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 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) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
96 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
103 # if defined(BUGGY_MSC6)
104 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
105 # pragma optimize("a",off)
106 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
107 # pragma optimize("w",on )
108 # endif /* BUGGY_MSC6 */
112 #define STATIC static
116 typedef struct RExC_state_t {
117 U32 flags; /* RXf_* are we folding, multilining? */
118 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
119 char *precomp; /* uncompiled string. */
120 REGEXP *rx_sv; /* The SV that is the regexp. */
121 regexp *rx; /* perl core regexp structure */
122 regexp_internal *rxi; /* internal data for regexp object pprivate field */
123 char *start; /* Start of input for compile */
124 char *end; /* End of input for compile */
125 char *parse; /* Input-scan pointer. */
126 SSize_t whilem_seen; /* number of WHILEM in this expr */
127 regnode *emit_start; /* Start of emitted-code area */
128 regnode *emit_bound; /* First regnode outside of the allocated space */
129 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
130 implies compiling, so don't emit */
131 regnode emit_dummy; /* placeholder for emit to point to */
132 I32 naughty; /* How bad is this pattern? */
133 I32 sawback; /* Did we see \1, ...? */
135 SSize_t size; /* Code size. */
136 I32 npar; /* Capture buffer count, (OPEN). */
137 I32 cpar; /* Capture buffer count, (CLOSE). */
138 I32 nestroot; /* root parens we are in - used by accept */
141 regnode **open_parens; /* pointers to open parens */
142 regnode **close_parens; /* pointers to close parens */
143 regnode *opend; /* END node in program */
144 I32 utf8; /* whether the pattern is utf8 or not */
145 I32 orig_utf8; /* whether the pattern was originally in utf8 */
146 /* XXX use this for future optimisation of case
147 * where pattern must be upgraded to utf8. */
148 I32 uni_semantics; /* If a d charset modifier should use unicode
149 rules, even if the pattern is not in
151 HV *paren_names; /* Paren names */
153 regnode **recurse; /* Recurse regops */
154 I32 recurse_count; /* Number of recurse regops */
157 I32 override_recoding;
158 I32 in_multi_char_class;
159 struct reg_code_block *code_blocks; /* positions of literal (?{})
161 int num_code_blocks; /* size of code_blocks[] */
162 int code_index; /* next code_blocks[] slot */
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 others */
191 #define RExC_emit (pRExC_state->emit)
192 #define RExC_emit_dummy (pRExC_state->emit_dummy)
193 #define RExC_emit_start (pRExC_state->emit_start)
194 #define RExC_emit_bound (pRExC_state->emit_bound)
195 #define RExC_naughty (pRExC_state->naughty)
196 #define RExC_sawback (pRExC_state->sawback)
197 #define RExC_seen (pRExC_state->seen)
198 #define RExC_size (pRExC_state->size)
199 #define RExC_npar (pRExC_state->npar)
200 #define RExC_nestroot (pRExC_state->nestroot)
201 #define RExC_extralen (pRExC_state->extralen)
202 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
203 #define RExC_utf8 (pRExC_state->utf8)
204 #define RExC_uni_semantics (pRExC_state->uni_semantics)
205 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
206 #define RExC_open_parens (pRExC_state->open_parens)
207 #define RExC_close_parens (pRExC_state->close_parens)
208 #define RExC_opend (pRExC_state->opend)
209 #define RExC_paren_names (pRExC_state->paren_names)
210 #define RExC_recurse (pRExC_state->recurse)
211 #define RExC_recurse_count (pRExC_state->recurse_count)
212 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
213 #define RExC_contains_locale (pRExC_state->contains_locale)
214 #define RExC_override_recoding (pRExC_state->override_recoding)
215 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
218 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
219 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
220 ((*s) == '{' && regcurly(s, FALSE)))
223 #undef SPSTART /* dratted cpp namespace... */
226 * Flags to be passed up and down.
228 #define WORST 0 /* Worst case. */
229 #define HASWIDTH 0x01 /* Known to match non-null strings. */
231 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
232 * character. (There needs to be a case: in the switch statement in regexec.c
233 * for any node marked SIMPLE.) Note that this is not the same thing as
236 #define SPSTART 0x04 /* Starts with * or + */
237 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
238 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
239 #define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
241 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
243 /* whether trie related optimizations are enabled */
244 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
245 #define TRIE_STUDY_OPT
246 #define FULL_TRIE_STUDY
252 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
253 #define PBITVAL(paren) (1 << ((paren) & 7))
254 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
255 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
256 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
258 #define REQUIRE_UTF8 STMT_START { \
260 *flagp = RESTART_UTF8; \
265 /* This converts the named class defined in regcomp.h to its equivalent class
266 * number defined in handy.h. */
267 #define namedclass_to_classnum(class) ((int) ((class) / 2))
268 #define classnum_to_namedclass(classnum) ((classnum) * 2)
270 /* About scan_data_t.
272 During optimisation we recurse through the regexp program performing
273 various inplace (keyhole style) optimisations. In addition study_chunk
274 and scan_commit populate this data structure with information about
275 what strings MUST appear in the pattern. We look for the longest
276 string that must appear at a fixed location, and we look for the
277 longest string that may appear at a floating location. So for instance
282 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
283 strings (because they follow a .* construct). study_chunk will identify
284 both FOO and BAR as being the longest fixed and floating strings respectively.
286 The strings can be composites, for instance
290 will result in a composite fixed substring 'foo'.
292 For each string some basic information is maintained:
294 - offset or min_offset
295 This is the position the string must appear at, or not before.
296 It also implicitly (when combined with minlenp) tells us how many
297 characters must match before the string we are searching for.
298 Likewise when combined with minlenp and the length of the string it
299 tells us how many characters must appear after the string we have
303 Only used for floating strings. This is the rightmost point that
304 the string can appear at. If set to SSize_t_MAX it indicates that the
305 string can occur infinitely far to the right.
308 A pointer to the minimum number of characters of the pattern that the
309 string was found inside. This is important as in the case of positive
310 lookahead or positive lookbehind we can have multiple patterns
315 The minimum length of the pattern overall is 3, the minimum length
316 of the lookahead part is 3, but the minimum length of the part that
317 will actually match is 1. So 'FOO's minimum length is 3, but the
318 minimum length for the F is 1. This is important as the minimum length
319 is used to determine offsets in front of and behind the string being
320 looked for. Since strings can be composites this is the length of the
321 pattern at the time it was committed with a scan_commit. Note that
322 the length is calculated by study_chunk, so that the minimum lengths
323 are not known until the full pattern has been compiled, thus the
324 pointer to the value.
328 In the case of lookbehind the string being searched for can be
329 offset past the start point of the final matching string.
330 If this value was just blithely removed from the min_offset it would
331 invalidate some of the calculations for how many chars must match
332 before or after (as they are derived from min_offset and minlen and
333 the length of the string being searched for).
334 When the final pattern is compiled and the data is moved from the
335 scan_data_t structure into the regexp structure the information
336 about lookbehind is factored in, with the information that would
337 have been lost precalculated in the end_shift field for the
340 The fields pos_min and pos_delta are used to store the minimum offset
341 and the delta to the maximum offset at the current point in the pattern.
345 typedef struct scan_data_t {
346 /*I32 len_min; unused */
347 /*I32 len_delta; unused */
351 SSize_t last_end; /* min value, <0 unless valid. */
352 SSize_t last_start_min;
353 SSize_t last_start_max;
354 SV **longest; /* Either &l_fixed, or &l_float. */
355 SV *longest_fixed; /* longest fixed string found in pattern */
356 SSize_t offset_fixed; /* offset where it starts */
357 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
358 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
359 SV *longest_float; /* longest floating string found in pattern */
360 SSize_t offset_float_min; /* earliest point in string it can appear */
361 SSize_t offset_float_max; /* latest point in string it can appear */
362 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
363 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
366 SSize_t *last_closep;
367 struct regnode_charclass_class *start_class;
370 /* The below is perhaps overboard, but this allows us to save a test at the
371 * expense of a mask. This is because on both EBCDIC and ASCII machines, 'A'
372 * and 'a' differ by a single bit; the same with the upper and lower case of
373 * all other ASCII-range alphabetics. On ASCII platforms, they are 32 apart;
374 * on EBCDIC, they are 64. This uses an exclusive 'or' to find that bit and
375 * then inverts it to form a mask, with just a single 0, in the bit position
376 * where the upper- and lowercase differ. XXX There are about 40 other
377 * instances in the Perl core where this micro-optimization could be used.
378 * Should decide if maintenance cost is worse, before changing those
380 * Returns a boolean as to whether or not 'v' is either a lowercase or
381 * uppercase instance of 'c', where 'c' is in [A-Za-z]. If 'c' is a
382 * compile-time constant, the generated code is better than some optimizing
383 * compilers figure out, amounting to a mask and test. The results are
384 * meaningless if 'c' is not one of [A-Za-z] */
385 #define isARG2_lower_or_UPPER_ARG1(c, v) \
386 (((v) & ~('A' ^ 'a')) == ((c) & ~('A' ^ 'a')))
389 * Forward declarations for pregcomp()'s friends.
392 static const scan_data_t zero_scan_data =
393 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
395 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
396 #define SF_BEFORE_SEOL 0x0001
397 #define SF_BEFORE_MEOL 0x0002
398 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
399 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
402 # define SF_FIX_SHIFT_EOL (0+2)
403 # define SF_FL_SHIFT_EOL (0+4)
405 # define SF_FIX_SHIFT_EOL (+2)
406 # define SF_FL_SHIFT_EOL (+4)
409 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
410 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
412 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
413 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
414 #define SF_IS_INF 0x0040
415 #define SF_HAS_PAR 0x0080
416 #define SF_IN_PAR 0x0100
417 #define SF_HAS_EVAL 0x0200
418 #define SCF_DO_SUBSTR 0x0400
419 #define SCF_DO_STCLASS_AND 0x0800
420 #define SCF_DO_STCLASS_OR 0x1000
421 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
422 #define SCF_WHILEM_VISITED_POS 0x2000
424 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
425 #define SCF_SEEN_ACCEPT 0x8000
426 #define SCF_TRIE_DOING_RESTUDY 0x10000
428 #define UTF cBOOL(RExC_utf8)
430 /* The enums for all these are ordered so things work out correctly */
431 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
432 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
433 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
434 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
435 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
436 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
437 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
439 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
441 #define OOB_NAMEDCLASS -1
443 /* There is no code point that is out-of-bounds, so this is problematic. But
444 * its only current use is to initialize a variable that is always set before
446 #define OOB_UNICODE 0xDEADBEEF
448 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
449 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
452 /* length of regex to show in messages that don't mark a position within */
453 #define RegexLengthToShowInErrorMessages 127
456 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
457 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
458 * op/pragma/warn/regcomp.
460 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
461 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
463 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
466 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
467 * arg. Show regex, up to a maximum length. If it's too long, chop and add
470 #define _FAIL(code) STMT_START { \
471 const char *ellipses = ""; \
472 IV len = RExC_end - RExC_precomp; \
475 SAVEFREESV(RExC_rx_sv); \
476 if (len > RegexLengthToShowInErrorMessages) { \
477 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
478 len = RegexLengthToShowInErrorMessages - 10; \
484 #define FAIL(msg) _FAIL( \
485 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
486 msg, (int)len, RExC_precomp, ellipses))
488 #define FAIL2(msg,arg) _FAIL( \
489 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
490 arg, (int)len, RExC_precomp, ellipses))
493 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
495 #define Simple_vFAIL(m) STMT_START { \
496 const IV offset = RExC_parse - RExC_precomp; \
497 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
498 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
502 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
504 #define vFAIL(m) STMT_START { \
506 SAVEFREESV(RExC_rx_sv); \
511 * Like Simple_vFAIL(), but accepts two arguments.
513 #define Simple_vFAIL2(m,a1) STMT_START { \
514 const IV offset = RExC_parse - RExC_precomp; \
515 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
516 (int)offset, RExC_precomp, RExC_precomp + offset); \
520 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
522 #define vFAIL2(m,a1) STMT_START { \
524 SAVEFREESV(RExC_rx_sv); \
525 Simple_vFAIL2(m, a1); \
530 * Like Simple_vFAIL(), but accepts three arguments.
532 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
533 const IV offset = RExC_parse - RExC_precomp; \
534 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
535 (int)offset, RExC_precomp, RExC_precomp + offset); \
539 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
541 #define vFAIL3(m,a1,a2) STMT_START { \
543 SAVEFREESV(RExC_rx_sv); \
544 Simple_vFAIL3(m, a1, a2); \
548 * Like Simple_vFAIL(), but accepts four arguments.
550 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
551 const IV offset = RExC_parse - RExC_precomp; \
552 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
553 (int)offset, RExC_precomp, RExC_precomp + offset); \
556 #define vFAIL4(m,a1,a2,a3) STMT_START { \
558 SAVEFREESV(RExC_rx_sv); \
559 Simple_vFAIL4(m, a1, a2, a3); \
562 /* m is not necessarily a "literal string", in this macro */
563 #define reg_warn_non_literal_string(loc, m) STMT_START { \
564 const IV offset = loc - RExC_precomp; \
565 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
566 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
569 #define ckWARNreg(loc,m) STMT_START { \
570 const IV offset = loc - RExC_precomp; \
571 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
572 (int)offset, RExC_precomp, RExC_precomp + offset); \
575 #define vWARN_dep(loc, m) STMT_START { \
576 const IV offset = loc - RExC_precomp; \
577 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
578 (int)offset, RExC_precomp, RExC_precomp + offset); \
581 #define ckWARNdep(loc,m) STMT_START { \
582 const IV offset = loc - RExC_precomp; \
583 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
585 (int)offset, RExC_precomp, RExC_precomp + offset); \
588 #define ckWARNregdep(loc,m) STMT_START { \
589 const IV offset = loc - RExC_precomp; \
590 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
592 (int)offset, RExC_precomp, RExC_precomp + offset); \
595 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
596 const IV offset = loc - RExC_precomp; \
597 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
599 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
602 #define ckWARN2reg(loc, m, a1) STMT_START { \
603 const IV offset = loc - RExC_precomp; \
604 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
605 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
608 #define vWARN3(loc, m, a1, a2) STMT_START { \
609 const IV offset = loc - RExC_precomp; \
610 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
611 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
614 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
615 const IV offset = loc - RExC_precomp; \
616 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
617 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
620 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
621 const IV offset = loc - RExC_precomp; \
622 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
623 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
626 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
627 const IV offset = loc - RExC_precomp; \
628 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
629 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
632 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
633 const IV offset = loc - RExC_precomp; \
634 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
635 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
639 /* Allow for side effects in s */
640 #define REGC(c,s) STMT_START { \
641 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
644 /* Macros for recording node offsets. 20001227 mjd@plover.com
645 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
646 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
647 * Element 0 holds the number n.
648 * Position is 1 indexed.
650 #ifndef RE_TRACK_PATTERN_OFFSETS
651 #define Set_Node_Offset_To_R(node,byte)
652 #define Set_Node_Offset(node,byte)
653 #define Set_Cur_Node_Offset
654 #define Set_Node_Length_To_R(node,len)
655 #define Set_Node_Length(node,len)
656 #define Set_Node_Cur_Length(node,start)
657 #define Node_Offset(n)
658 #define Node_Length(n)
659 #define Set_Node_Offset_Length(node,offset,len)
660 #define ProgLen(ri) ri->u.proglen
661 #define SetProgLen(ri,x) ri->u.proglen = x
663 #define ProgLen(ri) ri->u.offsets[0]
664 #define SetProgLen(ri,x) ri->u.offsets[0] = x
665 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
667 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
668 __LINE__, (int)(node), (int)(byte))); \
670 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
672 RExC_offsets[2*(node)-1] = (byte); \
677 #define Set_Node_Offset(node,byte) \
678 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
679 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
681 #define Set_Node_Length_To_R(node,len) STMT_START { \
683 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
684 __LINE__, (int)(node), (int)(len))); \
686 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
688 RExC_offsets[2*(node)] = (len); \
693 #define Set_Node_Length(node,len) \
694 Set_Node_Length_To_R((node)-RExC_emit_start, len)
695 #define Set_Node_Cur_Length(node, start) \
696 Set_Node_Length(node, RExC_parse - start)
698 /* Get offsets and lengths */
699 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
700 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
702 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
703 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
704 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
708 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
709 #define EXPERIMENTAL_INPLACESCAN
710 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
712 #define DEBUG_STUDYDATA(str,data,depth) \
713 DEBUG_OPTIMISE_MORE_r(if(data){ \
714 PerlIO_printf(Perl_debug_log, \
715 "%*s" str "Pos:%"IVdf"/%"IVdf \
716 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
717 (int)(depth)*2, "", \
718 (IV)((data)->pos_min), \
719 (IV)((data)->pos_delta), \
720 (UV)((data)->flags), \
721 (IV)((data)->whilem_c), \
722 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
723 is_inf ? "INF " : "" \
725 if ((data)->last_found) \
726 PerlIO_printf(Perl_debug_log, \
727 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
728 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
729 SvPVX_const((data)->last_found), \
730 (IV)((data)->last_end), \
731 (IV)((data)->last_start_min), \
732 (IV)((data)->last_start_max), \
733 ((data)->longest && \
734 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
735 SvPVX_const((data)->longest_fixed), \
736 (IV)((data)->offset_fixed), \
737 ((data)->longest && \
738 (data)->longest==&((data)->longest_float)) ? "*" : "", \
739 SvPVX_const((data)->longest_float), \
740 (IV)((data)->offset_float_min), \
741 (IV)((data)->offset_float_max) \
743 PerlIO_printf(Perl_debug_log,"\n"); \
746 /* Mark that we cannot extend a found fixed substring at this point.
747 Update the longest found anchored substring and the longest found
748 floating substrings if needed. */
751 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
752 SSize_t *minlenp, int is_inf)
754 const STRLEN l = CHR_SVLEN(data->last_found);
755 const STRLEN old_l = CHR_SVLEN(*data->longest);
756 GET_RE_DEBUG_FLAGS_DECL;
758 PERL_ARGS_ASSERT_SCAN_COMMIT;
760 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
761 SvSetMagicSV(*data->longest, data->last_found);
762 if (*data->longest == data->longest_fixed) {
763 data->offset_fixed = l ? data->last_start_min : data->pos_min;
764 if (data->flags & SF_BEFORE_EOL)
766 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
768 data->flags &= ~SF_FIX_BEFORE_EOL;
769 data->minlen_fixed=minlenp;
770 data->lookbehind_fixed=0;
772 else { /* *data->longest == data->longest_float */
773 data->offset_float_min = l ? data->last_start_min : data->pos_min;
774 data->offset_float_max = (l
775 ? data->last_start_max
776 : (data->pos_delta == SSize_t_MAX
778 : data->pos_min + data->pos_delta));
780 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
781 data->offset_float_max = SSize_t_MAX;
782 if (data->flags & SF_BEFORE_EOL)
784 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
786 data->flags &= ~SF_FL_BEFORE_EOL;
787 data->minlen_float=minlenp;
788 data->lookbehind_float=0;
791 SvCUR_set(data->last_found, 0);
793 SV * const sv = data->last_found;
794 if (SvUTF8(sv) && SvMAGICAL(sv)) {
795 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
801 data->flags &= ~SF_BEFORE_EOL;
802 DEBUG_STUDYDATA("commit: ",data,0);
805 /* These macros set, clear and test whether the synthetic start class ('ssc',
806 * given by the parameter) matches an empty string (EOS). This uses the
807 * 'next_off' field in the node, to save a bit in the flags field. The ssc
808 * stands alone, so there is never a next_off, so this field is otherwise
809 * unused. The EOS information is used only for compilation, but theoretically
810 * it could be passed on to the execution code. This could be used to store
811 * more than one bit of information, but only this one is currently used. */
812 #define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
813 #define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
814 #define TEST_SSC_EOS(node) cBOOL((node)->next_off)
816 /* Can match anything (initialization) */
818 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
820 PERL_ARGS_ASSERT_CL_ANYTHING;
822 ANYOF_BITMAP_SETALL(cl);
823 cl->flags = ANYOF_UNICODE_ALL;
826 /* If any portion of the regex is to operate under locale rules,
827 * initialization includes it. The reason this isn't done for all regexes
828 * is that the optimizer was written under the assumption that locale was
829 * all-or-nothing. Given the complexity and lack of documentation in the
830 * optimizer, and that there are inadequate test cases for locale, so many
831 * parts of it may not work properly, it is safest to avoid locale unless
833 if (RExC_contains_locale) {
834 ANYOF_CLASS_SETALL(cl); /* /l uses class */
835 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
838 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
842 /* Can match anything (initialization) */
844 S_cl_is_anything(const struct regnode_charclass_class *cl)
848 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
850 for (value = 0; value < ANYOF_MAX; value += 2)
851 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
853 if (!(cl->flags & ANYOF_UNICODE_ALL))
855 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
860 /* Can match anything (initialization) */
862 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
864 PERL_ARGS_ASSERT_CL_INIT;
866 Zero(cl, 1, struct regnode_charclass_class);
868 cl_anything(pRExC_state, cl);
869 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
872 /* These two functions currently do the exact same thing */
873 #define cl_init_zero cl_init
875 /* 'AND' a given class with another one. Can create false positives. 'cl'
876 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
877 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
879 S_cl_and(struct regnode_charclass_class *cl,
880 const struct regnode_charclass_class *and_with)
882 PERL_ARGS_ASSERT_CL_AND;
884 assert(PL_regkind[and_with->type] == ANYOF);
886 /* I (khw) am not sure all these restrictions are necessary XXX */
887 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
888 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
889 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
890 && !(and_with->flags & ANYOF_LOC_FOLD)
891 && !(cl->flags & ANYOF_LOC_FOLD)) {
894 if (and_with->flags & ANYOF_INVERT)
895 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
896 cl->bitmap[i] &= ~and_with->bitmap[i];
898 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
899 cl->bitmap[i] &= and_with->bitmap[i];
900 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
902 if (and_with->flags & ANYOF_INVERT) {
904 /* Here, the and'ed node is inverted. Get the AND of the flags that
905 * aren't affected by the inversion. Those that are affected are
906 * handled individually below */
907 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
908 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
909 cl->flags |= affected_flags;
911 /* We currently don't know how to deal with things that aren't in the
912 * bitmap, but we know that the intersection is no greater than what
913 * is already in cl, so let there be false positives that get sorted
914 * out after the synthetic start class succeeds, and the node is
915 * matched for real. */
917 /* The inversion of these two flags indicate that the resulting
918 * intersection doesn't have them */
919 if (and_with->flags & ANYOF_UNICODE_ALL) {
920 cl->flags &= ~ANYOF_UNICODE_ALL;
922 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
923 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
926 else { /* and'd node is not inverted */
927 U8 outside_bitmap_but_not_utf8; /* Temp variable */
929 if (! ANYOF_NONBITMAP(and_with)) {
931 /* Here 'and_with' doesn't match anything outside the bitmap
932 * (except possibly ANYOF_UNICODE_ALL), which means the
933 * intersection can't either, except for ANYOF_UNICODE_ALL, in
934 * which case we don't know what the intersection is, but it's no
935 * greater than what cl already has, so can just leave it alone,
936 * with possible false positives */
937 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
938 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
939 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
942 else if (! ANYOF_NONBITMAP(cl)) {
944 /* Here, 'and_with' does match something outside the bitmap, and cl
945 * doesn't have a list of things to match outside the bitmap. If
946 * cl can match all code points above 255, the intersection will
947 * be those above-255 code points that 'and_with' matches. If cl
948 * can't match all Unicode code points, it means that it can't
949 * match anything outside the bitmap (since the 'if' that got us
950 * into this block tested for that), so we leave the bitmap empty.
952 if (cl->flags & ANYOF_UNICODE_ALL) {
953 ARG_SET(cl, ARG(and_with));
955 /* and_with's ARG may match things that don't require UTF8.
956 * And now cl's will too, in spite of this being an 'and'. See
957 * the comments below about the kludge */
958 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
962 /* Here, both 'and_with' and cl match something outside the
963 * bitmap. Currently we do not do the intersection, so just match
964 * whatever cl had at the beginning. */
968 /* Take the intersection of the two sets of flags. However, the
969 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
970 * kludge around the fact that this flag is not treated like the others
971 * which are initialized in cl_anything(). The way the optimizer works
972 * is that the synthetic start class (SSC) is initialized to match
973 * anything, and then the first time a real node is encountered, its
974 * values are AND'd with the SSC's with the result being the values of
975 * the real node. However, there are paths through the optimizer where
976 * the AND never gets called, so those initialized bits are set
977 * inappropriately, which is not usually a big deal, as they just cause
978 * false positives in the SSC, which will just mean a probably
979 * imperceptible slow down in execution. However this bit has a
980 * higher false positive consequence in that it can cause utf8.pm,
981 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
982 * bigger slowdown and also causes significant extra memory to be used.
983 * In order to prevent this, the code now takes a different tack. The
984 * bit isn't set unless some part of the regular expression needs it,
985 * but once set it won't get cleared. This means that these extra
986 * modules won't get loaded unless there was some path through the
987 * pattern that would have required them anyway, and so any false
988 * positives that occur by not ANDing them out when they could be
989 * aren't as severe as they would be if we treated this bit like all
991 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
992 & ANYOF_NONBITMAP_NON_UTF8;
993 cl->flags &= and_with->flags;
994 cl->flags |= outside_bitmap_but_not_utf8;
998 /* 'OR' a given class with another one. Can create false positives. 'cl'
999 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
1000 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
1002 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
1004 PERL_ARGS_ASSERT_CL_OR;
1006 if (or_with->flags & ANYOF_INVERT) {
1008 /* Here, the or'd node is to be inverted. This means we take the
1009 * complement of everything not in the bitmap, but currently we don't
1010 * know what that is, so give up and match anything */
1011 if (ANYOF_NONBITMAP(or_with)) {
1012 cl_anything(pRExC_state, cl);
1015 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
1016 * <= (B1 | !B2) | (CL1 | !CL2)
1017 * which is wasteful if CL2 is small, but we ignore CL2:
1018 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
1019 * XXXX Can we handle case-fold? Unclear:
1020 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
1021 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
1023 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1024 && !(or_with->flags & ANYOF_LOC_FOLD)
1025 && !(cl->flags & ANYOF_LOC_FOLD) ) {
1028 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1029 cl->bitmap[i] |= ~or_with->bitmap[i];
1030 } /* XXXX: logic is complicated otherwise */
1032 cl_anything(pRExC_state, cl);
1035 /* And, we can just take the union of the flags that aren't affected
1036 * by the inversion */
1037 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1039 /* For the remaining flags:
1040 ANYOF_UNICODE_ALL and inverted means to not match anything above
1041 255, which means that the union with cl should just be
1042 what cl has in it, so can ignore this flag
1043 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
1044 is (ASCII) 127-255 to match them, but then invert that, so
1045 the union with cl should just be what cl has in it, so can
1048 } else { /* 'or_with' is not inverted */
1049 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1050 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1051 && (!(or_with->flags & ANYOF_LOC_FOLD)
1052 || (cl->flags & ANYOF_LOC_FOLD)) ) {
1055 /* OR char bitmap and class bitmap separately */
1056 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1057 cl->bitmap[i] |= or_with->bitmap[i];
1058 if (or_with->flags & ANYOF_CLASS) {
1059 ANYOF_CLASS_OR(or_with, cl);
1062 else { /* XXXX: logic is complicated, leave it along for a moment. */
1063 cl_anything(pRExC_state, cl);
1066 if (ANYOF_NONBITMAP(or_with)) {
1068 /* Use the added node's outside-the-bit-map match if there isn't a
1069 * conflict. If there is a conflict (both nodes match something
1070 * outside the bitmap, but what they match outside is not the same
1071 * pointer, and hence not easily compared until XXX we extend
1072 * inversion lists this far), give up and allow the start class to
1073 * match everything outside the bitmap. If that stuff is all above
1074 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1075 if (! ANYOF_NONBITMAP(cl)) {
1076 ARG_SET(cl, ARG(or_with));
1078 else if (ARG(cl) != ARG(or_with)) {
1080 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1081 cl_anything(pRExC_state, cl);
1084 cl->flags |= ANYOF_UNICODE_ALL;
1089 /* Take the union */
1090 cl->flags |= or_with->flags;
1094 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1095 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1096 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1097 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1102 dump_trie(trie,widecharmap,revcharmap)
1103 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1104 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1106 These routines dump out a trie in a somewhat readable format.
1107 The _interim_ variants are used for debugging the interim
1108 tables that are used to generate the final compressed
1109 representation which is what dump_trie expects.
1111 Part of the reason for their existence is to provide a form
1112 of documentation as to how the different representations function.
1117 Dumps the final compressed table form of the trie to Perl_debug_log.
1118 Used for debugging make_trie().
1122 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1123 AV *revcharmap, U32 depth)
1126 SV *sv=sv_newmortal();
1127 int colwidth= widecharmap ? 6 : 4;
1129 GET_RE_DEBUG_FLAGS_DECL;
1131 PERL_ARGS_ASSERT_DUMP_TRIE;
1133 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1134 (int)depth * 2 + 2,"",
1135 "Match","Base","Ofs" );
1137 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1138 SV ** const tmp = av_fetch( revcharmap, state, 0);
1140 PerlIO_printf( Perl_debug_log, "%*s",
1142 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1143 PL_colors[0], PL_colors[1],
1144 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1145 PERL_PV_ESCAPE_FIRSTCHAR
1150 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1151 (int)depth * 2 + 2,"");
1153 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1154 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1155 PerlIO_printf( Perl_debug_log, "\n");
1157 for( state = 1 ; state < trie->statecount ; state++ ) {
1158 const U32 base = trie->states[ state ].trans.base;
1160 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1162 if ( trie->states[ state ].wordnum ) {
1163 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1165 PerlIO_printf( Perl_debug_log, "%6s", "" );
1168 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1173 while( ( base + ofs < trie->uniquecharcount ) ||
1174 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1175 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1178 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1180 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1181 if ( ( base + ofs >= trie->uniquecharcount ) &&
1182 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1183 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1185 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1187 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1189 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1193 PerlIO_printf( Perl_debug_log, "]");
1196 PerlIO_printf( Perl_debug_log, "\n" );
1198 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1199 for (word=1; word <= trie->wordcount; word++) {
1200 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1201 (int)word, (int)(trie->wordinfo[word].prev),
1202 (int)(trie->wordinfo[word].len));
1204 PerlIO_printf(Perl_debug_log, "\n" );
1207 Dumps a fully constructed but uncompressed trie in list form.
1208 List tries normally only are used for construction when the number of
1209 possible chars (trie->uniquecharcount) is very high.
1210 Used for debugging make_trie().
1213 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1214 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1218 SV *sv=sv_newmortal();
1219 int colwidth= widecharmap ? 6 : 4;
1220 GET_RE_DEBUG_FLAGS_DECL;
1222 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1224 /* print out the table precompression. */
1225 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1226 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1227 "------:-----+-----------------\n" );
1229 for( state=1 ; state < next_alloc ; state ++ ) {
1232 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1233 (int)depth * 2 + 2,"", (UV)state );
1234 if ( ! trie->states[ state ].wordnum ) {
1235 PerlIO_printf( Perl_debug_log, "%5s| ","");
1237 PerlIO_printf( Perl_debug_log, "W%4x| ",
1238 trie->states[ state ].wordnum
1241 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1242 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1244 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1246 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1247 PL_colors[0], PL_colors[1],
1248 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1249 PERL_PV_ESCAPE_FIRSTCHAR
1251 TRIE_LIST_ITEM(state,charid).forid,
1252 (UV)TRIE_LIST_ITEM(state,charid).newstate
1255 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1256 (int)((depth * 2) + 14), "");
1259 PerlIO_printf( Perl_debug_log, "\n");
1264 Dumps a fully constructed but uncompressed trie in table form.
1265 This is the normal DFA style state transition table, with a few
1266 twists to facilitate compression later.
1267 Used for debugging make_trie().
1270 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1271 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1276 SV *sv=sv_newmortal();
1277 int colwidth= widecharmap ? 6 : 4;
1278 GET_RE_DEBUG_FLAGS_DECL;
1280 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1283 print out the table precompression so that we can do a visual check
1284 that they are identical.
1287 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1289 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1290 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1292 PerlIO_printf( Perl_debug_log, "%*s",
1294 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1295 PL_colors[0], PL_colors[1],
1296 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1297 PERL_PV_ESCAPE_FIRSTCHAR
1303 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1305 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1306 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1309 PerlIO_printf( Perl_debug_log, "\n" );
1311 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1313 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1314 (int)depth * 2 + 2,"",
1315 (UV)TRIE_NODENUM( state ) );
1317 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1318 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1320 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1322 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1324 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1325 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1327 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1328 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1336 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1337 startbranch: the first branch in the whole branch sequence
1338 first : start branch of sequence of branch-exact nodes.
1339 May be the same as startbranch
1340 last : Thing following the last branch.
1341 May be the same as tail.
1342 tail : item following the branch sequence
1343 count : words in the sequence
1344 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1345 depth : indent depth
1347 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1349 A trie is an N'ary tree where the branches are determined by digital
1350 decomposition of the key. IE, at the root node you look up the 1st character and
1351 follow that branch repeat until you find the end of the branches. Nodes can be
1352 marked as "accepting" meaning they represent a complete word. Eg:
1356 would convert into the following structure. Numbers represent states, letters
1357 following numbers represent valid transitions on the letter from that state, if
1358 the number is in square brackets it represents an accepting state, otherwise it
1359 will be in parenthesis.
1361 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1365 (1) +-i->(6)-+-s->[7]
1367 +-s->(3)-+-h->(4)-+-e->[5]
1369 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1371 This shows that when matching against the string 'hers' we will begin at state 1
1372 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1373 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1374 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1375 single traverse. We store a mapping from accepting to state to which word was
1376 matched, and then when we have multiple possibilities we try to complete the
1377 rest of the regex in the order in which they occured in the alternation.
1379 The only prior NFA like behaviour that would be changed by the TRIE support is
1380 the silent ignoring of duplicate alternations which are of the form:
1382 / (DUPE|DUPE) X? (?{ ... }) Y /x
1384 Thus EVAL blocks following a trie may be called a different number of times with
1385 and without the optimisation. With the optimisations dupes will be silently
1386 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1387 the following demonstrates:
1389 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1391 which prints out 'word' three times, but
1393 'words'=~/(word|word|word)(?{ print $1 })S/
1395 which doesnt print it out at all. This is due to other optimisations kicking in.
1397 Example of what happens on a structural level:
1399 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1401 1: CURLYM[1] {1,32767}(18)
1412 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1413 and should turn into:
1415 1: CURLYM[1] {1,32767}(18)
1417 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1425 Cases where tail != last would be like /(?foo|bar)baz/:
1435 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1436 and would end up looking like:
1439 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1446 d = uvchr_to_utf8_flags(d, uv, 0);
1448 is the recommended Unicode-aware way of saying
1453 #define TRIE_STORE_REVCHAR(val) \
1456 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1457 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1458 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
1459 SvCUR_set(zlopp, kapow - flrbbbbb); \
1462 av_push(revcharmap, zlopp); \
1464 char ooooff = (char)val; \
1465 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1469 /* This gets the next character from the input, folding it if not already
1471 #define TRIE_READ_CHAR STMT_START { \
1474 /* if it is UTF then it is either already folded, or does not need \
1476 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1478 else if (folder == PL_fold_latin1) { \
1479 /* This folder implies Unicode rules, which in the range expressible \
1480 * by not UTF is the lower case, with the two exceptions, one of \
1481 * which should have been taken care of before calling this */ \
1482 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
1483 uvc = toLOWER_L1(*uc); \
1484 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
1487 /* raw data, will be folded later if needed */ \
1495 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1496 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1497 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1498 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1500 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1501 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1502 TRIE_LIST_CUR( state )++; \
1505 #define TRIE_LIST_NEW(state) STMT_START { \
1506 Newxz( trie->states[ state ].trans.list, \
1507 4, reg_trie_trans_le ); \
1508 TRIE_LIST_CUR( state ) = 1; \
1509 TRIE_LIST_LEN( state ) = 4; \
1512 #define TRIE_HANDLE_WORD(state) STMT_START { \
1513 U16 dupe= trie->states[ state ].wordnum; \
1514 regnode * const noper_next = regnext( noper ); \
1517 /* store the word for dumping */ \
1519 if (OP(noper) != NOTHING) \
1520 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1522 tmp = newSVpvn_utf8( "", 0, UTF ); \
1523 av_push( trie_words, tmp ); \
1527 trie->wordinfo[curword].prev = 0; \
1528 trie->wordinfo[curword].len = wordlen; \
1529 trie->wordinfo[curword].accept = state; \
1531 if ( noper_next < tail ) { \
1533 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1534 trie->jump[curword] = (U16)(noper_next - convert); \
1536 jumper = noper_next; \
1538 nextbranch= regnext(cur); \
1542 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1543 /* chain, so that when the bits of chain are later */\
1544 /* linked together, the dups appear in the chain */\
1545 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1546 trie->wordinfo[dupe].prev = curword; \
1548 /* we haven't inserted this word yet. */ \
1549 trie->states[ state ].wordnum = curword; \
1554 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1555 ( ( base + charid >= ucharcount \
1556 && base + charid < ubound \
1557 && state == trie->trans[ base - ucharcount + charid ].check \
1558 && trie->trans[ base - ucharcount + charid ].next ) \
1559 ? trie->trans[ base - ucharcount + charid ].next \
1560 : ( state==1 ? special : 0 ) \
1564 #define MADE_JUMP_TRIE 2
1565 #define MADE_EXACT_TRIE 4
1568 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1571 /* first pass, loop through and scan words */
1572 reg_trie_data *trie;
1573 HV *widecharmap = NULL;
1574 AV *revcharmap = newAV();
1576 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1581 regnode *jumper = NULL;
1582 regnode *nextbranch = NULL;
1583 regnode *convert = NULL;
1584 U32 *prev_states; /* temp array mapping each state to previous one */
1585 /* we just use folder as a flag in utf8 */
1586 const U8 * folder = NULL;
1589 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1590 AV *trie_words = NULL;
1591 /* along with revcharmap, this only used during construction but both are
1592 * useful during debugging so we store them in the struct when debugging.
1595 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1596 STRLEN trie_charcount=0;
1598 SV *re_trie_maxbuff;
1599 GET_RE_DEBUG_FLAGS_DECL;
1601 PERL_ARGS_ASSERT_MAKE_TRIE;
1603 PERL_UNUSED_ARG(depth);
1610 case EXACTFU_TRICKYFOLD:
1611 case EXACTFU: folder = PL_fold_latin1; break;
1612 case EXACTF: folder = PL_fold; break;
1613 case EXACTFL: folder = PL_fold_locale; break;
1614 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1617 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1619 trie->startstate = 1;
1620 trie->wordcount = word_count;
1621 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1622 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1624 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1625 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1626 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1629 trie_words = newAV();
1632 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1633 if (!SvIOK(re_trie_maxbuff)) {
1634 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1636 DEBUG_TRIE_COMPILE_r({
1637 PerlIO_printf( Perl_debug_log,
1638 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1639 (int)depth * 2 + 2, "",
1640 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1641 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1645 /* Find the node we are going to overwrite */
1646 if ( first == startbranch && OP( last ) != BRANCH ) {
1647 /* whole branch chain */
1650 /* branch sub-chain */
1651 convert = NEXTOPER( first );
1654 /* -- First loop and Setup --
1656 We first traverse the branches and scan each word to determine if it
1657 contains widechars, and how many unique chars there are, this is
1658 important as we have to build a table with at least as many columns as we
1661 We use an array of integers to represent the character codes 0..255
1662 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1663 native representation of the character value as the key and IV's for the
1666 *TODO* If we keep track of how many times each character is used we can
1667 remap the columns so that the table compression later on is more
1668 efficient in terms of memory by ensuring the most common value is in the
1669 middle and the least common are on the outside. IMO this would be better
1670 than a most to least common mapping as theres a decent chance the most
1671 common letter will share a node with the least common, meaning the node
1672 will not be compressible. With a middle is most common approach the worst
1673 case is when we have the least common nodes twice.
1677 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1678 regnode *noper = NEXTOPER( cur );
1679 const U8 *uc = (U8*)STRING( noper );
1680 const U8 *e = uc + STR_LEN( noper );
1682 U32 wordlen = 0; /* required init */
1683 STRLEN minbytes = 0;
1684 STRLEN maxbytes = 0;
1685 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1687 if (OP(noper) == NOTHING) {
1688 regnode *noper_next= regnext(noper);
1689 if (noper_next != tail && OP(noper_next) == flags) {
1691 uc= (U8*)STRING(noper);
1692 e= uc + STR_LEN(noper);
1693 trie->minlen= STR_LEN(noper);
1700 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1701 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1702 regardless of encoding */
1703 if (OP( noper ) == EXACTFU_SS) {
1704 /* false positives are ok, so just set this */
1705 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
1708 for ( ; uc < e ; uc += len ) {
1709 TRIE_CHARCOUNT(trie)++;
1712 /* Acummulate to the current values, the range in the number of
1713 * bytes that this character could match. The max is presumed to
1714 * be the same as the folded input (which TRIE_READ_CHAR returns),
1715 * except that when this is not in UTF-8, it could be matched
1716 * against a string which is UTF-8, and the variant characters
1717 * could be 2 bytes instead of the 1 here. Likewise, for the
1718 * minimum number of bytes when not folded. When folding, the min
1719 * is assumed to be 1 byte could fold to match the single character
1720 * here, or in the case of a multi-char fold, 1 byte can fold to
1721 * the whole sequence. 'foldlen' is used to denote whether we are
1722 * in such a sequence, skipping the min setting if so. XXX TODO
1723 * Use the exact list of what folds to each character, from
1724 * PL_utf8_foldclosures */
1726 maxbytes += UTF8SKIP(uc);
1728 /* A non-UTF-8 string could be 1 byte to match our 2 */
1729 minbytes += (UTF8_IS_DOWNGRADEABLE_START(*uc))
1735 foldlen -= UTF8SKIP(uc);
1738 foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e);
1744 maxbytes += (UNI_IS_INVARIANT(*uc))
1755 foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e);
1762 U8 folded= folder[ (U8) uvc ];
1763 if ( !trie->charmap[ folded ] ) {
1764 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1765 TRIE_STORE_REVCHAR( folded );
1768 if ( !trie->charmap[ uvc ] ) {
1769 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1770 TRIE_STORE_REVCHAR( uvc );
1773 /* store the codepoint in the bitmap, and its folded
1775 TRIE_BITMAP_SET(trie, uvc);
1777 /* store the folded codepoint */
1778 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1781 /* store first byte of utf8 representation of
1782 variant codepoints */
1783 if (! NATIVE_IS_INVARIANT(uvc)) {
1784 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1787 set_bit = 0; /* We've done our bit :-) */
1792 widecharmap = newHV();
1794 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1797 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1799 if ( !SvTRUE( *svpp ) ) {
1800 sv_setiv( *svpp, ++trie->uniquecharcount );
1801 TRIE_STORE_REVCHAR(uvc);
1805 if( cur == first ) {
1806 trie->minlen = minbytes;
1807 trie->maxlen = maxbytes;
1808 } else if (minbytes < trie->minlen) {
1809 trie->minlen = minbytes;
1810 } else if (maxbytes > trie->maxlen) {
1811 trie->maxlen = maxbytes;
1813 if (OP( noper ) == EXACTFU_SS) {
1814 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1815 if (trie->minlen > 1)
1818 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1819 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1820 * - We assume that any such sequence might match a 2 byte string */
1821 if (trie->minlen > 2 )
1825 } /* end first pass */
1826 DEBUG_TRIE_COMPILE_r(
1827 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1828 (int)depth * 2 + 2,"",
1829 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1830 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1831 (int)trie->minlen, (int)trie->maxlen )
1835 We now know what we are dealing with in terms of unique chars and
1836 string sizes so we can calculate how much memory a naive
1837 representation using a flat table will take. If it's over a reasonable
1838 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1839 conservative but potentially much slower representation using an array
1842 At the end we convert both representations into the same compressed
1843 form that will be used in regexec.c for matching with. The latter
1844 is a form that cannot be used to construct with but has memory
1845 properties similar to the list form and access properties similar
1846 to the table form making it both suitable for fast searches and
1847 small enough that its feasable to store for the duration of a program.
1849 See the comment in the code where the compressed table is produced
1850 inplace from the flat tabe representation for an explanation of how
1851 the compression works.
1856 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1859 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1861 Second Pass -- Array Of Lists Representation
1863 Each state will be represented by a list of charid:state records
1864 (reg_trie_trans_le) the first such element holds the CUR and LEN
1865 points of the allocated array. (See defines above).
1867 We build the initial structure using the lists, and then convert
1868 it into the compressed table form which allows faster lookups
1869 (but cant be modified once converted).
1872 STRLEN transcount = 1;
1874 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1875 "%*sCompiling trie using list compiler\n",
1876 (int)depth * 2 + 2, ""));
1878 trie->states = (reg_trie_state *)
1879 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1880 sizeof(reg_trie_state) );
1884 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1886 regnode *noper = NEXTOPER( cur );
1887 U8 *uc = (U8*)STRING( noper );
1888 const U8 *e = uc + STR_LEN( noper );
1889 U32 state = 1; /* required init */
1890 U16 charid = 0; /* sanity init */
1891 U32 wordlen = 0; /* required init */
1893 if (OP(noper) == NOTHING) {
1894 regnode *noper_next= regnext(noper);
1895 if (noper_next != tail && OP(noper_next) == flags) {
1897 uc= (U8*)STRING(noper);
1898 e= uc + STR_LEN(noper);
1902 if (OP(noper) != NOTHING) {
1903 for ( ; uc < e ; uc += len ) {
1908 charid = trie->charmap[ uvc ];
1910 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1914 charid=(U16)SvIV( *svpp );
1917 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1924 if ( !trie->states[ state ].trans.list ) {
1925 TRIE_LIST_NEW( state );
1927 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1928 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1929 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1934 newstate = next_alloc++;
1935 prev_states[newstate] = state;
1936 TRIE_LIST_PUSH( state, charid, newstate );
1941 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1945 TRIE_HANDLE_WORD(state);
1947 } /* end second pass */
1949 /* next alloc is the NEXT state to be allocated */
1950 trie->statecount = next_alloc;
1951 trie->states = (reg_trie_state *)
1952 PerlMemShared_realloc( trie->states,
1954 * sizeof(reg_trie_state) );
1956 /* and now dump it out before we compress it */
1957 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1958 revcharmap, next_alloc,
1962 trie->trans = (reg_trie_trans *)
1963 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1970 for( state=1 ; state < next_alloc ; state ++ ) {
1974 DEBUG_TRIE_COMPILE_MORE_r(
1975 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1979 if (trie->states[state].trans.list) {
1980 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1984 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1985 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1986 if ( forid < minid ) {
1988 } else if ( forid > maxid ) {
1992 if ( transcount < tp + maxid - minid + 1) {
1994 trie->trans = (reg_trie_trans *)
1995 PerlMemShared_realloc( trie->trans,
1997 * sizeof(reg_trie_trans) );
1998 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
2000 base = trie->uniquecharcount + tp - minid;
2001 if ( maxid == minid ) {
2003 for ( ; zp < tp ; zp++ ) {
2004 if ( ! trie->trans[ zp ].next ) {
2005 base = trie->uniquecharcount + zp - minid;
2006 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
2007 trie->trans[ zp ].check = state;
2013 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
2014 trie->trans[ tp ].check = state;
2019 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2020 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
2021 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
2022 trie->trans[ tid ].check = state;
2024 tp += ( maxid - minid + 1 );
2026 Safefree(trie->states[ state ].trans.list);
2029 DEBUG_TRIE_COMPILE_MORE_r(
2030 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2033 trie->states[ state ].trans.base=base;
2035 trie->lasttrans = tp + 1;
2039 Second Pass -- Flat Table Representation.
2041 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
2042 We know that we will need Charcount+1 trans at most to store the data
2043 (one row per char at worst case) So we preallocate both structures
2044 assuming worst case.
2046 We then construct the trie using only the .next slots of the entry
2049 We use the .check field of the first entry of the node temporarily to
2050 make compression both faster and easier by keeping track of how many non
2051 zero fields are in the node.
2053 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2056 There are two terms at use here: state as a TRIE_NODEIDX() which is a
2057 number representing the first entry of the node, and state as a
2058 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
2059 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
2060 are 2 entrys per node. eg:
2068 The table is internally in the right hand, idx form. However as we also
2069 have to deal with the states array which is indexed by nodenum we have to
2070 use TRIE_NODENUM() to convert.
2073 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2074 "%*sCompiling trie using table compiler\n",
2075 (int)depth * 2 + 2, ""));
2077 trie->trans = (reg_trie_trans *)
2078 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2079 * trie->uniquecharcount + 1,
2080 sizeof(reg_trie_trans) );
2081 trie->states = (reg_trie_state *)
2082 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2083 sizeof(reg_trie_state) );
2084 next_alloc = trie->uniquecharcount + 1;
2087 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2089 regnode *noper = NEXTOPER( cur );
2090 const U8 *uc = (U8*)STRING( noper );
2091 const U8 *e = uc + STR_LEN( noper );
2093 U32 state = 1; /* required init */
2095 U16 charid = 0; /* sanity init */
2096 U32 accept_state = 0; /* sanity init */
2098 U32 wordlen = 0; /* required init */
2100 if (OP(noper) == NOTHING) {
2101 regnode *noper_next= regnext(noper);
2102 if (noper_next != tail && OP(noper_next) == flags) {
2104 uc= (U8*)STRING(noper);
2105 e= uc + STR_LEN(noper);
2109 if ( OP(noper) != NOTHING ) {
2110 for ( ; uc < e ; uc += len ) {
2115 charid = trie->charmap[ uvc ];
2117 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2118 charid = svpp ? (U16)SvIV(*svpp) : 0;
2122 if ( !trie->trans[ state + charid ].next ) {
2123 trie->trans[ state + charid ].next = next_alloc;
2124 trie->trans[ state ].check++;
2125 prev_states[TRIE_NODENUM(next_alloc)]
2126 = TRIE_NODENUM(state);
2127 next_alloc += trie->uniquecharcount;
2129 state = trie->trans[ state + charid ].next;
2131 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2133 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2136 accept_state = TRIE_NODENUM( state );
2137 TRIE_HANDLE_WORD(accept_state);
2139 } /* end second pass */
2141 /* and now dump it out before we compress it */
2142 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2144 next_alloc, depth+1));
2148 * Inplace compress the table.*
2150 For sparse data sets the table constructed by the trie algorithm will
2151 be mostly 0/FAIL transitions or to put it another way mostly empty.
2152 (Note that leaf nodes will not contain any transitions.)
2154 This algorithm compresses the tables by eliminating most such
2155 transitions, at the cost of a modest bit of extra work during lookup:
2157 - Each states[] entry contains a .base field which indicates the
2158 index in the state[] array wheres its transition data is stored.
2160 - If .base is 0 there are no valid transitions from that node.
2162 - If .base is nonzero then charid is added to it to find an entry in
2165 -If trans[states[state].base+charid].check!=state then the
2166 transition is taken to be a 0/Fail transition. Thus if there are fail
2167 transitions at the front of the node then the .base offset will point
2168 somewhere inside the previous nodes data (or maybe even into a node
2169 even earlier), but the .check field determines if the transition is
2173 The following process inplace converts the table to the compressed
2174 table: We first do not compress the root node 1,and mark all its
2175 .check pointers as 1 and set its .base pointer as 1 as well. This
2176 allows us to do a DFA construction from the compressed table later,
2177 and ensures that any .base pointers we calculate later are greater
2180 - We set 'pos' to indicate the first entry of the second node.
2182 - We then iterate over the columns of the node, finding the first and
2183 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2184 and set the .check pointers accordingly, and advance pos
2185 appropriately and repreat for the next node. Note that when we copy
2186 the next pointers we have to convert them from the original
2187 NODEIDX form to NODENUM form as the former is not valid post
2190 - If a node has no transitions used we mark its base as 0 and do not
2191 advance the pos pointer.
2193 - If a node only has one transition we use a second pointer into the
2194 structure to fill in allocated fail transitions from other states.
2195 This pointer is independent of the main pointer and scans forward
2196 looking for null transitions that are allocated to a state. When it
2197 finds one it writes the single transition into the "hole". If the
2198 pointer doesnt find one the single transition is appended as normal.
2200 - Once compressed we can Renew/realloc the structures to release the
2203 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2204 specifically Fig 3.47 and the associated pseudocode.
2208 const U32 laststate = TRIE_NODENUM( next_alloc );
2211 trie->statecount = laststate;
2213 for ( state = 1 ; state < laststate ; state++ ) {
2215 const U32 stateidx = TRIE_NODEIDX( state );
2216 const U32 o_used = trie->trans[ stateidx ].check;
2217 U32 used = trie->trans[ stateidx ].check;
2218 trie->trans[ stateidx ].check = 0;
2220 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2221 if ( flag || trie->trans[ stateidx + charid ].next ) {
2222 if ( trie->trans[ stateidx + charid ].next ) {
2224 for ( ; zp < pos ; zp++ ) {
2225 if ( ! trie->trans[ zp ].next ) {
2229 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2230 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2231 trie->trans[ zp ].check = state;
2232 if ( ++zp > pos ) pos = zp;
2239 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2241 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2242 trie->trans[ pos ].check = state;
2247 trie->lasttrans = pos + 1;
2248 trie->states = (reg_trie_state *)
2249 PerlMemShared_realloc( trie->states, laststate
2250 * sizeof(reg_trie_state) );
2251 DEBUG_TRIE_COMPILE_MORE_r(
2252 PerlIO_printf( Perl_debug_log,
2253 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2254 (int)depth * 2 + 2,"",
2255 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2258 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2261 } /* end table compress */
2263 DEBUG_TRIE_COMPILE_MORE_r(
2264 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2265 (int)depth * 2 + 2, "",
2266 (UV)trie->statecount,
2267 (UV)trie->lasttrans)
2269 /* resize the trans array to remove unused space */
2270 trie->trans = (reg_trie_trans *)
2271 PerlMemShared_realloc( trie->trans, trie->lasttrans
2272 * sizeof(reg_trie_trans) );
2274 { /* Modify the program and insert the new TRIE node */
2275 U8 nodetype =(U8)(flags & 0xFF);
2279 regnode *optimize = NULL;
2280 #ifdef RE_TRACK_PATTERN_OFFSETS
2283 U32 mjd_nodelen = 0;
2284 #endif /* RE_TRACK_PATTERN_OFFSETS */
2285 #endif /* DEBUGGING */
2287 This means we convert either the first branch or the first Exact,
2288 depending on whether the thing following (in 'last') is a branch
2289 or not and whther first is the startbranch (ie is it a sub part of
2290 the alternation or is it the whole thing.)
2291 Assuming its a sub part we convert the EXACT otherwise we convert
2292 the whole branch sequence, including the first.
2294 /* Find the node we are going to overwrite */
2295 if ( first != startbranch || OP( last ) == BRANCH ) {
2296 /* branch sub-chain */
2297 NEXT_OFF( first ) = (U16)(last - first);
2298 #ifdef RE_TRACK_PATTERN_OFFSETS
2300 mjd_offset= Node_Offset((convert));
2301 mjd_nodelen= Node_Length((convert));
2304 /* whole branch chain */
2306 #ifdef RE_TRACK_PATTERN_OFFSETS
2309 const regnode *nop = NEXTOPER( convert );
2310 mjd_offset= Node_Offset((nop));
2311 mjd_nodelen= Node_Length((nop));
2315 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2316 (int)depth * 2 + 2, "",
2317 (UV)mjd_offset, (UV)mjd_nodelen)
2320 /* But first we check to see if there is a common prefix we can
2321 split out as an EXACT and put in front of the TRIE node. */
2322 trie->startstate= 1;
2323 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2325 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2329 const U32 base = trie->states[ state ].trans.base;
2331 if ( trie->states[state].wordnum )
2334 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2335 if ( ( base + ofs >= trie->uniquecharcount ) &&
2336 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2337 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2339 if ( ++count > 1 ) {
2340 SV **tmp = av_fetch( revcharmap, ofs, 0);
2341 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2342 if ( state == 1 ) break;
2344 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2346 PerlIO_printf(Perl_debug_log,
2347 "%*sNew Start State=%"UVuf" Class: [",
2348 (int)depth * 2 + 2, "",
2351 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2352 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2354 TRIE_BITMAP_SET(trie,*ch);
2356 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2358 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2362 TRIE_BITMAP_SET(trie,*ch);
2364 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2365 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2371 SV **tmp = av_fetch( revcharmap, idx, 0);
2373 char *ch = SvPV( *tmp, len );
2375 SV *sv=sv_newmortal();
2376 PerlIO_printf( Perl_debug_log,
2377 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2378 (int)depth * 2 + 2, "",
2380 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2381 PL_colors[0], PL_colors[1],
2382 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2383 PERL_PV_ESCAPE_FIRSTCHAR
2388 OP( convert ) = nodetype;
2389 str=STRING(convert);
2392 STR_LEN(convert) += len;
2398 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2403 trie->prefixlen = (state-1);
2405 regnode *n = convert+NODE_SZ_STR(convert);
2406 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2407 trie->startstate = state;
2408 trie->minlen -= (state - 1);
2409 trie->maxlen -= (state - 1);
2411 /* At least the UNICOS C compiler choked on this
2412 * being argument to DEBUG_r(), so let's just have
2415 #ifdef PERL_EXT_RE_BUILD
2421 regnode *fix = convert;
2422 U32 word = trie->wordcount;
2424 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2425 while( ++fix < n ) {
2426 Set_Node_Offset_Length(fix, 0, 0);
2429 SV ** const tmp = av_fetch( trie_words, word, 0 );
2431 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2432 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2434 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2442 NEXT_OFF(convert) = (U16)(tail - convert);
2443 DEBUG_r(optimize= n);
2449 if ( trie->maxlen ) {
2450 NEXT_OFF( convert ) = (U16)(tail - convert);
2451 ARG_SET( convert, data_slot );
2452 /* Store the offset to the first unabsorbed branch in
2453 jump[0], which is otherwise unused by the jump logic.
2454 We use this when dumping a trie and during optimisation. */
2456 trie->jump[0] = (U16)(nextbranch - convert);
2458 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2459 * and there is a bitmap
2460 * and the first "jump target" node we found leaves enough room
2461 * then convert the TRIE node into a TRIEC node, with the bitmap
2462 * embedded inline in the opcode - this is hypothetically faster.
2464 if ( !trie->states[trie->startstate].wordnum
2466 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2468 OP( convert ) = TRIEC;
2469 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2470 PerlMemShared_free(trie->bitmap);
2473 OP( convert ) = TRIE;
2475 /* store the type in the flags */
2476 convert->flags = nodetype;
2480 + regarglen[ OP( convert ) ];
2482 /* XXX We really should free up the resource in trie now,
2483 as we won't use them - (which resources?) dmq */
2485 /* needed for dumping*/
2486 DEBUG_r(if (optimize) {
2487 regnode *opt = convert;
2489 while ( ++opt < optimize) {
2490 Set_Node_Offset_Length(opt,0,0);
2493 Try to clean up some of the debris left after the
2496 while( optimize < jumper ) {
2497 mjd_nodelen += Node_Length((optimize));
2498 OP( optimize ) = OPTIMIZED;
2499 Set_Node_Offset_Length(optimize,0,0);
2502 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2504 } /* end node insert */
2506 /* Finish populating the prev field of the wordinfo array. Walk back
2507 * from each accept state until we find another accept state, and if
2508 * so, point the first word's .prev field at the second word. If the
2509 * second already has a .prev field set, stop now. This will be the
2510 * case either if we've already processed that word's accept state,
2511 * or that state had multiple words, and the overspill words were
2512 * already linked up earlier.
2519 for (word=1; word <= trie->wordcount; word++) {
2521 if (trie->wordinfo[word].prev)
2523 state = trie->wordinfo[word].accept;
2525 state = prev_states[state];
2528 prev = trie->states[state].wordnum;
2532 trie->wordinfo[word].prev = prev;
2534 Safefree(prev_states);
2538 /* and now dump out the compressed format */
2539 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2541 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2543 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2544 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2546 SvREFCNT_dec_NN(revcharmap);
2550 : trie->startstate>1
2556 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2558 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2560 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2561 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2564 We find the fail state for each state in the trie, this state is the longest proper
2565 suffix of the current state's 'word' that is also a proper prefix of another word in our
2566 trie. State 1 represents the word '' and is thus the default fail state. This allows
2567 the DFA not to have to restart after its tried and failed a word at a given point, it
2568 simply continues as though it had been matching the other word in the first place.
2570 'abcdgu'=~/abcdefg|cdgu/
2571 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2572 fail, which would bring us to the state representing 'd' in the second word where we would
2573 try 'g' and succeed, proceeding to match 'cdgu'.
2575 /* add a fail transition */
2576 const U32 trie_offset = ARG(source);
2577 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2579 const U32 ucharcount = trie->uniquecharcount;
2580 const U32 numstates = trie->statecount;
2581 const U32 ubound = trie->lasttrans + ucharcount;
2585 U32 base = trie->states[ 1 ].trans.base;
2588 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2589 GET_RE_DEBUG_FLAGS_DECL;
2591 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2593 PERL_UNUSED_ARG(depth);
2597 ARG_SET( stclass, data_slot );
2598 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2599 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2600 aho->trie=trie_offset;
2601 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2602 Copy( trie->states, aho->states, numstates, reg_trie_state );
2603 Newxz( q, numstates, U32);
2604 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2607 /* initialize fail[0..1] to be 1 so that we always have
2608 a valid final fail state */
2609 fail[ 0 ] = fail[ 1 ] = 1;
2611 for ( charid = 0; charid < ucharcount ; charid++ ) {
2612 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2614 q[ q_write ] = newstate;
2615 /* set to point at the root */
2616 fail[ q[ q_write++ ] ]=1;
2619 while ( q_read < q_write) {
2620 const U32 cur = q[ q_read++ % numstates ];
2621 base = trie->states[ cur ].trans.base;
2623 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2624 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2626 U32 fail_state = cur;
2629 fail_state = fail[ fail_state ];
2630 fail_base = aho->states[ fail_state ].trans.base;
2631 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2633 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2634 fail[ ch_state ] = fail_state;
2635 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2637 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2639 q[ q_write++ % numstates] = ch_state;
2643 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2644 when we fail in state 1, this allows us to use the
2645 charclass scan to find a valid start char. This is based on the principle
2646 that theres a good chance the string being searched contains lots of stuff
2647 that cant be a start char.
2649 fail[ 0 ] = fail[ 1 ] = 0;
2650 DEBUG_TRIE_COMPILE_r({
2651 PerlIO_printf(Perl_debug_log,
2652 "%*sStclass Failtable (%"UVuf" states): 0",
2653 (int)(depth * 2), "", (UV)numstates
2655 for( q_read=1; q_read<numstates; q_read++ ) {
2656 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2658 PerlIO_printf(Perl_debug_log, "\n");
2661 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2666 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2667 * These need to be revisited when a newer toolchain becomes available.
2669 #if defined(__sparc64__) && defined(__GNUC__)
2670 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2671 # undef SPARC64_GCC_WORKAROUND
2672 # define SPARC64_GCC_WORKAROUND 1
2676 #define DEBUG_PEEP(str,scan,depth) \
2677 DEBUG_OPTIMISE_r({if (scan){ \
2678 SV * const mysv=sv_newmortal(); \
2679 regnode *Next = regnext(scan); \
2680 regprop(RExC_rx, mysv, scan); \
2681 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2682 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2683 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2687 /* The below joins as many adjacent EXACTish nodes as possible into a single
2688 * one. The regop may be changed if the node(s) contain certain sequences that
2689 * require special handling. The joining is only done if:
2690 * 1) there is room in the current conglomerated node to entirely contain the
2692 * 2) they are the exact same node type
2694 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2695 * these get optimized out
2697 * If a node is to match under /i (folded), the number of characters it matches
2698 * can be different than its character length if it contains a multi-character
2699 * fold. *min_subtract is set to the total delta of the input nodes.
2701 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2702 * and contains LATIN SMALL LETTER SHARP S
2704 * This is as good a place as any to discuss the design of handling these
2705 * multi-character fold sequences. It's been wrong in Perl for a very long
2706 * time. There are three code points in Unicode whose multi-character folds
2707 * were long ago discovered to mess things up. The previous designs for
2708 * dealing with these involved assigning a special node for them. This
2709 * approach doesn't work, as evidenced by this example:
2710 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2711 * Both these fold to "sss", but if the pattern is parsed to create a node that
2712 * would match just the \xDF, it won't be able to handle the case where a
2713 * successful match would have to cross the node's boundary. The new approach
2714 * that hopefully generally solves the problem generates an EXACTFU_SS node
2717 * It turns out that there are problems with all multi-character folds, and not
2718 * just these three. Now the code is general, for all such cases, but the
2719 * three still have some special handling. The approach taken is:
2720 * 1) This routine examines each EXACTFish node that could contain multi-
2721 * character fold sequences. It returns in *min_subtract how much to
2722 * subtract from the the actual length of the string to get a real minimum
2723 * match length; it is 0 if there are no multi-char folds. This delta is
2724 * used by the caller to adjust the min length of the match, and the delta
2725 * between min and max, so that the optimizer doesn't reject these
2726 * possibilities based on size constraints.
2727 * 2) Certain of these sequences require special handling by the trie code,
2728 * so, if found, this code changes the joined node type to special ops:
2729 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2730 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2731 * is used for an EXACTFU node that contains at least one "ss" sequence in
2732 * it. For non-UTF-8 patterns and strings, this is the only case where
2733 * there is a possible fold length change. That means that a regular
2734 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2735 * with length changes, and so can be processed faster. regexec.c takes
2736 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2737 * pre-folded by regcomp.c. This saves effort in regex matching.
2738 * However, the pre-folding isn't done for non-UTF8 patterns because the
2739 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2740 * down by forcing the pattern into UTF8 unless necessary. Also what
2741 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2742 * possibilities for the non-UTF8 patterns are quite simple, except for
2743 * the sharp s. All the ones that don't involve a UTF-8 target string are
2744 * members of a fold-pair, and arrays are set up for all of them so that
2745 * the other member of the pair can be found quickly. Code elsewhere in
2746 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2747 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2748 * described in the next item.
2749 * 4) A problem remains for the sharp s in EXACTF and EXACTFA nodes when the
2750 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2751 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2752 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2753 * character in the pattern corresponds to at most a single character in
2754 * the target string. (And I do mean character, and not byte here, unlike
2755 * other parts of the documentation that have never been updated to
2756 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2757 * two character string 'ss'; in EXACTFA nodes it can match
2758 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2759 * instances where it is violated. I'm reluctant to try to change the
2760 * assumption, as the code involved is impenetrable to me (khw), so
2761 * instead the code here punts. This routine examines (when the pattern
2762 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2763 * boolean indicating whether or not the node contains a sharp s. When it
2764 * is true, the caller sets a flag that later causes the optimizer in this
2765 * file to not set values for the floating and fixed string lengths, and
2766 * thus avoids the optimizer code in regexec.c that makes the invalid
2767 * assumption. Thus, there is no optimization based on string lengths for
2768 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2769 * (The reason the assumption is wrong only in these two cases is that all
2770 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2771 * other folds to their expanded versions. We can't prefold sharp s to
2772 * 'ss' in EXACTF nodes because we don't know at compile time if it
2773 * actually matches 'ss' or not. It will match iff the target string is
2774 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2775 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2776 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2777 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2778 * require the pattern to be forced into UTF-8, the overhead of which we
2782 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2783 if (PL_regkind[OP(scan)] == EXACT) \
2784 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2787 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, UV *min_subtract, bool *has_exactf_sharp_s, U32 flags,regnode *val, U32 depth) {
2788 /* Merge several consecutive EXACTish nodes into one. */
2789 regnode *n = regnext(scan);
2791 regnode *next = scan + NODE_SZ_STR(scan);
2795 regnode *stop = scan;
2796 GET_RE_DEBUG_FLAGS_DECL;
2798 PERL_UNUSED_ARG(depth);
2801 PERL_ARGS_ASSERT_JOIN_EXACT;
2802 #ifndef EXPERIMENTAL_INPLACESCAN
2803 PERL_UNUSED_ARG(flags);
2804 PERL_UNUSED_ARG(val);
2806 DEBUG_PEEP("join",scan,depth);
2808 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2809 * EXACT ones that are mergeable to the current one. */
2811 && (PL_regkind[OP(n)] == NOTHING
2812 || (stringok && OP(n) == OP(scan)))
2814 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2817 if (OP(n) == TAIL || n > next)
2819 if (PL_regkind[OP(n)] == NOTHING) {
2820 DEBUG_PEEP("skip:",n,depth);
2821 NEXT_OFF(scan) += NEXT_OFF(n);
2822 next = n + NODE_STEP_REGNODE;
2829 else if (stringok) {
2830 const unsigned int oldl = STR_LEN(scan);
2831 regnode * const nnext = regnext(n);
2833 /* XXX I (khw) kind of doubt that this works on platforms where
2834 * U8_MAX is above 255 because of lots of other assumptions */
2835 /* Don't join if the sum can't fit into a single node */
2836 if (oldl + STR_LEN(n) > U8_MAX)
2839 DEBUG_PEEP("merg",n,depth);
2842 NEXT_OFF(scan) += NEXT_OFF(n);
2843 STR_LEN(scan) += STR_LEN(n);
2844 next = n + NODE_SZ_STR(n);
2845 /* Now we can overwrite *n : */
2846 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2854 #ifdef EXPERIMENTAL_INPLACESCAN
2855 if (flags && !NEXT_OFF(n)) {
2856 DEBUG_PEEP("atch", val, depth);
2857 if (reg_off_by_arg[OP(n)]) {
2858 ARG_SET(n, val - n);
2861 NEXT_OFF(n) = val - n;
2869 *has_exactf_sharp_s = FALSE;
2871 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2872 * can now analyze for sequences of problematic code points. (Prior to
2873 * this final joining, sequences could have been split over boundaries, and
2874 * hence missed). The sequences only happen in folding, hence for any
2875 * non-EXACT EXACTish node */
2876 if (OP(scan) != EXACT) {
2877 const U8 * const s0 = (U8*) STRING(scan);
2879 const U8 * const s_end = s0 + STR_LEN(scan);
2881 /* One pass is made over the node's string looking for all the
2882 * possibilities. to avoid some tests in the loop, there are two main
2883 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2887 /* Examine the string for a multi-character fold sequence. UTF-8
2888 * patterns have all characters pre-folded by the time this code is
2890 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2891 length sequence we are looking for is 2 */
2894 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2895 if (! len) { /* Not a multi-char fold: get next char */
2900 /* Nodes with 'ss' require special handling, except for EXACTFL
2901 * and EXACTFA for which there is no multi-char fold to this */
2902 if (len == 2 && *s == 's' && *(s+1) == 's'
2903 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2906 OP(scan) = EXACTFU_SS;
2909 else if (len == 6 /* len is the same in both ASCII and EBCDIC
2911 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2912 COMBINING_DIAERESIS_UTF8
2913 COMBINING_ACUTE_ACCENT_UTF8,
2915 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2916 COMBINING_DIAERESIS_UTF8
2917 COMBINING_ACUTE_ACCENT_UTF8,
2922 /* These two folds require special handling by trie's, so
2923 * change the node type to indicate this. If EXACTFA and
2924 * EXACTFL were ever to be handled by trie's, this would
2925 * have to be changed. If this node has already been
2926 * changed to EXACTFU_SS in this loop, leave it as is. (I
2927 * (khw) think it doesn't matter in regexec.c for UTF
2928 * patterns, but no need to change it */
2929 if (OP(scan) == EXACTFU) {
2930 OP(scan) = EXACTFU_TRICKYFOLD;
2934 else { /* Here is a generic multi-char fold. */
2935 const U8* multi_end = s + len;
2937 /* Count how many characters in it. In the case of /l and
2938 * /aa, no folds which contain ASCII code points are
2939 * allowed, so check for those, and skip if found. (In
2940 * EXACTFL, no folds are allowed to any Latin1 code point,
2941 * not just ASCII. But there aren't any of these
2942 * currently, nor ever likely, so don't take the time to
2943 * test for them. The code that generates the
2944 * is_MULTI_foo() macros croaks should one actually get put
2945 * into Unicode .) */
2946 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2947 count = utf8_length(s, multi_end);
2951 while (s < multi_end) {
2954 goto next_iteration;
2964 /* The delta is how long the sequence is minus 1 (1 is how long
2965 * the character that folds to the sequence is) */
2966 *min_subtract += count - 1;
2970 else if (OP(scan) == EXACTFA) {
2972 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2973 * fold to the ASCII range (and there are no existing ones in the
2974 * upper latin1 range). But, as outlined in the comments preceding
2975 * this function, we need to flag any occurrences of the sharp s */
2977 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
2978 *has_exactf_sharp_s = TRUE;
2985 else if (OP(scan) != EXACTFL) {
2987 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2988 * multi-char folds that are all Latin1. (This code knows that
2989 * there are no current multi-char folds possible with EXACTFL,
2990 * relying on fold_grind.t to catch any errors if the very unlikely
2991 * event happens that some get added in future Unicode versions.)
2992 * As explained in the comments preceding this function, we look
2993 * also for the sharp s in EXACTF nodes; it can be in the final
2994 * position. Otherwise we can stop looking 1 byte earlier because
2995 * have to find at least two characters for a multi-fold */
2996 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2999 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3000 if (! len) { /* Not a multi-char fold. */
3001 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
3003 *has_exactf_sharp_s = TRUE;
3010 && isARG2_lower_or_UPPER_ARG1('s', *s)
3011 && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
3014 /* EXACTF nodes need to know that the minimum length
3015 * changed so that a sharp s in the string can match this
3016 * ss in the pattern, but they remain EXACTF nodes, as they
3017 * won't match this unless the target string is is UTF-8,
3018 * which we don't know until runtime */
3019 if (OP(scan) != EXACTF) {
3020 OP(scan) = EXACTFU_SS;
3024 *min_subtract += len - 1;
3031 /* Allow dumping but overwriting the collection of skipped
3032 * ops and/or strings with fake optimized ops */
3033 n = scan + NODE_SZ_STR(scan);
3041 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3045 /* REx optimizer. Converts nodes into quicker variants "in place".
3046 Finds fixed substrings. */
3048 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3049 to the position after last scanned or to NULL. */
3051 #define INIT_AND_WITHP \
3052 assert(!and_withp); \
3053 Newx(and_withp,1,struct regnode_charclass_class); \
3054 SAVEFREEPV(and_withp)
3056 /* this is a chain of data about sub patterns we are processing that
3057 need to be handled separately/specially in study_chunk. Its so
3058 we can simulate recursion without losing state. */
3060 typedef struct scan_frame {
3061 regnode *last; /* last node to process in this frame */
3062 regnode *next; /* next node to process when last is reached */
3063 struct scan_frame *prev; /*previous frame*/
3064 I32 stop; /* what stopparen do we use */
3068 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3071 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3072 SSize_t *minlenp, SSize_t *deltap,
3077 struct regnode_charclass_class *and_withp,
3078 U32 flags, U32 depth)
3079 /* scanp: Start here (read-write). */
3080 /* deltap: Write maxlen-minlen here. */
3081 /* last: Stop before this one. */
3082 /* data: string data about the pattern */
3083 /* stopparen: treat close N as END */
3084 /* recursed: which subroutines have we recursed into */
3085 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3088 /* There must be at least this number of characters to match */
3091 regnode *scan = *scanp, *next;
3093 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3094 int is_inf_internal = 0; /* The studied chunk is infinite */
3095 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3096 scan_data_t data_fake;
3097 SV *re_trie_maxbuff = NULL;
3098 regnode *first_non_open = scan;
3099 SSize_t stopmin = SSize_t_MAX;
3100 scan_frame *frame = NULL;
3101 GET_RE_DEBUG_FLAGS_DECL;
3103 PERL_ARGS_ASSERT_STUDY_CHUNK;
3106 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3110 while (first_non_open && OP(first_non_open) == OPEN)
3111 first_non_open=regnext(first_non_open);
3116 while ( scan && OP(scan) != END && scan < last ){
3117 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3118 node length to get a real minimum (because
3119 the folded version may be shorter) */
3120 bool has_exactf_sharp_s = FALSE;
3121 /* Peephole optimizer: */
3122 DEBUG_STUDYDATA("Peep:", data,depth);
3123 DEBUG_PEEP("Peep",scan,depth);
3125 /* Its not clear to khw or hv why this is done here, and not in the
3126 * clauses that deal with EXACT nodes. khw's guess is that it's
3127 * because of a previous design */
3128 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3130 /* Follow the next-chain of the current node and optimize
3131 away all the NOTHINGs from it. */
3132 if (OP(scan) != CURLYX) {
3133 const int max = (reg_off_by_arg[OP(scan)]
3135 /* I32 may be smaller than U16 on CRAYs! */
3136 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3137 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3141 /* Skip NOTHING and LONGJMP. */
3142 while ((n = regnext(n))
3143 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3144 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3145 && off + noff < max)
3147 if (reg_off_by_arg[OP(scan)])
3150 NEXT_OFF(scan) = off;
3155 /* The principal pseudo-switch. Cannot be a switch, since we
3156 look into several different things. */
3157 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3158 || OP(scan) == IFTHEN) {
3159 next = regnext(scan);
3161 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3163 if (OP(next) == code || code == IFTHEN) {
3164 /* NOTE - There is similar code to this block below for handling
3165 TRIE nodes on a re-study. If you change stuff here check there
3167 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3168 struct regnode_charclass_class accum;
3169 regnode * const startbranch=scan;
3171 if (flags & SCF_DO_SUBSTR)
3172 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3173 if (flags & SCF_DO_STCLASS)
3174 cl_init_zero(pRExC_state, &accum);
3176 while (OP(scan) == code) {
3177 SSize_t deltanext, minnext, fake;
3179 struct regnode_charclass_class this_class;
3182 data_fake.flags = 0;
3184 data_fake.whilem_c = data->whilem_c;
3185 data_fake.last_closep = data->last_closep;
3188 data_fake.last_closep = &fake;
3190 data_fake.pos_delta = delta;
3191 next = regnext(scan);
3192 scan = NEXTOPER(scan);
3194 scan = NEXTOPER(scan);
3195 if (flags & SCF_DO_STCLASS) {
3196 cl_init(pRExC_state, &this_class);
3197 data_fake.start_class = &this_class;
3198 f = SCF_DO_STCLASS_AND;
3200 if (flags & SCF_WHILEM_VISITED_POS)
3201 f |= SCF_WHILEM_VISITED_POS;
3203 /* we suppose the run is continuous, last=next...*/
3204 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3206 stopparen, recursed, NULL, f,depth+1);
3209 if (deltanext == SSize_t_MAX) {
3210 is_inf = is_inf_internal = 1;
3212 } else if (max1 < minnext + deltanext)
3213 max1 = minnext + deltanext;
3215 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3217 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3218 if ( stopmin > minnext)
3219 stopmin = min + min1;
3220 flags &= ~SCF_DO_SUBSTR;
3222 data->flags |= SCF_SEEN_ACCEPT;
3225 if (data_fake.flags & SF_HAS_EVAL)
3226 data->flags |= SF_HAS_EVAL;
3227 data->whilem_c = data_fake.whilem_c;
3229 if (flags & SCF_DO_STCLASS)
3230 cl_or(pRExC_state, &accum, &this_class);
3232 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3234 if (flags & SCF_DO_SUBSTR) {
3235 data->pos_min += min1;
3236 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3237 data->pos_delta = SSize_t_MAX;
3239 data->pos_delta += max1 - min1;
3240 if (max1 != min1 || is_inf)
3241 data->longest = &(data->longest_float);
3244 if (delta == SSize_t_MAX
3245 || SSize_t_MAX - delta - (max1 - min1) < 0)
3246 delta = SSize_t_MAX;
3248 delta += max1 - min1;
3249 if (flags & SCF_DO_STCLASS_OR) {
3250 cl_or(pRExC_state, data->start_class, &accum);
3252 cl_and(data->start_class, and_withp);
3253 flags &= ~SCF_DO_STCLASS;
3256 else if (flags & SCF_DO_STCLASS_AND) {
3258 cl_and(data->start_class, &accum);
3259 flags &= ~SCF_DO_STCLASS;
3262 /* Switch to OR mode: cache the old value of
3263 * data->start_class */
3265 StructCopy(data->start_class, and_withp,
3266 struct regnode_charclass_class);
3267 flags &= ~SCF_DO_STCLASS_AND;
3268 StructCopy(&accum, data->start_class,
3269 struct regnode_charclass_class);
3270 flags |= SCF_DO_STCLASS_OR;
3271 SET_SSC_EOS(data->start_class);
3275 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3278 Assuming this was/is a branch we are dealing with: 'scan' now
3279 points at the item that follows the branch sequence, whatever
3280 it is. We now start at the beginning of the sequence and look
3287 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3289 If we can find such a subsequence we need to turn the first
3290 element into a trie and then add the subsequent branch exact
3291 strings to the trie.
3295 1. patterns where the whole set of branches can be converted.
3297 2. patterns where only a subset can be converted.
3299 In case 1 we can replace the whole set with a single regop
3300 for the trie. In case 2 we need to keep the start and end
3303 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3304 becomes BRANCH TRIE; BRANCH X;
3306 There is an additional case, that being where there is a
3307 common prefix, which gets split out into an EXACT like node
3308 preceding the TRIE node.
3310 If x(1..n)==tail then we can do a simple trie, if not we make
3311 a "jump" trie, such that when we match the appropriate word
3312 we "jump" to the appropriate tail node. Essentially we turn
3313 a nested if into a case structure of sorts.
3318 if (!re_trie_maxbuff) {
3319 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3320 if (!SvIOK(re_trie_maxbuff))
3321 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3323 if ( SvIV(re_trie_maxbuff)>=0 ) {
3325 regnode *first = (regnode *)NULL;
3326 regnode *last = (regnode *)NULL;
3327 regnode *tail = scan;
3332 SV * const mysv = sv_newmortal(); /* for dumping */
3334 /* var tail is used because there may be a TAIL
3335 regop in the way. Ie, the exacts will point to the
3336 thing following the TAIL, but the last branch will
3337 point at the TAIL. So we advance tail. If we
3338 have nested (?:) we may have to move through several
3342 while ( OP( tail ) == TAIL ) {
3343 /* this is the TAIL generated by (?:) */
3344 tail = regnext( tail );
3348 DEBUG_TRIE_COMPILE_r({
3349 regprop(RExC_rx, mysv, tail );
3350 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3351 (int)depth * 2 + 2, "",
3352 "Looking for TRIE'able sequences. Tail node is: ",
3353 SvPV_nolen_const( mysv )
3359 Step through the branches
3360 cur represents each branch,
3361 noper is the first thing to be matched as part of that branch
3362 noper_next is the regnext() of that node.
3364 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3365 via a "jump trie" but we also support building with NOJUMPTRIE,
3366 which restricts the trie logic to structures like /FOO|BAR/.
3368 If noper is a trieable nodetype then the branch is a possible optimization
3369 target. If we are building under NOJUMPTRIE then we require that noper_next
3370 is the same as scan (our current position in the regex program).
3372 Once we have two or more consecutive such branches we can create a
3373 trie of the EXACT's contents and stitch it in place into the program.
3375 If the sequence represents all of the branches in the alternation we
3376 replace the entire thing with a single TRIE node.
3378 Otherwise when it is a subsequence we need to stitch it in place and
3379 replace only the relevant branches. This means the first branch has
3380 to remain as it is used by the alternation logic, and its next pointer,
3381 and needs to be repointed at the item on the branch chain following
3382 the last branch we have optimized away.
3384 This could be either a BRANCH, in which case the subsequence is internal,
3385 or it could be the item following the branch sequence in which case the
3386 subsequence is at the end (which does not necessarily mean the first node
3387 is the start of the alternation).
3389 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3392 ----------------+-----------
3396 EXACTFU_SS | EXACTFU
3397 EXACTFU_TRICKYFOLD | EXACTFU
3402 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3403 ( EXACT == (X) ) ? EXACT : \
3404 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3407 /* dont use tail as the end marker for this traverse */
3408 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3409 regnode * const noper = NEXTOPER( cur );
3410 U8 noper_type = OP( noper );
3411 U8 noper_trietype = TRIE_TYPE( noper_type );
3412 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3413 regnode * const noper_next = regnext( noper );
3414 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3415 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3418 DEBUG_TRIE_COMPILE_r({
3419 regprop(RExC_rx, mysv, cur);
3420 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3421 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3423 regprop(RExC_rx, mysv, noper);
3424 PerlIO_printf( Perl_debug_log, " -> %s",
3425 SvPV_nolen_const(mysv));
3428 regprop(RExC_rx, mysv, noper_next );
3429 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3430 SvPV_nolen_const(mysv));
3432 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3433 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3434 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3438 /* Is noper a trieable nodetype that can be merged with the
3439 * current trie (if there is one)? */
3443 ( noper_trietype == NOTHING)
3444 || ( trietype == NOTHING )
3445 || ( trietype == noper_trietype )
3448 && noper_next == tail
3452 /* Handle mergable triable node
3453 * Either we are the first node in a new trieable sequence,
3454 * in which case we do some bookkeeping, otherwise we update
3455 * the end pointer. */
3458 if ( noper_trietype == NOTHING ) {
3459 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3460 regnode * const noper_next = regnext( noper );
3461 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3462 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3465 if ( noper_next_trietype ) {
3466 trietype = noper_next_trietype;
3467 } else if (noper_next_type) {
3468 /* a NOTHING regop is 1 regop wide. We need at least two
3469 * for a trie so we can't merge this in */
3473 trietype = noper_trietype;
3476 if ( trietype == NOTHING )
3477 trietype = noper_trietype;
3482 } /* end handle mergable triable node */
3484 /* handle unmergable node -
3485 * noper may either be a triable node which can not be tried
3486 * together with the current trie, or a non triable node */
3488 /* If last is set and trietype is not NOTHING then we have found
3489 * at least two triable branch sequences in a row of a similar
3490 * trietype so we can turn them into a trie. If/when we
3491 * allow NOTHING to start a trie sequence this condition will be
3492 * required, and it isn't expensive so we leave it in for now. */
3493 if ( trietype && trietype != NOTHING )
3494 make_trie( pRExC_state,
3495 startbranch, first, cur, tail, count,
3496 trietype, depth+1 );
3497 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3501 && noper_next == tail
3504 /* noper is triable, so we can start a new trie sequence */
3507 trietype = noper_trietype;
3509 /* if we already saw a first but the current node is not triable then we have
3510 * to reset the first information. */
3515 } /* end handle unmergable node */
3516 } /* loop over branches */
3517 DEBUG_TRIE_COMPILE_r({
3518 regprop(RExC_rx, mysv, cur);
3519 PerlIO_printf( Perl_debug_log,
3520 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3521 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3524 if ( last && trietype ) {
3525 if ( trietype != NOTHING ) {
3526 /* the last branch of the sequence was part of a trie,
3527 * so we have to construct it here outside of the loop
3529 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3530 #ifdef TRIE_STUDY_OPT
3531 if ( ((made == MADE_EXACT_TRIE &&
3532 startbranch == first)
3533 || ( first_non_open == first )) &&
3535 flags |= SCF_TRIE_RESTUDY;
3536 if ( startbranch == first
3539 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3544 /* at this point we know whatever we have is a NOTHING sequence/branch
3545 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3547 if ( startbranch == first ) {
3549 /* the entire thing is a NOTHING sequence, something like this:
3550 * (?:|) So we can turn it into a plain NOTHING op. */
3551 DEBUG_TRIE_COMPILE_r({
3552 regprop(RExC_rx, mysv, cur);
3553 PerlIO_printf( Perl_debug_log,
3554 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3555 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3558 OP(startbranch)= NOTHING;
3559 NEXT_OFF(startbranch)= tail - startbranch;
3560 for ( opt= startbranch + 1; opt < tail ; opt++ )
3564 } /* end if ( last) */
3565 } /* TRIE_MAXBUF is non zero */
3570 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3571 scan = NEXTOPER(NEXTOPER(scan));
3572 } else /* single branch is optimized. */
3573 scan = NEXTOPER(scan);
3575 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3576 scan_frame *newframe = NULL;
3581 if (OP(scan) != SUSPEND) {
3582 /* set the pointer */
3583 if (OP(scan) == GOSUB) {
3585 RExC_recurse[ARG2L(scan)] = scan;
3586 start = RExC_open_parens[paren-1];
3587 end = RExC_close_parens[paren-1];
3590 start = RExC_rxi->program + 1;
3594 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3595 SAVEFREEPV(recursed);
3597 if (!PAREN_TEST(recursed,paren+1)) {
3598 PAREN_SET(recursed,paren+1);
3599 Newx(newframe,1,scan_frame);
3601 if (flags & SCF_DO_SUBSTR) {
3602 SCAN_COMMIT(pRExC_state,data,minlenp);
3603 data->longest = &(data->longest_float);
3605 is_inf = is_inf_internal = 1;
3606 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3607 cl_anything(pRExC_state, data->start_class);
3608 flags &= ~SCF_DO_STCLASS;
3611 Newx(newframe,1,scan_frame);
3614 end = regnext(scan);
3619 SAVEFREEPV(newframe);
3620 newframe->next = regnext(scan);
3621 newframe->last = last;
3622 newframe->stop = stopparen;
3623 newframe->prev = frame;
3633 else if (OP(scan) == EXACT) {
3634 SSize_t l = STR_LEN(scan);
3637 const U8 * const s = (U8*)STRING(scan);
3638 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3639 l = utf8_length(s, s + l);
3641 uc = *((U8*)STRING(scan));
3644 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3645 /* The code below prefers earlier match for fixed
3646 offset, later match for variable offset. */
3647 if (data->last_end == -1) { /* Update the start info. */
3648 data->last_start_min = data->pos_min;
3649 data->last_start_max = is_inf
3650 ? SSize_t_MAX : data->pos_min + data->pos_delta;
3652 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3654 SvUTF8_on(data->last_found);
3656 SV * const sv = data->last_found;
3657 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3658 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3659 if (mg && mg->mg_len >= 0)
3660 mg->mg_len += utf8_length((U8*)STRING(scan),
3661 (U8*)STRING(scan)+STR_LEN(scan));
3663 data->last_end = data->pos_min + l;
3664 data->pos_min += l; /* As in the first entry. */
3665 data->flags &= ~SF_BEFORE_EOL;
3667 if (flags & SCF_DO_STCLASS_AND) {
3668 /* Check whether it is compatible with what we know already! */
3672 /* If compatible, we or it in below. It is compatible if is
3673 * in the bitmp and either 1) its bit or its fold is set, or 2)
3674 * it's for a locale. Even if there isn't unicode semantics
3675 * here, at runtime there may be because of matching against a
3676 * utf8 string, so accept a possible false positive for
3677 * latin1-range folds */
3679 (!(data->start_class->flags & ANYOF_LOCALE)
3680 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3681 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3682 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3687 ANYOF_CLASS_ZERO(data->start_class);
3688 ANYOF_BITMAP_ZERO(data->start_class);
3690 ANYOF_BITMAP_SET(data->start_class, uc);
3691 else if (uc >= 0x100) {
3694 /* Some Unicode code points fold to the Latin1 range; as
3695 * XXX temporary code, instead of figuring out if this is
3696 * one, just assume it is and set all the start class bits
3697 * that could be some such above 255 code point's fold
3698 * which will generate fals positives. As the code
3699 * elsewhere that does compute the fold settles down, it
3700 * can be extracted out and re-used here */
3701 for (i = 0; i < 256; i++){
3702 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3703 ANYOF_BITMAP_SET(data->start_class, i);
3707 CLEAR_SSC_EOS(data->start_class);
3709 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3711 else if (flags & SCF_DO_STCLASS_OR) {
3712 /* false positive possible if the class is case-folded */
3714 ANYOF_BITMAP_SET(data->start_class, uc);
3716 data->start_class->flags |= ANYOF_UNICODE_ALL;
3717 CLEAR_SSC_EOS(data->start_class);
3718 cl_and(data->start_class, and_withp);
3720 flags &= ~SCF_DO_STCLASS;
3722 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3723 SSize_t l = STR_LEN(scan);
3724 UV uc = *((U8*)STRING(scan));
3726 /* Search for fixed substrings supports EXACT only. */
3727 if (flags & SCF_DO_SUBSTR) {
3729 SCAN_COMMIT(pRExC_state, data, minlenp);
3732 const U8 * const s = (U8 *)STRING(scan);
3733 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3734 l = utf8_length(s, s + l);
3736 if (has_exactf_sharp_s) {
3737 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3739 min += l - min_subtract;
3741 delta += min_subtract;
3742 if (flags & SCF_DO_SUBSTR) {
3743 data->pos_min += l - min_subtract;
3744 if (data->pos_min < 0) {
3747 data->pos_delta += min_subtract;
3749 data->longest = &(data->longest_float);
3752 if (flags & SCF_DO_STCLASS_AND) {
3753 /* Check whether it is compatible with what we know already! */
3756 (!(data->start_class->flags & ANYOF_LOCALE)
3757 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3758 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3762 ANYOF_CLASS_ZERO(data->start_class);
3763 ANYOF_BITMAP_ZERO(data->start_class);
3765 ANYOF_BITMAP_SET(data->start_class, uc);
3766 CLEAR_SSC_EOS(data->start_class);
3767 if (OP(scan) == EXACTFL) {
3768 /* XXX This set is probably no longer necessary, and
3769 * probably wrong as LOCALE now is on in the initial
3771 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3775 /* Also set the other member of the fold pair. In case
3776 * that unicode semantics is called for at runtime, use
3777 * the full latin1 fold. (Can't do this for locale,
3778 * because not known until runtime) */
3779 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3781 /* All other (EXACTFL handled above) folds except under
3782 * /iaa that include s, S, and sharp_s also may include
3784 if (OP(scan) != EXACTFA) {
3785 if (uc == 's' || uc == 'S') {
3786 ANYOF_BITMAP_SET(data->start_class,
3787 LATIN_SMALL_LETTER_SHARP_S);
3789 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3790 ANYOF_BITMAP_SET(data->start_class, 's');
3791 ANYOF_BITMAP_SET(data->start_class, 'S');
3796 else if (uc >= 0x100) {
3798 for (i = 0; i < 256; i++){
3799 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3800 ANYOF_BITMAP_SET(data->start_class, i);
3805 else if (flags & SCF_DO_STCLASS_OR) {
3806 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3807 /* false positive possible if the class is case-folded.
3808 Assume that the locale settings are the same... */
3810 ANYOF_BITMAP_SET(data->start_class, uc);
3811 if (OP(scan) != EXACTFL) {
3813 /* And set the other member of the fold pair, but
3814 * can't do that in locale because not known until
3816 ANYOF_BITMAP_SET(data->start_class,
3817 PL_fold_latin1[uc]);
3819 /* All folds except under /iaa that include s, S,
3820 * and sharp_s also may include the others */
3821 if (OP(scan) != EXACTFA) {
3822 if (uc == 's' || uc == 'S') {
3823 ANYOF_BITMAP_SET(data->start_class,
3824 LATIN_SMALL_LETTER_SHARP_S);
3826 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3827 ANYOF_BITMAP_SET(data->start_class, 's');
3828 ANYOF_BITMAP_SET(data->start_class, 'S');
3833 CLEAR_SSC_EOS(data->start_class);
3835 cl_and(data->start_class, and_withp);
https://perl5.git.perl.org / perl5.git / blob