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 127-255 to match them, but then invert that, so the
1045 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 = uvuni_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 = uvuni_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 #define TRIE_READ_CHAR STMT_START { \
1472 /* if it is UTF then it is either already folded, or does not need folding */ \
1473 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1475 else if (folder == PL_fold_latin1) { \
1476 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1477 if ( foldlen > 0 ) { \
1478 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1484 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, FOLD_FLAGS_FULL); \
1485 skiplen = UNISKIP(uvc); \
1486 foldlen -= skiplen; \
1487 scan = foldbuf + skiplen; \
1490 /* raw data, will be folded later if needed */ \
1498 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1499 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1500 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1501 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1503 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1504 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1505 TRIE_LIST_CUR( state )++; \
1508 #define TRIE_LIST_NEW(state) STMT_START { \
1509 Newxz( trie->states[ state ].trans.list, \
1510 4, reg_trie_trans_le ); \
1511 TRIE_LIST_CUR( state ) = 1; \
1512 TRIE_LIST_LEN( state ) = 4; \
1515 #define TRIE_HANDLE_WORD(state) STMT_START { \
1516 U16 dupe= trie->states[ state ].wordnum; \
1517 regnode * const noper_next = regnext( noper ); \
1520 /* store the word for dumping */ \
1522 if (OP(noper) != NOTHING) \
1523 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1525 tmp = newSVpvn_utf8( "", 0, UTF ); \
1526 av_push( trie_words, tmp ); \
1530 trie->wordinfo[curword].prev = 0; \
1531 trie->wordinfo[curword].len = wordlen; \
1532 trie->wordinfo[curword].accept = state; \
1534 if ( noper_next < tail ) { \
1536 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1537 trie->jump[curword] = (U16)(noper_next - convert); \
1539 jumper = noper_next; \
1541 nextbranch= regnext(cur); \
1545 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1546 /* chain, so that when the bits of chain are later */\
1547 /* linked together, the dups appear in the chain */\
1548 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1549 trie->wordinfo[dupe].prev = curword; \
1551 /* we haven't inserted this word yet. */ \
1552 trie->states[ state ].wordnum = curword; \
1557 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1558 ( ( base + charid >= ucharcount \
1559 && base + charid < ubound \
1560 && state == trie->trans[ base - ucharcount + charid ].check \
1561 && trie->trans[ base - ucharcount + charid ].next ) \
1562 ? trie->trans[ base - ucharcount + charid ].next \
1563 : ( state==1 ? special : 0 ) \
1567 #define MADE_JUMP_TRIE 2
1568 #define MADE_EXACT_TRIE 4
1571 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1574 /* first pass, loop through and scan words */
1575 reg_trie_data *trie;
1576 HV *widecharmap = NULL;
1577 AV *revcharmap = newAV();
1579 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1584 regnode *jumper = NULL;
1585 regnode *nextbranch = NULL;
1586 regnode *convert = NULL;
1587 U32 *prev_states; /* temp array mapping each state to previous one */
1588 /* we just use folder as a flag in utf8 */
1589 const U8 * folder = NULL;
1592 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1593 AV *trie_words = NULL;
1594 /* along with revcharmap, this only used during construction but both are
1595 * useful during debugging so we store them in the struct when debugging.
1598 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1599 STRLEN trie_charcount=0;
1601 SV *re_trie_maxbuff;
1602 GET_RE_DEBUG_FLAGS_DECL;
1604 PERL_ARGS_ASSERT_MAKE_TRIE;
1606 PERL_UNUSED_ARG(depth);
1613 case EXACTFU_TRICKYFOLD:
1614 case EXACTFU: folder = PL_fold_latin1; break;
1615 case EXACTF: folder = PL_fold; break;
1616 case EXACTFL: folder = PL_fold_locale; break;
1617 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1620 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1622 trie->startstate = 1;
1623 trie->wordcount = word_count;
1624 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1625 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1627 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1628 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1629 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1632 trie_words = newAV();
1635 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1636 if (!SvIOK(re_trie_maxbuff)) {
1637 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1639 DEBUG_TRIE_COMPILE_r({
1640 PerlIO_printf( Perl_debug_log,
1641 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1642 (int)depth * 2 + 2, "",
1643 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1644 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1648 /* Find the node we are going to overwrite */
1649 if ( first == startbranch && OP( last ) != BRANCH ) {
1650 /* whole branch chain */
1653 /* branch sub-chain */
1654 convert = NEXTOPER( first );
1657 /* -- First loop and Setup --
1659 We first traverse the branches and scan each word to determine if it
1660 contains widechars, and how many unique chars there are, this is
1661 important as we have to build a table with at least as many columns as we
1664 We use an array of integers to represent the character codes 0..255
1665 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1666 native representation of the character value as the key and IV's for the
1669 *TODO* If we keep track of how many times each character is used we can
1670 remap the columns so that the table compression later on is more
1671 efficient in terms of memory by ensuring the most common value is in the
1672 middle and the least common are on the outside. IMO this would be better
1673 than a most to least common mapping as theres a decent chance the most
1674 common letter will share a node with the least common, meaning the node
1675 will not be compressible. With a middle is most common approach the worst
1676 case is when we have the least common nodes twice.
1680 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1681 regnode *noper = NEXTOPER( cur );
1682 const U8 *uc = (U8*)STRING( noper );
1683 const U8 *e = uc + STR_LEN( noper );
1685 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1687 const U8 *scan = (U8*)NULL;
1688 U32 wordlen = 0; /* required init */
1690 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1692 if (OP(noper) == NOTHING) {
1693 regnode *noper_next= regnext(noper);
1694 if (noper_next != tail && OP(noper_next) == flags) {
1696 uc= (U8*)STRING(noper);
1697 e= uc + STR_LEN(noper);
1698 trie->minlen= STR_LEN(noper);
1705 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1706 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1707 regardless of encoding */
1708 if (OP( noper ) == EXACTFU_SS) {
1709 /* false positives are ok, so just set this */
1710 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
1713 for ( ; uc < e ; uc += len ) {
1714 TRIE_CHARCOUNT(trie)++;
1719 U8 folded= folder[ (U8) uvc ];
1720 if ( !trie->charmap[ folded ] ) {
1721 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1722 TRIE_STORE_REVCHAR( folded );
1725 if ( !trie->charmap[ uvc ] ) {
1726 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1727 TRIE_STORE_REVCHAR( uvc );
1730 /* store the codepoint in the bitmap, and its folded
1732 TRIE_BITMAP_SET(trie, uvc);
1734 /* store the folded codepoint */
1735 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1738 /* store first byte of utf8 representation of
1739 variant codepoints */
1740 if (! UNI_IS_INVARIANT(uvc)) {
1741 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1744 set_bit = 0; /* We've done our bit :-) */
1749 widecharmap = newHV();
1751 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1754 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1756 if ( !SvTRUE( *svpp ) ) {
1757 sv_setiv( *svpp, ++trie->uniquecharcount );
1758 TRIE_STORE_REVCHAR(uvc);
1762 if( cur == first ) {
1763 trie->minlen = chars;
1764 trie->maxlen = chars;
1765 } else if (chars < trie->minlen) {
1766 trie->minlen = chars;
1767 } else if (chars > trie->maxlen) {
1768 trie->maxlen = chars;
1770 if (OP( noper ) == EXACTFU_SS) {
1771 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1772 if (trie->minlen > 1)
1775 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1776 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1777 * - We assume that any such sequence might match a 2 byte string */
1778 if (trie->minlen > 2 )
1782 } /* end first pass */
1783 DEBUG_TRIE_COMPILE_r(
1784 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1785 (int)depth * 2 + 2,"",
1786 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1787 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1788 (int)trie->minlen, (int)trie->maxlen )
1792 We now know what we are dealing with in terms of unique chars and
1793 string sizes so we can calculate how much memory a naive
1794 representation using a flat table will take. If it's over a reasonable
1795 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1796 conservative but potentially much slower representation using an array
1799 At the end we convert both representations into the same compressed
1800 form that will be used in regexec.c for matching with. The latter
1801 is a form that cannot be used to construct with but has memory
1802 properties similar to the list form and access properties similar
1803 to the table form making it both suitable for fast searches and
1804 small enough that its feasable to store for the duration of a program.
1806 See the comment in the code where the compressed table is produced
1807 inplace from the flat tabe representation for an explanation of how
1808 the compression works.
1813 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1816 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1818 Second Pass -- Array Of Lists Representation
1820 Each state will be represented by a list of charid:state records
1821 (reg_trie_trans_le) the first such element holds the CUR and LEN
1822 points of the allocated array. (See defines above).
1824 We build the initial structure using the lists, and then convert
1825 it into the compressed table form which allows faster lookups
1826 (but cant be modified once converted).
1829 STRLEN transcount = 1;
1831 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1832 "%*sCompiling trie using list compiler\n",
1833 (int)depth * 2 + 2, ""));
1835 trie->states = (reg_trie_state *)
1836 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1837 sizeof(reg_trie_state) );
1841 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1843 regnode *noper = NEXTOPER( cur );
1844 U8 *uc = (U8*)STRING( noper );
1845 const U8 *e = uc + STR_LEN( noper );
1846 U32 state = 1; /* required init */
1847 U16 charid = 0; /* sanity init */
1848 U8 *scan = (U8*)NULL; /* sanity init */
1849 STRLEN foldlen = 0; /* required init */
1850 U32 wordlen = 0; /* required init */
1851 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1854 if (OP(noper) == NOTHING) {
1855 regnode *noper_next= regnext(noper);
1856 if (noper_next != tail && OP(noper_next) == flags) {
1858 uc= (U8*)STRING(noper);
1859 e= uc + STR_LEN(noper);
1863 if (OP(noper) != NOTHING) {
1864 for ( ; uc < e ; uc += len ) {
1869 charid = trie->charmap[ uvc ];
1871 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1875 charid=(U16)SvIV( *svpp );
1878 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1885 if ( !trie->states[ state ].trans.list ) {
1886 TRIE_LIST_NEW( state );
1888 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1889 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1890 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1895 newstate = next_alloc++;
1896 prev_states[newstate] = state;
1897 TRIE_LIST_PUSH( state, charid, newstate );
1902 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1906 TRIE_HANDLE_WORD(state);
1908 } /* end second pass */
1910 /* next alloc is the NEXT state to be allocated */
1911 trie->statecount = next_alloc;
1912 trie->states = (reg_trie_state *)
1913 PerlMemShared_realloc( trie->states,
1915 * sizeof(reg_trie_state) );
1917 /* and now dump it out before we compress it */
1918 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1919 revcharmap, next_alloc,
1923 trie->trans = (reg_trie_trans *)
1924 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1931 for( state=1 ; state < next_alloc ; state ++ ) {
1935 DEBUG_TRIE_COMPILE_MORE_r(
1936 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1940 if (trie->states[state].trans.list) {
1941 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1945 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1946 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1947 if ( forid < minid ) {
1949 } else if ( forid > maxid ) {
1953 if ( transcount < tp + maxid - minid + 1) {
1955 trie->trans = (reg_trie_trans *)
1956 PerlMemShared_realloc( trie->trans,
1958 * sizeof(reg_trie_trans) );
1959 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1961 base = trie->uniquecharcount + tp - minid;
1962 if ( maxid == minid ) {
1964 for ( ; zp < tp ; zp++ ) {
1965 if ( ! trie->trans[ zp ].next ) {
1966 base = trie->uniquecharcount + zp - minid;
1967 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1968 trie->trans[ zp ].check = state;
1974 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1975 trie->trans[ tp ].check = state;
1980 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1981 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1982 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1983 trie->trans[ tid ].check = state;
1985 tp += ( maxid - minid + 1 );
1987 Safefree(trie->states[ state ].trans.list);
1990 DEBUG_TRIE_COMPILE_MORE_r(
1991 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1994 trie->states[ state ].trans.base=base;
1996 trie->lasttrans = tp + 1;
2000 Second Pass -- Flat Table Representation.
2002 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
2003 We know that we will need Charcount+1 trans at most to store the data
2004 (one row per char at worst case) So we preallocate both structures
2005 assuming worst case.
2007 We then construct the trie using only the .next slots of the entry
2010 We use the .check field of the first entry of the node temporarily to
2011 make compression both faster and easier by keeping track of how many non
2012 zero fields are in the node.
2014 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2017 There are two terms at use here: state as a TRIE_NODEIDX() which is a
2018 number representing the first entry of the node, and state as a
2019 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
2020 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
2021 are 2 entrys per node. eg:
2029 The table is internally in the right hand, idx form. However as we also
2030 have to deal with the states array which is indexed by nodenum we have to
2031 use TRIE_NODENUM() to convert.
2034 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2035 "%*sCompiling trie using table compiler\n",
2036 (int)depth * 2 + 2, ""));
2038 trie->trans = (reg_trie_trans *)
2039 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2040 * trie->uniquecharcount + 1,
2041 sizeof(reg_trie_trans) );
2042 trie->states = (reg_trie_state *)
2043 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2044 sizeof(reg_trie_state) );
2045 next_alloc = trie->uniquecharcount + 1;
2048 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2050 regnode *noper = NEXTOPER( cur );
2051 const U8 *uc = (U8*)STRING( noper );
2052 const U8 *e = uc + STR_LEN( noper );
2054 U32 state = 1; /* required init */
2056 U16 charid = 0; /* sanity init */
2057 U32 accept_state = 0; /* sanity init */
2058 U8 *scan = (U8*)NULL; /* sanity init */
2060 STRLEN foldlen = 0; /* required init */
2061 U32 wordlen = 0; /* required init */
2063 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2065 if (OP(noper) == NOTHING) {
2066 regnode *noper_next= regnext(noper);
2067 if (noper_next != tail && OP(noper_next) == flags) {
2069 uc= (U8*)STRING(noper);
2070 e= uc + STR_LEN(noper);
2074 if ( OP(noper) != NOTHING ) {
2075 for ( ; uc < e ; uc += len ) {
2080 charid = trie->charmap[ uvc ];
2082 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2083 charid = svpp ? (U16)SvIV(*svpp) : 0;
2087 if ( !trie->trans[ state + charid ].next ) {
2088 trie->trans[ state + charid ].next = next_alloc;
2089 trie->trans[ state ].check++;
2090 prev_states[TRIE_NODENUM(next_alloc)]
2091 = TRIE_NODENUM(state);
2092 next_alloc += trie->uniquecharcount;
2094 state = trie->trans[ state + charid ].next;
2096 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2098 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2101 accept_state = TRIE_NODENUM( state );
2102 TRIE_HANDLE_WORD(accept_state);
2104 } /* end second pass */
2106 /* and now dump it out before we compress it */
2107 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2109 next_alloc, depth+1));
2113 * Inplace compress the table.*
2115 For sparse data sets the table constructed by the trie algorithm will
2116 be mostly 0/FAIL transitions or to put it another way mostly empty.
2117 (Note that leaf nodes will not contain any transitions.)
2119 This algorithm compresses the tables by eliminating most such
2120 transitions, at the cost of a modest bit of extra work during lookup:
2122 - Each states[] entry contains a .base field which indicates the
2123 index in the state[] array wheres its transition data is stored.
2125 - If .base is 0 there are no valid transitions from that node.
2127 - If .base is nonzero then charid is added to it to find an entry in
2130 -If trans[states[state].base+charid].check!=state then the
2131 transition is taken to be a 0/Fail transition. Thus if there are fail
2132 transitions at the front of the node then the .base offset will point
2133 somewhere inside the previous nodes data (or maybe even into a node
2134 even earlier), but the .check field determines if the transition is
2138 The following process inplace converts the table to the compressed
2139 table: We first do not compress the root node 1,and mark all its
2140 .check pointers as 1 and set its .base pointer as 1 as well. This
2141 allows us to do a DFA construction from the compressed table later,
2142 and ensures that any .base pointers we calculate later are greater
2145 - We set 'pos' to indicate the first entry of the second node.
2147 - We then iterate over the columns of the node, finding the first and
2148 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2149 and set the .check pointers accordingly, and advance pos
2150 appropriately and repreat for the next node. Note that when we copy
2151 the next pointers we have to convert them from the original
2152 NODEIDX form to NODENUM form as the former is not valid post
2155 - If a node has no transitions used we mark its base as 0 and do not
2156 advance the pos pointer.
2158 - If a node only has one transition we use a second pointer into the
2159 structure to fill in allocated fail transitions from other states.
2160 This pointer is independent of the main pointer and scans forward
2161 looking for null transitions that are allocated to a state. When it
2162 finds one it writes the single transition into the "hole". If the
2163 pointer doesnt find one the single transition is appended as normal.
2165 - Once compressed we can Renew/realloc the structures to release the
2168 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2169 specifically Fig 3.47 and the associated pseudocode.
2173 const U32 laststate = TRIE_NODENUM( next_alloc );
2176 trie->statecount = laststate;
2178 for ( state = 1 ; state < laststate ; state++ ) {
2180 const U32 stateidx = TRIE_NODEIDX( state );
2181 const U32 o_used = trie->trans[ stateidx ].check;
2182 U32 used = trie->trans[ stateidx ].check;
2183 trie->trans[ stateidx ].check = 0;
2185 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2186 if ( flag || trie->trans[ stateidx + charid ].next ) {
2187 if ( trie->trans[ stateidx + charid ].next ) {
2189 for ( ; zp < pos ; zp++ ) {
2190 if ( ! trie->trans[ zp ].next ) {
2194 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2195 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2196 trie->trans[ zp ].check = state;
2197 if ( ++zp > pos ) pos = zp;
2204 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2206 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2207 trie->trans[ pos ].check = state;
2212 trie->lasttrans = pos + 1;
2213 trie->states = (reg_trie_state *)
2214 PerlMemShared_realloc( trie->states, laststate
2215 * sizeof(reg_trie_state) );
2216 DEBUG_TRIE_COMPILE_MORE_r(
2217 PerlIO_printf( Perl_debug_log,
2218 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2219 (int)depth * 2 + 2,"",
2220 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2223 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2226 } /* end table compress */
2228 DEBUG_TRIE_COMPILE_MORE_r(
2229 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2230 (int)depth * 2 + 2, "",
2231 (UV)trie->statecount,
2232 (UV)trie->lasttrans)
2234 /* resize the trans array to remove unused space */
2235 trie->trans = (reg_trie_trans *)
2236 PerlMemShared_realloc( trie->trans, trie->lasttrans
2237 * sizeof(reg_trie_trans) );
2239 { /* Modify the program and insert the new TRIE node */
2240 U8 nodetype =(U8)(flags & 0xFF);
2244 regnode *optimize = NULL;
2245 #ifdef RE_TRACK_PATTERN_OFFSETS
2248 U32 mjd_nodelen = 0;
2249 #endif /* RE_TRACK_PATTERN_OFFSETS */
2250 #endif /* DEBUGGING */
2252 This means we convert either the first branch or the first Exact,
2253 depending on whether the thing following (in 'last') is a branch
2254 or not and whther first is the startbranch (ie is it a sub part of
2255 the alternation or is it the whole thing.)
2256 Assuming its a sub part we convert the EXACT otherwise we convert
2257 the whole branch sequence, including the first.
2259 /* Find the node we are going to overwrite */
2260 if ( first != startbranch || OP( last ) == BRANCH ) {
2261 /* branch sub-chain */
2262 NEXT_OFF( first ) = (U16)(last - first);
2263 #ifdef RE_TRACK_PATTERN_OFFSETS
2265 mjd_offset= Node_Offset((convert));
2266 mjd_nodelen= Node_Length((convert));
2269 /* whole branch chain */
2271 #ifdef RE_TRACK_PATTERN_OFFSETS
2274 const regnode *nop = NEXTOPER( convert );
2275 mjd_offset= Node_Offset((nop));
2276 mjd_nodelen= Node_Length((nop));
2280 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2281 (int)depth * 2 + 2, "",
2282 (UV)mjd_offset, (UV)mjd_nodelen)
2285 /* But first we check to see if there is a common prefix we can
2286 split out as an EXACT and put in front of the TRIE node. */
2287 trie->startstate= 1;
2288 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2290 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2294 const U32 base = trie->states[ state ].trans.base;
2296 if ( trie->states[state].wordnum )
2299 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2300 if ( ( base + ofs >= trie->uniquecharcount ) &&
2301 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2302 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2304 if ( ++count > 1 ) {
2305 SV **tmp = av_fetch( revcharmap, ofs, 0);
2306 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2307 if ( state == 1 ) break;
2309 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2311 PerlIO_printf(Perl_debug_log,
2312 "%*sNew Start State=%"UVuf" Class: [",
2313 (int)depth * 2 + 2, "",
2316 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2317 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2319 TRIE_BITMAP_SET(trie,*ch);
2321 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2323 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2327 TRIE_BITMAP_SET(trie,*ch);
2329 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2330 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2336 SV **tmp = av_fetch( revcharmap, idx, 0);
2338 char *ch = SvPV( *tmp, len );
2340 SV *sv=sv_newmortal();
2341 PerlIO_printf( Perl_debug_log,
2342 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2343 (int)depth * 2 + 2, "",
2345 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2346 PL_colors[0], PL_colors[1],
2347 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2348 PERL_PV_ESCAPE_FIRSTCHAR
2353 OP( convert ) = nodetype;
2354 str=STRING(convert);
2357 STR_LEN(convert) += len;
2363 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2368 trie->prefixlen = (state-1);
2370 regnode *n = convert+NODE_SZ_STR(convert);
2371 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2372 trie->startstate = state;
2373 trie->minlen -= (state - 1);
2374 trie->maxlen -= (state - 1);
2376 /* At least the UNICOS C compiler choked on this
2377 * being argument to DEBUG_r(), so let's just have
2380 #ifdef PERL_EXT_RE_BUILD
2386 regnode *fix = convert;
2387 U32 word = trie->wordcount;
2389 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2390 while( ++fix < n ) {
2391 Set_Node_Offset_Length(fix, 0, 0);
2394 SV ** const tmp = av_fetch( trie_words, word, 0 );
2396 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2397 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2399 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2407 NEXT_OFF(convert) = (U16)(tail - convert);
2408 DEBUG_r(optimize= n);
2414 if ( trie->maxlen ) {
2415 NEXT_OFF( convert ) = (U16)(tail - convert);
2416 ARG_SET( convert, data_slot );
2417 /* Store the offset to the first unabsorbed branch in
2418 jump[0], which is otherwise unused by the jump logic.
2419 We use this when dumping a trie and during optimisation. */
2421 trie->jump[0] = (U16)(nextbranch - convert);
2423 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2424 * and there is a bitmap
2425 * and the first "jump target" node we found leaves enough room
2426 * then convert the TRIE node into a TRIEC node, with the bitmap
2427 * embedded inline in the opcode - this is hypothetically faster.
2429 if ( !trie->states[trie->startstate].wordnum
2431 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2433 OP( convert ) = TRIEC;
2434 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2435 PerlMemShared_free(trie->bitmap);
2438 OP( convert ) = TRIE;
2440 /* store the type in the flags */
2441 convert->flags = nodetype;
2445 + regarglen[ OP( convert ) ];
2447 /* XXX We really should free up the resource in trie now,
2448 as we won't use them - (which resources?) dmq */
2450 /* needed for dumping*/
2451 DEBUG_r(if (optimize) {
2452 regnode *opt = convert;
2454 while ( ++opt < optimize) {
2455 Set_Node_Offset_Length(opt,0,0);
2458 Try to clean up some of the debris left after the
2461 while( optimize < jumper ) {
2462 mjd_nodelen += Node_Length((optimize));
2463 OP( optimize ) = OPTIMIZED;
2464 Set_Node_Offset_Length(optimize,0,0);
2467 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2469 } /* end node insert */
2471 /* Finish populating the prev field of the wordinfo array. Walk back
2472 * from each accept state until we find another accept state, and if
2473 * so, point the first word's .prev field at the second word. If the
2474 * second already has a .prev field set, stop now. This will be the
2475 * case either if we've already processed that word's accept state,
2476 * or that state had multiple words, and the overspill words were
2477 * already linked up earlier.
2484 for (word=1; word <= trie->wordcount; word++) {
2486 if (trie->wordinfo[word].prev)
2488 state = trie->wordinfo[word].accept;
2490 state = prev_states[state];
2493 prev = trie->states[state].wordnum;
2497 trie->wordinfo[word].prev = prev;
2499 Safefree(prev_states);
2503 /* and now dump out the compressed format */
2504 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2506 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2508 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2509 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2511 SvREFCNT_dec_NN(revcharmap);
2515 : trie->startstate>1
2521 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2523 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2525 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2526 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2529 We find the fail state for each state in the trie, this state is the longest proper
2530 suffix of the current state's 'word' that is also a proper prefix of another word in our
2531 trie. State 1 represents the word '' and is thus the default fail state. This allows
2532 the DFA not to have to restart after its tried and failed a word at a given point, it
2533 simply continues as though it had been matching the other word in the first place.
2535 'abcdgu'=~/abcdefg|cdgu/
2536 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2537 fail, which would bring us to the state representing 'd' in the second word where we would
2538 try 'g' and succeed, proceeding to match 'cdgu'.
2540 /* add a fail transition */
2541 const U32 trie_offset = ARG(source);
2542 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2544 const U32 ucharcount = trie->uniquecharcount;
2545 const U32 numstates = trie->statecount;
2546 const U32 ubound = trie->lasttrans + ucharcount;
2550 U32 base = trie->states[ 1 ].trans.base;
2553 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2554 GET_RE_DEBUG_FLAGS_DECL;
2556 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2558 PERL_UNUSED_ARG(depth);
2562 ARG_SET( stclass, data_slot );
2563 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2564 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2565 aho->trie=trie_offset;
2566 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2567 Copy( trie->states, aho->states, numstates, reg_trie_state );
2568 Newxz( q, numstates, U32);
2569 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2572 /* initialize fail[0..1] to be 1 so that we always have
2573 a valid final fail state */
2574 fail[ 0 ] = fail[ 1 ] = 1;
2576 for ( charid = 0; charid < ucharcount ; charid++ ) {
2577 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2579 q[ q_write ] = newstate;
2580 /* set to point at the root */
2581 fail[ q[ q_write++ ] ]=1;
2584 while ( q_read < q_write) {
2585 const U32 cur = q[ q_read++ % numstates ];
2586 base = trie->states[ cur ].trans.base;
2588 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2589 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2591 U32 fail_state = cur;
2594 fail_state = fail[ fail_state ];
2595 fail_base = aho->states[ fail_state ].trans.base;
2596 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2598 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2599 fail[ ch_state ] = fail_state;
2600 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2602 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2604 q[ q_write++ % numstates] = ch_state;
2608 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2609 when we fail in state 1, this allows us to use the
2610 charclass scan to find a valid start char. This is based on the principle
2611 that theres a good chance the string being searched contains lots of stuff
2612 that cant be a start char.
2614 fail[ 0 ] = fail[ 1 ] = 0;
2615 DEBUG_TRIE_COMPILE_r({
2616 PerlIO_printf(Perl_debug_log,
2617 "%*sStclass Failtable (%"UVuf" states): 0",
2618 (int)(depth * 2), "", (UV)numstates
2620 for( q_read=1; q_read<numstates; q_read++ ) {
2621 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2623 PerlIO_printf(Perl_debug_log, "\n");
2626 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2631 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2632 * These need to be revisited when a newer toolchain becomes available.
2634 #if defined(__sparc64__) && defined(__GNUC__)
2635 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2636 # undef SPARC64_GCC_WORKAROUND
2637 # define SPARC64_GCC_WORKAROUND 1
2641 #define DEBUG_PEEP(str,scan,depth) \
2642 DEBUG_OPTIMISE_r({if (scan){ \
2643 SV * const mysv=sv_newmortal(); \
2644 regnode *Next = regnext(scan); \
2645 regprop(RExC_rx, mysv, scan); \
2646 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2647 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2648 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2652 /* The below joins as many adjacent EXACTish nodes as possible into a single
2653 * one. The regop may be changed if the node(s) contain certain sequences that
2654 * require special handling. The joining is only done if:
2655 * 1) there is room in the current conglomerated node to entirely contain the
2657 * 2) they are the exact same node type
2659 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2660 * these get optimized out
2662 * If a node is to match under /i (folded), the number of characters it matches
2663 * can be different than its character length if it contains a multi-character
2664 * fold. *min_subtract is set to the total delta of the input nodes.
2666 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2667 * and contains LATIN SMALL LETTER SHARP S
2669 * This is as good a place as any to discuss the design of handling these
2670 * multi-character fold sequences. It's been wrong in Perl for a very long
2671 * time. There are three code points in Unicode whose multi-character folds
2672 * were long ago discovered to mess things up. The previous designs for
2673 * dealing with these involved assigning a special node for them. This
2674 * approach doesn't work, as evidenced by this example:
2675 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2676 * Both these fold to "sss", but if the pattern is parsed to create a node that
2677 * would match just the \xDF, it won't be able to handle the case where a
2678 * successful match would have to cross the node's boundary. The new approach
2679 * that hopefully generally solves the problem generates an EXACTFU_SS node
2682 * It turns out that there are problems with all multi-character folds, and not
2683 * just these three. Now the code is general, for all such cases, but the
2684 * three still have some special handling. The approach taken is:
2685 * 1) This routine examines each EXACTFish node that could contain multi-
2686 * character fold sequences. It returns in *min_subtract how much to
2687 * subtract from the the actual length of the string to get a real minimum
2688 * match length; it is 0 if there are no multi-char folds. This delta is
2689 * used by the caller to adjust the min length of the match, and the delta
2690 * between min and max, so that the optimizer doesn't reject these
2691 * possibilities based on size constraints.
2692 * 2) Certain of these sequences require special handling by the trie code,
2693 * so, if found, this code changes the joined node type to special ops:
2694 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2695 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2696 * is used for an EXACTFU node that contains at least one "ss" sequence in
2697 * it. For non-UTF-8 patterns and strings, this is the only case where
2698 * there is a possible fold length change. That means that a regular
2699 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2700 * with length changes, and so can be processed faster. regexec.c takes
2701 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2702 * pre-folded by regcomp.c. This saves effort in regex matching.
2703 * However, the pre-folding isn't done for non-UTF8 patterns because the
2704 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2705 * down by forcing the pattern into UTF8 unless necessary. Also what
2706 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2707 * possibilities for the non-UTF8 patterns are quite simple, except for
2708 * the sharp s. All the ones that don't involve a UTF-8 target string are
2709 * members of a fold-pair, and arrays are set up for all of them so that
2710 * the other member of the pair can be found quickly. Code elsewhere in
2711 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2712 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2713 * described in the next item.
2714 * 4) A problem remains for the sharp s in EXACTF and EXACTFA nodes when the
2715 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2716 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2717 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2718 * character in the pattern corresponds to at most a single character in
2719 * the target string. (And I do mean character, and not byte here, unlike
2720 * other parts of the documentation that have never been updated to
2721 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2722 * two character string 'ss'; in EXACTFA nodes it can match
2723 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2724 * instances where it is violated. I'm reluctant to try to change the
2725 * assumption, as the code involved is impenetrable to me (khw), so
2726 * instead the code here punts. This routine examines (when the pattern
2727 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2728 * boolean indicating whether or not the node contains a sharp s. When it
2729 * is true, the caller sets a flag that later causes the optimizer in this
2730 * file to not set values for the floating and fixed string lengths, and
2731 * thus avoids the optimizer code in regexec.c that makes the invalid
2732 * assumption. Thus, there is no optimization based on string lengths for
2733 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2734 * (The reason the assumption is wrong only in these two cases is that all
2735 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2736 * other folds to their expanded versions. We can't prefold sharp s to
2737 * 'ss' in EXACTF nodes because we don't know at compile time if it
2738 * actually matches 'ss' or not. It will match iff the target string is
2739 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2740 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2741 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2742 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2743 * require the pattern to be forced into UTF-8, the overhead of which we
2747 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2748 if (PL_regkind[OP(scan)] == EXACT) \
2749 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2752 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) {
2753 /* Merge several consecutive EXACTish nodes into one. */
2754 regnode *n = regnext(scan);
2756 regnode *next = scan + NODE_SZ_STR(scan);
2760 regnode *stop = scan;
2761 GET_RE_DEBUG_FLAGS_DECL;
2763 PERL_UNUSED_ARG(depth);
2766 PERL_ARGS_ASSERT_JOIN_EXACT;
2767 #ifndef EXPERIMENTAL_INPLACESCAN
2768 PERL_UNUSED_ARG(flags);
2769 PERL_UNUSED_ARG(val);
2771 DEBUG_PEEP("join",scan,depth);
2773 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2774 * EXACT ones that are mergeable to the current one. */
2776 && (PL_regkind[OP(n)] == NOTHING
2777 || (stringok && OP(n) == OP(scan)))
2779 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2782 if (OP(n) == TAIL || n > next)
2784 if (PL_regkind[OP(n)] == NOTHING) {
2785 DEBUG_PEEP("skip:",n,depth);
2786 NEXT_OFF(scan) += NEXT_OFF(n);
2787 next = n + NODE_STEP_REGNODE;
2794 else if (stringok) {
2795 const unsigned int oldl = STR_LEN(scan);
2796 regnode * const nnext = regnext(n);
2798 /* XXX I (khw) kind of doubt that this works on platforms where
2799 * U8_MAX is above 255 because of lots of other assumptions */
2800 /* Don't join if the sum can't fit into a single node */
2801 if (oldl + STR_LEN(n) > U8_MAX)
2804 DEBUG_PEEP("merg",n,depth);
2807 NEXT_OFF(scan) += NEXT_OFF(n);
2808 STR_LEN(scan) += STR_LEN(n);
2809 next = n + NODE_SZ_STR(n);
2810 /* Now we can overwrite *n : */
2811 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2819 #ifdef EXPERIMENTAL_INPLACESCAN
2820 if (flags && !NEXT_OFF(n)) {
2821 DEBUG_PEEP("atch", val, depth);
2822 if (reg_off_by_arg[OP(n)]) {
2823 ARG_SET(n, val - n);
2826 NEXT_OFF(n) = val - n;
2834 *has_exactf_sharp_s = FALSE;
2836 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2837 * can now analyze for sequences of problematic code points. (Prior to
2838 * this final joining, sequences could have been split over boundaries, and
2839 * hence missed). The sequences only happen in folding, hence for any
2840 * non-EXACT EXACTish node */
2841 if (OP(scan) != EXACT) {
2842 const U8 * const s0 = (U8*) STRING(scan);
2844 const U8 * const s_end = s0 + STR_LEN(scan);
2846 /* One pass is made over the node's string looking for all the
2847 * possibilities. to avoid some tests in the loop, there are two main
2848 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2852 /* Examine the string for a multi-character fold sequence. UTF-8
2853 * patterns have all characters pre-folded by the time this code is
2855 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2856 length sequence we are looking for is 2 */
2859 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2860 if (! len) { /* Not a multi-char fold: get next char */
2865 /* Nodes with 'ss' require special handling, except for EXACTFL
2866 * and EXACTFA for which there is no multi-char fold to this */
2867 if (len == 2 && *s == 's' && *(s+1) == 's'
2868 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2871 OP(scan) = EXACTFU_SS;
2874 else if (len == 6 /* len is the same in both ASCII and EBCDIC
2876 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2877 COMBINING_DIAERESIS_UTF8
2878 COMBINING_ACUTE_ACCENT_UTF8,
2880 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2881 COMBINING_DIAERESIS_UTF8
2882 COMBINING_ACUTE_ACCENT_UTF8,
2887 /* These two folds require special handling by trie's, so
2888 * change the node type to indicate this. If EXACTFA and
2889 * EXACTFL were ever to be handled by trie's, this would
2890 * have to be changed. If this node has already been
2891 * changed to EXACTFU_SS in this loop, leave it as is. (I
2892 * (khw) think it doesn't matter in regexec.c for UTF
2893 * patterns, but no need to change it */
2894 if (OP(scan) == EXACTFU) {
2895 OP(scan) = EXACTFU_TRICKYFOLD;
2899 else { /* Here is a generic multi-char fold. */
2900 const U8* multi_end = s + len;
2902 /* Count how many characters in it. In the case of /l and
2903 * /aa, no folds which contain ASCII code points are
2904 * allowed, so check for those, and skip if found. (In
2905 * EXACTFL, no folds are allowed to any Latin1 code point,
2906 * not just ASCII. But there aren't any of these
2907 * currently, nor ever likely, so don't take the time to
2908 * test for them. The code that generates the
2909 * is_MULTI_foo() macros croaks should one actually get put
2910 * into Unicode .) */
2911 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2912 count = utf8_length(s, multi_end);
2916 while (s < multi_end) {
2919 goto next_iteration;
2929 /* The delta is how long the sequence is minus 1 (1 is how long
2930 * the character that folds to the sequence is) */
2931 *min_subtract += count - 1;
2935 else if (OP(scan) == EXACTFA) {
2937 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2938 * fold to the ASCII range (and there are no existing ones in the
2939 * upper latin1 range). But, as outlined in the comments preceding
2940 * this function, we need to flag any occurrences of the sharp s */
2942 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
2943 *has_exactf_sharp_s = TRUE;
2950 else if (OP(scan) != EXACTFL) {
2952 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2953 * multi-char folds that are all Latin1. (This code knows that
2954 * there are no current multi-char folds possible with EXACTFL,
2955 * relying on fold_grind.t to catch any errors if the very unlikely
2956 * event happens that some get added in future Unicode versions.)
2957 * As explained in the comments preceding this function, we look
2958 * also for the sharp s in EXACTF nodes; it can be in the final
2959 * position. Otherwise we can stop looking 1 byte earlier because
2960 * have to find at least two characters for a multi-fold */
2961 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2964 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2965 if (! len) { /* Not a multi-char fold. */
2966 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2968 *has_exactf_sharp_s = TRUE;
2975 && isARG2_lower_or_UPPER_ARG1('s', *s)
2976 && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
2979 /* EXACTF nodes need to know that the minimum length
2980 * changed so that a sharp s in the string can match this
2981 * ss in the pattern, but they remain EXACTF nodes, as they
2982 * won't match this unless the target string is is UTF-8,
2983 * which we don't know until runtime */
2984 if (OP(scan) != EXACTF) {
2985 OP(scan) = EXACTFU_SS;
2989 *min_subtract += len - 1;
2996 /* Allow dumping but overwriting the collection of skipped
2997 * ops and/or strings with fake optimized ops */
2998 n = scan + NODE_SZ_STR(scan);
3006 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3010 /* REx optimizer. Converts nodes into quicker variants "in place".
3011 Finds fixed substrings. */
3013 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3014 to the position after last scanned or to NULL. */
3016 #define INIT_AND_WITHP \
3017 assert(!and_withp); \
3018 Newx(and_withp,1,struct regnode_charclass_class); \
3019 SAVEFREEPV(and_withp)
3021 /* this is a chain of data about sub patterns we are processing that
3022 need to be handled separately/specially in study_chunk. Its so
3023 we can simulate recursion without losing state. */
3025 typedef struct scan_frame {
3026 regnode *last; /* last node to process in this frame */
3027 regnode *next; /* next node to process when last is reached */
3028 struct scan_frame *prev; /*previous frame*/
3029 I32 stop; /* what stopparen do we use */
3033 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3036 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3037 SSize_t *minlenp, SSize_t *deltap,
3042 struct regnode_charclass_class *and_withp,
3043 U32 flags, U32 depth)
3044 /* scanp: Start here (read-write). */
3045 /* deltap: Write maxlen-minlen here. */
3046 /* last: Stop before this one. */
3047 /* data: string data about the pattern */
3048 /* stopparen: treat close N as END */
3049 /* recursed: which subroutines have we recursed into */
3050 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3053 /* There must be at least this number of characters to match */
3056 regnode *scan = *scanp, *next;
3058 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3059 int is_inf_internal = 0; /* The studied chunk is infinite */
3060 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3061 scan_data_t data_fake;
3062 SV *re_trie_maxbuff = NULL;
3063 regnode *first_non_open = scan;
3064 SSize_t stopmin = SSize_t_MAX;
3065 scan_frame *frame = NULL;
3066 GET_RE_DEBUG_FLAGS_DECL;
3068 PERL_ARGS_ASSERT_STUDY_CHUNK;
3071 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3075 while (first_non_open && OP(first_non_open) == OPEN)
3076 first_non_open=regnext(first_non_open);
3081 while ( scan && OP(scan) != END && scan < last ){
3082 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3083 node length to get a real minimum (because
3084 the folded version may be shorter) */
3085 bool has_exactf_sharp_s = FALSE;
3086 /* Peephole optimizer: */
3087 DEBUG_STUDYDATA("Peep:", data,depth);
3088 DEBUG_PEEP("Peep",scan,depth);
3090 /* Its not clear to khw or hv why this is done here, and not in the
3091 * clauses that deal with EXACT nodes. khw's guess is that it's
3092 * because of a previous design */
3093 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3095 /* Follow the next-chain of the current node and optimize
3096 away all the NOTHINGs from it. */
3097 if (OP(scan) != CURLYX) {
3098 const int max = (reg_off_by_arg[OP(scan)]
3100 /* I32 may be smaller than U16 on CRAYs! */
3101 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3102 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3106 /* Skip NOTHING and LONGJMP. */
3107 while ((n = regnext(n))
3108 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3109 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3110 && off + noff < max)
3112 if (reg_off_by_arg[OP(scan)])
3115 NEXT_OFF(scan) = off;
3120 /* The principal pseudo-switch. Cannot be a switch, since we
3121 look into several different things. */
3122 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3123 || OP(scan) == IFTHEN) {
3124 next = regnext(scan);
3126 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3128 if (OP(next) == code || code == IFTHEN) {
3129 /* NOTE - There is similar code to this block below for handling
3130 TRIE nodes on a re-study. If you change stuff here check there
3132 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3133 struct regnode_charclass_class accum;
3134 regnode * const startbranch=scan;
3136 if (flags & SCF_DO_SUBSTR)
3137 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3138 if (flags & SCF_DO_STCLASS)
3139 cl_init_zero(pRExC_state, &accum);
3141 while (OP(scan) == code) {
3142 SSize_t deltanext, minnext, fake;
3144 struct regnode_charclass_class this_class;
3147 data_fake.flags = 0;
3149 data_fake.whilem_c = data->whilem_c;
3150 data_fake.last_closep = data->last_closep;
3153 data_fake.last_closep = &fake;
3155 data_fake.pos_delta = delta;
3156 next = regnext(scan);
3157 scan = NEXTOPER(scan);
3159 scan = NEXTOPER(scan);
3160 if (flags & SCF_DO_STCLASS) {
3161 cl_init(pRExC_state, &this_class);
3162 data_fake.start_class = &this_class;
3163 f = SCF_DO_STCLASS_AND;
3165 if (flags & SCF_WHILEM_VISITED_POS)
3166 f |= SCF_WHILEM_VISITED_POS;
3168 /* we suppose the run is continuous, last=next...*/
3169 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3171 stopparen, recursed, NULL, f,depth+1);
3174 if (deltanext == SSize_t_MAX) {
3175 is_inf = is_inf_internal = 1;
3177 } else if (max1 < minnext + deltanext)
3178 max1 = minnext + deltanext;
3180 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3182 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3183 if ( stopmin > minnext)
3184 stopmin = min + min1;
3185 flags &= ~SCF_DO_SUBSTR;
3187 data->flags |= SCF_SEEN_ACCEPT;
3190 if (data_fake.flags & SF_HAS_EVAL)
3191 data->flags |= SF_HAS_EVAL;
3192 data->whilem_c = data_fake.whilem_c;
3194 if (flags & SCF_DO_STCLASS)
3195 cl_or(pRExC_state, &accum, &this_class);
3197 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3199 if (flags & SCF_DO_SUBSTR) {
3200 data->pos_min += min1;
3201 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3202 data->pos_delta = SSize_t_MAX;
3204 data->pos_delta += max1 - min1;
3205 if (max1 != min1 || is_inf)
3206 data->longest = &(data->longest_float);
3209 if (delta == SSize_t_MAX
3210 || SSize_t_MAX - delta - (max1 - min1) < 0)
3211 delta = SSize_t_MAX;
3213 delta += max1 - min1;
3214 if (flags & SCF_DO_STCLASS_OR) {
3215 cl_or(pRExC_state, data->start_class, &accum);
3217 cl_and(data->start_class, and_withp);
3218 flags &= ~SCF_DO_STCLASS;
3221 else if (flags & SCF_DO_STCLASS_AND) {
3223 cl_and(data->start_class, &accum);
3224 flags &= ~SCF_DO_STCLASS;
3227 /* Switch to OR mode: cache the old value of
3228 * data->start_class */
3230 StructCopy(data->start_class, and_withp,
3231 struct regnode_charclass_class);
3232 flags &= ~SCF_DO_STCLASS_AND;
3233 StructCopy(&accum, data->start_class,
3234 struct regnode_charclass_class);
3235 flags |= SCF_DO_STCLASS_OR;
3236 SET_SSC_EOS(data->start_class);
3240 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3243 Assuming this was/is a branch we are dealing with: 'scan' now
3244 points at the item that follows the branch sequence, whatever
3245 it is. We now start at the beginning of the sequence and look
3252 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3254 If we can find such a subsequence we need to turn the first
3255 element into a trie and then add the subsequent branch exact
3256 strings to the trie.
3260 1. patterns where the whole set of branches can be converted.
3262 2. patterns where only a subset can be converted.
3264 In case 1 we can replace the whole set with a single regop
3265 for the trie. In case 2 we need to keep the start and end
3268 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3269 becomes BRANCH TRIE; BRANCH X;
3271 There is an additional case, that being where there is a
3272 common prefix, which gets split out into an EXACT like node
3273 preceding the TRIE node.
3275 If x(1..n)==tail then we can do a simple trie, if not we make
3276 a "jump" trie, such that when we match the appropriate word
3277 we "jump" to the appropriate tail node. Essentially we turn
3278 a nested if into a case structure of sorts.
3283 if (!re_trie_maxbuff) {
3284 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3285 if (!SvIOK(re_trie_maxbuff))
3286 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3288 if ( SvIV(re_trie_maxbuff)>=0 ) {
3290 regnode *first = (regnode *)NULL;
3291 regnode *last = (regnode *)NULL;
3292 regnode *tail = scan;
3297 SV * const mysv = sv_newmortal(); /* for dumping */
3299 /* var tail is used because there may be a TAIL
3300 regop in the way. Ie, the exacts will point to the
3301 thing following the TAIL, but the last branch will
3302 point at the TAIL. So we advance tail. If we
3303 have nested (?:) we may have to move through several
3307 while ( OP( tail ) == TAIL ) {
3308 /* this is the TAIL generated by (?:) */
3309 tail = regnext( tail );
3313 DEBUG_TRIE_COMPILE_r({
3314 regprop(RExC_rx, mysv, tail );
3315 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3316 (int)depth * 2 + 2, "",
3317 "Looking for TRIE'able sequences. Tail node is: ",
3318 SvPV_nolen_const( mysv )
3324 Step through the branches
3325 cur represents each branch,
3326 noper is the first thing to be matched as part of that branch
3327 noper_next is the regnext() of that node.
3329 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3330 via a "jump trie" but we also support building with NOJUMPTRIE,
3331 which restricts the trie logic to structures like /FOO|BAR/.
3333 If noper is a trieable nodetype then the branch is a possible optimization
3334 target. If we are building under NOJUMPTRIE then we require that noper_next
3335 is the same as scan (our current position in the regex program).
3337 Once we have two or more consecutive such branches we can create a
3338 trie of the EXACT's contents and stitch it in place into the program.
3340 If the sequence represents all of the branches in the alternation we
3341 replace the entire thing with a single TRIE node.
3343 Otherwise when it is a subsequence we need to stitch it in place and
3344 replace only the relevant branches. This means the first branch has
3345 to remain as it is used by the alternation logic, and its next pointer,
3346 and needs to be repointed at the item on the branch chain following
3347 the last branch we have optimized away.
3349 This could be either a BRANCH, in which case the subsequence is internal,
3350 or it could be the item following the branch sequence in which case the
3351 subsequence is at the end (which does not necessarily mean the first node
3352 is the start of the alternation).
3354 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3357 ----------------+-----------
3361 EXACTFU_SS | EXACTFU
3362 EXACTFU_TRICKYFOLD | EXACTFU
3367 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3368 ( EXACT == (X) ) ? EXACT : \
3369 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3372 /* dont use tail as the end marker for this traverse */
3373 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3374 regnode * const noper = NEXTOPER( cur );
3375 U8 noper_type = OP( noper );
3376 U8 noper_trietype = TRIE_TYPE( noper_type );
3377 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3378 regnode * const noper_next = regnext( noper );
3379 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3380 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3383 DEBUG_TRIE_COMPILE_r({
3384 regprop(RExC_rx, mysv, cur);
3385 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3386 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3388 regprop(RExC_rx, mysv, noper);
3389 PerlIO_printf( Perl_debug_log, " -> %s",
3390 SvPV_nolen_const(mysv));
3393 regprop(RExC_rx, mysv, noper_next );
3394 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3395 SvPV_nolen_const(mysv));
3397 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3398 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3399 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3403 /* Is noper a trieable nodetype that can be merged with the
3404 * current trie (if there is one)? */
3408 ( noper_trietype == NOTHING)
3409 || ( trietype == NOTHING )
3410 || ( trietype == noper_trietype )
3413 && noper_next == tail
3417 /* Handle mergable triable node
3418 * Either we are the first node in a new trieable sequence,
3419 * in which case we do some bookkeeping, otherwise we update
3420 * the end pointer. */
3423 if ( noper_trietype == NOTHING ) {
3424 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3425 regnode * const noper_next = regnext( noper );
3426 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3427 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3430 if ( noper_next_trietype ) {
3431 trietype = noper_next_trietype;
3432 } else if (noper_next_type) {
3433 /* a NOTHING regop is 1 regop wide. We need at least two
3434 * for a trie so we can't merge this in */
3438 trietype = noper_trietype;
3441 if ( trietype == NOTHING )
3442 trietype = noper_trietype;
3447 } /* end handle mergable triable node */
3449 /* handle unmergable node -
3450 * noper may either be a triable node which can not be tried
3451 * together with the current trie, or a non triable node */
3453 /* If last is set and trietype is not NOTHING then we have found
3454 * at least two triable branch sequences in a row of a similar
3455 * trietype so we can turn them into a trie. If/when we
3456 * allow NOTHING to start a trie sequence this condition will be
3457 * required, and it isn't expensive so we leave it in for now. */
3458 if ( trietype && trietype != NOTHING )
3459 make_trie( pRExC_state,
3460 startbranch, first, cur, tail, count,
3461 trietype, depth+1 );
3462 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3466 && noper_next == tail
3469 /* noper is triable, so we can start a new trie sequence */
3472 trietype = noper_trietype;
3474 /* if we already saw a first but the current node is not triable then we have
3475 * to reset the first information. */
3480 } /* end handle unmergable node */
3481 } /* loop over branches */
3482 DEBUG_TRIE_COMPILE_r({
3483 regprop(RExC_rx, mysv, cur);
3484 PerlIO_printf( Perl_debug_log,
3485 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3486 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3489 if ( last && trietype ) {
3490 if ( trietype != NOTHING ) {
3491 /* the last branch of the sequence was part of a trie,
3492 * so we have to construct it here outside of the loop
3494 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3495 #ifdef TRIE_STUDY_OPT
3496 if ( ((made == MADE_EXACT_TRIE &&
3497 startbranch == first)
3498 || ( first_non_open == first )) &&
3500 flags |= SCF_TRIE_RESTUDY;
3501 if ( startbranch == first
3504 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3509 /* at this point we know whatever we have is a NOTHING sequence/branch
3510 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3512 if ( startbranch == first ) {
3514 /* the entire thing is a NOTHING sequence, something like this:
3515 * (?:|) So we can turn it into a plain NOTHING op. */
3516 DEBUG_TRIE_COMPILE_r({
3517 regprop(RExC_rx, mysv, cur);
3518 PerlIO_printf( Perl_debug_log,
3519 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3520 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3523 OP(startbranch)= NOTHING;
3524 NEXT_OFF(startbranch)= tail - startbranch;
3525 for ( opt= startbranch + 1; opt < tail ; opt++ )
3529 } /* end if ( last) */
3530 } /* TRIE_MAXBUF is non zero */
3535 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3536 scan = NEXTOPER(NEXTOPER(scan));
3537 } else /* single branch is optimized. */
3538 scan = NEXTOPER(scan);
3540 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3541 scan_frame *newframe = NULL;
3546 if (OP(scan) != SUSPEND) {
3547 /* set the pointer */
3548 if (OP(scan) == GOSUB) {
3550 RExC_recurse[ARG2L(scan)] = scan;
3551 start = RExC_open_parens[paren-1];
3552 end = RExC_close_parens[paren-1];
3555 start = RExC_rxi->program + 1;
3559 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3560 SAVEFREEPV(recursed);
3562 if (!PAREN_TEST(recursed,paren+1)) {
3563 PAREN_SET(recursed,paren+1);
3564 Newx(newframe,1,scan_frame);
3566 if (flags & SCF_DO_SUBSTR) {
3567 SCAN_COMMIT(pRExC_state,data,minlenp);
3568 data->longest = &(data->longest_float);
3570 is_inf = is_inf_internal = 1;
3571 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3572 cl_anything(pRExC_state, data->start_class);
3573 flags &= ~SCF_DO_STCLASS;
3576 Newx(newframe,1,scan_frame);
3579 end = regnext(scan);
3584 SAVEFREEPV(newframe);
3585 newframe->next = regnext(scan);
3586 newframe->last = last;
3587 newframe->stop = stopparen;
3588 newframe->prev = frame;
3598 else if (OP(scan) == EXACT) {
3599 SSize_t l = STR_LEN(scan);
3602 const U8 * const s = (U8*)STRING(scan);
3603 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3604 l = utf8_length(s, s + l);
3606 uc = *((U8*)STRING(scan));
3609 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3610 /* The code below prefers earlier match for fixed
3611 offset, later match for variable offset. */
3612 if (data->last_end == -1) { /* Update the start info. */
3613 data->last_start_min = data->pos_min;
3614 data->last_start_max = is_inf
3615 ? SSize_t_MAX : data->pos_min + data->pos_delta;
3617 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3619 SvUTF8_on(data->last_found);
3621 SV * const sv = data->last_found;
3622 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3623 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3624 if (mg && mg->mg_len >= 0)
3625 mg->mg_len += utf8_length((U8*)STRING(scan),
3626 (U8*)STRING(scan)+STR_LEN(scan));
3628 data->last_end = data->pos_min + l;
3629 data->pos_min += l; /* As in the first entry. */
3630 data->flags &= ~SF_BEFORE_EOL;
3632 if (flags & SCF_DO_STCLASS_AND) {
3633 /* Check whether it is compatible with what we know already! */
3637 /* If compatible, we or it in below. It is compatible if is
3638 * in the bitmp and either 1) its bit or its fold is set, or 2)
3639 * it's for a locale. Even if there isn't unicode semantics
3640 * here, at runtime there may be because of matching against a
3641 * utf8 string, so accept a possible false positive for
3642 * latin1-range folds */
3644 (!(data->start_class->flags & ANYOF_LOCALE)
3645 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3646 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3647 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3652 ANYOF_CLASS_ZERO(data->start_class);
3653 ANYOF_BITMAP_ZERO(data->start_class);
3655 ANYOF_BITMAP_SET(data->start_class, uc);
3656 else if (uc >= 0x100) {
3659 /* Some Unicode code points fold to the Latin1 range; as
3660 * XXX temporary code, instead of figuring out if this is
3661 * one, just assume it is and set all the start class bits
3662 * that could be some such above 255 code point's fold
3663 * which will generate fals positives. As the code
3664 * elsewhere that does compute the fold settles down, it
3665 * can be extracted out and re-used here */
3666 for (i = 0; i < 256; i++){
3667 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3668 ANYOF_BITMAP_SET(data->start_class, i);
3672 CLEAR_SSC_EOS(data->start_class);
3674 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3676 else if (flags & SCF_DO_STCLASS_OR) {
3677 /* false positive possible if the class is case-folded */
3679 ANYOF_BITMAP_SET(data->start_class, uc);
3681 data->start_class->flags |= ANYOF_UNICODE_ALL;
3682 CLEAR_SSC_EOS(data->start_class);
3683 cl_and(data->start_class, and_withp);
3685 flags &= ~SCF_DO_STCLASS;
3687 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3688 SSize_t l = STR_LEN(scan);
3689 UV uc = *((U8*)STRING(scan));
3691 /* Search for fixed substrings supports EXACT only. */
3692 if (flags & SCF_DO_SUBSTR) {
3694 SCAN_COMMIT(pRExC_state, data, minlenp);
3697 const U8 * const s = (U8 *)STRING(scan);
3698 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3699 l = utf8_length(s, s + l);
3701 if (has_exactf_sharp_s) {
3702 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3704 min += l - min_subtract;
3706 delta += min_subtract;
3707 if (flags & SCF_DO_SUBSTR) {
3708 data->pos_min += l - min_subtract;
3709 if (data->pos_min < 0) {
3712 data->pos_delta += min_subtract;
3714 data->longest = &(data->longest_float);
3717 if (flags & SCF_DO_STCLASS_AND) {
3718 /* Check whether it is compatible with what we know already! */
3721 (!(data->start_class->flags & ANYOF_LOCALE)
3722 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3723 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3727 ANYOF_CLASS_ZERO(data->start_class);
3728 ANYOF_BITMAP_ZERO(data->start_class);
3730 ANYOF_BITMAP_SET(data->start_class, uc);
3731 CLEAR_SSC_EOS(data->start_class);
3732 if (OP(scan) == EXACTFL) {
3733 /* XXX This set is probably no longer necessary, and
3734 * probably wrong as LOCALE now is on in the initial
3736 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3740 /* Also set the other member of the fold pair. In case
3741 * that unicode semantics is called for at runtime, use
3742 * the full latin1 fold. (Can't do this for locale,
3743 * because not known until runtime) */
3744 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3746 /* All other (EXACTFL handled above) folds except under
3747 * /iaa that include s, S, and sharp_s also may include
3749 if (OP(scan) != EXACTFA) {
3750 if (uc == 's' || uc == 'S') {
3751 ANYOF_BITMAP_SET(data->start_class,
3752 LATIN_SMALL_LETTER_SHARP_S);
3754 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3755 ANYOF_BITMAP_SET(data->start_class, 's');
3756 ANYOF_BITMAP_SET(data->start_class, 'S');
3761 else if (uc >= 0x100) {
3763 for (i = 0; i < 256; i++){
3764 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3765 ANYOF_BITMAP_SET(data->start_class, i);
3770 else if (flags & SCF_DO_STCLASS_OR) {
3771 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3772 /* false positive possible if the class is case-folded.
3773 Assume that the locale settings are the same... */
3775 ANYOF_BITMAP_SET(data->start_class, uc);
3776 if (OP(scan) != EXACTFL) {
3778 /* And set the other member of the fold pair, but
3779 * can't do that in locale because not known until
3781 ANYOF_BITMAP_SET(data->start_class,
3782 PL_fold_latin1[uc]);
3784 /* All folds except under /iaa that include s, S,
3785 * and sharp_s also may include the others */
3786 if (OP(scan) != EXACTFA) {
3787 if (uc == 's' || uc == 'S') {
3788 ANYOF_BITMAP_SET(data->start_class,
3789 LATIN_SMALL_LETTER_SHARP_S);
3791 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3792 ANYOF_BITMAP_SET(data->start_class, 's');
3793 ANYOF_BITMAP_SET(data->start_class, 'S');
3798 CLEAR_SSC_EOS(data->start_class);
3800 cl_and(data->start_class, and_withp);
3802 flags &= ~SCF_DO_STCLASS;
3804 else if (REGNODE_VARIES(OP(scan))) {
3805 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
3806 I32 fl = 0, f = flags;
3807 regnode * const oscan = scan;
3808 struct regnode_charclass_class this_class;
3809 struct regnode_charclass_class *oclass = NULL;
3810 I32 next_is_eval = 0;
3812 switch (PL_regkind[OP(scan)]) {
3813 case WHILEM: /* End of (?:...)* . */
3814 scan = NEXTOPER(scan);
3817 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3818 next = NEXTOPER(scan);
3819 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3821 maxcount = REG_INFTY;
3822 next = regnext(scan);
3823 scan = NEXTOPER(scan);
3827 if (flags & SCF_DO_SUBSTR)
3832 if (flags & SCF_DO_STCLASS) {
3834 maxcount = REG_INFTY;
3835 next = regnext(scan);
3836 scan = NEXTOPER(scan);