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 I32 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 I32 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 I32 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 I32 last_end; /* min value, <0 unless valid. */
354 SV **longest; /* Either &l_fixed, or &l_float. */
355 SV *longest_fixed; /* longest fixed string found in pattern */
356 I32 offset_fixed; /* offset where it starts */
357 I32 *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 I32 offset_float_min; /* earliest point in string it can appear */
361 I32 offset_float_max; /* latest point in string it can appear */
362 I32 *minlen_float; /* pointer to the minlen relevant to the string */
363 I32 lookbehind_float; /* is the position of the string modified by LB */
367 struct regnode_charclass_class *start_class;
371 * Forward declarations for pregcomp()'s friends.
374 static const scan_data_t zero_scan_data =
375 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
377 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
378 #define SF_BEFORE_SEOL 0x0001
379 #define SF_BEFORE_MEOL 0x0002
380 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
381 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
384 # define SF_FIX_SHIFT_EOL (0+2)
385 # define SF_FL_SHIFT_EOL (0+4)
387 # define SF_FIX_SHIFT_EOL (+2)
388 # define SF_FL_SHIFT_EOL (+4)
391 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
392 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
394 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
395 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
396 #define SF_IS_INF 0x0040
397 #define SF_HAS_PAR 0x0080
398 #define SF_IN_PAR 0x0100
399 #define SF_HAS_EVAL 0x0200
400 #define SCF_DO_SUBSTR 0x0400
401 #define SCF_DO_STCLASS_AND 0x0800
402 #define SCF_DO_STCLASS_OR 0x1000
403 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
404 #define SCF_WHILEM_VISITED_POS 0x2000
406 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
407 #define SCF_SEEN_ACCEPT 0x8000
409 #define UTF cBOOL(RExC_utf8)
411 /* The enums for all these are ordered so things work out correctly */
412 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
413 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
414 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
415 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
416 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
417 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
418 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
420 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
422 #define OOB_NAMEDCLASS -1
424 /* There is no code point that is out-of-bounds, so this is problematic. But
425 * its only current use is to initialize a variable that is always set before
427 #define OOB_UNICODE 0xDEADBEEF
429 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
430 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
433 /* length of regex to show in messages that don't mark a position within */
434 #define RegexLengthToShowInErrorMessages 127
437 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
438 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
439 * op/pragma/warn/regcomp.
441 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
442 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
444 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
447 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
448 * arg. Show regex, up to a maximum length. If it's too long, chop and add
451 #define _FAIL(code) STMT_START { \
452 const char *ellipses = ""; \
453 IV len = RExC_end - RExC_precomp; \
456 SAVEFREESV(RExC_rx_sv); \
457 if (len > RegexLengthToShowInErrorMessages) { \
458 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
459 len = RegexLengthToShowInErrorMessages - 10; \
465 #define FAIL(msg) _FAIL( \
466 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
467 msg, (int)len, RExC_precomp, ellipses))
469 #define FAIL2(msg,arg) _FAIL( \
470 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
471 arg, (int)len, RExC_precomp, ellipses))
474 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
476 #define Simple_vFAIL(m) STMT_START { \
477 const IV offset = RExC_parse - RExC_precomp; \
478 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
479 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
483 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
485 #define vFAIL(m) STMT_START { \
487 SAVEFREESV(RExC_rx_sv); \
492 * Like Simple_vFAIL(), but accepts two arguments.
494 #define Simple_vFAIL2(m,a1) STMT_START { \
495 const IV offset = RExC_parse - RExC_precomp; \
496 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
497 (int)offset, RExC_precomp, RExC_precomp + offset); \
501 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
503 #define vFAIL2(m,a1) STMT_START { \
505 SAVEFREESV(RExC_rx_sv); \
506 Simple_vFAIL2(m, a1); \
511 * Like Simple_vFAIL(), but accepts three arguments.
513 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
514 const IV offset = RExC_parse - RExC_precomp; \
515 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
516 (int)offset, RExC_precomp, RExC_precomp + offset); \
520 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
522 #define vFAIL3(m,a1,a2) STMT_START { \
524 SAVEFREESV(RExC_rx_sv); \
525 Simple_vFAIL3(m, a1, a2); \
529 * Like Simple_vFAIL(), but accepts four arguments.
531 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
532 const IV offset = RExC_parse - RExC_precomp; \
533 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
534 (int)offset, RExC_precomp, RExC_precomp + offset); \
537 #define vFAIL4(m,a1,a2,a3) STMT_START { \
539 SAVEFREESV(RExC_rx_sv); \
540 Simple_vFAIL4(m, a1, a2, a3); \
543 /* m is not necessarily a "literal string", in this macro */
544 #define reg_warn_non_literal_string(loc, m) STMT_START { \
545 const IV offset = loc - RExC_precomp; \
546 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
547 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
550 #define ckWARNreg(loc,m) STMT_START { \
551 const IV offset = loc - RExC_precomp; \
552 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
553 (int)offset, RExC_precomp, RExC_precomp + offset); \
556 #define vWARN_dep(loc, m) STMT_START { \
557 const IV offset = loc - RExC_precomp; \
558 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
559 (int)offset, RExC_precomp, RExC_precomp + offset); \
562 #define ckWARNdep(loc,m) STMT_START { \
563 const IV offset = loc - RExC_precomp; \
564 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
566 (int)offset, RExC_precomp, RExC_precomp + offset); \
569 #define ckWARNregdep(loc,m) STMT_START { \
570 const IV offset = loc - RExC_precomp; \
571 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
573 (int)offset, RExC_precomp, RExC_precomp + offset); \
576 #define ckWARN2regdep(loc,m, a1) STMT_START { \
577 const IV offset = loc - RExC_precomp; \
578 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
580 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
583 #define ckWARN2reg(loc, m, a1) STMT_START { \
584 const IV offset = loc - RExC_precomp; \
585 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
586 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
589 #define vWARN3(loc, m, a1, a2) STMT_START { \
590 const IV offset = loc - RExC_precomp; \
591 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
592 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
595 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
596 const IV offset = loc - RExC_precomp; \
597 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
598 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
601 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
602 const IV offset = loc - RExC_precomp; \
603 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
604 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
607 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
608 const IV offset = loc - RExC_precomp; \
609 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
610 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
613 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
614 const IV offset = loc - RExC_precomp; \
615 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
616 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
620 /* Allow for side effects in s */
621 #define REGC(c,s) STMT_START { \
622 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
625 /* Macros for recording node offsets. 20001227 mjd@plover.com
626 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
627 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
628 * Element 0 holds the number n.
629 * Position is 1 indexed.
631 #ifndef RE_TRACK_PATTERN_OFFSETS
632 #define Set_Node_Offset_To_R(node,byte)
633 #define Set_Node_Offset(node,byte)
634 #define Set_Cur_Node_Offset
635 #define Set_Node_Length_To_R(node,len)
636 #define Set_Node_Length(node,len)
637 #define Set_Node_Cur_Length(node,start)
638 #define Node_Offset(n)
639 #define Node_Length(n)
640 #define Set_Node_Offset_Length(node,offset,len)
641 #define ProgLen(ri) ri->u.proglen
642 #define SetProgLen(ri,x) ri->u.proglen = x
644 #define ProgLen(ri) ri->u.offsets[0]
645 #define SetProgLen(ri,x) ri->u.offsets[0] = x
646 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
648 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
649 __LINE__, (int)(node), (int)(byte))); \
651 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
653 RExC_offsets[2*(node)-1] = (byte); \
658 #define Set_Node_Offset(node,byte) \
659 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
660 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
662 #define Set_Node_Length_To_R(node,len) STMT_START { \
664 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
665 __LINE__, (int)(node), (int)(len))); \
667 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
669 RExC_offsets[2*(node)] = (len); \
674 #define Set_Node_Length(node,len) \
675 Set_Node_Length_To_R((node)-RExC_emit_start, len)
676 #define Set_Node_Cur_Length(node, start) \
677 Set_Node_Length(node, RExC_parse - start)
679 /* Get offsets and lengths */
680 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
681 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
683 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
684 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
685 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
689 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
690 #define EXPERIMENTAL_INPLACESCAN
691 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
693 #define DEBUG_STUDYDATA(str,data,depth) \
694 DEBUG_OPTIMISE_MORE_r(if(data){ \
695 PerlIO_printf(Perl_debug_log, \
696 "%*s" str "Pos:%"IVdf"/%"IVdf \
697 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
698 (int)(depth)*2, "", \
699 (IV)((data)->pos_min), \
700 (IV)((data)->pos_delta), \
701 (UV)((data)->flags), \
702 (IV)((data)->whilem_c), \
703 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
704 is_inf ? "INF " : "" \
706 if ((data)->last_found) \
707 PerlIO_printf(Perl_debug_log, \
708 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
709 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
710 SvPVX_const((data)->last_found), \
711 (IV)((data)->last_end), \
712 (IV)((data)->last_start_min), \
713 (IV)((data)->last_start_max), \
714 ((data)->longest && \
715 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
716 SvPVX_const((data)->longest_fixed), \
717 (IV)((data)->offset_fixed), \
718 ((data)->longest && \
719 (data)->longest==&((data)->longest_float)) ? "*" : "", \
720 SvPVX_const((data)->longest_float), \
721 (IV)((data)->offset_float_min), \
722 (IV)((data)->offset_float_max) \
724 PerlIO_printf(Perl_debug_log,"\n"); \
727 /* Mark that we cannot extend a found fixed substring at this point.
728 Update the longest found anchored substring and the longest found
729 floating substrings if needed. */
732 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
734 const STRLEN l = CHR_SVLEN(data->last_found);
735 const STRLEN old_l = CHR_SVLEN(*data->longest);
736 GET_RE_DEBUG_FLAGS_DECL;
738 PERL_ARGS_ASSERT_SCAN_COMMIT;
740 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
741 SvSetMagicSV(*data->longest, data->last_found);
742 if (*data->longest == data->longest_fixed) {
743 data->offset_fixed = l ? data->last_start_min : data->pos_min;
744 if (data->flags & SF_BEFORE_EOL)
746 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
748 data->flags &= ~SF_FIX_BEFORE_EOL;
749 data->minlen_fixed=minlenp;
750 data->lookbehind_fixed=0;
752 else { /* *data->longest == data->longest_float */
753 data->offset_float_min = l ? data->last_start_min : data->pos_min;
754 data->offset_float_max = (l
755 ? data->last_start_max
756 : (data->pos_delta == I32_MAX ? I32_MAX : data->pos_min + data->pos_delta));
757 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
758 data->offset_float_max = I32_MAX;
759 if (data->flags & SF_BEFORE_EOL)
761 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
763 data->flags &= ~SF_FL_BEFORE_EOL;
764 data->minlen_float=minlenp;
765 data->lookbehind_float=0;
768 SvCUR_set(data->last_found, 0);
770 SV * const sv = data->last_found;
771 if (SvUTF8(sv) && SvMAGICAL(sv)) {
772 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
778 data->flags &= ~SF_BEFORE_EOL;
779 DEBUG_STUDYDATA("commit: ",data,0);
782 /* These macros set, clear and test whether the synthetic start class ('ssc',
783 * given by the parameter) matches an empty string (EOS). This uses the
784 * 'next_off' field in the node, to save a bit in the flags field. The ssc
785 * stands alone, so there is never a next_off, so this field is otherwise
786 * unused. The EOS information is used only for compilation, but theoretically
787 * it could be passed on to the execution code. This could be used to store
788 * more than one bit of information, but only this one is currently used. */
789 #define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
790 #define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
791 #define TEST_SSC_EOS(node) cBOOL((node)->next_off)
793 /* Can match anything (initialization) */
795 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
797 PERL_ARGS_ASSERT_CL_ANYTHING;
799 ANYOF_BITMAP_SETALL(cl);
800 cl->flags = ANYOF_UNICODE_ALL;
803 /* If any portion of the regex is to operate under locale rules,
804 * initialization includes it. The reason this isn't done for all regexes
805 * is that the optimizer was written under the assumption that locale was
806 * all-or-nothing. Given the complexity and lack of documentation in the
807 * optimizer, and that there are inadequate test cases for locale, so many
808 * parts of it may not work properly, it is safest to avoid locale unless
810 if (RExC_contains_locale) {
811 ANYOF_CLASS_SETALL(cl); /* /l uses class */
812 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
815 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
819 /* Can match anything (initialization) */
821 S_cl_is_anything(const struct regnode_charclass_class *cl)
825 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
827 for (value = 0; value < ANYOF_MAX; value += 2)
828 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
830 if (!(cl->flags & ANYOF_UNICODE_ALL))
832 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
837 /* Can match anything (initialization) */
839 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
841 PERL_ARGS_ASSERT_CL_INIT;
843 Zero(cl, 1, struct regnode_charclass_class);
845 cl_anything(pRExC_state, cl);
846 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
849 /* These two functions currently do the exact same thing */
850 #define cl_init_zero S_cl_init
852 /* 'AND' a given class with another one. Can create false positives. 'cl'
853 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
854 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
856 S_cl_and(struct regnode_charclass_class *cl,
857 const struct regnode_charclass_class *and_with)
859 PERL_ARGS_ASSERT_CL_AND;
861 assert(PL_regkind[and_with->type] == ANYOF);
863 /* I (khw) am not sure all these restrictions are necessary XXX */
864 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
865 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
866 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
867 && !(and_with->flags & ANYOF_LOC_FOLD)
868 && !(cl->flags & ANYOF_LOC_FOLD)) {
871 if (and_with->flags & ANYOF_INVERT)
872 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
873 cl->bitmap[i] &= ~and_with->bitmap[i];
875 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
876 cl->bitmap[i] &= and_with->bitmap[i];
877 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
879 if (and_with->flags & ANYOF_INVERT) {
881 /* Here, the and'ed node is inverted. Get the AND of the flags that
882 * aren't affected by the inversion. Those that are affected are
883 * handled individually below */
884 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
885 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
886 cl->flags |= affected_flags;
888 /* We currently don't know how to deal with things that aren't in the
889 * bitmap, but we know that the intersection is no greater than what
890 * is already in cl, so let there be false positives that get sorted
891 * out after the synthetic start class succeeds, and the node is
892 * matched for real. */
894 /* The inversion of these two flags indicate that the resulting
895 * intersection doesn't have them */
896 if (and_with->flags & ANYOF_UNICODE_ALL) {
897 cl->flags &= ~ANYOF_UNICODE_ALL;
899 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
900 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
903 else { /* and'd node is not inverted */
904 U8 outside_bitmap_but_not_utf8; /* Temp variable */
906 if (! ANYOF_NONBITMAP(and_with)) {
908 /* Here 'and_with' doesn't match anything outside the bitmap
909 * (except possibly ANYOF_UNICODE_ALL), which means the
910 * intersection can't either, except for ANYOF_UNICODE_ALL, in
911 * which case we don't know what the intersection is, but it's no
912 * greater than what cl already has, so can just leave it alone,
913 * with possible false positives */
914 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
915 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
916 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
919 else if (! ANYOF_NONBITMAP(cl)) {
921 /* Here, 'and_with' does match something outside the bitmap, and cl
922 * doesn't have a list of things to match outside the bitmap. If
923 * cl can match all code points above 255, the intersection will
924 * be those above-255 code points that 'and_with' matches. If cl
925 * can't match all Unicode code points, it means that it can't
926 * match anything outside the bitmap (since the 'if' that got us
927 * into this block tested for that), so we leave the bitmap empty.
929 if (cl->flags & ANYOF_UNICODE_ALL) {
930 ARG_SET(cl, ARG(and_with));
932 /* and_with's ARG may match things that don't require UTF8.
933 * And now cl's will too, in spite of this being an 'and'. See
934 * the comments below about the kludge */
935 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
939 /* Here, both 'and_with' and cl match something outside the
940 * bitmap. Currently we do not do the intersection, so just match
941 * whatever cl had at the beginning. */
945 /* Take the intersection of the two sets of flags. However, the
946 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
947 * kludge around the fact that this flag is not treated like the others
948 * which are initialized in cl_anything(). The way the optimizer works
949 * is that the synthetic start class (SSC) is initialized to match
950 * anything, and then the first time a real node is encountered, its
951 * values are AND'd with the SSC's with the result being the values of
952 * the real node. However, there are paths through the optimizer where
953 * the AND never gets called, so those initialized bits are set
954 * inappropriately, which is not usually a big deal, as they just cause
955 * false positives in the SSC, which will just mean a probably
956 * imperceptible slow down in execution. However this bit has a
957 * higher false positive consequence in that it can cause utf8.pm,
958 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
959 * bigger slowdown and also causes significant extra memory to be used.
960 * In order to prevent this, the code now takes a different tack. The
961 * bit isn't set unless some part of the regular expression needs it,
962 * but once set it won't get cleared. This means that these extra
963 * modules won't get loaded unless there was some path through the
964 * pattern that would have required them anyway, and so any false
965 * positives that occur by not ANDing them out when they could be
966 * aren't as severe as they would be if we treated this bit like all
968 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
969 & ANYOF_NONBITMAP_NON_UTF8;
970 cl->flags &= and_with->flags;
971 cl->flags |= outside_bitmap_but_not_utf8;
975 /* 'OR' a given class with another one. Can create false positives. 'cl'
976 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
977 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
979 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
981 PERL_ARGS_ASSERT_CL_OR;
983 if (or_with->flags & ANYOF_INVERT) {
985 /* Here, the or'd node is to be inverted. This means we take the
986 * complement of everything not in the bitmap, but currently we don't
987 * know what that is, so give up and match anything */
988 if (ANYOF_NONBITMAP(or_with)) {
989 cl_anything(pRExC_state, cl);
992 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
993 * <= (B1 | !B2) | (CL1 | !CL2)
994 * which is wasteful if CL2 is small, but we ignore CL2:
995 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
996 * XXXX Can we handle case-fold? Unclear:
997 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
998 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
1000 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1001 && !(or_with->flags & ANYOF_LOC_FOLD)
1002 && !(cl->flags & ANYOF_LOC_FOLD) ) {
1005 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1006 cl->bitmap[i] |= ~or_with->bitmap[i];
1007 } /* XXXX: logic is complicated otherwise */
1009 cl_anything(pRExC_state, cl);
1012 /* And, we can just take the union of the flags that aren't affected
1013 * by the inversion */
1014 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1016 /* For the remaining flags:
1017 ANYOF_UNICODE_ALL and inverted means to not match anything above
1018 255, which means that the union with cl should just be
1019 what cl has in it, so can ignore this flag
1020 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
1021 is 127-255 to match them, but then invert that, so the
1022 union with cl should just be what cl has in it, so can
1025 } else { /* 'or_with' is not inverted */
1026 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1027 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1028 && (!(or_with->flags & ANYOF_LOC_FOLD)
1029 || (cl->flags & ANYOF_LOC_FOLD)) ) {
1032 /* OR char bitmap and class bitmap separately */
1033 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1034 cl->bitmap[i] |= or_with->bitmap[i];
1035 if (or_with->flags & ANYOF_CLASS) {
1036 ANYOF_CLASS_OR(or_with, cl);
1039 else { /* XXXX: logic is complicated, leave it along for a moment. */
1040 cl_anything(pRExC_state, cl);
1043 if (ANYOF_NONBITMAP(or_with)) {
1045 /* Use the added node's outside-the-bit-map match if there isn't a
1046 * conflict. If there is a conflict (both nodes match something
1047 * outside the bitmap, but what they match outside is not the same
1048 * pointer, and hence not easily compared until XXX we extend
1049 * inversion lists this far), give up and allow the start class to
1050 * match everything outside the bitmap. If that stuff is all above
1051 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1052 if (! ANYOF_NONBITMAP(cl)) {
1053 ARG_SET(cl, ARG(or_with));
1055 else if (ARG(cl) != ARG(or_with)) {
1057 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1058 cl_anything(pRExC_state, cl);
1061 cl->flags |= ANYOF_UNICODE_ALL;
1066 /* Take the union */
1067 cl->flags |= or_with->flags;
1071 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1072 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1073 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1074 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1079 dump_trie(trie,widecharmap,revcharmap)
1080 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1081 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1083 These routines dump out a trie in a somewhat readable format.
1084 The _interim_ variants are used for debugging the interim
1085 tables that are used to generate the final compressed
1086 representation which is what dump_trie expects.
1088 Part of the reason for their existence is to provide a form
1089 of documentation as to how the different representations function.
1094 Dumps the final compressed table form of the trie to Perl_debug_log.
1095 Used for debugging make_trie().
1099 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1100 AV *revcharmap, U32 depth)
1103 SV *sv=sv_newmortal();
1104 int colwidth= widecharmap ? 6 : 4;
1106 GET_RE_DEBUG_FLAGS_DECL;
1108 PERL_ARGS_ASSERT_DUMP_TRIE;
1110 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1111 (int)depth * 2 + 2,"",
1112 "Match","Base","Ofs" );
1114 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1115 SV ** const tmp = av_fetch( revcharmap, state, 0);
1117 PerlIO_printf( Perl_debug_log, "%*s",
1119 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1120 PL_colors[0], PL_colors[1],
1121 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1122 PERL_PV_ESCAPE_FIRSTCHAR
1127 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1128 (int)depth * 2 + 2,"");
1130 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1131 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1132 PerlIO_printf( Perl_debug_log, "\n");
1134 for( state = 1 ; state < trie->statecount ; state++ ) {
1135 const U32 base = trie->states[ state ].trans.base;
1137 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1139 if ( trie->states[ state ].wordnum ) {
1140 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1142 PerlIO_printf( Perl_debug_log, "%6s", "" );
1145 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1150 while( ( base + ofs < trie->uniquecharcount ) ||
1151 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1152 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1155 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1157 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1158 if ( ( base + ofs >= trie->uniquecharcount ) &&
1159 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1160 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1162 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1164 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1166 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1170 PerlIO_printf( Perl_debug_log, "]");
1173 PerlIO_printf( Perl_debug_log, "\n" );
1175 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1176 for (word=1; word <= trie->wordcount; word++) {
1177 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1178 (int)word, (int)(trie->wordinfo[word].prev),
1179 (int)(trie->wordinfo[word].len));
1181 PerlIO_printf(Perl_debug_log, "\n" );
1184 Dumps a fully constructed but uncompressed trie in list form.
1185 List tries normally only are used for construction when the number of
1186 possible chars (trie->uniquecharcount) is very high.
1187 Used for debugging make_trie().
1190 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1191 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1195 SV *sv=sv_newmortal();
1196 int colwidth= widecharmap ? 6 : 4;
1197 GET_RE_DEBUG_FLAGS_DECL;
1199 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1201 /* print out the table precompression. */
1202 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1203 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1204 "------:-----+-----------------\n" );
1206 for( state=1 ; state < next_alloc ; state ++ ) {
1209 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1210 (int)depth * 2 + 2,"", (UV)state );
1211 if ( ! trie->states[ state ].wordnum ) {
1212 PerlIO_printf( Perl_debug_log, "%5s| ","");
1214 PerlIO_printf( Perl_debug_log, "W%4x| ",
1215 trie->states[ state ].wordnum
1218 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1219 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1221 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1223 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1224 PL_colors[0], PL_colors[1],
1225 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1226 PERL_PV_ESCAPE_FIRSTCHAR
1228 TRIE_LIST_ITEM(state,charid).forid,
1229 (UV)TRIE_LIST_ITEM(state,charid).newstate
1232 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1233 (int)((depth * 2) + 14), "");
1236 PerlIO_printf( Perl_debug_log, "\n");
1241 Dumps a fully constructed but uncompressed trie in table form.
1242 This is the normal DFA style state transition table, with a few
1243 twists to facilitate compression later.
1244 Used for debugging make_trie().
1247 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1248 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1253 SV *sv=sv_newmortal();
1254 int colwidth= widecharmap ? 6 : 4;
1255 GET_RE_DEBUG_FLAGS_DECL;
1257 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1260 print out the table precompression so that we can do a visual check
1261 that they are identical.
1264 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1266 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1267 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1269 PerlIO_printf( Perl_debug_log, "%*s",
1271 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1272 PL_colors[0], PL_colors[1],
1273 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1274 PERL_PV_ESCAPE_FIRSTCHAR
1280 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1282 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1283 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1286 PerlIO_printf( Perl_debug_log, "\n" );
1288 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1290 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1291 (int)depth * 2 + 2,"",
1292 (UV)TRIE_NODENUM( state ) );
1294 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1295 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1297 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1299 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1301 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1302 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1304 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1305 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1313 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1314 startbranch: the first branch in the whole branch sequence
1315 first : start branch of sequence of branch-exact nodes.
1316 May be the same as startbranch
1317 last : Thing following the last branch.
1318 May be the same as tail.
1319 tail : item following the branch sequence
1320 count : words in the sequence
1321 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1322 depth : indent depth
1324 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1326 A trie is an N'ary tree where the branches are determined by digital
1327 decomposition of the key. IE, at the root node you look up the 1st character and
1328 follow that branch repeat until you find the end of the branches. Nodes can be
1329 marked as "accepting" meaning they represent a complete word. Eg:
1333 would convert into the following structure. Numbers represent states, letters
1334 following numbers represent valid transitions on the letter from that state, if
1335 the number is in square brackets it represents an accepting state, otherwise it
1336 will be in parenthesis.
1338 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1342 (1) +-i->(6)-+-s->[7]
1344 +-s->(3)-+-h->(4)-+-e->[5]
1346 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1348 This shows that when matching against the string 'hers' we will begin at state 1
1349 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1350 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1351 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1352 single traverse. We store a mapping from accepting to state to which word was
1353 matched, and then when we have multiple possibilities we try to complete the
1354 rest of the regex in the order in which they occured in the alternation.
1356 The only prior NFA like behaviour that would be changed by the TRIE support is
1357 the silent ignoring of duplicate alternations which are of the form:
1359 / (DUPE|DUPE) X? (?{ ... }) Y /x
1361 Thus EVAL blocks following a trie may be called a different number of times with
1362 and without the optimisation. With the optimisations dupes will be silently
1363 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1364 the following demonstrates:
1366 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1368 which prints out 'word' three times, but
1370 'words'=~/(word|word|word)(?{ print $1 })S/
1372 which doesnt print it out at all. This is due to other optimisations kicking in.
1374 Example of what happens on a structural level:
1376 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1378 1: CURLYM[1] {1,32767}(18)
1389 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1390 and should turn into:
1392 1: CURLYM[1] {1,32767}(18)
1394 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1402 Cases where tail != last would be like /(?foo|bar)baz/:
1412 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1413 and would end up looking like:
1416 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1423 d = uvuni_to_utf8_flags(d, uv, 0);
1425 is the recommended Unicode-aware way of saying
1430 #define TRIE_STORE_REVCHAR(val) \
1433 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1434 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1435 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1436 SvCUR_set(zlopp, kapow - flrbbbbb); \
1439 av_push(revcharmap, zlopp); \
1441 char ooooff = (char)val; \
1442 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1446 #define TRIE_READ_CHAR STMT_START { \
1449 /* if it is UTF then it is either already folded, or does not need folding */ \
1450 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1452 else if (folder == PL_fold_latin1) { \
1453 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1454 if ( foldlen > 0 ) { \
1455 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1461 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, FOLD_FLAGS_FULL); \
1462 skiplen = UNISKIP(uvc); \
1463 foldlen -= skiplen; \
1464 scan = foldbuf + skiplen; \
1467 /* raw data, will be folded later if needed */ \
1475 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1476 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1477 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1478 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1480 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1481 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1482 TRIE_LIST_CUR( state )++; \
1485 #define TRIE_LIST_NEW(state) STMT_START { \
1486 Newxz( trie->states[ state ].trans.list, \
1487 4, reg_trie_trans_le ); \
1488 TRIE_LIST_CUR( state ) = 1; \
1489 TRIE_LIST_LEN( state ) = 4; \
1492 #define TRIE_HANDLE_WORD(state) STMT_START { \
1493 U16 dupe= trie->states[ state ].wordnum; \
1494 regnode * const noper_next = regnext( noper ); \
1497 /* store the word for dumping */ \
1499 if (OP(noper) != NOTHING) \
1500 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1502 tmp = newSVpvn_utf8( "", 0, UTF ); \
1503 av_push( trie_words, tmp ); \
1507 trie->wordinfo[curword].prev = 0; \
1508 trie->wordinfo[curword].len = wordlen; \
1509 trie->wordinfo[curword].accept = state; \
1511 if ( noper_next < tail ) { \
1513 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1514 trie->jump[curword] = (U16)(noper_next - convert); \
1516 jumper = noper_next; \
1518 nextbranch= regnext(cur); \
1522 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1523 /* chain, so that when the bits of chain are later */\
1524 /* linked together, the dups appear in the chain */\
1525 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1526 trie->wordinfo[dupe].prev = curword; \
1528 /* we haven't inserted this word yet. */ \
1529 trie->states[ state ].wordnum = curword; \
1534 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1535 ( ( base + charid >= ucharcount \
1536 && base + charid < ubound \
1537 && state == trie->trans[ base - ucharcount + charid ].check \
1538 && trie->trans[ base - ucharcount + charid ].next ) \
1539 ? trie->trans[ base - ucharcount + charid ].next \
1540 : ( state==1 ? special : 0 ) \
1544 #define MADE_JUMP_TRIE 2
1545 #define MADE_EXACT_TRIE 4
1548 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1551 /* first pass, loop through and scan words */
1552 reg_trie_data *trie;
1553 HV *widecharmap = NULL;
1554 AV *revcharmap = newAV();
1556 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1561 regnode *jumper = NULL;
1562 regnode *nextbranch = NULL;
1563 regnode *convert = NULL;
1564 U32 *prev_states; /* temp array mapping each state to previous one */
1565 /* we just use folder as a flag in utf8 */
1566 const U8 * folder = NULL;
1569 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1570 AV *trie_words = NULL;
1571 /* along with revcharmap, this only used during construction but both are
1572 * useful during debugging so we store them in the struct when debugging.
1575 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1576 STRLEN trie_charcount=0;
1578 SV *re_trie_maxbuff;
1579 GET_RE_DEBUG_FLAGS_DECL;
1581 PERL_ARGS_ASSERT_MAKE_TRIE;
1583 PERL_UNUSED_ARG(depth);
1590 case EXACTFU_TRICKYFOLD:
1591 case EXACTFU: folder = PL_fold_latin1; break;
1592 case EXACTF: folder = PL_fold; break;
1593 case EXACTFL: folder = PL_fold_locale; break;
1594 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1597 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1599 trie->startstate = 1;
1600 trie->wordcount = word_count;
1601 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1602 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1604 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1605 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1606 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1609 trie_words = newAV();
1612 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1613 if (!SvIOK(re_trie_maxbuff)) {
1614 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1616 DEBUG_TRIE_COMPILE_r({
1617 PerlIO_printf( Perl_debug_log,
1618 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1619 (int)depth * 2 + 2, "",
1620 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1621 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1625 /* Find the node we are going to overwrite */
1626 if ( first == startbranch && OP( last ) != BRANCH ) {
1627 /* whole branch chain */
1630 /* branch sub-chain */
1631 convert = NEXTOPER( first );
1634 /* -- First loop and Setup --
1636 We first traverse the branches and scan each word to determine if it
1637 contains widechars, and how many unique chars there are, this is
1638 important as we have to build a table with at least as many columns as we
1641 We use an array of integers to represent the character codes 0..255
1642 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1643 native representation of the character value as the key and IV's for the
1646 *TODO* If we keep track of how many times each character is used we can
1647 remap the columns so that the table compression later on is more
1648 efficient in terms of memory by ensuring the most common value is in the
1649 middle and the least common are on the outside. IMO this would be better
1650 than a most to least common mapping as theres a decent chance the most
1651 common letter will share a node with the least common, meaning the node
1652 will not be compressible. With a middle is most common approach the worst
1653 case is when we have the least common nodes twice.
1657 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1658 regnode *noper = NEXTOPER( cur );
1659 const U8 *uc = (U8*)STRING( noper );
1660 const U8 *e = uc + STR_LEN( noper );
1662 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1664 const U8 *scan = (U8*)NULL;
1665 U32 wordlen = 0; /* required init */
1667 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1669 if (OP(noper) == NOTHING) {
1670 regnode *noper_next= regnext(noper);
1671 if (noper_next != tail && OP(noper_next) == flags) {
1673 uc= (U8*)STRING(noper);
1674 e= uc + STR_LEN(noper);
1675 trie->minlen= STR_LEN(noper);
1682 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1683 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1684 regardless of encoding */
1685 if (OP( noper ) == EXACTFU_SS) {
1686 /* false positives are ok, so just set this */
1687 TRIE_BITMAP_SET(trie,0xDF);
1690 for ( ; uc < e ; uc += len ) {
1691 TRIE_CHARCOUNT(trie)++;
1696 U8 folded= folder[ (U8) uvc ];
1697 if ( !trie->charmap[ folded ] ) {
1698 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1699 TRIE_STORE_REVCHAR( folded );
1702 if ( !trie->charmap[ uvc ] ) {
1703 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1704 TRIE_STORE_REVCHAR( uvc );
1707 /* store the codepoint in the bitmap, and its folded
1709 TRIE_BITMAP_SET(trie, uvc);
1711 /* store the folded codepoint */
1712 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1715 /* store first byte of utf8 representation of
1716 variant codepoints */
1717 if (! UNI_IS_INVARIANT(uvc)) {
1718 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1721 set_bit = 0; /* We've done our bit :-) */
1726 widecharmap = newHV();
1728 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1731 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1733 if ( !SvTRUE( *svpp ) ) {
1734 sv_setiv( *svpp, ++trie->uniquecharcount );
1735 TRIE_STORE_REVCHAR(uvc);
1739 if( cur == first ) {
1740 trie->minlen = chars;
1741 trie->maxlen = chars;
1742 } else if (chars < trie->minlen) {
1743 trie->minlen = chars;
1744 } else if (chars > trie->maxlen) {
1745 trie->maxlen = chars;
1747 if (OP( noper ) == EXACTFU_SS) {
1748 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1749 if (trie->minlen > 1)
1752 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1753 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1754 * - We assume that any such sequence might match a 2 byte string */
1755 if (trie->minlen > 2 )
1759 } /* end first pass */
1760 DEBUG_TRIE_COMPILE_r(
1761 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1762 (int)depth * 2 + 2,"",
1763 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1764 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1765 (int)trie->minlen, (int)trie->maxlen )
1769 We now know what we are dealing with in terms of unique chars and
1770 string sizes so we can calculate how much memory a naive
1771 representation using a flat table will take. If it's over a reasonable
1772 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1773 conservative but potentially much slower representation using an array
1776 At the end we convert both representations into the same compressed
1777 form that will be used in regexec.c for matching with. The latter
1778 is a form that cannot be used to construct with but has memory
1779 properties similar to the list form and access properties similar
1780 to the table form making it both suitable for fast searches and
1781 small enough that its feasable to store for the duration of a program.
1783 See the comment in the code where the compressed table is produced
1784 inplace from the flat tabe representation for an explanation of how
1785 the compression works.
1790 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1793 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1795 Second Pass -- Array Of Lists Representation
1797 Each state will be represented by a list of charid:state records
1798 (reg_trie_trans_le) the first such element holds the CUR and LEN
1799 points of the allocated array. (See defines above).
1801 We build the initial structure using the lists, and then convert
1802 it into the compressed table form which allows faster lookups
1803 (but cant be modified once converted).
1806 STRLEN transcount = 1;
1808 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1809 "%*sCompiling trie using list compiler\n",
1810 (int)depth * 2 + 2, ""));
1812 trie->states = (reg_trie_state *)
1813 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1814 sizeof(reg_trie_state) );
1818 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1820 regnode *noper = NEXTOPER( cur );
1821 U8 *uc = (U8*)STRING( noper );
1822 const U8 *e = uc + STR_LEN( noper );
1823 U32 state = 1; /* required init */
1824 U16 charid = 0; /* sanity init */
1825 U8 *scan = (U8*)NULL; /* sanity init */
1826 STRLEN foldlen = 0; /* required init */
1827 U32 wordlen = 0; /* required init */
1828 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1831 if (OP(noper) == NOTHING) {
1832 regnode *noper_next= regnext(noper);
1833 if (noper_next != tail && OP(noper_next) == flags) {
1835 uc= (U8*)STRING(noper);
1836 e= uc + STR_LEN(noper);
1840 if (OP(noper) != NOTHING) {
1841 for ( ; uc < e ; uc += len ) {
1846 charid = trie->charmap[ uvc ];
1848 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1852 charid=(U16)SvIV( *svpp );
1855 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1862 if ( !trie->states[ state ].trans.list ) {
1863 TRIE_LIST_NEW( state );
1865 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1866 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1867 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1872 newstate = next_alloc++;
1873 prev_states[newstate] = state;
1874 TRIE_LIST_PUSH( state, charid, newstate );
1879 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1883 TRIE_HANDLE_WORD(state);
1885 } /* end second pass */
1887 /* next alloc is the NEXT state to be allocated */
1888 trie->statecount = next_alloc;
1889 trie->states = (reg_trie_state *)
1890 PerlMemShared_realloc( trie->states,
1892 * sizeof(reg_trie_state) );
1894 /* and now dump it out before we compress it */
1895 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1896 revcharmap, next_alloc,
1900 trie->trans = (reg_trie_trans *)
1901 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1908 for( state=1 ; state < next_alloc ; state ++ ) {
1912 DEBUG_TRIE_COMPILE_MORE_r(
1913 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1917 if (trie->states[state].trans.list) {
1918 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1922 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1923 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1924 if ( forid < minid ) {
1926 } else if ( forid > maxid ) {
1930 if ( transcount < tp + maxid - minid + 1) {
1932 trie->trans = (reg_trie_trans *)
1933 PerlMemShared_realloc( trie->trans,
1935 * sizeof(reg_trie_trans) );
1936 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1938 base = trie->uniquecharcount + tp - minid;
1939 if ( maxid == minid ) {
1941 for ( ; zp < tp ; zp++ ) {
1942 if ( ! trie->trans[ zp ].next ) {
1943 base = trie->uniquecharcount + zp - minid;
1944 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1945 trie->trans[ zp ].check = state;
1951 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1952 trie->trans[ tp ].check = state;
1957 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1958 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1959 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1960 trie->trans[ tid ].check = state;
1962 tp += ( maxid - minid + 1 );
1964 Safefree(trie->states[ state ].trans.list);
1967 DEBUG_TRIE_COMPILE_MORE_r(
1968 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1971 trie->states[ state ].trans.base=base;
1973 trie->lasttrans = tp + 1;
1977 Second Pass -- Flat Table Representation.
1979 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1980 We know that we will need Charcount+1 trans at most to store the data
1981 (one row per char at worst case) So we preallocate both structures
1982 assuming worst case.
1984 We then construct the trie using only the .next slots of the entry
1987 We use the .check field of the first entry of the node temporarily to
1988 make compression both faster and easier by keeping track of how many non
1989 zero fields are in the node.
1991 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1994 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1995 number representing the first entry of the node, and state as a
1996 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1997 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1998 are 2 entrys per node. eg:
2006 The table is internally in the right hand, idx form. However as we also
2007 have to deal with the states array which is indexed by nodenum we have to
2008 use TRIE_NODENUM() to convert.
2011 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2012 "%*sCompiling trie using table compiler\n",
2013 (int)depth * 2 + 2, ""));
2015 trie->trans = (reg_trie_trans *)
2016 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2017 * trie->uniquecharcount + 1,
2018 sizeof(reg_trie_trans) );
2019 trie->states = (reg_trie_state *)
2020 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2021 sizeof(reg_trie_state) );
2022 next_alloc = trie->uniquecharcount + 1;
2025 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2027 regnode *noper = NEXTOPER( cur );
2028 const U8 *uc = (U8*)STRING( noper );
2029 const U8 *e = uc + STR_LEN( noper );
2031 U32 state = 1; /* required init */
2033 U16 charid = 0; /* sanity init */
2034 U32 accept_state = 0; /* sanity init */
2035 U8 *scan = (U8*)NULL; /* sanity init */
2037 STRLEN foldlen = 0; /* required init */
2038 U32 wordlen = 0; /* required init */
2040 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2042 if (OP(noper) == NOTHING) {
2043 regnode *noper_next= regnext(noper);
2044 if (noper_next != tail && OP(noper_next) == flags) {
2046 uc= (U8*)STRING(noper);
2047 e= uc + STR_LEN(noper);
2051 if ( OP(noper) != NOTHING ) {
2052 for ( ; uc < e ; uc += len ) {
2057 charid = trie->charmap[ uvc ];
2059 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2060 charid = svpp ? (U16)SvIV(*svpp) : 0;
2064 if ( !trie->trans[ state + charid ].next ) {
2065 trie->trans[ state + charid ].next = next_alloc;
2066 trie->trans[ state ].check++;
2067 prev_states[TRIE_NODENUM(next_alloc)]
2068 = TRIE_NODENUM(state);
2069 next_alloc += trie->uniquecharcount;
2071 state = trie->trans[ state + charid ].next;
2073 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2075 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2078 accept_state = TRIE_NODENUM( state );
2079 TRIE_HANDLE_WORD(accept_state);
2081 } /* end second pass */
2083 /* and now dump it out before we compress it */
2084 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2086 next_alloc, depth+1));
2090 * Inplace compress the table.*
2092 For sparse data sets the table constructed by the trie algorithm will
2093 be mostly 0/FAIL transitions or to put it another way mostly empty.
2094 (Note that leaf nodes will not contain any transitions.)
2096 This algorithm compresses the tables by eliminating most such
2097 transitions, at the cost of a modest bit of extra work during lookup:
2099 - Each states[] entry contains a .base field which indicates the
2100 index in the state[] array wheres its transition data is stored.
2102 - If .base is 0 there are no valid transitions from that node.
2104 - If .base is nonzero then charid is added to it to find an entry in
2107 -If trans[states[state].base+charid].check!=state then the
2108 transition is taken to be a 0/Fail transition. Thus if there are fail
2109 transitions at the front of the node then the .base offset will point
2110 somewhere inside the previous nodes data (or maybe even into a node
2111 even earlier), but the .check field determines if the transition is
2115 The following process inplace converts the table to the compressed
2116 table: We first do not compress the root node 1,and mark all its
2117 .check pointers as 1 and set its .base pointer as 1 as well. This
2118 allows us to do a DFA construction from the compressed table later,
2119 and ensures that any .base pointers we calculate later are greater
2122 - We set 'pos' to indicate the first entry of the second node.
2124 - We then iterate over the columns of the node, finding the first and
2125 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2126 and set the .check pointers accordingly, and advance pos
2127 appropriately and repreat for the next node. Note that when we copy
2128 the next pointers we have to convert them from the original
2129 NODEIDX form to NODENUM form as the former is not valid post
2132 - If a node has no transitions used we mark its base as 0 and do not
2133 advance the pos pointer.
2135 - If a node only has one transition we use a second pointer into the
2136 structure to fill in allocated fail transitions from other states.
2137 This pointer is independent of the main pointer and scans forward
2138 looking for null transitions that are allocated to a state. When it
2139 finds one it writes the single transition into the "hole". If the
2140 pointer doesnt find one the single transition is appended as normal.
2142 - Once compressed we can Renew/realloc the structures to release the
2145 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2146 specifically Fig 3.47 and the associated pseudocode.
2150 const U32 laststate = TRIE_NODENUM( next_alloc );
2153 trie->statecount = laststate;
2155 for ( state = 1 ; state < laststate ; state++ ) {
2157 const U32 stateidx = TRIE_NODEIDX( state );
2158 const U32 o_used = trie->trans[ stateidx ].check;
2159 U32 used = trie->trans[ stateidx ].check;
2160 trie->trans[ stateidx ].check = 0;
2162 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2163 if ( flag || trie->trans[ stateidx + charid ].next ) {
2164 if ( trie->trans[ stateidx + charid ].next ) {
2166 for ( ; zp < pos ; zp++ ) {
2167 if ( ! trie->trans[ zp ].next ) {
2171 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2172 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2173 trie->trans[ zp ].check = state;
2174 if ( ++zp > pos ) pos = zp;
2181 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2183 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2184 trie->trans[ pos ].check = state;
2189 trie->lasttrans = pos + 1;
2190 trie->states = (reg_trie_state *)
2191 PerlMemShared_realloc( trie->states, laststate
2192 * sizeof(reg_trie_state) );
2193 DEBUG_TRIE_COMPILE_MORE_r(
2194 PerlIO_printf( Perl_debug_log,
2195 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2196 (int)depth * 2 + 2,"",
2197 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2200 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2203 } /* end table compress */
2205 DEBUG_TRIE_COMPILE_MORE_r(
2206 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2207 (int)depth * 2 + 2, "",
2208 (UV)trie->statecount,
2209 (UV)trie->lasttrans)
2211 /* resize the trans array to remove unused space */
2212 trie->trans = (reg_trie_trans *)
2213 PerlMemShared_realloc( trie->trans, trie->lasttrans
2214 * sizeof(reg_trie_trans) );
2216 { /* Modify the program and insert the new TRIE node */
2217 U8 nodetype =(U8)(flags & 0xFF);
2221 regnode *optimize = NULL;
2222 #ifdef RE_TRACK_PATTERN_OFFSETS
2225 U32 mjd_nodelen = 0;
2226 #endif /* RE_TRACK_PATTERN_OFFSETS */
2227 #endif /* DEBUGGING */
2229 This means we convert either the first branch or the first Exact,
2230 depending on whether the thing following (in 'last') is a branch
2231 or not and whther first is the startbranch (ie is it a sub part of
2232 the alternation or is it the whole thing.)
2233 Assuming its a sub part we convert the EXACT otherwise we convert
2234 the whole branch sequence, including the first.
2236 /* Find the node we are going to overwrite */
2237 if ( first != startbranch || OP( last ) == BRANCH ) {
2238 /* branch sub-chain */
2239 NEXT_OFF( first ) = (U16)(last - first);
2240 #ifdef RE_TRACK_PATTERN_OFFSETS
2242 mjd_offset= Node_Offset((convert));
2243 mjd_nodelen= Node_Length((convert));
2246 /* whole branch chain */
2248 #ifdef RE_TRACK_PATTERN_OFFSETS
2251 const regnode *nop = NEXTOPER( convert );
2252 mjd_offset= Node_Offset((nop));
2253 mjd_nodelen= Node_Length((nop));
2257 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2258 (int)depth * 2 + 2, "",
2259 (UV)mjd_offset, (UV)mjd_nodelen)
2262 /* But first we check to see if there is a common prefix we can
2263 split out as an EXACT and put in front of the TRIE node. */
2264 trie->startstate= 1;
2265 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2267 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2271 const U32 base = trie->states[ state ].trans.base;
2273 if ( trie->states[state].wordnum )
2276 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2277 if ( ( base + ofs >= trie->uniquecharcount ) &&
2278 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2279 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2281 if ( ++count > 1 ) {
2282 SV **tmp = av_fetch( revcharmap, ofs, 0);
2283 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2284 if ( state == 1 ) break;
2286 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2288 PerlIO_printf(Perl_debug_log,
2289 "%*sNew Start State=%"UVuf" Class: [",
2290 (int)depth * 2 + 2, "",
2293 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2294 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2296 TRIE_BITMAP_SET(trie,*ch);
2298 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2300 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2304 TRIE_BITMAP_SET(trie,*ch);
2306 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2307 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2313 SV **tmp = av_fetch( revcharmap, idx, 0);
2315 char *ch = SvPV( *tmp, len );
2317 SV *sv=sv_newmortal();
2318 PerlIO_printf( Perl_debug_log,
2319 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2320 (int)depth * 2 + 2, "",
2322 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2323 PL_colors[0], PL_colors[1],
2324 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2325 PERL_PV_ESCAPE_FIRSTCHAR
2330 OP( convert ) = nodetype;
2331 str=STRING(convert);
2334 STR_LEN(convert) += len;
2340 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2345 trie->prefixlen = (state-1);
2347 regnode *n = convert+NODE_SZ_STR(convert);
2348 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2349 trie->startstate = state;
2350 trie->minlen -= (state - 1);
2351 trie->maxlen -= (state - 1);
2353 /* At least the UNICOS C compiler choked on this
2354 * being argument to DEBUG_r(), so let's just have
2357 #ifdef PERL_EXT_RE_BUILD
2363 regnode *fix = convert;
2364 U32 word = trie->wordcount;
2366 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2367 while( ++fix < n ) {
2368 Set_Node_Offset_Length(fix, 0, 0);
2371 SV ** const tmp = av_fetch( trie_words, word, 0 );
2373 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2374 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2376 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2384 NEXT_OFF(convert) = (U16)(tail - convert);
2385 DEBUG_r(optimize= n);
2391 if ( trie->maxlen ) {
2392 NEXT_OFF( convert ) = (U16)(tail - convert);
2393 ARG_SET( convert, data_slot );
2394 /* Store the offset to the first unabsorbed branch in
2395 jump[0], which is otherwise unused by the jump logic.
2396 We use this when dumping a trie and during optimisation. */
2398 trie->jump[0] = (U16)(nextbranch - convert);
2400 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2401 * and there is a bitmap
2402 * and the first "jump target" node we found leaves enough room
2403 * then convert the TRIE node into a TRIEC node, with the bitmap
2404 * embedded inline in the opcode - this is hypothetically faster.
2406 if ( !trie->states[trie->startstate].wordnum
2408 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2410 OP( convert ) = TRIEC;
2411 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2412 PerlMemShared_free(trie->bitmap);
2415 OP( convert ) = TRIE;
2417 /* store the type in the flags */
2418 convert->flags = nodetype;
2422 + regarglen[ OP( convert ) ];
2424 /* XXX We really should free up the resource in trie now,
2425 as we won't use them - (which resources?) dmq */
2427 /* needed for dumping*/
2428 DEBUG_r(if (optimize) {
2429 regnode *opt = convert;
2431 while ( ++opt < optimize) {
2432 Set_Node_Offset_Length(opt,0,0);
2435 Try to clean up some of the debris left after the
2438 while( optimize < jumper ) {
2439 mjd_nodelen += Node_Length((optimize));
2440 OP( optimize ) = OPTIMIZED;
2441 Set_Node_Offset_Length(optimize,0,0);
2444 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2446 } /* end node insert */
2448 /* Finish populating the prev field of the wordinfo array. Walk back
2449 * from each accept state until we find another accept state, and if
2450 * so, point the first word's .prev field at the second word. If the
2451 * second already has a .prev field set, stop now. This will be the
2452 * case either if we've already processed that word's accept state,
2453 * or that state had multiple words, and the overspill words were
2454 * already linked up earlier.
2461 for (word=1; word <= trie->wordcount; word++) {
2463 if (trie->wordinfo[word].prev)
2465 state = trie->wordinfo[word].accept;
2467 state = prev_states[state];
2470 prev = trie->states[state].wordnum;
2474 trie->wordinfo[word].prev = prev;
2476 Safefree(prev_states);
2480 /* and now dump out the compressed format */
2481 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2483 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2485 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2486 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2488 SvREFCNT_dec_NN(revcharmap);
2492 : trie->startstate>1
2498 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2500 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2502 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2503 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2506 We find the fail state for each state in the trie, this state is the longest proper
2507 suffix of the current state's 'word' that is also a proper prefix of another word in our
2508 trie. State 1 represents the word '' and is thus the default fail state. This allows
2509 the DFA not to have to restart after its tried and failed a word at a given point, it
2510 simply continues as though it had been matching the other word in the first place.
2512 'abcdgu'=~/abcdefg|cdgu/
2513 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2514 fail, which would bring us to the state representing 'd' in the second word where we would
2515 try 'g' and succeed, proceeding to match 'cdgu'.
2517 /* add a fail transition */
2518 const U32 trie_offset = ARG(source);
2519 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2521 const U32 ucharcount = trie->uniquecharcount;
2522 const U32 numstates = trie->statecount;
2523 const U32 ubound = trie->lasttrans + ucharcount;
2527 U32 base = trie->states[ 1 ].trans.base;
2530 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2531 GET_RE_DEBUG_FLAGS_DECL;
2533 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2535 PERL_UNUSED_ARG(depth);
2539 ARG_SET( stclass, data_slot );
2540 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2541 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2542 aho->trie=trie_offset;
2543 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2544 Copy( trie->states, aho->states, numstates, reg_trie_state );
2545 Newxz( q, numstates, U32);
2546 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2549 /* initialize fail[0..1] to be 1 so that we always have
2550 a valid final fail state */
2551 fail[ 0 ] = fail[ 1 ] = 1;
2553 for ( charid = 0; charid < ucharcount ; charid++ ) {
2554 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2556 q[ q_write ] = newstate;
2557 /* set to point at the root */
2558 fail[ q[ q_write++ ] ]=1;
2561 while ( q_read < q_write) {
2562 const U32 cur = q[ q_read++ % numstates ];
2563 base = trie->states[ cur ].trans.base;
2565 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2566 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2568 U32 fail_state = cur;
2571 fail_state = fail[ fail_state ];
2572 fail_base = aho->states[ fail_state ].trans.base;
2573 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2575 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2576 fail[ ch_state ] = fail_state;
2577 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2579 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2581 q[ q_write++ % numstates] = ch_state;
2585 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2586 when we fail in state 1, this allows us to use the
2587 charclass scan to find a valid start char. This is based on the principle
2588 that theres a good chance the string being searched contains lots of stuff
2589 that cant be a start char.
2591 fail[ 0 ] = fail[ 1 ] = 0;
2592 DEBUG_TRIE_COMPILE_r({
2593 PerlIO_printf(Perl_debug_log,
2594 "%*sStclass Failtable (%"UVuf" states): 0",
2595 (int)(depth * 2), "", (UV)numstates
2597 for( q_read=1; q_read<numstates; q_read++ ) {
2598 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2600 PerlIO_printf(Perl_debug_log, "\n");
2603 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2608 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2609 * These need to be revisited when a newer toolchain becomes available.
2611 #if defined(__sparc64__) && defined(__GNUC__)
2612 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2613 # undef SPARC64_GCC_WORKAROUND
2614 # define SPARC64_GCC_WORKAROUND 1
2618 #define DEBUG_PEEP(str,scan,depth) \
2619 DEBUG_OPTIMISE_r({if (scan){ \
2620 SV * const mysv=sv_newmortal(); \
2621 regnode *Next = regnext(scan); \
2622 regprop(RExC_rx, mysv, scan); \
2623 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2624 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2625 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2629 /* The below joins as many adjacent EXACTish nodes as possible into a single
2630 * one. The regop may be changed if the node(s) contain certain sequences that
2631 * require special handling. The joining is only done if:
2632 * 1) there is room in the current conglomerated node to entirely contain the
2634 * 2) they are the exact same node type
2636 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2637 * these get optimized out
2639 * If a node is to match under /i (folded), the number of characters it matches
2640 * can be different than its character length if it contains a multi-character
2641 * fold. *min_subtract is set to the total delta of the input nodes.
2643 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2644 * and contains LATIN SMALL LETTER SHARP S
2646 * This is as good a place as any to discuss the design of handling these
2647 * multi-character fold sequences. It's been wrong in Perl for a very long
2648 * time. There are three code points in Unicode whose multi-character folds
2649 * were long ago discovered to mess things up. The previous designs for
2650 * dealing with these involved assigning a special node for them. This
2651 * approach doesn't work, as evidenced by this example:
2652 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2653 * Both these fold to "sss", but if the pattern is parsed to create a node that
2654 * would match just the \xDF, it won't be able to handle the case where a
2655 * successful match would have to cross the node's boundary. The new approach
2656 * that hopefully generally solves the problem generates an EXACTFU_SS node
2659 * It turns out that there are problems with all multi-character folds, and not
2660 * just these three. Now the code is general, for all such cases, but the
2661 * three still have some special handling. The approach taken is:
2662 * 1) This routine examines each EXACTFish node that could contain multi-
2663 * character fold sequences. It returns in *min_subtract how much to
2664 * subtract from the the actual length of the string to get a real minimum
2665 * match length; it is 0 if there are no multi-char folds. This delta is
2666 * used by the caller to adjust the min length of the match, and the delta
2667 * between min and max, so that the optimizer doesn't reject these
2668 * possibilities based on size constraints.
2669 * 2) Certain of these sequences require special handling by the trie code,
2670 * so, if found, this code changes the joined node type to special ops:
2671 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2672 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2673 * is used for an EXACTFU node that contains at least one "ss" sequence in
2674 * it. For non-UTF-8 patterns and strings, this is the only case where
2675 * there is a possible fold length change. That means that a regular
2676 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2677 * with length changes, and so can be processed faster. regexec.c takes
2678 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2679 * pre-folded by regcomp.c. This saves effort in regex matching.
2680 * However, the pre-folding isn't done for non-UTF8 patterns because the
2681 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2682 * down by forcing the pattern into UTF8 unless necessary. Also what
2683 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2684 * possibilities for the non-UTF8 patterns are quite simple, except for
2685 * the sharp s. All the ones that don't involve a UTF-8 target string are
2686 * members of a fold-pair, and arrays are set up for all of them so that
2687 * the other member of the pair can be found quickly. Code elsewhere in
2688 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2689 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2690 * described in the next item.
2691 * 4) A problem remains for the sharp s in EXACTF and EXACTFA nodes when the
2692 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2693 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2694 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2695 * character in the pattern corresponds to at most a single character in
2696 * the target string. (And I do mean character, and not byte here, unlike
2697 * other parts of the documentation that have never been updated to
2698 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2699 * two character string 'ss'; in EXACTFA nodes it can match
2700 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2701 * instances where it is violated. I'm reluctant to try to change the
2702 * assumption, as the code involved is impenetrable to me (khw), so
2703 * instead the code here punts. This routine examines (when the pattern
2704 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2705 * boolean indicating whether or not the node contains a sharp s. When it
2706 * is true, the caller sets a flag that later causes the optimizer in this
2707 * file to not set values for the floating and fixed string lengths, and
2708 * thus avoids the optimizer code in regexec.c that makes the invalid
2709 * assumption. Thus, there is no optimization based on string lengths for
2710 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2711 * (The reason the assumption is wrong only in these two cases is that all
2712 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2713 * other folds to their expanded versions. We can't prefold sharp s to
2714 * 'ss' in EXACTF nodes because we don't know at compile time if it
2715 * actually matches 'ss' or not. It will match iff the target string is
2716 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2717 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2718 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2719 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2720 * require the pattern to be forced into UTF-8, the overhead of which we
2724 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2725 if (PL_regkind[OP(scan)] == EXACT) \
2726 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2729 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) {
2730 /* Merge several consecutive EXACTish nodes into one. */
2731 regnode *n = regnext(scan);
2733 regnode *next = scan + NODE_SZ_STR(scan);
2737 regnode *stop = scan;
2738 GET_RE_DEBUG_FLAGS_DECL;
2740 PERL_UNUSED_ARG(depth);
2743 PERL_ARGS_ASSERT_JOIN_EXACT;
2744 #ifndef EXPERIMENTAL_INPLACESCAN
2745 PERL_UNUSED_ARG(flags);
2746 PERL_UNUSED_ARG(val);
2748 DEBUG_PEEP("join",scan,depth);
2750 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2751 * EXACT ones that are mergeable to the current one. */
2753 && (PL_regkind[OP(n)] == NOTHING
2754 || (stringok && OP(n) == OP(scan)))
2756 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2759 if (OP(n) == TAIL || n > next)
2761 if (PL_regkind[OP(n)] == NOTHING) {
2762 DEBUG_PEEP("skip:",n,depth);
2763 NEXT_OFF(scan) += NEXT_OFF(n);
2764 next = n + NODE_STEP_REGNODE;
2771 else if (stringok) {
2772 const unsigned int oldl = STR_LEN(scan);
2773 regnode * const nnext = regnext(n);
2775 /* XXX I (khw) kind of doubt that this works on platforms where
2776 * U8_MAX is above 255 because of lots of other assumptions */
2777 /* Don't join if the sum can't fit into a single node */
2778 if (oldl + STR_LEN(n) > U8_MAX)
2781 DEBUG_PEEP("merg",n,depth);
2784 NEXT_OFF(scan) += NEXT_OFF(n);
2785 STR_LEN(scan) += STR_LEN(n);
2786 next = n + NODE_SZ_STR(n);
2787 /* Now we can overwrite *n : */
2788 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2796 #ifdef EXPERIMENTAL_INPLACESCAN
2797 if (flags && !NEXT_OFF(n)) {
2798 DEBUG_PEEP("atch", val, depth);
2799 if (reg_off_by_arg[OP(n)]) {
2800 ARG_SET(n, val - n);
2803 NEXT_OFF(n) = val - n;
2811 *has_exactf_sharp_s = FALSE;
2813 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2814 * can now analyze for sequences of problematic code points. (Prior to
2815 * this final joining, sequences could have been split over boundaries, and
2816 * hence missed). The sequences only happen in folding, hence for any
2817 * non-EXACT EXACTish node */
2818 if (OP(scan) != EXACT) {
2819 const U8 * const s0 = (U8*) STRING(scan);
2821 const U8 * const s_end = s0 + STR_LEN(scan);
2823 /* One pass is made over the node's string looking for all the
2824 * possibilities. to avoid some tests in the loop, there are two main
2825 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2829 /* Examine the string for a multi-character fold sequence. UTF-8
2830 * patterns have all characters pre-folded by the time this code is
2832 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2833 length sequence we are looking for is 2 */
2836 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2837 if (! len) { /* Not a multi-char fold: get next char */
2842 /* Nodes with 'ss' require special handling, except for EXACTFL
2843 * and EXACTFA for which there is no multi-char fold to this */
2844 if (len == 2 && *s == 's' && *(s+1) == 's'
2845 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2848 OP(scan) = EXACTFU_SS;
2851 else if (len == 6 /* len is the same in both ASCII and EBCDIC
2853 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2854 COMBINING_DIAERESIS_UTF8
2855 COMBINING_ACUTE_ACCENT_UTF8,
2857 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2858 COMBINING_DIAERESIS_UTF8
2859 COMBINING_ACUTE_ACCENT_UTF8,
2864 /* These two folds require special handling by trie's, so
2865 * change the node type to indicate this. If EXACTFA and
2866 * EXACTFL were ever to be handled by trie's, this would
2867 * have to be changed. If this node has already been
2868 * changed to EXACTFU_SS in this loop, leave it as is. (I
2869 * (khw) think it doesn't matter in regexec.c for UTF
2870 * patterns, but no need to change it */
2871 if (OP(scan) == EXACTFU) {
2872 OP(scan) = EXACTFU_TRICKYFOLD;
2876 else { /* Here is a generic multi-char fold. */
2877 const U8* multi_end = s + len;
2879 /* Count how many characters in it. In the case of /l and
2880 * /aa, no folds which contain ASCII code points are
2881 * allowed, so check for those, and skip if found. (In
2882 * EXACTFL, no folds are allowed to any Latin1 code point,
2883 * not just ASCII. But there aren't any of these
2884 * currently, nor ever likely, so don't take the time to
2885 * test for them. The code that generates the
2886 * is_MULTI_foo() macros croaks should one actually get put
2887 * into Unicode .) */
2888 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2889 count = utf8_length(s, multi_end);
2893 while (s < multi_end) {
2896 goto next_iteration;
2906 /* The delta is how long the sequence is minus 1 (1 is how long
2907 * the character that folds to the sequence is) */
2908 *min_subtract += count - 1;
2912 else if (OP(scan) == EXACTFA) {
2914 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2915 * fold to the ASCII range (and there are no existing ones in the
2916 * upper latin1 range). But, as outlined in the comments preceding
2917 * this function, we need to flag any occurrences of the sharp s */
2919 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
2920 *has_exactf_sharp_s = TRUE;
2927 else if (OP(scan) != EXACTFL) {
2929 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2930 * multi-char folds that are all Latin1. (This code knows that
2931 * there are no current multi-char folds possible with EXACTFL,
2932 * relying on fold_grind.t to catch any errors if the very unlikely
2933 * event happens that some get added in future Unicode versions.)
2934 * As explained in the comments preceding this function, we look
2935 * also for the sharp s in EXACTF nodes; it can be in the final
2936 * position. Otherwise we can stop looking 1 byte earlier because
2937 * have to find at least two characters for a multi-fold */
2938 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2940 /* The below is perhaps overboard, but this allows us to save a
2941 * test each time through the loop at the expense of a mask. This
2942 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2943 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2944 * are 64. This uses an exclusive 'or' to find that bit and then
2945 * inverts it to form a mask, with just a single 0, in the bit
2946 * position where 'S' and 's' differ. */
2947 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2948 const U8 s_masked = 's' & S_or_s_mask;
2951 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2952 if (! len) { /* Not a multi-char fold. */
2953 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2955 *has_exactf_sharp_s = TRUE;
2962 && ((*s & S_or_s_mask) == s_masked)
2963 && ((*(s+1) & S_or_s_mask) == s_masked))
2966 /* EXACTF nodes need to know that the minimum length
2967 * changed so that a sharp s in the string can match this
2968 * ss in the pattern, but they remain EXACTF nodes, as they
2969 * won't match this unless the target string is is UTF-8,
2970 * which we don't know until runtime */
2971 if (OP(scan) != EXACTF) {
2972 OP(scan) = EXACTFU_SS;
2976 *min_subtract += len - 1;
2983 /* Allow dumping but overwriting the collection of skipped
2984 * ops and/or strings with fake optimized ops */
2985 n = scan + NODE_SZ_STR(scan);
2993 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2997 /* REx optimizer. Converts nodes into quicker variants "in place".
2998 Finds fixed substrings. */
3000 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3001 to the position after last scanned or to NULL. */
3003 #define INIT_AND_WITHP \
3004 assert(!and_withp); \
3005 Newx(and_withp,1,struct regnode_charclass_class); \
3006 SAVEFREEPV(and_withp)
3008 /* this is a chain of data about sub patterns we are processing that
3009 need to be handled separately/specially in study_chunk. Its so
3010 we can simulate recursion without losing state. */
3012 typedef struct scan_frame {
3013 regnode *last; /* last node to process in this frame */
3014 regnode *next; /* next node to process when last is reached */
3015 struct scan_frame *prev; /*previous frame*/
3016 I32 stop; /* what stopparen do we use */
3020 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3023 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3024 I32 *minlenp, I32 *deltap,
3029 struct regnode_charclass_class *and_withp,
3030 U32 flags, U32 depth)
3031 /* scanp: Start here (read-write). */
3032 /* deltap: Write maxlen-minlen here. */
3033 /* last: Stop before this one. */
3034 /* data: string data about the pattern */
3035 /* stopparen: treat close N as END */
3036 /* recursed: which subroutines have we recursed into */
3037 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3040 I32 min = 0; /* There must be at least this number of characters to match */
3042 regnode *scan = *scanp, *next;
3044 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3045 int is_inf_internal = 0; /* The studied chunk is infinite */
3046 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3047 scan_data_t data_fake;
3048 SV *re_trie_maxbuff = NULL;
3049 regnode *first_non_open = scan;
3050 I32 stopmin = I32_MAX;
3051 scan_frame *frame = NULL;
3052 GET_RE_DEBUG_FLAGS_DECL;
3054 PERL_ARGS_ASSERT_STUDY_CHUNK;
3057 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3061 while (first_non_open && OP(first_non_open) == OPEN)
3062 first_non_open=regnext(first_non_open);
3067 while ( scan && OP(scan) != END && scan < last ){
3068 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3069 node length to get a real minimum (because
3070 the folded version may be shorter) */
3071 bool has_exactf_sharp_s = FALSE;
3072 /* Peephole optimizer: */
3073 DEBUG_STUDYDATA("Peep:", data,depth);
3074 DEBUG_PEEP("Peep",scan,depth);
3076 /* Its not clear to khw or hv why this is done here, and not in the
3077 * clauses that deal with EXACT nodes. khw's guess is that it's
3078 * because of a previous design */
3079 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3081 /* Follow the next-chain of the current node and optimize
3082 away all the NOTHINGs from it. */
3083 if (OP(scan) != CURLYX) {
3084 const int max = (reg_off_by_arg[OP(scan)]
3086 /* I32 may be smaller than U16 on CRAYs! */
3087 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3088 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3092 /* Skip NOTHING and LONGJMP. */
3093 while ((n = regnext(n))
3094 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3095 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3096 && off + noff < max)
3098 if (reg_off_by_arg[OP(scan)])
3101 NEXT_OFF(scan) = off;
3106 /* The principal pseudo-switch. Cannot be a switch, since we
3107 look into several different things. */
3108 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3109 || OP(scan) == IFTHEN) {
3110 next = regnext(scan);
3112 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3114 if (OP(next) == code || code == IFTHEN) {
3115 /* NOTE - There is similar code to this block below for handling
3116 TRIE nodes on a re-study. If you change stuff here check there
3118 I32 max1 = 0, min1 = I32_MAX, num = 0;
3119 struct regnode_charclass_class accum;
3120 regnode * const startbranch=scan;
3122 if (flags & SCF_DO_SUBSTR)
3123 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3124 if (flags & SCF_DO_STCLASS)
3125 cl_init_zero(pRExC_state, &accum);
3127 while (OP(scan) == code) {
3128 I32 deltanext, minnext, f = 0, fake;
3129 struct regnode_charclass_class this_class;
3132 data_fake.flags = 0;
3134 data_fake.whilem_c = data->whilem_c;
3135 data_fake.last_closep = data->last_closep;
3138 data_fake.last_closep = &fake;
3140 data_fake.pos_delta = delta;
3141 next = regnext(scan);
3142 scan = NEXTOPER(scan);
3144 scan = NEXTOPER(scan);
3145 if (flags & SCF_DO_STCLASS) {
3146 cl_init(pRExC_state, &this_class);
3147 data_fake.start_class = &this_class;
3148 f = SCF_DO_STCLASS_AND;
3150 if (flags & SCF_WHILEM_VISITED_POS)
3151 f |= SCF_WHILEM_VISITED_POS;
3153 /* we suppose the run is continuous, last=next...*/
3154 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3156 stopparen, recursed, NULL, f,depth+1);
3159 if (deltanext == I32_MAX) {
3160 is_inf = is_inf_internal = 1;
3162 } else if (max1 < minnext + deltanext)
3163 max1 = minnext + deltanext;
3165 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3167 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3168 if ( stopmin > minnext)
3169 stopmin = min + min1;
3170 flags &= ~SCF_DO_SUBSTR;
3172 data->flags |= SCF_SEEN_ACCEPT;
3175 if (data_fake.flags & SF_HAS_EVAL)
3176 data->flags |= SF_HAS_EVAL;
3177 data->whilem_c = data_fake.whilem_c;
3179 if (flags & SCF_DO_STCLASS)
3180 cl_or(pRExC_state, &accum, &this_class);
3182 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3184 if (flags & SCF_DO_SUBSTR) {
3185 data->pos_min += min1;
3186 if (data->pos_delta >= I32_MAX - (max1 - min1))
3187 data->pos_delta = I32_MAX;
3189 data->pos_delta += max1 - min1;
3190 if (max1 != min1 || is_inf)
3191 data->longest = &(data->longest_float);
3194 if (delta == I32_MAX || I32_MAX - delta - (max1 - min1) < 0)
3197 delta += max1 - min1;
3198 if (flags & SCF_DO_STCLASS_OR) {
3199 cl_or(pRExC_state, data->start_class, &accum);
3201 cl_and(data->start_class, and_withp);
3202 flags &= ~SCF_DO_STCLASS;
3205 else if (flags & SCF_DO_STCLASS_AND) {
3207 cl_and(data->start_class, &accum);
3208 flags &= ~SCF_DO_STCLASS;
3211 /* Switch to OR mode: cache the old value of
3212 * data->start_class */
3214 StructCopy(data->start_class, and_withp,
3215 struct regnode_charclass_class);
3216 flags &= ~SCF_DO_STCLASS_AND;
3217 StructCopy(&accum, data->start_class,
3218 struct regnode_charclass_class);
3219 flags |= SCF_DO_STCLASS_OR;
3220 SET_SSC_EOS(data->start_class);
3224 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3227 Assuming this was/is a branch we are dealing with: 'scan' now
3228 points at the item that follows the branch sequence, whatever
3229 it is. We now start at the beginning of the sequence and look
3236 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3238 If we can find such a subsequence we need to turn the first
3239 element into a trie and then add the subsequent branch exact
3240 strings to the trie.
3244 1. patterns where the whole set of branches can be converted.
3246 2. patterns where only a subset can be converted.
3248 In case 1 we can replace the whole set with a single regop
3249 for the trie. In case 2 we need to keep the start and end
3252 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3253 becomes BRANCH TRIE; BRANCH X;
3255 There is an additional case, that being where there is a
3256 common prefix, which gets split out into an EXACT like node
3257 preceding the TRIE node.
3259 If x(1..n)==tail then we can do a simple trie, if not we make
3260 a "jump" trie, such that when we match the appropriate word
3261 we "jump" to the appropriate tail node. Essentially we turn
3262 a nested if into a case structure of sorts.
3267 if (!re_trie_maxbuff) {
3268 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3269 if (!SvIOK(re_trie_maxbuff))
3270 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3272 if ( SvIV(re_trie_maxbuff)>=0 ) {
3274 regnode *first = (regnode *)NULL;
3275 regnode *last = (regnode *)NULL;
3276 regnode *tail = scan;
3281 SV * const mysv = sv_newmortal(); /* for dumping */
3283 /* var tail is used because there may be a TAIL
3284 regop in the way. Ie, the exacts will point to the
3285 thing following the TAIL, but the last branch will
3286 point at the TAIL. So we advance tail. If we
3287 have nested (?:) we may have to move through several
3291 while ( OP( tail ) == TAIL ) {
3292 /* this is the TAIL generated by (?:) */
3293 tail = regnext( tail );
3297 DEBUG_TRIE_COMPILE_r({
3298 regprop(RExC_rx, mysv, tail );
3299 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3300 (int)depth * 2 + 2, "",
3301 "Looking for TRIE'able sequences. Tail node is: ",
3302 SvPV_nolen_const( mysv )
3308 Step through the branches
3309 cur represents each branch,
3310 noper is the first thing to be matched as part of that branch
3311 noper_next is the regnext() of that node.
3313 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3314 via a "jump trie" but we also support building with NOJUMPTRIE,
3315 which restricts the trie logic to structures like /FOO|BAR/.
3317 If noper is a trieable nodetype then the branch is a possible optimization
3318 target. If we are building under NOJUMPTRIE then we require that noper_next
3319 is the same as scan (our current position in the regex program).
3321 Once we have two or more consecutive such branches we can create a
3322 trie of the EXACT's contents and stitch it in place into the program.
3324 If the sequence represents all of the branches in the alternation we
3325 replace the entire thing with a single TRIE node.
3327 Otherwise when it is a subsequence we need to stitch it in place and
3328 replace only the relevant branches. This means the first branch has
3329 to remain as it is used by the alternation logic, and its next pointer,
3330 and needs to be repointed at the item on the branch chain following
3331 the last branch we have optimized away.
3333 This could be either a BRANCH, in which case the subsequence is internal,
3334 or it could be the item following the branch sequence in which case the
3335 subsequence is at the end (which does not necessarily mean the first node
3336 is the start of the alternation).
3338 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3341 ----------------+-----------
3345 EXACTFU_SS | EXACTFU
3346 EXACTFU_TRICKYFOLD | EXACTFU
3351 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3352 ( EXACT == (X) ) ? EXACT : \
3353 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3356 /* dont use tail as the end marker for this traverse */
3357 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3358 regnode * const noper = NEXTOPER( cur );
3359 U8 noper_type = OP( noper );
3360 U8 noper_trietype = TRIE_TYPE( noper_type );
3361 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3362 regnode * const noper_next = regnext( noper );
3363 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3364 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3367 DEBUG_TRIE_COMPILE_r({
3368 regprop(RExC_rx, mysv, cur);
3369 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3370 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3372 regprop(RExC_rx, mysv, noper);
3373 PerlIO_printf( Perl_debug_log, " -> %s",
3374 SvPV_nolen_const(mysv));
3377 regprop(RExC_rx, mysv, noper_next );
3378 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3379 SvPV_nolen_const(mysv));
3381 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3382 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3383 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3387 /* Is noper a trieable nodetype that can be merged with the
3388 * current trie (if there is one)? */
3392 ( noper_trietype == NOTHING)
3393 || ( trietype == NOTHING )
3394 || ( trietype == noper_trietype )
3397 && noper_next == tail
3401 /* Handle mergable triable node
3402 * Either we are the first node in a new trieable sequence,
3403 * in which case we do some bookkeeping, otherwise we update
3404 * the end pointer. */
3407 if ( noper_trietype == NOTHING ) {
3408 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3409 regnode * const noper_next = regnext( noper );
3410 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3411 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3414 if ( noper_next_trietype ) {
3415 trietype = noper_next_trietype;
3416 } else if (noper_next_type) {
3417 /* a NOTHING regop is 1 regop wide. We need at least two
3418 * for a trie so we can't merge this in */
3422 trietype = noper_trietype;
3425 if ( trietype == NOTHING )
3426 trietype = noper_trietype;
3431 } /* end handle mergable triable node */
3433 /* handle unmergable node -
3434 * noper may either be a triable node which can not be tried
3435 * together with the current trie, or a non triable node */
3437 /* If last is set and trietype is not NOTHING then we have found
3438 * at least two triable branch sequences in a row of a similar
3439 * trietype so we can turn them into a trie. If/when we
3440 * allow NOTHING to start a trie sequence this condition will be
3441 * required, and it isn't expensive so we leave it in for now. */
3442 if ( trietype && trietype != NOTHING )
3443 make_trie( pRExC_state,
3444 startbranch, first, cur, tail, count,
3445 trietype, depth+1 );
3446 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3450 && noper_next == tail
3453 /* noper is triable, so we can start a new trie sequence */
3456 trietype = noper_trietype;
3458 /* if we already saw a first but the current node is not triable then we have
3459 * to reset the first information. */
3464 } /* end handle unmergable node */
3465 } /* loop over branches */
3466 DEBUG_TRIE_COMPILE_r({
3467 regprop(RExC_rx, mysv, cur);
3468 PerlIO_printf( Perl_debug_log,
3469 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3470 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3473 if ( last && trietype ) {
3474 if ( trietype != NOTHING ) {
3475 /* the last branch of the sequence was part of a trie,
3476 * so we have to construct it here outside of the loop
3478 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3479 #ifdef TRIE_STUDY_OPT
3480 if ( ((made == MADE_EXACT_TRIE &&
3481 startbranch == first)
3482 || ( first_non_open == first )) &&
3484 flags |= SCF_TRIE_RESTUDY;
3485 if ( startbranch == first
3488 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3493 /* at this point we know whatever we have is a NOTHING sequence/branch
3494 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3496 if ( startbranch == first ) {
3498 /* the entire thing is a NOTHING sequence, something like this:
3499 * (?:|) So we can turn it into a plain NOTHING op. */
3500 DEBUG_TRIE_COMPILE_r({
3501 regprop(RExC_rx, mysv, cur);
3502 PerlIO_printf( Perl_debug_log,
3503 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3504 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3507 OP(startbranch)= NOTHING;
3508 NEXT_OFF(startbranch)= tail - startbranch;
3509 for ( opt= startbranch + 1; opt < tail ; opt++ )
3513 } /* end if ( last) */
3514 } /* TRIE_MAXBUF is non zero */
3519 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3520 scan = NEXTOPER(NEXTOPER(scan));
3521 } else /* single branch is optimized. */
3522 scan = NEXTOPER(scan);
3524 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3525 scan_frame *newframe = NULL;
3530 if (OP(scan) != SUSPEND) {
3531 /* set the pointer */
3532 if (OP(scan) == GOSUB) {
3534 RExC_recurse[ARG2L(scan)] = scan;
3535 start = RExC_open_parens[paren-1];
3536 end = RExC_close_parens[paren-1];
3539 start = RExC_rxi->program + 1;
3543 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3544 SAVEFREEPV(recursed);
3546 if (!PAREN_TEST(recursed,paren+1)) {
3547 PAREN_SET(recursed,paren+1);
3548 Newx(newframe,1,scan_frame);
3550 if (flags & SCF_DO_SUBSTR) {
3551 SCAN_COMMIT(pRExC_state,data,minlenp);
3552 data->longest = &(data->longest_float);
3554 is_inf = is_inf_internal = 1;
3555 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3556 cl_anything(pRExC_state, data->start_class);
3557 flags &= ~SCF_DO_STCLASS;
3560 Newx(newframe,1,scan_frame);
3563 end = regnext(scan);
3568 SAVEFREEPV(newframe);
3569 newframe->next = regnext(scan);
3570 newframe->last = last;
3571 newframe->stop = stopparen;
3572 newframe->prev = frame;
3582 else if (OP(scan) == EXACT) {
3583 I32 l = STR_LEN(scan);
3586 const U8 * const s = (U8*)STRING(scan);
3587 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3588 l = utf8_length(s, s + l);
3590 uc = *((U8*)STRING(scan));
3593 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3594 /* The code below prefers earlier match for fixed
3595 offset, later match for variable offset. */
3596 if (data->last_end == -1) { /* Update the start info. */
3597 data->last_start_min = data->pos_min;
3598 data->last_start_max = is_inf
3599 ? I32_MAX : data->pos_min + data->pos_delta;
3601 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3603 SvUTF8_on(data->last_found);
3605 SV * const sv = data->last_found;
3606 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3607 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3608 if (mg && mg->mg_len >= 0)
3609 mg->mg_len += utf8_length((U8*)STRING(scan),
3610 (U8*)STRING(scan)+STR_LEN(scan));
3612 data->last_end = data->pos_min + l;
3613 data->pos_min += l; /* As in the first entry. */
3614 data->flags &= ~SF_BEFORE_EOL;
3616 if (flags & SCF_DO_STCLASS_AND) {
3617 /* Check whether it is compatible with what we know already! */
3621 /* If compatible, we or it in below. It is compatible if is
3622 * in the bitmp and either 1) its bit or its fold is set, or 2)
3623 * it's for a locale. Even if there isn't unicode semantics
3624 * here, at runtime there may be because of matching against a
3625 * utf8 string, so accept a possible false positive for
3626 * latin1-range folds */
3628 (!(data->start_class->flags & ANYOF_LOCALE)
3629 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3630 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3631 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3636 ANYOF_CLASS_ZERO(data->start_class);
3637 ANYOF_BITMAP_ZERO(data->start_class);
3639 ANYOF_BITMAP_SET(data->start_class, uc);
3640 else if (uc >= 0x100) {
3643 /* Some Unicode code points fold to the Latin1 range; as
3644 * XXX temporary code, instead of figuring out if this is
3645 * one, just assume it is and set all the start class bits
3646 * that could be some such above 255 code point's fold
3647 * which will generate fals positives. As the code
3648 * elsewhere that does compute the fold settles down, it
3649 * can be extracted out and re-used here */
3650 for (i = 0; i < 256; i++){
3651 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3652 ANYOF_BITMAP_SET(data->start_class, i);
3656 CLEAR_SSC_EOS(data->start_class);
3658 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3660 else if (flags & SCF_DO_STCLASS_OR) {
3661 /* false positive possible if the class is case-folded */
3663 ANYOF_BITMAP_SET(data->start_class, uc);
3665 data->start_class->flags |= ANYOF_UNICODE_ALL;
3666 CLEAR_SSC_EOS(data->start_class);
3667 cl_and(data->start_class, and_withp);
3669 flags &= ~SCF_DO_STCLASS;
3671 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3672 I32 l = STR_LEN(scan);
3673 UV uc = *((U8*)STRING(scan));
3675 /* Search for fixed substrings supports EXACT only. */
3676 if (flags & SCF_DO_SUBSTR) {
3678 SCAN_COMMIT(pRExC_state, data, minlenp);
3681 const U8 * const s = (U8 *)STRING(scan);
3682 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3683 l = utf8_length(s, s + l);
3685 if (has_exactf_sharp_s) {
3686 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3688 min += l - min_subtract;
3690 delta += min_subtract;
3691 if (flags & SCF_DO_SUBSTR) {
3692 data->pos_min += l - min_subtract;
3693 if (data->pos_min < 0) {
3696 data->pos_delta += min_subtract;
3698 data->longest = &(data->longest_float);
3701 if (flags & SCF_DO_STCLASS_AND) {
3702 /* Check whether it is compatible with what we know already! */
3705 (!(data->start_class->flags & ANYOF_LOCALE)
3706 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3707 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3711 ANYOF_CLASS_ZERO(data->start_class);
3712 ANYOF_BITMAP_ZERO(data->start_class);
3714 ANYOF_BITMAP_SET(data->start_class, uc);
3715 CLEAR_SSC_EOS(data->start_class);
3716 if (OP(scan) == EXACTFL) {
3717 /* XXX This set is probably no longer necessary, and
3718 * probably wrong as LOCALE now is on in the initial
3720 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3724 /* Also set the other member of the fold pair. In case
3725 * that unicode semantics is called for at runtime, use
3726 * the full latin1 fold. (Can't do this for locale,
3727 * because not known until runtime) */
3728 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3730 /* All other (EXACTFL handled above) folds except under
3731 * /iaa that include s, S, and sharp_s also may include
3733 if (OP(scan) != EXACTFA) {
3734 if (uc == 's' || uc == 'S') {
3735 ANYOF_BITMAP_SET(data->start_class,
3736 LATIN_SMALL_LETTER_SHARP_S);
3738 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3739 ANYOF_BITMAP_SET(data->start_class, 's');
3740 ANYOF_BITMAP_SET(data->start_class, 'S');
3745 else if (uc >= 0x100) {
3747 for (i = 0; i < 256; i++){
3748 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3749 ANYOF_BITMAP_SET(data->start_class, i);
3754 else if (flags & SCF_DO_STCLASS_OR) {
3755 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3756 /* false positive possible if the class is case-folded.
3757 Assume that the locale settings are the same... */
3759 ANYOF_BITMAP_SET(data->start_class, uc);
3760 if (OP(scan) != EXACTFL) {
3762 /* And set the other member of the fold pair, but
3763 * can't do that in locale because not known until
3765 ANYOF_BITMAP_SET(data->start_class,
3766 PL_fold_latin1[uc]);
3768 /* All folds except under /iaa that include s, S,
3769 * and sharp_s also may include the others */
3770 if (OP(scan) != EXACTFA) {
3771 if (uc == 's' || uc == 'S') {
3772 ANYOF_BITMAP_SET(data->start_class,
3773 LATIN_SMALL_LETTER_SHARP_S);
3775 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3776 ANYOF_BITMAP_SET(data->start_class, 's');
3777 ANYOF_BITMAP_SET(data->start_class, 'S');
3782 CLEAR_SSC_EOS(data->start_class);
3784 cl_and(data->start_class, and_withp);
3786 flags &= ~SCF_DO_STCLASS;
3788 else if (REGNODE_VARIES(OP(scan))) {
3789 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3790 I32 f = flags, pos_before = 0;
3791 regnode * const oscan = scan;
3792 struct regnode_charclass_class this_class;
3793 struct regnode_charclass_class *oclass = NULL;
3794 I32 next_is_eval = 0;
3796 switch (PL_regkind[OP(scan)]) {
3797 case WHILEM: /* End of (?:...)* . */
3798 scan = NEXTOPER(scan);
3801 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3802 next = NEXTOPER(scan);
3803 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3805 maxcount = REG_INFTY;
3806 next = regnext(scan);
3807 scan = NEXTOPER(scan);
3811 if (flags & SCF_DO_SUBSTR)
3816 if (flags & SCF_DO_STCLASS) {
3818 maxcount = REG_INFTY;
3819 next = regnext(scan);
3820 scan = NEXTOPER(scan);
3823 is_inf = is_inf_internal = 1;
3824 scan = regnext(scan);
3825 if (flags & SCF_DO_SUBSTR) {
3826 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3827 data->longest = &(data->longest_float);
3829 goto optimize_curly_tail;
3831 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3832 && (scan->flags == stopparen))
https://perl5.git.perl.org / perl5.git / blob