010 assert(re_trie_maxbuff);
2011 if (!SvIOK(re_trie_maxbuff)) {
2012 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2014 DEBUG_TRIE_COMPILE_r({
2015 PerlIO_printf( Perl_debug_log,
2016 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2017 (int)depth * 2 + 2, "",
2018 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2019 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2022 /* Find the node we are going to overwrite */
2023 if ( first == startbranch && OP( last ) != BRANCH ) {
2024 /* whole branch chain */
2027 /* branch sub-chain */
2028 convert = NEXTOPER( first );
2031 /* -- First loop and Setup --
2033 We first traverse the branches and scan each word to determine if it
2034 contains widechars, and how many unique chars there are, this is
2035 important as we have to build a table with at least as many columns as we
2038 We use an array of integers to represent the character codes 0..255
2039 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2040 the native representation of the character value as the key and IV's for
2043 *TODO* If we keep track of how many times each character is used we can
2044 remap the columns so that the table compression later on is more
2045 efficient in terms of memory by ensuring the most common value is in the
2046 middle and the least common are on the outside. IMO this would be better
2047 than a most to least common mapping as theres a decent chance the most
2048 common letter will share a node with the least common, meaning the node
2049 will not be compressible. With a middle is most common approach the worst
2050 case is when we have the least common nodes twice.
2054 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2055 regnode *noper = NEXTOPER( cur );
2056 const U8 *uc = (U8*)STRING( noper );
2057 const U8 *e = uc + STR_LEN( noper );
2059 U32 wordlen = 0; /* required init */
2060 STRLEN minchars = 0;
2061 STRLEN maxchars = 0;
2062 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2065 if (OP(noper) == NOTHING) {
2066 regnode *noper_next= regnext(noper);
2067 if (noper_next != tail && OP(noper_next) == flags) {
2069 uc= (U8*)STRING(noper);
2070 e= uc + STR_LEN(noper);
2071 trie->minlen= STR_LEN(noper);
2078 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2079 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2080 regardless of encoding */
2081 if (OP( noper ) == EXACTFU_SS) {
2082 /* false positives are ok, so just set this */
2083 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2086 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2088 TRIE_CHARCOUNT(trie)++;
2091 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2092 * is in effect. Under /i, this character can match itself, or
2093 * anything that folds to it. If not under /i, it can match just
2094 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2095 * all fold to k, and all are single characters. But some folds
2096 * expand to more than one character, so for example LATIN SMALL
2097 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2098 * the string beginning at 'uc' is 'ffi', it could be matched by
2099 * three characters, or just by the one ligature character. (It
2100 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2101 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2102 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2103 * match.) The trie needs to know the minimum and maximum number
2104 * of characters that could match so that it can use size alone to
2105 * quickly reject many match attempts. The max is simple: it is
2106 * the number of folded characters in this branch (since a fold is
2107 * never shorter than what folds to it. */
2111 /* And the min is equal to the max if not under /i (indicated by
2112 * 'folder' being NULL), or there are no multi-character folds. If
2113 * there is a multi-character fold, the min is incremented just
2114 * once, for the character that folds to the sequence. Each
2115 * character in the sequence needs to be added to the list below of
2116 * characters in the trie, but we count only the first towards the
2117 * min number of characters needed. This is done through the
2118 * variable 'foldlen', which is returned by the macros that look
2119 * for these sequences as the number of bytes the sequence
2120 * occupies. Each time through the loop, we decrement 'foldlen' by
2121 * how many bytes the current char occupies. Only when it reaches
2122 * 0 do we increment 'minchars' or look for another multi-character
2124 if (folder == NULL) {
2127 else if (foldlen > 0) {
2128 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2133 /* See if *uc is the beginning of a multi-character fold. If
2134 * so, we decrement the length remaining to look at, to account
2135 * for the current character this iteration. (We can use 'uc'
2136 * instead of the fold returned by TRIE_READ_CHAR because for
2137 * non-UTF, the latin1_safe macro is smart enough to account
2138 * for all the unfolded characters, and because for UTF, the
2139 * string will already have been folded earlier in the
2140 * compilation process */
2142 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2143 foldlen -= UTF8SKIP(uc);
2146 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2151 /* The current character (and any potential folds) should be added
2152 * to the possible matching characters for this position in this
2156 U8 folded= folder[ (U8) uvc ];
2157 if ( !trie->charmap[ folded ] ) {
2158 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2159 TRIE_STORE_REVCHAR( folded );
2162 if ( !trie->charmap[ uvc ] ) {
2163 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2164 TRIE_STORE_REVCHAR( uvc );
2167 /* store the codepoint in the bitmap, and its folded
2169 TRIE_BITMAP_SET(trie, uvc);
2171 /* store the folded codepoint */
2172 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2175 /* store first byte of utf8 representation of
2176 variant codepoints */
2177 if (! UVCHR_IS_INVARIANT(uvc)) {
2178 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2181 set_bit = 0; /* We've done our bit :-) */
2185 /* XXX We could come up with the list of code points that fold
2186 * to this using PL_utf8_foldclosures, except not for
2187 * multi-char folds, as there may be multiple combinations
2188 * there that could work, which needs to wait until runtime to
2189 * resolve (The comment about LIGATURE FFI above is such an
2194 widecharmap = newHV();
2196 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2199 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2201 if ( !SvTRUE( *svpp ) ) {
2202 sv_setiv( *svpp, ++trie->uniquecharcount );
2203 TRIE_STORE_REVCHAR(uvc);
2206 } /* end loop through characters in this branch of the trie */
2208 /* We take the min and max for this branch and combine to find the min
2209 * and max for all branches processed so far */
2210 if( cur == first ) {
2211 trie->minlen = minchars;
2212 trie->maxlen = maxchars;
2213 } else if (minchars < trie->minlen) {
2214 trie->minlen = minchars;
2215 } else if (maxchars > trie->maxlen) {
2216 trie->maxlen = maxchars;
2218 } /* end first pass */
2219 DEBUG_TRIE_COMPILE_r(
2220 PerlIO_printf( Perl_debug_log,
2221 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2222 (int)depth * 2 + 2,"",
2223 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2224 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2225 (int)trie->minlen, (int)trie->maxlen )
2229 We now know what we are dealing with in terms of unique chars and
2230 string sizes so we can calculate how much memory a naive
2231 representation using a flat table will take. If it's over a reasonable
2232 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2233 conservative but potentially much slower representation using an array
2236 At the end we convert both representations into the same compressed
2237 form that will be used in regexec.c for matching with. The latter
2238 is a form that cannot be used to construct with but has memory
2239 properties similar to the list form and access properties similar
2240 to the table form making it both suitable for fast searches and
2241 small enough that its feasable to store for the duration of a program.
2243 See the comment in the code where the compressed table is produced
2244 inplace from the flat tabe representation for an explanation of how
2245 the compression works.
2250 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2253 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2254 > SvIV(re_trie_maxbuff) )
2257 Second Pass -- Array Of Lists Representation
2259 Each state will be represented by a list of charid:state records
2260 (reg_trie_trans_le) the first such element holds the CUR and LEN
2261 points of the allocated array. (See defines above).
2263 We build the initial structure using the lists, and then convert
2264 it into the compressed table form which allows faster lookups
2265 (but cant be modified once converted).
2268 STRLEN transcount = 1;
2270 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2271 "%*sCompiling trie using list compiler\n",
2272 (int)depth * 2 + 2, ""));
2274 trie->states = (reg_trie_state *)
2275 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2276 sizeof(reg_trie_state) );
2280 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2282 regnode *noper = NEXTOPER( cur );
2283 U8 *uc = (U8*)STRING( noper );
2284 const U8 *e = uc + STR_LEN( noper );
2285 U32 state = 1; /* required init */
2286 U16 charid = 0; /* sanity init */
2287 U32 wordlen = 0; /* required init */
2289 if (OP(noper) == NOTHING) {
2290 regnode *noper_next= regnext(noper);
2291 if (noper_next != tail && OP(noper_next) == flags) {
2293 uc= (U8*)STRING(noper);
2294 e= uc + STR_LEN(noper);
2298 if (OP(noper) != NOTHING) {
2299 for ( ; uc < e ; uc += len ) {
2304 charid = trie->charmap[ uvc ];
2306 SV** const svpp = hv_fetch( widecharmap,
2313 charid=(U16)SvIV( *svpp );
2316 /* charid is now 0 if we dont know the char read, or
2317 * nonzero if we do */
2324 if ( !trie->states[ state ].trans.list ) {
2325 TRIE_LIST_NEW( state );
2328 check <= TRIE_LIST_USED( state );
2331 if ( TRIE_LIST_ITEM( state, check ).forid
2334 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2339 newstate = next_alloc++;
2340 prev_states[newstate] = state;
2341 TRIE_LIST_PUSH( state, charid, newstate );
2346 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2350 TRIE_HANDLE_WORD(state);
2352 } /* end second pass */
2354 /* next alloc is the NEXT state to be allocated */
2355 trie->statecount = next_alloc;
2356 trie->states = (reg_trie_state *)
2357 PerlMemShared_realloc( trie->states,
2359 * sizeof(reg_trie_state) );
2361 /* and now dump it out before we compress it */
2362 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2363 revcharmap, next_alloc,
2367 trie->trans = (reg_trie_trans *)
2368 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2375 for( state=1 ; state < next_alloc ; state ++ ) {
2379 DEBUG_TRIE_COMPILE_MORE_r(
2380 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2384 if (trie->states[state].trans.list) {
2385 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2389 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2390 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2391 if ( forid < minid ) {
2393 } else if ( forid > maxid ) {
2397 if ( transcount < tp + maxid - minid + 1) {
2399 trie->trans = (reg_trie_trans *)
2400 PerlMemShared_realloc( trie->trans,
2402 * sizeof(reg_trie_trans) );
2403 Zero( trie->trans + (transcount / 2),
2407 base = trie->uniquecharcount + tp - minid;
2408 if ( maxid == minid ) {
2410 for ( ; zp < tp ; zp++ ) {
2411 if ( ! trie->trans[ zp ].next ) {
2412 base = trie->uniquecharcount + zp - minid;
2413 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2415 trie->trans[ zp ].check = state;
2421 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2423 trie->trans[ tp ].check = state;
2428 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2429 const U32 tid = base
2430 - trie->uniquecharcount
2431 + TRIE_LIST_ITEM( state, idx ).forid;
2432 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2434 trie->trans[ tid ].check = state;
2436 tp += ( maxid - minid + 1 );
2438 Safefree(trie->states[ state ].trans.list);
2441 DEBUG_TRIE_COMPILE_MORE_r(
2442 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2445 trie->states[ state ].trans.base=base;
2447 trie->lasttrans = tp + 1;
2451 Second Pass -- Flat Table Representation.
2453 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2454 each. We know that we will need Charcount+1 trans at most to store
2455 the data (one row per char at worst case) So we preallocate both
2456 structures assuming worst case.
2458 We then construct the trie using only the .next slots of the entry
2461 We use the .check field of the first entry of the node temporarily
2462 to make compression both faster and easier by keeping track of how
2463 many non zero fields are in the node.
2465 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2468 There are two terms at use here: state as a TRIE_NODEIDX() which is
2469 a number representing the first entry of the node, and state as a
2470 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2471 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2472 if there are 2 entrys per node. eg:
2480 The table is internally in the right hand, idx form. However as we
2481 also have to deal with the states array which is indexed by nodenum
2482 we have to use TRIE_NODENUM() to convert.
2485 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2486 "%*sCompiling trie using table compiler\n",
2487 (int)depth * 2 + 2, ""));
2489 trie->trans = (reg_trie_trans *)
2490 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2491 * trie->uniquecharcount + 1,
2492 sizeof(reg_trie_trans) );
2493 trie->states = (reg_trie_state *)
2494 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2495 sizeof(reg_trie_state) );
2496 next_alloc = trie->uniquecharcount + 1;
2499 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2501 regnode *noper = NEXTOPER( cur );
2502 const U8 *uc = (U8*)STRING( noper );
2503 const U8 *e = uc + STR_LEN( noper );
2505 U32 state = 1; /* required init */
2507 U16 charid = 0; /* sanity init */
2508 U32 accept_state = 0; /* sanity init */
2510 U32 wordlen = 0; /* required init */
2512 if (OP(noper) == NOTHING) {
2513 regnode *noper_next= regnext(noper);
2514 if (noper_next != tail && OP(noper_next) == flags) {
2516 uc= (U8*)STRING(noper);
2517 e= uc + STR_LEN(noper);
2521 if ( OP(noper) != NOTHING ) {
2522 for ( ; uc < e ; uc += len ) {
2527 charid = trie->charmap[ uvc ];
2529 SV* const * const svpp = hv_fetch( widecharmap,
2533 charid = svpp ? (U16)SvIV(*svpp) : 0;
2537 if ( !trie->trans[ state + charid ].next ) {
2538 trie->trans[ state + charid ].next = next_alloc;
2539 trie->trans[ state ].check++;
2540 prev_states[TRIE_NODENUM(next_alloc)]
2541 = TRIE_NODENUM(state);
2542 next_alloc += trie->uniquecharcount;
2544 state = trie->trans[ state + charid ].next;
2546 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2548 /* charid is now 0 if we dont know the char read, or
2549 * nonzero if we do */
2552 accept_state = TRIE_NODENUM( state );
2553 TRIE_HANDLE_WORD(accept_state);
2555 } /* end second pass */
2557 /* and now dump it out before we compress it */
2558 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2560 next_alloc, depth+1));
2564 * Inplace compress the table.*
2566 For sparse data sets the table constructed by the trie algorithm will
2567 be mostly 0/FAIL transitions or to put it another way mostly empty.
2568 (Note that leaf nodes will not contain any transitions.)
2570 This algorithm compresses the tables by eliminating most such
2571 transitions, at the cost of a modest bit of extra work during lookup:
2573 - Each states[] entry contains a .base field which indicates the
2574 index in the state[] array wheres its transition data is stored.
2576 - If .base is 0 there are no valid transitions from that node.
2578 - If .base is nonzero then charid is added to it to find an entry in
2581 -If trans[states[state].base+charid].check!=state then the
2582 transition is taken to be a 0/Fail transition. Thus if there are fail
2583 transitions at the front of the node then the .base offset will point
2584 somewhere inside the previous nodes data (or maybe even into a node
2585 even earlier), but the .check field determines if the transition is
2589 The following process inplace converts the table to the compressed
2590 table: We first do not compress the root node 1,and mark all its
2591 .check pointers as 1 and set its .base pointer as 1 as well. This
2592 allows us to do a DFA construction from the compressed table later,
2593 and ensures that any .base pointers we calculate later are greater
2596 - We set 'pos' to indicate the first entry of the second node.
2598 - We then iterate over the columns of the node, finding the first and
2599 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2600 and set the .check pointers accordingly, and advance pos
2601 appropriately and repreat for the next node. Note that when we copy
2602 the next pointers we have to convert them from the original
2603 NODEIDX form to NODENUM form as the former is not valid post
2606 - If a node has no transitions used we mark its base as 0 and do not
2607 advance the pos pointer.
2609 - If a node only has one transition we use a second pointer into the
2610 structure to fill in allocated fail transitions from other states.
2611 This pointer is independent of the main pointer and scans forward
2612 looking for null transitions that are allocated to a state. When it
2613 finds one it writes the single transition into the "hole". If the
2614 pointer doesnt find one the single transition is appended as normal.
2616 - Once compressed we can Renew/realloc the structures to release the
2619 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2620 specifically Fig 3.47 and the associated pseudocode.
2624 const U32 laststate = TRIE_NODENUM( next_alloc );
2627 trie->statecount = laststate;
2629 for ( state = 1 ; state < laststate ; state++ ) {
2631 const U32 stateidx = TRIE_NODEIDX( state );
2632 const U32 o_used = trie->trans[ stateidx ].check;
2633 U32 used = trie->trans[ stateidx ].check;
2634 trie->trans[ stateidx ].check = 0;
2637 used && charid < trie->uniquecharcount;
2640 if ( flag || trie->trans[ stateidx + charid ].next ) {
2641 if ( trie->trans[ stateidx + charid ].next ) {
2643 for ( ; zp < pos ; zp++ ) {
2644 if ( ! trie->trans[ zp ].next ) {
2648 trie->states[ state ].trans.base
2650 + trie->uniquecharcount
2652 trie->trans[ zp ].next
2653 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2655 trie->trans[ zp ].check = state;
2656 if ( ++zp > pos ) pos = zp;
2663 trie->states[ state ].trans.base
2664 = pos + trie->uniquecharcount - charid ;
2666 trie->trans[ pos ].next
2667 = SAFE_TRIE_NODENUM(
2668 trie->trans[ stateidx + charid ].next );
2669 trie->trans[ pos ].check = state;
2674 trie->lasttrans = pos + 1;
2675 trie->states = (reg_trie_state *)
2676 PerlMemShared_realloc( trie->states, laststate
2677 * sizeof(reg_trie_state) );
2678 DEBUG_TRIE_COMPILE_MORE_r(
2679 PerlIO_printf( Perl_debug_log,
2680 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2681 (int)depth * 2 + 2,"",
2682 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2686 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2689 } /* end table compress */
2691 DEBUG_TRIE_COMPILE_MORE_r(
2692 PerlIO_printf(Perl_debug_log,
2693 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2694 (int)depth * 2 + 2, "",
2695 (UV)trie->statecount,
2696 (UV)trie->lasttrans)
2698 /* resize the trans array to remove unused space */
2699 trie->trans = (reg_trie_trans *)
2700 PerlMemShared_realloc( trie->trans, trie->lasttrans
2701 * sizeof(reg_trie_trans) );
2703 { /* Modify the program and insert the new TRIE node */
2704 U8 nodetype =(U8)(flags & 0xFF);
2708 regnode *optimize = NULL;
2709 #ifdef RE_TRACK_PATTERN_OFFSETS
2712 U32 mjd_nodelen = 0;
2713 #endif /* RE_TRACK_PATTERN_OFFSETS */
2714 #endif /* DEBUGGING */
2716 This means we convert either the first branch or the first Exact,
2717 depending on whether the thing following (in 'last') is a branch
2718 or not and whther first is the startbranch (ie is it a sub part of
2719 the alternation or is it the whole thing.)
2720 Assuming its a sub part we convert the EXACT otherwise we convert
2721 the whole branch sequence, including the first.
2723 /* Find the node we are going to overwrite */
2724 if ( first != startbranch || OP( last ) == BRANCH ) {
2725 /* branch sub-chain */
2726 NEXT_OFF( first ) = (U16)(last - first);
2727 #ifdef RE_TRACK_PATTERN_OFFSETS
2729 mjd_offset= Node_Offset((convert));
2730 mjd_nodelen= Node_Length((convert));
2733 /* whole branch chain */
2735 #ifdef RE_TRACK_PATTERN_OFFSETS
2738 const regnode *nop = NEXTOPER( convert );
2739 mjd_offset= Node_Offset((nop));
2740 mjd_nodelen= Node_Length((nop));
2744 PerlIO_printf(Perl_debug_log,
2745 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2746 (int)depth * 2 + 2, "",
2747 (UV)mjd_offset, (UV)mjd_nodelen)
2750 /* But first we check to see if there is a common prefix we can
2751 split out as an EXACT and put in front of the TRIE node. */
2752 trie->startstate= 1;
2753 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2755 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2759 const U32 base = trie->states[ state ].trans.base;
2761 if ( trie->states[state].wordnum )
2764 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2765 if ( ( base + ofs >= trie->uniquecharcount ) &&
2766 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2767 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2769 if ( ++count > 1 ) {
2770 SV **tmp = av_fetch( revcharmap, ofs, 0);
2771 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2772 if ( state == 1 ) break;
2774 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2776 PerlIO_printf(Perl_debug_log,
2777 "%*sNew Start State=%"UVuf" Class: [",
2778 (int)depth * 2 + 2, "",
2781 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2782 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2784 TRIE_BITMAP_SET(trie,*ch);
2786 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2788 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2792 TRIE_BITMAP_SET(trie,*ch);
2794 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2795 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2801 SV **tmp = av_fetch( revcharmap, idx, 0);
2803 char *ch = SvPV( *tmp, len );
2805 SV *sv=sv_newmortal();
2806 PerlIO_printf( Perl_debug_log,
2807 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2808 (int)depth * 2 + 2, "",
2810 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2811 PL_colors[0], PL_colors[1],
2812 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2813 PERL_PV_ESCAPE_FIRSTCHAR
2818 OP( convert ) = nodetype;
2819 str=STRING(convert);
2822 STR_LEN(convert) += len;
2828 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2833 trie->prefixlen = (state-1);
2835 regnode *n = convert+NODE_SZ_STR(convert);
2836 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2837 trie->startstate = state;
2838 trie->minlen -= (state - 1);
2839 trie->maxlen -= (state - 1);
2841 /* At least the UNICOS C compiler choked on this
2842 * being argument to DEBUG_r(), so let's just have
2845 #ifdef PERL_EXT_RE_BUILD
2851 regnode *fix = convert;
2852 U32 word = trie->wordcount;
2854 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2855 while( ++fix < n ) {
2856 Set_Node_Offset_Length(fix, 0, 0);
2859 SV ** const tmp = av_fetch( trie_words, word, 0 );
2861 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2862 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2864 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2872 NEXT_OFF(convert) = (U16)(tail - convert);
2873 DEBUG_r(optimize= n);
2879 if ( trie->maxlen ) {
2880 NEXT_OFF( convert ) = (U16)(tail - convert);
2881 ARG_SET( convert, data_slot );
2882 /* Store the offset to the first unabsorbed branch in
2883 jump[0], which is otherwise unused by the jump logic.
2884 We use this when dumping a trie and during optimisation. */
2886 trie->jump[0] = (U16)(nextbranch - convert);
2888 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2889 * and there is a bitmap
2890 * and the first "jump target" node we found leaves enough room
2891 * then convert the TRIE node into a TRIEC node, with the bitmap
2892 * embedded inline in the opcode - this is hypothetically faster.
2894 if ( !trie->states[trie->startstate].wordnum
2896 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2898 OP( convert ) = TRIEC;
2899 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2900 PerlMemShared_free(trie->bitmap);
2903 OP( convert ) = TRIE;
2905 /* store the type in the flags */
2906 convert->flags = nodetype;
2910 + regarglen[ OP( convert ) ];
2912 /* XXX We really should free up the resource in trie now,
2913 as we won't use them - (which resources?) dmq */
2915 /* needed for dumping*/
2916 DEBUG_r(if (optimize) {
2917 regnode *opt = convert;
2919 while ( ++opt < optimize) {
2920 Set_Node_Offset_Length(opt,0,0);
2923 Try to clean up some of the debris left after the
2926 while( optimize < jumper ) {
2927 mjd_nodelen += Node_Length((optimize));
2928 OP( optimize ) = OPTIMIZED;
2929 Set_Node_Offset_Length(optimize,0,0);
2932 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2934 } /* end node insert */
2936 /* Finish populating the prev field of the wordinfo array. Walk back
2937 * from each accept state until we find another accept state, and if
2938 * so, point the first word's .prev field at the second word. If the
2939 * second already has a .prev field set, stop now. This will be the
2940 * case either if we've already processed that word's accept state,
2941 * or that state had multiple words, and the overspill words were
2942 * already linked up earlier.
2949 for (word=1; word <= trie->wordcount; word++) {
2951 if (trie->wordinfo[word].prev)
2953 state = trie->wordinfo[word].accept;
2955 state = prev_states[state];
2958 prev = trie->states[state].wordnum;
2962 trie->wordinfo[word].prev = prev;
2964 Safefree(prev_states);
2968 /* and now dump out the compressed format */
2969 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2971 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2973 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2974 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2976 SvREFCNT_dec_NN(revcharmap);
2980 : trie->startstate>1
2986 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2988 /* The Trie is constructed and compressed now so we can build a fail array if
2991 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
2993 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
2997 We find the fail state for each state in the trie, this state is the longest
2998 proper suffix of the current state's 'word' that is also a proper prefix of
2999 another word in our trie. State 1 represents the word '' and is thus the
3000 default fail state. This allows the DFA not to have to restart after its
3001 tried and failed a word at a given point, it simply continues as though it
3002 had been matching the other word in the first place.
3004 'abcdgu'=~/abcdefg|cdgu/
3005 When we get to 'd' we are still matching the first word, we would encounter
3006 'g' which would fail, which would bring us to the state representing 'd' in
3007 the second word where we would try 'g' and succeed, proceeding to match
3010 /* add a fail transition */
3011 const U32 trie_offset = ARG(source);
3012 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3014 const U32 ucharcount = trie->uniquecharcount;
3015 const U32 numstates = trie->statecount;
3016 const U32 ubound = trie->lasttrans + ucharcount;
3020 U32 base = trie->states[ 1 ].trans.base;
3023 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3024 GET_RE_DEBUG_FLAGS_DECL;
3026 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
3028 PERL_UNUSED_ARG(depth);
3032 ARG_SET( stclass, data_slot );
3033 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3034 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3035 aho->trie=trie_offset;
3036 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3037 Copy( trie->states, aho->states, numstates, reg_trie_state );
3038 Newxz( q, numstates, U32);
3039 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3042 /* initialize fail[0..1] to be 1 so that we always have
3043 a valid final fail state */
3044 fail[ 0 ] = fail[ 1 ] = 1;
3046 for ( charid = 0; charid < ucharcount ; charid++ ) {
3047 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3049 q[ q_write ] = newstate;
3050 /* set to point at the root */
3051 fail[ q[ q_write++ ] ]=1;
3054 while ( q_read < q_write) {
3055 const U32 cur = q[ q_read++ % numstates ];
3056 base = trie->states[ cur ].trans.base;
3058 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3059 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3061 U32 fail_state = cur;
3064 fail_state = fail[ fail_state ];
3065 fail_base = aho->states[ fail_state ].trans.base;
3066 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3068 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3069 fail[ ch_state ] = fail_state;
3070 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3072 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3074 q[ q_write++ % numstates] = ch_state;
3078 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3079 when we fail in state 1, this allows us to use the
3080 charclass scan to find a valid start char. This is based on the principle
3081 that theres a good chance the string being searched contains lots of stuff
3082 that cant be a start char.
3084 fail[ 0 ] = fail[ 1 ] = 0;
3085 DEBUG_TRIE_COMPILE_r({
3086 PerlIO_printf(Perl_debug_log,
3087 "%*sStclass Failtable (%"UVuf" states): 0",
3088 (int)(depth * 2), "", (UV)numstates
3090 for( q_read=1; q_read<numstates; q_read++ ) {
3091 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3093 PerlIO_printf(Perl_debug_log, "\n");
3096 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3100 #define DEBUG_PEEP(str,scan,depth) \
3101 DEBUG_OPTIMISE_r({if (scan){ \
3102 SV * const mysv=sv_newmortal(); \
3103 regnode *Next = regnext(scan); \
3104 regprop(RExC_rx, mysv, scan, NULL); \
3105 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
3106 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
3107 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3111 /* The below joins as many adjacent EXACTish nodes as possible into a single
3112 * one. The regop may be changed if the node(s) contain certain sequences that
3113 * require special handling. The joining is only done if:
3114 * 1) there is room in the current conglomerated node to entirely contain the
3116 * 2) they are the exact same node type
3118 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3119 * these get optimized out
3121 * If a node is to match under /i (folded), the number of characters it matches
3122 * can be different than its character length if it contains a multi-character
3123 * fold. *min_subtract is set to the total delta number of characters of the
3126 * And *unfolded_multi_char is set to indicate whether or not the node contains
3127 * an unfolded multi-char fold. This happens when whether the fold is valid or
3128 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3129 * SMALL LETTER SHARP S, as only if the target string being matched against
3130 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3131 * folding rules depend on the locale in force at runtime. (Multi-char folds
3132 * whose components are all above the Latin1 range are not run-time locale
3133 * dependent, and have already been folded by the time this function is
3136 * This is as good a place as any to discuss the design of handling these
3137 * multi-character fold sequences. It's been wrong in Perl for a very long
3138 * time. There are three code points in Unicode whose multi-character folds
3139 * were long ago discovered to mess things up. The previous designs for
3140 * dealing with these involved assigning a special node for them. This
3141 * approach doesn't always work, as evidenced by this example:
3142 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3143 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3144 * would match just the \xDF, it won't be able to handle the case where a
3145 * successful match would have to cross the node's boundary. The new approach
3146 * that hopefully generally solves the problem generates an EXACTFU_SS node
3147 * that is "sss" in this case.
3149 * It turns out that there are problems with all multi-character folds, and not
3150 * just these three. Now the code is general, for all such cases. The
3151 * approach taken is:
3152 * 1) This routine examines each EXACTFish node that could contain multi-
3153 * character folded sequences. Since a single character can fold into
3154 * such a sequence, the minimum match length for this node is less than
3155 * the number of characters in the node. This routine returns in
3156 * *min_subtract how many characters to subtract from the the actual
3157 * length of the string to get a real minimum match length; it is 0 if
3158 * there are no multi-char foldeds. This delta is used by the caller to
3159 * adjust the min length of the match, and the delta between min and max,
3160 * so that the optimizer doesn't reject these possibilities based on size
3162 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3163 * is used for an EXACTFU node that contains at least one "ss" sequence in
3164 * it. For non-UTF-8 patterns and strings, this is the only case where
3165 * there is a possible fold length change. That means that a regular
3166 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3167 * with length changes, and so can be processed faster. regexec.c takes
3168 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3169 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3170 * known until runtime). This saves effort in regex matching. However,
3171 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3172 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3173 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3174 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3175 * possibilities for the non-UTF8 patterns are quite simple, except for
3176 * the sharp s. All the ones that don't involve a UTF-8 target string are
3177 * members of a fold-pair, and arrays are set up for all of them so that
3178 * the other member of the pair can be found quickly. Code elsewhere in
3179 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3180 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3181 * described in the next item.
3182 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3183 * validity of the fold won't be known until runtime, and so must remain
3184 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3185 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3186 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3187 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3188 * The reason this is a problem is that the optimizer part of regexec.c
3189 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3190 * that a character in the pattern corresponds to at most a single
3191 * character in the target string. (And I do mean character, and not byte
3192 * here, unlike other parts of the documentation that have never been
3193 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3194 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3195 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3196 * nodes, violate the assumption, and they are the only instances where it
3197 * is violated. I'm reluctant to try to change the assumption, as the
3198 * code involved is impenetrable to me (khw), so instead the code here
3199 * punts. This routine examines EXACTFL nodes, and (when the pattern
3200 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3201 * boolean indicating whether or not the node contains such a fold. When
3202 * it is true, the caller sets a flag that later causes the optimizer in
3203 * this file to not set values for the floating and fixed string lengths,
3204 * and thus avoids the optimizer code in regexec.c that makes the invalid
3205 * assumption. Thus, there is no optimization based on string lengths for
3206 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3207 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3208 * assumption is wrong only in these cases is that all other non-UTF-8
3209 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3210 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3211 * EXACTF nodes because we don't know at compile time if it actually
3212 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3213 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3214 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3215 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3216 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3217 * string would require the pattern to be forced into UTF-8, the overhead
3218 * of which we want to avoid. Similarly the unfolded multi-char folds in
3219 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3222 * Similarly, the code that generates tries doesn't currently handle
3223 * not-already-folded multi-char folds, and it looks like a pain to change
3224 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3225 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3226 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3227 * using /iaa matching will be doing so almost entirely with ASCII
3228 * strings, so this should rarely be encountered in practice */
3230 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3231 if (PL_regkind[OP(scan)] == EXACT) \
3232 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3235 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3236 UV *min_subtract, bool *unfolded_multi_char,
3237 U32 flags,regnode *val, U32 depth)
3239 /* Merge several consecutive EXACTish nodes into one. */
3240 regnode *n = regnext(scan);
3242 regnode *next = scan + NODE_SZ_STR(scan);
3246 regnode *stop = scan;
3247 GET_RE_DEBUG_FLAGS_DECL;
3249 PERL_UNUSED_ARG(depth);
3252 PERL_ARGS_ASSERT_JOIN_EXACT;
3253 #ifndef EXPERIMENTAL_INPLACESCAN
3254 PERL_UNUSED_ARG(flags);
3255 PERL_UNUSED_ARG(val);
3257 DEBUG_PEEP("join",scan,depth);
3259 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3260 * EXACT ones that are mergeable to the current one. */
3262 && (PL_regkind[OP(n)] == NOTHING
3263 || (stringok && OP(n) == OP(scan)))
3265 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3268 if (OP(n) == TAIL || n > next)
3270 if (PL_regkind[OP(n)] == NOTHING) {
3271 DEBUG_PEEP("skip:",n,depth);
3272 NEXT_OFF(scan) += NEXT_OFF(n);
3273 next = n + NODE_STEP_REGNODE;
3280 else if (stringok) {
3281 const unsigned int oldl = STR_LEN(scan);
3282 regnode * const nnext = regnext(n);
3284 /* XXX I (khw) kind of doubt that this works on platforms (should
3285 * Perl ever run on one) where U8_MAX is above 255 because of lots
3286 * of other assumptions */
3287 /* Don't join if the sum can't fit into a single node */
3288 if (oldl + STR_LEN(n) > U8_MAX)
3291 DEBUG_PEEP("merg",n,depth);
3294 NEXT_OFF(scan) += NEXT_OFF(n);
3295 STR_LEN(scan) += STR_LEN(n);
3296 next = n + NODE_SZ_STR(n);
3297 /* Now we can overwrite *n : */
3298 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3306 #ifdef EXPERIMENTAL_INPLACESCAN
3307 if (flags && !NEXT_OFF(n)) {
3308 DEBUG_PEEP("atch", val, depth);
3309 if (reg_off_by_arg[OP(n)]) {
3310 ARG_SET(n, val - n);
3313 NEXT_OFF(n) = val - n;
3321 *unfolded_multi_char = FALSE;
3323 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3324 * can now analyze for sequences of problematic code points. (Prior to
3325 * this final joining, sequences could have been split over boundaries, and
3326 * hence missed). The sequences only happen in folding, hence for any
3327 * non-EXACT EXACTish node */
3328 if (OP(scan) != EXACT) {
3329 U8* s0 = (U8*) STRING(scan);
3331 U8* s_end = s0 + STR_LEN(scan);
3333 int total_count_delta = 0; /* Total delta number of characters that
3334 multi-char folds expand to */
3336 /* One pass is made over the node's string looking for all the
3337 * possibilities. To avoid some tests in the loop, there are two main
3338 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3343 if (OP(scan) == EXACTFL) {
3346 /* An EXACTFL node would already have been changed to another
3347 * node type unless there is at least one character in it that
3348 * is problematic; likely a character whose fold definition
3349 * won't be known until runtime, and so has yet to be folded.
3350 * For all but the UTF-8 locale, folds are 1-1 in length, but
3351 * to handle the UTF-8 case, we need to create a temporary
3352 * folded copy using UTF-8 locale rules in order to analyze it.
3353 * This is because our macros that look to see if a sequence is
3354 * a multi-char fold assume everything is folded (otherwise the
3355 * tests in those macros would be too complicated and slow).
3356 * Note that here, the non-problematic folds will have already
3357 * been done, so we can just copy such characters. We actually
3358 * don't completely fold the EXACTFL string. We skip the
3359 * unfolded multi-char folds, as that would just create work
3360 * below to figure out the size they already are */
3362 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3365 STRLEN s_len = UTF8SKIP(s);
3366 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3367 Copy(s, d, s_len, U8);
3370 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3371 *unfolded_multi_char = TRUE;
3372 Copy(s, d, s_len, U8);
3375 else if (isASCII(*s)) {
3376 *(d++) = toFOLD(*s);
3380 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3386 /* Point the remainder of the routine to look at our temporary
3390 } /* End of creating folded copy of EXACTFL string */
3392 /* Examine the string for a multi-character fold sequence. UTF-8
3393 * patterns have all characters pre-folded by the time this code is
3395 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3396 length sequence we are looking for is 2 */
3398 int count = 0; /* How many characters in a multi-char fold */
3399 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3400 if (! len) { /* Not a multi-char fold: get next char */
3405 /* Nodes with 'ss' require special handling, except for
3406 * EXACTFA-ish for which there is no multi-char fold to this */
3407 if (len == 2 && *s == 's' && *(s+1) == 's'
3408 && OP(scan) != EXACTFA
3409 && OP(scan) != EXACTFA_NO_TRIE)
3412 if (OP(scan) != EXACTFL) {
3413 OP(scan) = EXACTFU_SS;
3417 else { /* Here is a generic multi-char fold. */
3418 U8* multi_end = s + len;
3420 /* Count how many characters are in it. In the case of
3421 * /aa, no folds which contain ASCII code points are
3422 * allowed, so check for those, and skip if found. */
3423 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3424 count = utf8_length(s, multi_end);
3428 while (s < multi_end) {
3431 goto next_iteration;
3441 /* The delta is how long the sequence is minus 1 (1 is how long
3442 * the character that folds to the sequence is) */
3443 total_count_delta += count - 1;
3447 /* We created a temporary folded copy of the string in EXACTFL
3448 * nodes. Therefore we need to be sure it doesn't go below zero,
3449 * as the real string could be shorter */
3450 if (OP(scan) == EXACTFL) {
3451 int total_chars = utf8_length((U8*) STRING(scan),
3452 (U8*) STRING(scan) + STR_LEN(scan));
3453 if (total_count_delta > total_chars) {
3454 total_count_delta = total_chars;
3458 *min_subtract += total_count_delta;
3461 else if (OP(scan) == EXACTFA) {
3463 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3464 * fold to the ASCII range (and there are no existing ones in the
3465 * upper latin1 range). But, as outlined in the comments preceding
3466 * this function, we need to flag any occurrences of the sharp s.
3467 * This character forbids trie formation (because of added
3470 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3471 OP(scan) = EXACTFA_NO_TRIE;
3472 *unfolded_multi_char = TRUE;
3481 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3482 * folds that are all Latin1. As explained in the comments
3483 * preceding this function, we look also for the sharp s in EXACTF
3484 * and EXACTFL nodes; it can be in the final position. Otherwise
3485 * we can stop looking 1 byte earlier because have to find at least
3486 * two characters for a multi-fold */
3487 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3492 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3493 if (! len) { /* Not a multi-char fold. */
3494 if (*s == LATIN_SMALL_LETTER_SHARP_S
3495 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3497 *unfolded_multi_char = TRUE;
3504 && isARG2_lower_or_UPPER_ARG1('s', *s)
3505 && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
3508 /* EXACTF nodes need to know that the minimum length
3509 * changed so that a sharp s in the string can match this
3510 * ss in the pattern, but they remain EXACTF nodes, as they
3511 * won't match this unless the target string is is UTF-8,
3512 * which we don't know until runtime. EXACTFL nodes can't
3513 * transform into EXACTFU nodes */
3514 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
3515 OP(scan) = EXACTFU_SS;
3519 *min_subtract += len - 1;
3526 /* Allow dumping but overwriting the collection of skipped
3527 * ops and/or strings with fake optimized ops */
3528 n = scan + NODE_SZ_STR(scan);
3536 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3540 /* REx optimizer. Converts nodes into quicker variants "in place".
3541 Finds fixed substrings. */
3543 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3544 to the position after last scanned or to NULL. */
3546 #define INIT_AND_WITHP \
3547 assert(!and_withp); \
3548 Newx(and_withp,1, regnode_ssc); \
3549 SAVEFREEPV(and_withp)
3551 /* this is a chain of data about sub patterns we are processing that
3552 need to be handled separately/specially in study_chunk. Its so
3553 we can simulate recursion without losing state. */
3555 typedef struct scan_frame {
3556 regnode *last; /* last node to process in this frame */
3557 regnode *next; /* next node to process when last is reached */
3558 struct scan_frame *prev; /*previous frame*/
3559 U32 prev_recursed_depth;
3560 I32 stop; /* what stopparen do we use */
3565 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3566 SSize_t *minlenp, SSize_t *deltap,
3571 regnode_ssc *and_withp,
3572 U32 flags, U32 depth)
3573 /* scanp: Start here (read-write). */
3574 /* deltap: Write maxlen-minlen here. */
3575 /* last: Stop before this one. */
3576 /* data: string data about the pattern */
3577 /* stopparen: treat close N as END */
3578 /* recursed: which subroutines have we recursed into */
3579 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3582 /* There must be at least this number of characters to match */
3585 regnode *scan = *scanp, *next;
3587 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3588 int is_inf_internal = 0; /* The studied chunk is infinite */
3589 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3590 scan_data_t data_fake;
3591 SV *re_trie_maxbuff = NULL;
3592 regnode *first_non_open = scan;
3593 SSize_t stopmin = SSize_t_MAX;
3594 scan_frame *frame = NULL;
3595 GET_RE_DEBUG_FLAGS_DECL;
3597 PERL_ARGS_ASSERT_STUDY_CHUNK;
3600 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3603 while (first_non_open && OP(first_non_open) == OPEN)
3604 first_non_open=regnext(first_non_open);
3609 while ( scan && OP(scan) != END && scan < last ){
3610 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3611 node length to get a real minimum (because
3612 the folded version may be shorter) */
3613 bool unfolded_multi_char = FALSE;
3614 /* Peephole optimizer: */
3615 DEBUG_OPTIMISE_MORE_r(
3617 PerlIO_printf(Perl_debug_log,
3618 "%*sstudy_chunk stopparen=%ld depth=%lu recursed_depth=%lu ",
3619 ((int) depth*2), "", (long)stopparen,
3620 (unsigned long)depth, (unsigned long)recursed_depth);
3621 if (recursed_depth) {
3624 for ( j = 0 ; j < recursed_depth ; j++ ) {
3625 PerlIO_printf(Perl_debug_log,"[");
3626 for ( i = 0 ; i < (U32)RExC_npar ; i++ )
3627 PerlIO_printf(Perl_debug_log,"%d",
3628 PAREN_TEST(RExC_study_chunk_recursed +
3629 (j * RExC_study_chunk_recursed_bytes), i)
3632 PerlIO_printf(Perl_debug_log,"]");
3635 PerlIO_printf(Perl_debug_log,"\n");
3638 DEBUG_STUDYDATA("Peep:", data, depth);
3639 DEBUG_PEEP("Peep", scan, depth);
3642 /* The reason we do this here we need to deal with things like /(?:f)(?:o)(?:o)/
3643 * which cant be dealt with by the normal EXACT parsing code, as each (?:..) is handled
3644 * by a different invocation of reg() -- Yves
3646 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
3648 /* Follow the next-chain of the current node and optimize
3649 away all the NOTHINGs from it. */
3650 if (OP(scan) != CURLYX) {
3651 const int max = (reg_off_by_arg[OP(scan)]
3653 /* I32 may be smaller than U16 on CRAYs! */
3654 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3655 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3659 /* Skip NOTHING and LONGJMP. */
3660 while ((n = regnext(n))
3661 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3662 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3663 && off + noff < max)
3665 if (reg_off_by_arg[OP(scan)])
3668 NEXT_OFF(scan) = off;
3673 /* The principal pseudo-switch. Cannot be a switch, since we
3674 look into several different things. */
3675 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3676 || OP(scan) == IFTHEN) {
3677 next = regnext(scan);
3679 /* demq: the op(next)==code check is to see if we have
3680 * "branch-branch" AFAICT */
3682 if (OP(next) == code || code == IFTHEN) {
3683 /* NOTE - There is similar code to this block below for
3684 * handling TRIE nodes on a re-study. If you change stuff here
3685 * check there too. */
3686 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3688 regnode * const startbranch=scan;
3690 if (flags & SCF_DO_SUBSTR) {
3691 /* Cannot merge strings after this. */
3692 scan_commit(pRExC_state, data, minlenp, is_inf);
3695 if (flags & SCF_DO_STCLASS)
3696 ssc_init_zero(pRExC_state, &accum);
3698 while (OP(scan) == code) {
3699 SSize_t deltanext, minnext, fake;
3701 regnode_ssc this_class;
3704 data_fake.flags = 0;
3706 data_fake.whilem_c = data->whilem_c;
3707 data_fake.last_closep = data->last_closep;
3710 data_fake.last_closep = &fake;
3712 data_fake.pos_delta = delta;
3713 next = regnext(scan);
3714 scan = NEXTOPER(scan);
3716 scan = NEXTOPER(scan);
3717 if (flags & SCF_DO_STCLASS) {
3718 ssc_init(pRExC_state, &this_class);
3719 data_fake.start_class = &this_class;
3720 f = SCF_DO_STCLASS_AND;
3722 if (flags & SCF_WHILEM_VISITED_POS)
3723 f |= SCF_WHILEM_VISITED_POS;
3725 /* we suppose the run is continuous, last=next...*/
3726 minnext = study_chunk(pRExC_state, &scan, minlenp,
3727 &deltanext, next, &data_fake, stopparen,
3728 recursed_depth, NULL, f,depth+1);
3731 if (deltanext == SSize_t_MAX) {
3732 is_inf = is_inf_internal = 1;
3734 } else if (max1 < minnext + deltanext)
3735 max1 = minnext + deltanext;
3737 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3739 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3740 if ( stopmin > minnext)
3741 stopmin = min + min1;
3742 flags &= ~SCF_DO_SUBSTR;
3744 data->flags |= SCF_SEEN_ACCEPT;
3747 if (data_fake.flags & SF_HAS_EVAL)
3748 data->flags |= SF_HAS_EVAL;
3749 data->whilem_c = data_fake.whilem_c;
3751 if (flags & SCF_DO_STCLASS)
3752 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
3754 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3756 if (flags & SCF_DO_SUBSTR) {
3757 data->pos_min += min1;
3758 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3759 data->pos_delta = SSize_t_MAX;
3761 data->pos_delta += max1 - min1;
3762 if (max1 != min1 || is_inf)
3763 data->longest = &(data->longest_float);
3766 if (delta == SSize_t_MAX
3767 || SSize_t_MAX - delta - (max1 - min1) < 0)
3768 delta = SSize_t_MAX;
3770 delta += max1 - min1;
3771 if (flags & SCF_DO_STCLASS_OR) {
3772 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
3774 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
3775 flags &= ~SCF_DO_STCLASS;
3778 else if (flags & SCF_DO_STCLASS_AND) {
3780 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
3781 flags &= ~SCF_DO_STCLASS;
3784 /* Switch to OR mode: cache the old value of
3785 * data->start_class */
3787 StructCopy(data->start_class, and_withp, regnode_ssc);
3788 flags &= ~SCF_DO_STCLASS_AND;
3789 StructCopy(&accum, data->start_class, regnode_ssc);
3790 flags |= SCF_DO_STCLASS_OR;
3794 if (PERL_ENABLE_TRIE_OPTIMISATION &&
3795 OP( startbranch ) == BRANCH )
3799 Assuming this was/is a branch we are dealing with: 'scan'
3800 now points at the item that follows the branch sequence,
3801 whatever it is. We now start at the beginning of the
3802 sequence and look for subsequences of
3808 which would be constructed from a pattern like
3811 If we can find such a subsequence we need to turn the first
3812 element into a trie and then add the subsequent branch exact
3813 strings to the trie.
3817 1. patterns where the whole set of branches can be
3820 2. patterns where only a subset can be converted.
3822 In case 1 we can replace the whole set with a single regop
3823 for the trie. In case 2 we need to keep the start and end
3826 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3827 becomes BRANCH TRIE; BRANCH X;
3829 There is an additional case, that being where there is a
3830 common prefix, which gets split out into an EXACT like node
3831 preceding the TRIE node.
3833 If x(1..n)==tail then we can do a simple trie, if not we make
3834 a "jump" trie, such that when we match the appropriate word
3835 we "jump" to the appropriate tail node. Essentially we turn
3836 a nested if into a case structure of sorts.
3841 if (!re_trie_maxbuff) {
3842 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3843 if (!SvIOK(re_trie_maxbuff))
3844 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3846 if ( SvIV(re_trie_maxbuff)>=0 ) {
3848 regnode *first = (regnode *)NULL;
3849 regnode *last = (regnode *)NULL;
3850 regnode *tail = scan;
3855 SV * const mysv = sv_newmortal(); /* for dumping */
3857 /* var tail is used because there may be a TAIL
3858 regop in the way. Ie, the exacts will point to the
3859 thing following the TAIL, but the last branch will
3860 point at the TAIL. So we advance tail. If we
3861 have nested (?:) we may have to move through several
3865 while ( OP( tail ) == TAIL ) {
3866 /* this is the TAIL generated by (?:) */
3867 tail = regnext( tail );
3871 DEBUG_TRIE_COMPILE_r({
3872 regprop(RExC_rx, mysv, tail, NULL);
3873 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3874 (int)depth * 2 + 2, "",
3875 "Looking for TRIE'able sequences. Tail node is: ",
3876 SvPV_nolen_const( mysv )
3882 Step through the branches
3883 cur represents each branch,
3884 noper is the first thing to be matched as part
3886 noper_next is the regnext() of that node.
3888 We normally handle a case like this
3889 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
3890 support building with NOJUMPTRIE, which restricts
3891 the trie logic to structures like /FOO|BAR/.
3893 If noper is a trieable nodetype then the branch is
3894 a possible optimization target. If we are building
3895 under NOJUMPTRIE then we require that noper_next is
3896 the same as scan (our current position in the regex
3899 Once we have two or more consecutive such branches
3900 we can create a trie of the EXACT's contents and
3901 stitch it in place into the program.
3903 If the sequence represents all of the branches in
3904 the alternation we replace the entire thing with a
3907 Otherwise when it is a subsequence we need to
3908 stitch it in place and replace only the relevant
3909 branches. This means the first branch has to remain
3910 as it is used by the alternation logic, and its
3911 next pointer, and needs to be repointed at the item
3912 on the branch chain following the last branch we
3913 have optimized away.
3915 This could be either a BRANCH, in which case the
3916 subsequence is internal, or it could be the item
3917 following the branch sequence in which case the
3918 subsequence is at the end (which does not
3919 necessarily mean the first node is the start of the
3922 TRIE_TYPE(X) is a define which maps the optype to a
3926 ----------------+-----------
https://perl5.git.perl.org / perl5.git / blob