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
408 #define SCF_TRIE_DOING_RESTUDY 0x10000
410 #define UTF cBOOL(RExC_utf8)
412 /* The enums for all these are ordered so things work out correctly */
413 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
414 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
415 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
416 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
417 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
418 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
419 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
421 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
423 #define OOB_NAMEDCLASS -1
425 /* There is no code point that is out-of-bounds, so this is problematic. But
426 * its only current use is to initialize a variable that is always set before
428 #define OOB_UNICODE 0xDEADBEEF
430 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
431 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
434 /* length of regex to show in messages that don't mark a position within */
435 #define RegexLengthToShowInErrorMessages 127
438 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
439 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
440 * op/pragma/warn/regcomp.
442 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
443 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
445 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
448 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
449 * arg. Show regex, up to a maximum length. If it's too long, chop and add
452 #define _FAIL(code) STMT_START { \
453 const char *ellipses = ""; \
454 IV len = RExC_end - RExC_precomp; \
457 SAVEFREESV(RExC_rx_sv); \
458 if (len > RegexLengthToShowInErrorMessages) { \
459 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
460 len = RegexLengthToShowInErrorMessages - 10; \
466 #define FAIL(msg) _FAIL( \
467 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
468 msg, (int)len, RExC_precomp, ellipses))
470 #define FAIL2(msg,arg) _FAIL( \
471 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
472 arg, (int)len, RExC_precomp, ellipses))
475 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
477 #define Simple_vFAIL(m) STMT_START { \
478 const IV offset = RExC_parse - RExC_precomp; \
479 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
480 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
484 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
486 #define vFAIL(m) STMT_START { \
488 SAVEFREESV(RExC_rx_sv); \
493 * Like Simple_vFAIL(), but accepts two arguments.
495 #define Simple_vFAIL2(m,a1) STMT_START { \
496 const IV offset = RExC_parse - RExC_precomp; \
497 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
498 (int)offset, RExC_precomp, RExC_precomp + offset); \
502 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
504 #define vFAIL2(m,a1) STMT_START { \
506 SAVEFREESV(RExC_rx_sv); \
507 Simple_vFAIL2(m, a1); \
512 * Like Simple_vFAIL(), but accepts three arguments.
514 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
515 const IV offset = RExC_parse - RExC_precomp; \
516 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
517 (int)offset, RExC_precomp, RExC_precomp + offset); \
521 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
523 #define vFAIL3(m,a1,a2) STMT_START { \
525 SAVEFREESV(RExC_rx_sv); \
526 Simple_vFAIL3(m, a1, a2); \
530 * Like Simple_vFAIL(), but accepts four arguments.
532 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
533 const IV offset = RExC_parse - RExC_precomp; \
534 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
535 (int)offset, RExC_precomp, RExC_precomp + offset); \
538 #define vFAIL4(m,a1,a2,a3) STMT_START { \
540 SAVEFREESV(RExC_rx_sv); \
541 Simple_vFAIL4(m, a1, a2, a3); \
544 /* m is not necessarily a "literal string", in this macro */
545 #define reg_warn_non_literal_string(loc, m) STMT_START { \
546 const IV offset = loc - RExC_precomp; \
547 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
548 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
551 #define ckWARNreg(loc,m) STMT_START { \
552 const IV offset = loc - RExC_precomp; \
553 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
554 (int)offset, RExC_precomp, RExC_precomp + offset); \
557 #define vWARN_dep(loc, m) STMT_START { \
558 const IV offset = loc - RExC_precomp; \
559 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
560 (int)offset, RExC_precomp, RExC_precomp + offset); \
563 #define ckWARNdep(loc,m) STMT_START { \
564 const IV offset = loc - RExC_precomp; \
565 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
567 (int)offset, RExC_precomp, RExC_precomp + offset); \
570 #define ckWARNregdep(loc,m) STMT_START { \
571 const IV offset = loc - RExC_precomp; \
572 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
574 (int)offset, RExC_precomp, RExC_precomp + offset); \
577 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
578 const IV offset = loc - RExC_precomp; \
579 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
581 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
584 #define ckWARN2reg(loc, m, a1) STMT_START { \
585 const IV offset = loc - RExC_precomp; \
586 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
587 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
590 #define vWARN3(loc, m, a1, a2) STMT_START { \
591 const IV offset = loc - RExC_precomp; \
592 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
593 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
596 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
597 const IV offset = loc - RExC_precomp; \
598 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
599 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
602 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
603 const IV offset = loc - RExC_precomp; \
604 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
605 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
608 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
609 const IV offset = loc - RExC_precomp; \
610 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
611 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
614 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
615 const IV offset = loc - RExC_precomp; \
616 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
617 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
621 /* Allow for side effects in s */
622 #define REGC(c,s) STMT_START { \
623 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
626 /* Macros for recording node offsets. 20001227 mjd@plover.com
627 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
628 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
629 * Element 0 holds the number n.
630 * Position is 1 indexed.
632 #ifndef RE_TRACK_PATTERN_OFFSETS
633 #define Set_Node_Offset_To_R(node,byte)
634 #define Set_Node_Offset(node,byte)
635 #define Set_Cur_Node_Offset
636 #define Set_Node_Length_To_R(node,len)
637 #define Set_Node_Length(node,len)
638 #define Set_Node_Cur_Length(node,start)
639 #define Node_Offset(n)
640 #define Node_Length(n)
641 #define Set_Node_Offset_Length(node,offset,len)
642 #define ProgLen(ri) ri->u.proglen
643 #define SetProgLen(ri,x) ri->u.proglen = x
645 #define ProgLen(ri) ri->u.offsets[0]
646 #define SetProgLen(ri,x) ri->u.offsets[0] = x
647 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
649 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
650 __LINE__, (int)(node), (int)(byte))); \
652 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
654 RExC_offsets[2*(node)-1] = (byte); \
659 #define Set_Node_Offset(node,byte) \
660 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
661 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
663 #define Set_Node_Length_To_R(node,len) STMT_START { \
665 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
666 __LINE__, (int)(node), (int)(len))); \
668 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
670 RExC_offsets[2*(node)] = (len); \
675 #define Set_Node_Length(node,len) \
676 Set_Node_Length_To_R((node)-RExC_emit_start, len)
677 #define Set_Node_Cur_Length(node, start) \
678 Set_Node_Length(node, RExC_parse - start)
680 /* Get offsets and lengths */
681 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
682 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
684 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
685 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
686 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
690 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
691 #define EXPERIMENTAL_INPLACESCAN
692 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
694 #define DEBUG_STUDYDATA(str,data,depth) \
695 DEBUG_OPTIMISE_MORE_r(if(data){ \
696 PerlIO_printf(Perl_debug_log, \
697 "%*s" str "Pos:%"IVdf"/%"IVdf \
698 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
699 (int)(depth)*2, "", \
700 (IV)((data)->pos_min), \
701 (IV)((data)->pos_delta), \
702 (UV)((data)->flags), \
703 (IV)((data)->whilem_c), \
704 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
705 is_inf ? "INF " : "" \
707 if ((data)->last_found) \
708 PerlIO_printf(Perl_debug_log, \
709 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
710 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
711 SvPVX_const((data)->last_found), \
712 (IV)((data)->last_end), \
713 (IV)((data)->last_start_min), \
714 (IV)((data)->last_start_max), \
715 ((data)->longest && \
716 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
717 SvPVX_const((data)->longest_fixed), \
718 (IV)((data)->offset_fixed), \
719 ((data)->longest && \
720 (data)->longest==&((data)->longest_float)) ? "*" : "", \
721 SvPVX_const((data)->longest_float), \
722 (IV)((data)->offset_float_min), \
723 (IV)((data)->offset_float_max) \
725 PerlIO_printf(Perl_debug_log,"\n"); \
728 /* Mark that we cannot extend a found fixed substring at this point.
729 Update the longest found anchored substring and the longest found
730 floating substrings if needed. */
733 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
735 const STRLEN l = CHR_SVLEN(data->last_found);
736 const STRLEN old_l = CHR_SVLEN(*data->longest);
737 GET_RE_DEBUG_FLAGS_DECL;
739 PERL_ARGS_ASSERT_SCAN_COMMIT;
741 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
742 SvSetMagicSV(*data->longest, data->last_found);
743 if (*data->longest == data->longest_fixed) {
744 data->offset_fixed = l ? data->last_start_min : data->pos_min;
745 if (data->flags & SF_BEFORE_EOL)
747 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
749 data->flags &= ~SF_FIX_BEFORE_EOL;
750 data->minlen_fixed=minlenp;
751 data->lookbehind_fixed=0;
753 else { /* *data->longest == data->longest_float */
754 data->offset_float_min = l ? data->last_start_min : data->pos_min;
755 data->offset_float_max = (l
756 ? data->last_start_max
757 : (data->pos_delta == I32_MAX ? I32_MAX : data->pos_min + data->pos_delta));
758 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
759 data->offset_float_max = I32_MAX;
760 if (data->flags & SF_BEFORE_EOL)
762 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
764 data->flags &= ~SF_FL_BEFORE_EOL;
765 data->minlen_float=minlenp;
766 data->lookbehind_float=0;
769 SvCUR_set(data->last_found, 0);
771 SV * const sv = data->last_found;
772 if (SvUTF8(sv) && SvMAGICAL(sv)) {
773 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
779 data->flags &= ~SF_BEFORE_EOL;
780 DEBUG_STUDYDATA("commit: ",data,0);
783 /* These macros set, clear and test whether the synthetic start class ('ssc',
784 * given by the parameter) matches an empty string (EOS). This uses the
785 * 'next_off' field in the node, to save a bit in the flags field. The ssc
786 * stands alone, so there is never a next_off, so this field is otherwise
787 * unused. The EOS information is used only for compilation, but theoretically
788 * it could be passed on to the execution code. This could be used to store
789 * more than one bit of information, but only this one is currently used. */
790 #define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
791 #define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
792 #define TEST_SSC_EOS(node) cBOOL((node)->next_off)
794 /* Can match anything (initialization) */
796 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
798 PERL_ARGS_ASSERT_CL_ANYTHING;
800 ANYOF_BITMAP_SETALL(cl);
801 cl->flags = ANYOF_UNICODE_ALL;
804 /* If any portion of the regex is to operate under locale rules,
805 * initialization includes it. The reason this isn't done for all regexes
806 * is that the optimizer was written under the assumption that locale was
807 * all-or-nothing. Given the complexity and lack of documentation in the
808 * optimizer, and that there are inadequate test cases for locale, so many
809 * parts of it may not work properly, it is safest to avoid locale unless
811 if (RExC_contains_locale) {
812 ANYOF_CLASS_SETALL(cl); /* /l uses class */
813 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
816 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
820 /* Can match anything (initialization) */
822 S_cl_is_anything(const struct regnode_charclass_class *cl)
826 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
828 for (value = 0; value < ANYOF_MAX; value += 2)
829 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
831 if (!(cl->flags & ANYOF_UNICODE_ALL))
833 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
838 /* Can match anything (initialization) */
840 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
842 PERL_ARGS_ASSERT_CL_INIT;
844 Zero(cl, 1, struct regnode_charclass_class);
846 cl_anything(pRExC_state, cl);
847 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
850 /* These two functions currently do the exact same thing */
851 #define cl_init_zero cl_init
853 /* 'AND' a given class with another one. Can create false positives. 'cl'
854 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
855 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
857 S_cl_and(struct regnode_charclass_class *cl,
858 const struct regnode_charclass_class *and_with)
860 PERL_ARGS_ASSERT_CL_AND;
862 assert(PL_regkind[and_with->type] == ANYOF);
864 /* I (khw) am not sure all these restrictions are necessary XXX */
865 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
866 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
867 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
868 && !(and_with->flags & ANYOF_LOC_FOLD)
869 && !(cl->flags & ANYOF_LOC_FOLD)) {
872 if (and_with->flags & ANYOF_INVERT)
873 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
874 cl->bitmap[i] &= ~and_with->bitmap[i];
876 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
877 cl->bitmap[i] &= and_with->bitmap[i];
878 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
880 if (and_with->flags & ANYOF_INVERT) {
882 /* Here, the and'ed node is inverted. Get the AND of the flags that
883 * aren't affected by the inversion. Those that are affected are
884 * handled individually below */
885 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
886 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
887 cl->flags |= affected_flags;
889 /* We currently don't know how to deal with things that aren't in the
890 * bitmap, but we know that the intersection is no greater than what
891 * is already in cl, so let there be false positives that get sorted
892 * out after the synthetic start class succeeds, and the node is
893 * matched for real. */
895 /* The inversion of these two flags indicate that the resulting
896 * intersection doesn't have them */
897 if (and_with->flags & ANYOF_UNICODE_ALL) {
898 cl->flags &= ~ANYOF_UNICODE_ALL;
900 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
901 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
904 else { /* and'd node is not inverted */
905 U8 outside_bitmap_but_not_utf8; /* Temp variable */
907 if (! ANYOF_NONBITMAP(and_with)) {
909 /* Here 'and_with' doesn't match anything outside the bitmap
910 * (except possibly ANYOF_UNICODE_ALL), which means the
911 * intersection can't either, except for ANYOF_UNICODE_ALL, in
912 * which case we don't know what the intersection is, but it's no
913 * greater than what cl already has, so can just leave it alone,
914 * with possible false positives */
915 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
916 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
917 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
920 else if (! ANYOF_NONBITMAP(cl)) {
922 /* Here, 'and_with' does match something outside the bitmap, and cl
923 * doesn't have a list of things to match outside the bitmap. If
924 * cl can match all code points above 255, the intersection will
925 * be those above-255 code points that 'and_with' matches. If cl
926 * can't match all Unicode code points, it means that it can't
927 * match anything outside the bitmap (since the 'if' that got us
928 * into this block tested for that), so we leave the bitmap empty.
930 if (cl->flags & ANYOF_UNICODE_ALL) {
931 ARG_SET(cl, ARG(and_with));
933 /* and_with's ARG may match things that don't require UTF8.
934 * And now cl's will too, in spite of this being an 'and'. See
935 * the comments below about the kludge */
936 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
940 /* Here, both 'and_with' and cl match something outside the
941 * bitmap. Currently we do not do the intersection, so just match
942 * whatever cl had at the beginning. */
946 /* Take the intersection of the two sets of flags. However, the
947 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
948 * kludge around the fact that this flag is not treated like the others
949 * which are initialized in cl_anything(). The way the optimizer works
950 * is that the synthetic start class (SSC) is initialized to match
951 * anything, and then the first time a real node is encountered, its
952 * values are AND'd with the SSC's with the result being the values of
953 * the real node. However, there are paths through the optimizer where
954 * the AND never gets called, so those initialized bits are set
955 * inappropriately, which is not usually a big deal, as they just cause
956 * false positives in the SSC, which will just mean a probably
957 * imperceptible slow down in execution. However this bit has a
958 * higher false positive consequence in that it can cause utf8.pm,
959 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
960 * bigger slowdown and also causes significant extra memory to be used.
961 * In order to prevent this, the code now takes a different tack. The
962 * bit isn't set unless some part of the regular expression needs it,
963 * but once set it won't get cleared. This means that these extra
964 * modules won't get loaded unless there was some path through the
965 * pattern that would have required them anyway, and so any false
966 * positives that occur by not ANDing them out when they could be
967 * aren't as severe as they would be if we treated this bit like all
969 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
970 & ANYOF_NONBITMAP_NON_UTF8;
971 cl->flags &= and_with->flags;
972 cl->flags |= outside_bitmap_but_not_utf8;
976 /* 'OR' a given class with another one. Can create false positives. 'cl'
977 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
978 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
980 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
982 PERL_ARGS_ASSERT_CL_OR;
984 if (or_with->flags & ANYOF_INVERT) {
986 /* Here, the or'd node is to be inverted. This means we take the
987 * complement of everything not in the bitmap, but currently we don't
988 * know what that is, so give up and match anything */
989 if (ANYOF_NONBITMAP(or_with)) {
990 cl_anything(pRExC_state, cl);
993 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
994 * <= (B1 | !B2) | (CL1 | !CL2)
995 * which is wasteful if CL2 is small, but we ignore CL2:
996 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
997 * XXXX Can we handle case-fold? Unclear:
998 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
999 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
1001 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1002 && !(or_with->flags & ANYOF_LOC_FOLD)
1003 && !(cl->flags & ANYOF_LOC_FOLD) ) {
1006 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1007 cl->bitmap[i] |= ~or_with->bitmap[i];
1008 } /* XXXX: logic is complicated otherwise */
1010 cl_anything(pRExC_state, cl);
1013 /* And, we can just take the union of the flags that aren't affected
1014 * by the inversion */
1015 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1017 /* For the remaining flags:
1018 ANYOF_UNICODE_ALL and inverted means to not match anything above
1019 255, which means that the union with cl should just be
1020 what cl has in it, so can ignore this flag
1021 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
1022 is 127-255 to match them, but then invert that, so the
1023 union with cl should just be what cl has in it, so can
1026 } else { /* 'or_with' is not inverted */
1027 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1028 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1029 && (!(or_with->flags & ANYOF_LOC_FOLD)
1030 || (cl->flags & ANYOF_LOC_FOLD)) ) {
1033 /* OR char bitmap and class bitmap separately */
1034 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1035 cl->bitmap[i] |= or_with->bitmap[i];
1036 if (or_with->flags & ANYOF_CLASS) {
1037 ANYOF_CLASS_OR(or_with, cl);
1040 else { /* XXXX: logic is complicated, leave it along for a moment. */
1041 cl_anything(pRExC_state, cl);
1044 if (ANYOF_NONBITMAP(or_with)) {
1046 /* Use the added node's outside-the-bit-map match if there isn't a
1047 * conflict. If there is a conflict (both nodes match something
1048 * outside the bitmap, but what they match outside is not the same
1049 * pointer, and hence not easily compared until XXX we extend
1050 * inversion lists this far), give up and allow the start class to
1051 * match everything outside the bitmap. If that stuff is all above
1052 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1053 if (! ANYOF_NONBITMAP(cl)) {
1054 ARG_SET(cl, ARG(or_with));
1056 else if (ARG(cl) != ARG(or_with)) {
1058 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1059 cl_anything(pRExC_state, cl);
1062 cl->flags |= ANYOF_UNICODE_ALL;
1067 /* Take the union */
1068 cl->flags |= or_with->flags;
1072 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1073 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1074 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1075 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1080 dump_trie(trie,widecharmap,revcharmap)
1081 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1082 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1084 These routines dump out a trie in a somewhat readable format.
1085 The _interim_ variants are used for debugging the interim
1086 tables that are used to generate the final compressed
1087 representation which is what dump_trie expects.
1089 Part of the reason for their existence is to provide a form
1090 of documentation as to how the different representations function.
1095 Dumps the final compressed table form of the trie to Perl_debug_log.
1096 Used for debugging make_trie().
1100 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1101 AV *revcharmap, U32 depth)
1104 SV *sv=sv_newmortal();
1105 int colwidth= widecharmap ? 6 : 4;
1107 GET_RE_DEBUG_FLAGS_DECL;
1109 PERL_ARGS_ASSERT_DUMP_TRIE;
1111 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1112 (int)depth * 2 + 2,"",
1113 "Match","Base","Ofs" );
1115 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1116 SV ** const tmp = av_fetch( revcharmap, state, 0);
1118 PerlIO_printf( Perl_debug_log, "%*s",
1120 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1121 PL_colors[0], PL_colors[1],
1122 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1123 PERL_PV_ESCAPE_FIRSTCHAR
1128 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1129 (int)depth * 2 + 2,"");
1131 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1132 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1133 PerlIO_printf( Perl_debug_log, "\n");
1135 for( state = 1 ; state < trie->statecount ; state++ ) {
1136 const U32 base = trie->states[ state ].trans.base;
1138 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1140 if ( trie->states[ state ].wordnum ) {
1141 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1143 PerlIO_printf( Perl_debug_log, "%6s", "" );
1146 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1151 while( ( base + ofs < trie->uniquecharcount ) ||
1152 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1153 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1156 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1158 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1159 if ( ( base + ofs >= trie->uniquecharcount ) &&
1160 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1161 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1163 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1165 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1167 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1171 PerlIO_printf( Perl_debug_log, "]");
1174 PerlIO_printf( Perl_debug_log, "\n" );
1176 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1177 for (word=1; word <= trie->wordcount; word++) {
1178 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1179 (int)word, (int)(trie->wordinfo[word].prev),
1180 (int)(trie->wordinfo[word].len));
1182 PerlIO_printf(Perl_debug_log, "\n" );
1185 Dumps a fully constructed but uncompressed trie in list form.
1186 List tries normally only are used for construction when the number of
1187 possible chars (trie->uniquecharcount) is very high.
1188 Used for debugging make_trie().
1191 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1192 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1196 SV *sv=sv_newmortal();
1197 int colwidth= widecharmap ? 6 : 4;
1198 GET_RE_DEBUG_FLAGS_DECL;
1200 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1202 /* print out the table precompression. */
1203 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1204 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1205 "------:-----+-----------------\n" );
1207 for( state=1 ; state < next_alloc ; state ++ ) {
1210 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1211 (int)depth * 2 + 2,"", (UV)state );
1212 if ( ! trie->states[ state ].wordnum ) {
1213 PerlIO_printf( Perl_debug_log, "%5s| ","");
1215 PerlIO_printf( Perl_debug_log, "W%4x| ",
1216 trie->states[ state ].wordnum
1219 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1220 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1222 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1224 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1225 PL_colors[0], PL_colors[1],
1226 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1227 PERL_PV_ESCAPE_FIRSTCHAR
1229 TRIE_LIST_ITEM(state,charid).forid,
1230 (UV)TRIE_LIST_ITEM(state,charid).newstate
1233 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1234 (int)((depth * 2) + 14), "");
1237 PerlIO_printf( Perl_debug_log, "\n");
1242 Dumps a fully constructed but uncompressed trie in table form.
1243 This is the normal DFA style state transition table, with a few
1244 twists to facilitate compression later.
1245 Used for debugging make_trie().
1248 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1249 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1254 SV *sv=sv_newmortal();
1255 int colwidth= widecharmap ? 6 : 4;
1256 GET_RE_DEBUG_FLAGS_DECL;
1258 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1261 print out the table precompression so that we can do a visual check
1262 that they are identical.
1265 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1267 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1268 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1270 PerlIO_printf( Perl_debug_log, "%*s",
1272 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1273 PL_colors[0], PL_colors[1],
1274 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1275 PERL_PV_ESCAPE_FIRSTCHAR
1281 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1283 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1284 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1287 PerlIO_printf( Perl_debug_log, "\n" );
1289 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1291 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1292 (int)depth * 2 + 2,"",
1293 (UV)TRIE_NODENUM( state ) );
1295 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1296 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1298 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1300 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1302 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1303 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1305 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1306 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1314 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1315 startbranch: the first branch in the whole branch sequence
1316 first : start branch of sequence of branch-exact nodes.
1317 May be the same as startbranch
1318 last : Thing following the last branch.
1319 May be the same as tail.
1320 tail : item following the branch sequence
1321 count : words in the sequence
1322 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1323 depth : indent depth
1325 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1327 A trie is an N'ary tree where the branches are determined by digital
1328 decomposition of the key. IE, at the root node you look up the 1st character and
1329 follow that branch repeat until you find the end of the branches. Nodes can be
1330 marked as "accepting" meaning they represent a complete word. Eg:
1334 would convert into the following structure. Numbers represent states, letters
1335 following numbers represent valid transitions on the letter from that state, if
1336 the number is in square brackets it represents an accepting state, otherwise it
1337 will be in parenthesis.
1339 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1343 (1) +-i->(6)-+-s->[7]
1345 +-s->(3)-+-h->(4)-+-e->[5]
1347 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1349 This shows that when matching against the string 'hers' we will begin at state 1
1350 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1351 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1352 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1353 single traverse. We store a mapping from accepting to state to which word was
1354 matched, and then when we have multiple possibilities we try to complete the
1355 rest of the regex in the order in which they occured in the alternation.
1357 The only prior NFA like behaviour that would be changed by the TRIE support is
1358 the silent ignoring of duplicate alternations which are of the form:
1360 / (DUPE|DUPE) X? (?{ ... }) Y /x
1362 Thus EVAL blocks following a trie may be called a different number of times with
1363 and without the optimisation. With the optimisations dupes will be silently
1364 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1365 the following demonstrates:
1367 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1369 which prints out 'word' three times, but
1371 'words'=~/(word|word|word)(?{ print $1 })S/
1373 which doesnt print it out at all. This is due to other optimisations kicking in.
1375 Example of what happens on a structural level:
1377 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1379 1: CURLYM[1] {1,32767}(18)
1390 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1391 and should turn into:
1393 1: CURLYM[1] {1,32767}(18)
1395 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1403 Cases where tail != last would be like /(?foo|bar)baz/:
1413 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1414 and would end up looking like:
1417 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1424 d = uvuni_to_utf8_flags(d, uv, 0);
1426 is the recommended Unicode-aware way of saying
1431 #define TRIE_STORE_REVCHAR(val) \
1434 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1435 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1436 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1437 SvCUR_set(zlopp, kapow - flrbbbbb); \
1440 av_push(revcharmap, zlopp); \
1442 char ooooff = (char)val; \
1443 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1447 #define TRIE_READ_CHAR STMT_START { \
1450 /* if it is UTF then it is either already folded, or does not need folding */ \
1451 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1453 else if (folder == PL_fold_latin1) { \
1454 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1455 if ( foldlen > 0 ) { \
1456 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1462 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, FOLD_FLAGS_FULL); \
1463 skiplen = UNISKIP(uvc); \
1464 foldlen -= skiplen; \
1465 scan = foldbuf + skiplen; \
1468 /* raw data, will be folded later if needed */ \
1476 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1477 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1478 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1479 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1481 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1482 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1483 TRIE_LIST_CUR( state )++; \
1486 #define TRIE_LIST_NEW(state) STMT_START { \
1487 Newxz( trie->states[ state ].trans.list, \
1488 4, reg_trie_trans_le ); \
1489 TRIE_LIST_CUR( state ) = 1; \
1490 TRIE_LIST_LEN( state ) = 4; \
1493 #define TRIE_HANDLE_WORD(state) STMT_START { \
1494 U16 dupe= trie->states[ state ].wordnum; \
1495 regnode * const noper_next = regnext( noper ); \
1498 /* store the word for dumping */ \
1500 if (OP(noper) != NOTHING) \
1501 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1503 tmp = newSVpvn_utf8( "", 0, UTF ); \
1504 av_push( trie_words, tmp ); \
1508 trie->wordinfo[curword].prev = 0; \
1509 trie->wordinfo[curword].len = wordlen; \
1510 trie->wordinfo[curword].accept = state; \
1512 if ( noper_next < tail ) { \
1514 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1515 trie->jump[curword] = (U16)(noper_next - convert); \
1517 jumper = noper_next; \
1519 nextbranch= regnext(cur); \
1523 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1524 /* chain, so that when the bits of chain are later */\
1525 /* linked together, the dups appear in the chain */\
1526 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1527 trie->wordinfo[dupe].prev = curword; \
1529 /* we haven't inserted this word yet. */ \
1530 trie->states[ state ].wordnum = curword; \
1535 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1536 ( ( base + charid >= ucharcount \
1537 && base + charid < ubound \
1538 && state == trie->trans[ base - ucharcount + charid ].check \
1539 && trie->trans[ base - ucharcount + charid ].next ) \
1540 ? trie->trans[ base - ucharcount + charid ].next \
1541 : ( state==1 ? special : 0 ) \
1545 #define MADE_JUMP_TRIE 2
1546 #define MADE_EXACT_TRIE 4
1549 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1552 /* first pass, loop through and scan words */
1553 reg_trie_data *trie;
1554 HV *widecharmap = NULL;
1555 AV *revcharmap = newAV();
1557 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1562 regnode *jumper = NULL;
1563 regnode *nextbranch = NULL;
1564 regnode *convert = NULL;
1565 U32 *prev_states; /* temp array mapping each state to previous one */
1566 /* we just use folder as a flag in utf8 */
1567 const U8 * folder = NULL;
1570 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1571 AV *trie_words = NULL;
1572 /* along with revcharmap, this only used during construction but both are
1573 * useful during debugging so we store them in the struct when debugging.
1576 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1577 STRLEN trie_charcount=0;
1579 SV *re_trie_maxbuff;
1580 GET_RE_DEBUG_FLAGS_DECL;
1582 PERL_ARGS_ASSERT_MAKE_TRIE;
1584 PERL_UNUSED_ARG(depth);
1591 case EXACTFU_TRICKYFOLD:
1592 case EXACTFU: folder = PL_fold_latin1; break;
1593 case EXACTF: folder = PL_fold; break;
1594 case EXACTFL: folder = PL_fold_locale; break;
1595 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1598 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1600 trie->startstate = 1;
1601 trie->wordcount = word_count;
1602 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1603 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1605 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1606 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1607 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1610 trie_words = newAV();
1613 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1614 if (!SvIOK(re_trie_maxbuff)) {
1615 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1617 DEBUG_TRIE_COMPILE_r({
1618 PerlIO_printf( Perl_debug_log,
1619 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1620 (int)depth * 2 + 2, "",
1621 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1622 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1626 /* Find the node we are going to overwrite */
1627 if ( first == startbranch && OP( last ) != BRANCH ) {
1628 /* whole branch chain */
1631 /* branch sub-chain */
1632 convert = NEXTOPER( first );
1635 /* -- First loop and Setup --
1637 We first traverse the branches and scan each word to determine if it
1638 contains widechars, and how many unique chars there are, this is
1639 important as we have to build a table with at least as many columns as we
1642 We use an array of integers to represent the character codes 0..255
1643 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1644 native representation of the character value as the key and IV's for the
1647 *TODO* If we keep track of how many times each character is used we can
1648 remap the columns so that the table compression later on is more
1649 efficient in terms of memory by ensuring the most common value is in the
1650 middle and the least common are on the outside. IMO this would be better
1651 than a most to least common mapping as theres a decent chance the most
1652 common letter will share a node with the least common, meaning the node
1653 will not be compressible. With a middle is most common approach the worst
1654 case is when we have the least common nodes twice.
1658 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1659 regnode *noper = NEXTOPER( cur );
1660 const U8 *uc = (U8*)STRING( noper );
1661 const U8 *e = uc + STR_LEN( noper );
1663 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1665 const U8 *scan = (U8*)NULL;
1666 U32 wordlen = 0; /* required init */
1668 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1670 if (OP(noper) == NOTHING) {
1671 regnode *noper_next= regnext(noper);
1672 if (noper_next != tail && OP(noper_next) == flags) {
1674 uc= (U8*)STRING(noper);
1675 e= uc + STR_LEN(noper);
1676 trie->minlen= STR_LEN(noper);
1683 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1684 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1685 regardless of encoding */
1686 if (OP( noper ) == EXACTFU_SS) {
1687 /* false positives are ok, so just set this */
1688 TRIE_BITMAP_SET(trie,0xDF);
1691 for ( ; uc < e ; uc += len ) {
1692 TRIE_CHARCOUNT(trie)++;
1697 U8 folded= folder[ (U8) uvc ];
1698 if ( !trie->charmap[ folded ] ) {
1699 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1700 TRIE_STORE_REVCHAR( folded );
1703 if ( !trie->charmap[ uvc ] ) {
1704 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1705 TRIE_STORE_REVCHAR( uvc );
1708 /* store the codepoint in the bitmap, and its folded
1710 TRIE_BITMAP_SET(trie, uvc);
1712 /* store the folded codepoint */
1713 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1716 /* store first byte of utf8 representation of
1717 variant codepoints */
1718 if (! UNI_IS_INVARIANT(uvc)) {
1719 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1722 set_bit = 0; /* We've done our bit :-) */
1727 widecharmap = newHV();
1729 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1732 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1734 if ( !SvTRUE( *svpp ) ) {
1735 sv_setiv( *svpp, ++trie->uniquecharcount );
1736 TRIE_STORE_REVCHAR(uvc);
1740 if( cur == first ) {
1741 trie->minlen = chars;
1742 trie->maxlen = chars;
1743 } else if (chars < trie->minlen) {
1744 trie->minlen = chars;
1745 } else if (chars > trie->maxlen) {
1746 trie->maxlen = chars;
1748 if (OP( noper ) == EXACTFU_SS) {
1749 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1750 if (trie->minlen > 1)
1753 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1754 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1755 * - We assume that any such sequence might match a 2 byte string */
1756 if (trie->minlen > 2 )
1760 } /* end first pass */
1761 DEBUG_TRIE_COMPILE_r(
1762 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1763 (int)depth * 2 + 2,"",
1764 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1765 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1766 (int)trie->minlen, (int)trie->maxlen )
1770 We now know what we are dealing with in terms of unique chars and
1771 string sizes so we can calculate how much memory a naive
1772 representation using a flat table will take. If it's over a reasonable
1773 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1774 conservative but potentially much slower representation using an array
1777 At the end we convert both representations into the same compressed
1778 form that will be used in regexec.c for matching with. The latter
1779 is a form that cannot be used to construct with but has memory
1780 properties similar to the list form and access properties similar
1781 to the table form making it both suitable for fast searches and
1782 small enough that its feasable to store for the duration of a program.
1784 See the comment in the code where the compressed table is produced
1785 inplace from the flat tabe representation for an explanation of how
1786 the compression works.
1791 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1794 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1796 Second Pass -- Array Of Lists Representation
1798 Each state will be represented by a list of charid:state records
1799 (reg_trie_trans_le) the first such element holds the CUR and LEN
1800 points of the allocated array. (See defines above).
1802 We build the initial structure using the lists, and then convert
1803 it into the compressed table form which allows faster lookups
1804 (but cant be modified once converted).
1807 STRLEN transcount = 1;
1809 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1810 "%*sCompiling trie using list compiler\n",
1811 (int)depth * 2 + 2, ""));
1813 trie->states = (reg_trie_state *)
1814 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1815 sizeof(reg_trie_state) );
1819 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1821 regnode *noper = NEXTOPER( cur );
1822 U8 *uc = (U8*)STRING( noper );
1823 const U8 *e = uc + STR_LEN( noper );
1824 U32 state = 1; /* required init */
1825 U16 charid = 0; /* sanity init */
1826 U8 *scan = (U8*)NULL; /* sanity init */
1827 STRLEN foldlen = 0; /* required init */
1828 U32 wordlen = 0; /* required init */
1829 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1832 if (OP(noper) == NOTHING) {
1833 regnode *noper_next= regnext(noper);
1834 if (noper_next != tail && OP(noper_next) == flags) {
1836 uc= (U8*)STRING(noper);
1837 e= uc + STR_LEN(noper);
1841 if (OP(noper) != NOTHING) {
1842 for ( ; uc < e ; uc += len ) {
1847 charid = trie->charmap[ uvc ];
1849 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1853 charid=(U16)SvIV( *svpp );
1856 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1863 if ( !trie->states[ state ].trans.list ) {
1864 TRIE_LIST_NEW( state );
1866 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1867 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1868 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1873 newstate = next_alloc++;
1874 prev_states[newstate] = state;
1875 TRIE_LIST_PUSH( state, charid, newstate );
1880 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1884 TRIE_HANDLE_WORD(state);
1886 } /* end second pass */
1888 /* next alloc is the NEXT state to be allocated */
1889 trie->statecount = next_alloc;
1890 trie->states = (reg_trie_state *)
1891 PerlMemShared_realloc( trie->states,
1893 * sizeof(reg_trie_state) );
1895 /* and now dump it out before we compress it */
1896 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1897 revcharmap, next_alloc,
1901 trie->trans = (reg_trie_trans *)
1902 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1909 for( state=1 ; state < next_alloc ; state ++ ) {
1913 DEBUG_TRIE_COMPILE_MORE_r(
1914 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1918 if (trie->states[state].trans.list) {
1919 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1923 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1924 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1925 if ( forid < minid ) {
1927 } else if ( forid > maxid ) {
1931 if ( transcount < tp + maxid - minid + 1) {
1933 trie->trans = (reg_trie_trans *)
1934 PerlMemShared_realloc( trie->trans,
1936 * sizeof(reg_trie_trans) );
1937 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1939 base = trie->uniquecharcount + tp - minid;
1940 if ( maxid == minid ) {
1942 for ( ; zp < tp ; zp++ ) {
1943 if ( ! trie->trans[ zp ].next ) {
1944 base = trie->uniquecharcount + zp - minid;
1945 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1946 trie->trans[ zp ].check = state;
1952 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1953 trie->trans[ tp ].check = state;
1958 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1959 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1960 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1961 trie->trans[ tid ].check = state;
1963 tp += ( maxid - minid + 1 );
1965 Safefree(trie->states[ state ].trans.list);
1968 DEBUG_TRIE_COMPILE_MORE_r(
1969 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1972 trie->states[ state ].trans.base=base;
1974 trie->lasttrans = tp + 1;
1978 Second Pass -- Flat Table Representation.
1980 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1981 We know that we will need Charcount+1 trans at most to store the data
1982 (one row per char at worst case) So we preallocate both structures
1983 assuming worst case.
1985 We then construct the trie using only the .next slots of the entry
1988 We use the .check field of the first entry of the node temporarily to
1989 make compression both faster and easier by keeping track of how many non
1990 zero fields are in the node.
1992 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1995 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1996 number representing the first entry of the node, and state as a
1997 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1998 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1999 are 2 entrys per node. eg:
2007 The table is internally in the right hand, idx form. However as we also
2008 have to deal with the states array which is indexed by nodenum we have to
2009 use TRIE_NODENUM() to convert.
2012 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2013 "%*sCompiling trie using table compiler\n",
2014 (int)depth * 2 + 2, ""));
2016 trie->trans = (reg_trie_trans *)
2017 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2018 * trie->uniquecharcount + 1,
2019 sizeof(reg_trie_trans) );
2020 trie->states = (reg_trie_state *)
2021 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2022 sizeof(reg_trie_state) );
2023 next_alloc = trie->uniquecharcount + 1;
2026 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2028 regnode *noper = NEXTOPER( cur );
2029 const U8 *uc = (U8*)STRING( noper );
2030 const U8 *e = uc + STR_LEN( noper );
2032 U32 state = 1; /* required init */
2034 U16 charid = 0; /* sanity init */
2035 U32 accept_state = 0; /* sanity init */
2036 U8 *scan = (U8*)NULL; /* sanity init */
2038 STRLEN foldlen = 0; /* required init */
2039 U32 wordlen = 0; /* required init */
2041 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2043 if (OP(noper) == NOTHING) {
2044 regnode *noper_next= regnext(noper);
2045 if (noper_next != tail && OP(noper_next) == flags) {
2047 uc= (U8*)STRING(noper);
2048 e= uc + STR_LEN(noper);
2052 if ( OP(noper) != NOTHING ) {
2053 for ( ; uc < e ; uc += len ) {
2058 charid = trie->charmap[ uvc ];
2060 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2061 charid = svpp ? (U16)SvIV(*svpp) : 0;
2065 if ( !trie->trans[ state + charid ].next ) {
2066 trie->trans[ state + charid ].next = next_alloc;
2067 trie->trans[ state ].check++;
2068 prev_states[TRIE_NODENUM(next_alloc)]
2069 = TRIE_NODENUM(state);
2070 next_alloc += trie->uniquecharcount;
2072 state = trie->trans[ state + charid ].next;
2074 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2076 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2079 accept_state = TRIE_NODENUM( state );
2080 TRIE_HANDLE_WORD(accept_state);
2082 } /* end second pass */
2084 /* and now dump it out before we compress it */
2085 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2087 next_alloc, depth+1));
2091 * Inplace compress the table.*
2093 For sparse data sets the table constructed by the trie algorithm will
2094 be mostly 0/FAIL transitions or to put it another way mostly empty.
2095 (Note that leaf nodes will not contain any transitions.)
2097 This algorithm compresses the tables by eliminating most such
2098 transitions, at the cost of a modest bit of extra work during lookup:
2100 - Each states[] entry contains a .base field which indicates the
2101 index in the state[] array wheres its transition data is stored.
2103 - If .base is 0 there are no valid transitions from that node.
2105 - If .base is nonzero then charid is added to it to find an entry in
2108 -If trans[states[state].base+charid].check!=state then the
2109 transition is taken to be a 0/Fail transition. Thus if there are fail
2110 transitions at the front of the node then the .base offset will point
2111 somewhere inside the previous nodes data (or maybe even into a node
2112 even earlier), but the .check field determines if the transition is
2116 The following process inplace converts the table to the compressed
2117 table: We first do not compress the root node 1,and mark all its
2118 .check pointers as 1 and set its .base pointer as 1 as well. This
2119 allows us to do a DFA construction from the compressed table later,
2120 and ensures that any .base pointers we calculate later are greater
2123 - We set 'pos' to indicate the first entry of the second node.
2125 - We then iterate over the columns of the node, finding the first and
2126 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2127 and set the .check pointers accordingly, and advance pos
2128 appropriately and repreat for the next node. Note that when we copy
2129 the next pointers we have to convert them from the original
2130 NODEIDX form to NODENUM form as the former is not valid post
2133 - If a node has no transitions used we mark its base as 0 and do not
2134 advance the pos pointer.
2136 - If a node only has one transition we use a second pointer into the
2137 structure to fill in allocated fail transitions from other states.
2138 This pointer is independent of the main pointer and scans forward
2139 looking for null transitions that are allocated to a state. When it
2140 finds one it writes the single transition into the "hole". If the
2141 pointer doesnt find one the single transition is appended as normal.
2143 - Once compressed we can Renew/realloc the structures to release the
2146 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2147 specifically Fig 3.47 and the associated pseudocode.
2151 const U32 laststate = TRIE_NODENUM( next_alloc );
2154 trie->statecount = laststate;
2156 for ( state = 1 ; state < laststate ; state++ ) {
2158 const U32 stateidx = TRIE_NODEIDX( state );
2159 const U32 o_used = trie->trans[ stateidx ].check;
2160 U32 used = trie->trans[ stateidx ].check;
2161 trie->trans[ stateidx ].check = 0;
2163 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2164 if ( flag || trie->trans[ stateidx + charid ].next ) {
2165 if ( trie->trans[ stateidx + charid ].next ) {
2167 for ( ; zp < pos ; zp++ ) {
2168 if ( ! trie->trans[ zp ].next ) {
2172 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2173 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2174 trie->trans[ zp ].check = state;
2175 if ( ++zp > pos ) pos = zp;
2182 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2184 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2185 trie->trans[ pos ].check = state;
2190 trie->lasttrans = pos + 1;
2191 trie->states = (reg_trie_state *)
2192 PerlMemShared_realloc( trie->states, laststate
2193 * sizeof(reg_trie_state) );
2194 DEBUG_TRIE_COMPILE_MORE_r(
2195 PerlIO_printf( Perl_debug_log,
2196 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2197 (int)depth * 2 + 2,"",
2198 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2201 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2204 } /* end table compress */
2206 DEBUG_TRIE_COMPILE_MORE_r(
2207 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2208 (int)depth * 2 + 2, "",
2209 (UV)trie->statecount,
2210 (UV)trie->lasttrans)
2212 /* resize the trans array to remove unused space */
2213 trie->trans = (reg_trie_trans *)
2214 PerlMemShared_realloc( trie->trans, trie->lasttrans
2215 * sizeof(reg_trie_trans) );
2217 { /* Modify the program and insert the new TRIE node */
2218 U8 nodetype =(U8)(flags & 0xFF);
2222 regnode *optimize = NULL;
2223 #ifdef RE_TRACK_PATTERN_OFFSETS
2226 U32 mjd_nodelen = 0;
2227 #endif /* RE_TRACK_PATTERN_OFFSETS */
2228 #endif /* DEBUGGING */
2230 This means we convert either the first branch or the first Exact,
2231 depending on whether the thing following (in 'last') is a branch
2232 or not and whther first is the startbranch (ie is it a sub part of
2233 the alternation or is it the whole thing.)
2234 Assuming its a sub part we convert the EXACT otherwise we convert
2235 the whole branch sequence, including the first.
2237 /* Find the node we are going to overwrite */
2238 if ( first != startbranch || OP( last ) == BRANCH ) {
2239 /* branch sub-chain */
2240 NEXT_OFF( first ) = (U16)(last - first);
2241 #ifdef RE_TRACK_PATTERN_OFFSETS
2243 mjd_offset= Node_Offset((convert));
2244 mjd_nodelen= Node_Length((convert));
2247 /* whole branch chain */
2249 #ifdef RE_TRACK_PATTERN_OFFSETS
2252 const regnode *nop = NEXTOPER( convert );
2253 mjd_offset= Node_Offset((nop));
2254 mjd_nodelen= Node_Length((nop));
2258 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2259 (int)depth * 2 + 2, "",
2260 (UV)mjd_offset, (UV)mjd_nodelen)
2263 /* But first we check to see if there is a common prefix we can
2264 split out as an EXACT and put in front of the TRIE node. */
2265 trie->startstate= 1;
2266 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2268 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2272 const U32 base = trie->states[ state ].trans.base;
2274 if ( trie->states[state].wordnum )
2277 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2278 if ( ( base + ofs >= trie->uniquecharcount ) &&
2279 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2280 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2282 if ( ++count > 1 ) {
2283 SV **tmp = av_fetch( revcharmap, ofs, 0);
2284 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2285 if ( state == 1 ) break;
2287 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2289 PerlIO_printf(Perl_debug_log,
2290 "%*sNew Start State=%"UVuf" Class: [",
2291 (int)depth * 2 + 2, "",
2294 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2295 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2297 TRIE_BITMAP_SET(trie,*ch);
2299 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2301 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2305 TRIE_BITMAP_SET(trie,*ch);
2307 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2308 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2314 SV **tmp = av_fetch( revcharmap, idx, 0);
2316 char *ch = SvPV( *tmp, len );
2318 SV *sv=sv_newmortal();
2319 PerlIO_printf( Perl_debug_log,
2320 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2321 (int)depth * 2 + 2, "",
2323 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2324 PL_colors[0], PL_colors[1],
2325 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2326 PERL_PV_ESCAPE_FIRSTCHAR
2331 OP( convert ) = nodetype;
2332 str=STRING(convert);
2335 STR_LEN(convert) += len;
2341 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2346 trie->prefixlen = (state-1);
2348 regnode *n = convert+NODE_SZ_STR(convert);
2349 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2350 trie->startstate = state;
2351 trie->minlen -= (state - 1);
2352 trie->maxlen -= (state - 1);
2354 /* At least the UNICOS C compiler choked on this
2355 * being argument to DEBUG_r(), so let's just have
2358 #ifdef PERL_EXT_RE_BUILD
2364 regnode *fix = convert;
2365 U32 word = trie->wordcount;
2367 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2368 while( ++fix < n ) {
2369 Set_Node_Offset_Length(fix, 0, 0);
2372 SV ** const tmp = av_fetch( trie_words, word, 0 );
2374 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2375 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2377 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2385 NEXT_OFF(convert) = (U16)(tail - convert);
2386 DEBUG_r(optimize= n);
2392 if ( trie->maxlen ) {
2393 NEXT_OFF( convert ) = (U16)(tail - convert);
2394 ARG_SET( convert, data_slot );
2395 /* Store the offset to the first unabsorbed branch in
2396 jump[0], which is otherwise unused by the jump logic.
2397 We use this when dumping a trie and during optimisation. */
2399 trie->jump[0] = (U16)(nextbranch - convert);
2401 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2402 * and there is a bitmap
2403 * and the first "jump target" node we found leaves enough room
2404 * then convert the TRIE node into a TRIEC node, with the bitmap
2405 * embedded inline in the opcode - this is hypothetically faster.
2407 if ( !trie->states[trie->startstate].wordnum
2409 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2411 OP( convert ) = TRIEC;
2412 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2413 PerlMemShared_free(trie->bitmap);
2416 OP( convert ) = TRIE;
2418 /* store the type in the flags */
2419 convert->flags = nodetype;
2423 + regarglen[ OP( convert ) ];
2425 /* XXX We really should free up the resource in trie now,
2426 as we won't use them - (which resources?) dmq */
2428 /* needed for dumping*/
2429 DEBUG_r(if (optimize) {
2430 regnode *opt = convert;
2432 while ( ++opt < optimize) {
2433 Set_Node_Offset_Length(opt,0,0);
2436 Try to clean up some of the debris left after the
2439 while( optimize < jumper ) {
2440 mjd_nodelen += Node_Length((optimize));
2441 OP( optimize ) = OPTIMIZED;
2442 Set_Node_Offset_Length(optimize,0,0);
2445 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2447 } /* end node insert */
2449 /* Finish populating the prev field of the wordinfo array. Walk back
2450 * from each accept state until we find another accept state, and if
2451 * so, point the first word's .prev field at the second word. If the
2452 * second already has a .prev field set, stop now. This will be the
2453 * case either if we've already processed that word's accept state,
2454 * or that state had multiple words, and the overspill words were
2455 * already linked up earlier.
2462 for (word=1; word <= trie->wordcount; word++) {
2464 if (trie->wordinfo[word].prev)
2466 state = trie->wordinfo[word].accept;
2468 state = prev_states[state];
2471 prev = trie->states[state].wordnum;
2475 trie->wordinfo[word].prev = prev;
2477 Safefree(prev_states);
2481 /* and now dump out the compressed format */
2482 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2484 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2486 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2487 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2489 SvREFCNT_dec_NN(revcharmap);
2493 : trie->startstate>1
2499 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2501 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2503 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2504 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2507 We find the fail state for each state in the trie, this state is the longest proper
2508 suffix of the current state's 'word' that is also a proper prefix of another word in our
2509 trie. State 1 represents the word '' and is thus the default fail state. This allows
2510 the DFA not to have to restart after its tried and failed a word at a given point, it
2511 simply continues as though it had been matching the other word in the first place.
2513 'abcdgu'=~/abcdefg|cdgu/
2514 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2515 fail, which would bring us to the state representing 'd' in the second word where we would
2516 try 'g' and succeed, proceeding to match 'cdgu'.
2518 /* add a fail transition */
2519 const U32 trie_offset = ARG(source);
2520 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2522 const U32 ucharcount = trie->uniquecharcount;
2523 const U32 numstates = trie->statecount;
2524 const U32 ubound = trie->lasttrans + ucharcount;
2528 U32 base = trie->states[ 1 ].trans.base;
2531 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2532 GET_RE_DEBUG_FLAGS_DECL;
2534 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2536 PERL_UNUSED_ARG(depth);
2540 ARG_SET( stclass, data_slot );
2541 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2542 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2543 aho->trie=trie_offset;
2544 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2545 Copy( trie->states, aho->states, numstates, reg_trie_state );
2546 Newxz( q, numstates, U32);
2547 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2550 /* initialize fail[0..1] to be 1 so that we always have
2551 a valid final fail state */
2552 fail[ 0 ] = fail[ 1 ] = 1;
2554 for ( charid = 0; charid < ucharcount ; charid++ ) {
2555 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2557 q[ q_write ] = newstate;
2558 /* set to point at the root */
2559 fail[ q[ q_write++ ] ]=1;
2562 while ( q_read < q_write) {
2563 const U32 cur = q[ q_read++ % numstates ];
2564 base = trie->states[ cur ].trans.base;
2566 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2567 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2569 U32 fail_state = cur;
2572 fail_state = fail[ fail_state ];
2573 fail_base = aho->states[ fail_state ].trans.base;
2574 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2576 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2577 fail[ ch_state ] = fail_state;
2578 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2580 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2582 q[ q_write++ % numstates] = ch_state;
2586 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2587 when we fail in state 1, this allows us to use the
2588 charclass scan to find a valid start char. This is based on the principle
2589 that theres a good chance the string being searched contains lots of stuff
2590 that cant be a start char.
2592 fail[ 0 ] = fail[ 1 ] = 0;
2593 DEBUG_TRIE_COMPILE_r({
2594 PerlIO_printf(Perl_debug_log,
2595 "%*sStclass Failtable (%"UVuf" states): 0",
2596 (int)(depth * 2), "", (UV)numstates
2598 for( q_read=1; q_read<numstates; q_read++ ) {
2599 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2601 PerlIO_printf(Perl_debug_log, "\n");
2604 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2609 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2610 * These need to be revisited when a newer toolchain becomes available.
2612 #if defined(__sparc64__) && defined(__GNUC__)
2613 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2614 # undef SPARC64_GCC_WORKAROUND
2615 # define SPARC64_GCC_WORKAROUND 1
2619 #define DEBUG_PEEP(str,scan,depth) \
2620 DEBUG_OPTIMISE_r({if (scan){ \
2621 SV * const mysv=sv_newmortal(); \
2622 regnode *Next = regnext(scan); \
2623 regprop(RExC_rx, mysv, scan); \
2624 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2625 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2626 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2630 /* The below joins as many adjacent EXACTish nodes as possible into a single
2631 * one. The regop may be changed if the node(s) contain certain sequences that
2632 * require special handling. The joining is only done if:
2633 * 1) there is room in the current conglomerated node to entirely contain the
2635 * 2) they are the exact same node type
2637 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2638 * these get optimized out
2640 * If a node is to match under /i (folded), the number of characters it matches
2641 * can be different than its character length if it contains a multi-character
2642 * fold. *min_subtract is set to the total delta of the input nodes.
2644 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2645 * and contains LATIN SMALL LETTER SHARP S
2647 * This is as good a place as any to discuss the design of handling these
2648 * multi-character fold sequences. It's been wrong in Perl for a very long
2649 * time. There are three code points in Unicode whose multi-character folds
2650 * were long ago discovered to mess things up. The previous designs for
2651 * dealing with these involved assigning a special node for them. This
2652 * approach doesn't work, as evidenced by this example:
2653 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2654 * Both these fold to "sss", but if the pattern is parsed to create a node that
2655 * would match just the \xDF, it won't be able to handle the case where a
2656 * successful match would have to cross the node's boundary. The new approach
2657 * that hopefully generally solves the problem generates an EXACTFU_SS node
2660 * It turns out that there are problems with all multi-character folds, and not
2661 * just these three. Now the code is general, for all such cases, but the
2662 * three still have some special handling. The approach taken is:
2663 * 1) This routine examines each EXACTFish node that could contain multi-
2664 * character fold sequences. It returns in *min_subtract how much to
2665 * subtract from the the actual length of the string to get a real minimum
2666 * match length; it is 0 if there are no multi-char folds. This delta is
2667 * used by the caller to adjust the min length of the match, and the delta
2668 * between min and max, so that the optimizer doesn't reject these
2669 * possibilities based on size constraints.
2670 * 2) Certain of these sequences require special handling by the trie code,
2671 * so, if found, this code changes the joined node type to special ops:
2672 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2673 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2674 * is used for an EXACTFU node that contains at least one "ss" sequence in
2675 * it. For non-UTF-8 patterns and strings, this is the only case where
2676 * there is a possible fold length change. That means that a regular
2677 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2678 * with length changes, and so can be processed faster. regexec.c takes
2679 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2680 * pre-folded by regcomp.c. This saves effort in regex matching.
2681 * However, the pre-folding isn't done for non-UTF8 patterns because the
2682 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2683 * down by forcing the pattern into UTF8 unless necessary. Also what
2684 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2685 * possibilities for the non-UTF8 patterns are quite simple, except for
2686 * the sharp s. All the ones that don't involve a UTF-8 target string are
2687 * members of a fold-pair, and arrays are set up for all of them so that
2688 * the other member of the pair can be found quickly. Code elsewhere in
2689 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2690 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2691 * described in the next item.
2692 * 4) A problem remains for the sharp s in EXACTF and EXACTFA nodes when the
2693 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2694 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2695 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2696 * character in the pattern corresponds to at most a single character in
2697 * the target string. (And I do mean character, and not byte here, unlike
2698 * other parts of the documentation that have never been updated to
2699 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2700 * two character string 'ss'; in EXACTFA nodes it can match
2701 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2702 * instances where it is violated. I'm reluctant to try to change the
2703 * assumption, as the code involved is impenetrable to me (khw), so
2704 * instead the code here punts. This routine examines (when the pattern
2705 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2706 * boolean indicating whether or not the node contains a sharp s. When it
2707 * is true, the caller sets a flag that later causes the optimizer in this
2708 * file to not set values for the floating and fixed string lengths, and
2709 * thus avoids the optimizer code in regexec.c that makes the invalid
2710 * assumption. Thus, there is no optimization based on string lengths for
2711 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2712 * (The reason the assumption is wrong only in these two cases is that all
2713 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2714 * other folds to their expanded versions. We can't prefold sharp s to
2715 * 'ss' in EXACTF nodes because we don't know at compile time if it
2716 * actually matches 'ss' or not. It will match iff the target string is
2717 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2718 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2719 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2720 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2721 * require the pattern to be forced into UTF-8, the overhead of which we
2725 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2726 if (PL_regkind[OP(scan)] == EXACT) \
2727 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2730 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) {
2731 /* Merge several consecutive EXACTish nodes into one. */
2732 regnode *n = regnext(scan);
2734 regnode *next = scan + NODE_SZ_STR(scan);
2738 regnode *stop = scan;
2739 GET_RE_DEBUG_FLAGS_DECL;
2741 PERL_UNUSED_ARG(depth);
2744 PERL_ARGS_ASSERT_JOIN_EXACT;
2745 #ifndef EXPERIMENTAL_INPLACESCAN
2746 PERL_UNUSED_ARG(flags);
2747 PERL_UNUSED_ARG(val);
2749 DEBUG_PEEP("join",scan,depth);
2751 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2752 * EXACT ones that are mergeable to the current one. */
2754 && (PL_regkind[OP(n)] == NOTHING
2755 || (stringok && OP(n) == OP(scan)))
2757 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2760 if (OP(n) == TAIL || n > next)
2762 if (PL_regkind[OP(n)] == NOTHING) {
2763 DEBUG_PEEP("skip:",n,depth);
2764 NEXT_OFF(scan) += NEXT_OFF(n);
2765 next = n + NODE_STEP_REGNODE;
2772 else if (stringok) {
2773 const unsigned int oldl = STR_LEN(scan);
2774 regnode * const nnext = regnext(n);
2776 /* XXX I (khw) kind of doubt that this works on platforms where
2777 * U8_MAX is above 255 because of lots of other assumptions */
2778 /* Don't join if the sum can't fit into a single node */
2779 if (oldl + STR_LEN(n) > U8_MAX)
2782 DEBUG_PEEP("merg",n,depth);
2785 NEXT_OFF(scan) += NEXT_OFF(n);
2786 STR_LEN(scan) += STR_LEN(n);
2787 next = n + NODE_SZ_STR(n);
2788 /* Now we can overwrite *n : */
2789 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2797 #ifdef EXPERIMENTAL_INPLACESCAN
2798 if (flags && !NEXT_OFF(n)) {
2799 DEBUG_PEEP("atch", val, depth);
2800 if (reg_off_by_arg[OP(n)]) {
2801 ARG_SET(n, val - n);
2804 NEXT_OFF(n) = val - n;
2812 *has_exactf_sharp_s = FALSE;
2814 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2815 * can now analyze for sequences of problematic code points. (Prior to
2816 * this final joining, sequences could have been split over boundaries, and
2817 * hence missed). The sequences only happen in folding, hence for any
2818 * non-EXACT EXACTish node */
2819 if (OP(scan) != EXACT) {
2820 const U8 * const s0 = (U8*) STRING(scan);
2822 const U8 * const s_end = s0 + STR_LEN(scan);
2824 /* One pass is made over the node's string looking for all the
2825 * possibilities. to avoid some tests in the loop, there are two main
2826 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2830 /* Examine the string for a multi-character fold sequence. UTF-8
2831 * patterns have all characters pre-folded by the time this code is
2833 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2834 length sequence we are looking for is 2 */
2837 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2838 if (! len) { /* Not a multi-char fold: get next char */
2843 /* Nodes with 'ss' require special handling, except for EXACTFL
2844 * and EXACTFA for which there is no multi-char fold to this */
2845 if (len == 2 && *s == 's' && *(s+1) == 's'
2846 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2849 OP(scan) = EXACTFU_SS;
2852 else if (len == 6 /* len is the same in both ASCII and EBCDIC
2854 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2855 COMBINING_DIAERESIS_UTF8
2856 COMBINING_ACUTE_ACCENT_UTF8,
2858 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2859 COMBINING_DIAERESIS_UTF8
2860 COMBINING_ACUTE_ACCENT_UTF8,
2865 /* These two folds require special handling by trie's, so
2866 * change the node type to indicate this. If EXACTFA and
2867 * EXACTFL were ever to be handled by trie's, this would
2868 * have to be changed. If this node has already been
2869 * changed to EXACTFU_SS in this loop, leave it as is. (I
2870 * (khw) think it doesn't matter in regexec.c for UTF
2871 * patterns, but no need to change it */
2872 if (OP(scan) == EXACTFU) {
2873 OP(scan) = EXACTFU_TRICKYFOLD;
2877 else { /* Here is a generic multi-char fold. */
2878 const U8* multi_end = s + len;
2880 /* Count how many characters in it. In the case of /l and
2881 * /aa, no folds which contain ASCII code points are
2882 * allowed, so check for those, and skip if found. (In
2883 * EXACTFL, no folds are allowed to any Latin1 code point,
2884 * not just ASCII. But there aren't any of these
2885 * currently, nor ever likely, so don't take the time to
2886 * test for them. The code that generates the
2887 * is_MULTI_foo() macros croaks should one actually get put
2888 * into Unicode .) */
2889 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2890 count = utf8_length(s, multi_end);
2894 while (s < multi_end) {
2897 goto next_iteration;
2907 /* The delta is how long the sequence is minus 1 (1 is how long
2908 * the character that folds to the sequence is) */
2909 *min_subtract += count - 1;
2913 else if (OP(scan) == EXACTFA) {
2915 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2916 * fold to the ASCII range (and there are no existing ones in the
2917 * upper latin1 range). But, as outlined in the comments preceding
2918 * this function, we need to flag any occurrences of the sharp s */
2920 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
2921 *has_exactf_sharp_s = TRUE;
2928 else if (OP(scan) != EXACTFL) {
2930 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2931 * multi-char folds that are all Latin1. (This code knows that
2932 * there are no current multi-char folds possible with EXACTFL,
2933 * relying on fold_grind.t to catch any errors if the very unlikely
2934 * event happens that some get added in future Unicode versions.)
2935 * As explained in the comments preceding this function, we look
2936 * also for the sharp s in EXACTF nodes; it can be in the final
2937 * position. Otherwise we can stop looking 1 byte earlier because
2938 * have to find at least two characters for a multi-fold */
2939 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2941 /* The below is perhaps overboard, but this allows us to save a
2942 * test each time through the loop at the expense of a mask. This
2943 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2944 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2945 * are 64. This uses an exclusive 'or' to find that bit and then
2946 * inverts it to form a mask, with just a single 0, in the bit
2947 * position where 'S' and 's' differ. */
2948 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2949 const U8 s_masked = 's' & S_or_s_mask;
2952 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2953 if (! len) { /* Not a multi-char fold. */
2954 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2956 *has_exactf_sharp_s = TRUE;
2963 && ((*s & S_or_s_mask) == s_masked)
2964 && ((*(s+1) & S_or_s_mask) == s_masked))
2967 /* EXACTF nodes need to know that the minimum length
2968 * changed so that a sharp s in the string can match this
2969 * ss in the pattern, but they remain EXACTF nodes, as they
2970 * won't match this unless the target string is is UTF-8,
2971 * which we don't know until runtime */
2972 if (OP(scan) != EXACTF) {
2973 OP(scan) = EXACTFU_SS;
2977 *min_subtract += len - 1;
2984 /* Allow dumping but overwriting the collection of skipped
2985 * ops and/or strings with fake optimized ops */
2986 n = scan + NODE_SZ_STR(scan);
2994 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2998 /* REx optimizer. Converts nodes into quicker variants "in place".
2999 Finds fixed substrings. */
3001 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3002 to the position after last scanned or to NULL. */
3004 #define INIT_AND_WITHP \
3005 assert(!and_withp); \
3006 Newx(and_withp,1,struct regnode_charclass_class); \
3007 SAVEFREEPV(and_withp)
3009 /* this is a chain of data about sub patterns we are processing that
3010 need to be handled separately/specially in study_chunk. Its so
3011 we can simulate recursion without losing state. */
3013 typedef struct scan_frame {
3014 regnode *last; /* last node to process in this frame */
3015 regnode *next; /* next node to process when last is reached */
3016 struct scan_frame *prev; /*previous frame*/
3017 I32 stop; /* what stopparen do we use */
3021 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3024 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3025 I32 *minlenp, I32 *deltap,
3030 struct regnode_charclass_class *and_withp,
3031 U32 flags, U32 depth)
3032 /* scanp: Start here (read-write). */
3033 /* deltap: Write maxlen-minlen here. */
3034 /* last: Stop before this one. */
3035 /* data: string data about the pattern */
3036 /* stopparen: treat close N as END */
3037 /* recursed: which subroutines have we recursed into */
3038 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3041 I32 min = 0; /* There must be at least this number of characters to match */
3043 regnode *scan = *scanp, *next;
3045 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3046 int is_inf_internal = 0; /* The studied chunk is infinite */
3047 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3048 scan_data_t data_fake;
3049 SV *re_trie_maxbuff = NULL;
3050 regnode *first_non_open = scan;
3051 I32 stopmin = I32_MAX;
3052 scan_frame *frame = NULL;
3053 GET_RE_DEBUG_FLAGS_DECL;
3055 PERL_ARGS_ASSERT_STUDY_CHUNK;
3058 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3062 while (first_non_open && OP(first_non_open) == OPEN)
3063 first_non_open=regnext(first_non_open);
3068 while ( scan && OP(scan) != END && scan < last ){
3069 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3070 node length to get a real minimum (because
3071 the folded version may be shorter) */
3072 bool has_exactf_sharp_s = FALSE;
3073 /* Peephole optimizer: */
3074 DEBUG_STUDYDATA("Peep:", data,depth);
3075 DEBUG_PEEP("Peep",scan,depth);
3077 /* Its not clear to khw or hv why this is done here, and not in the
3078 * clauses that deal with EXACT nodes. khw's guess is that it's
3079 * because of a previous design */
3080 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3082 /* Follow the next-chain of the current node and optimize
3083 away all the NOTHINGs from it. */
3084 if (OP(scan) != CURLYX) {
3085 const int max = (reg_off_by_arg[OP(scan)]
3087 /* I32 may be smaller than U16 on CRAYs! */
3088 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3089 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3093 /* Skip NOTHING and LONGJMP. */
3094 while ((n = regnext(n))
3095 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3096 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3097 && off + noff < max)
3099 if (reg_off_by_arg[OP(scan)])
3102 NEXT_OFF(scan) = off;
3107 /* The principal pseudo-switch. Cannot be a switch, since we
3108 look into several different things. */
3109 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3110 || OP(scan) == IFTHEN) {
3111 next = regnext(scan);
3113 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3115 if (OP(next) == code || code == IFTHEN) {
3116 /* NOTE - There is similar code to this block below for handling
3117 TRIE nodes on a re-study. If you change stuff here check there
3119 I32 max1 = 0, min1 = I32_MAX, num = 0;
3120 struct regnode_charclass_class accum;
3121 regnode * const startbranch=scan;
3123 if (flags & SCF_DO_SUBSTR)
3124 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3125 if (flags & SCF_DO_STCLASS)
3126 cl_init_zero(pRExC_state, &accum);
3128 while (OP(scan) == code) {
3129 I32 deltanext, minnext, f = 0, fake;
3130 struct regnode_charclass_class this_class;
3133 data_fake.flags = 0;
3135 data_fake.whilem_c = data->whilem_c;
3136 data_fake.last_closep = data->last_closep;
3139 data_fake.last_closep = &fake;
3141 data_fake.pos_delta = delta;
3142 next = regnext(scan);
3143 scan = NEXTOPER(scan);
3145 scan = NEXTOPER(scan);
3146 if (flags & SCF_DO_STCLASS) {
3147 cl_init(pRExC_state, &this_class);
3148 data_fake.start_class = &this_class;
3149 f = SCF_DO_STCLASS_AND;
3151 if (flags & SCF_WHILEM_VISITED_POS)
3152 f |= SCF_WHILEM_VISITED_POS;
3154 /* we suppose the run is continuous, last=next...*/
3155 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3157 stopparen, recursed, NULL, f,depth+1);
3160 if (deltanext == I32_MAX) {
3161 is_inf = is_inf_internal = 1;
3163 } else if (max1 < minnext + deltanext)
3164 max1 = minnext + deltanext;
3166 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3168 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3169 if ( stopmin > minnext)
3170 stopmin = min + min1;
3171 flags &= ~SCF_DO_SUBSTR;
3173 data->flags |= SCF_SEEN_ACCEPT;
3176 if (data_fake.flags & SF_HAS_EVAL)
3177 data->flags |= SF_HAS_EVAL;
3178 data->whilem_c = data_fake.whilem_c;
3180 if (flags & SCF_DO_STCLASS)
3181 cl_or(pRExC_state, &accum, &this_class);
3183 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3185 if (flags & SCF_DO_SUBSTR) {
3186 data->pos_min += min1;
3187 if (data->pos_delta >= I32_MAX - (max1 - min1))
3188 data->pos_delta = I32_MAX;
3190 data->pos_delta += max1 - min1;
3191 if (max1 != min1 || is_inf)
3192 data->longest = &(data->longest_float);
3195 if (delta == I32_MAX || I32_MAX - delta - (max1 - min1) < 0)
3198 delta += max1 - min1;
3199 if (flags & SCF_DO_STCLASS_OR) {
3200 cl_or(pRExC_state, data->start_class, &accum);
3202 cl_and(data->start_class, and_withp);
3203 flags &= ~SCF_DO_STCLASS;
3206 else if (flags & SCF_DO_STCLASS_AND) {
3208 cl_and(data->start_class, &accum);
3209 flags &= ~SCF_DO_STCLASS;
3212 /* Switch to OR mode: cache the old value of
3213 * data->start_class */
3215 StructCopy(data->start_class, and_withp,
3216 struct regnode_charclass_class);
3217 flags &= ~SCF_DO_STCLASS_AND;
3218 StructCopy(&accum, data->start_class,
3219 struct regnode_charclass_class);
3220 flags |= SCF_DO_STCLASS_OR;
3221 SET_SSC_EOS(data->start_class);
3225 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3228 Assuming this was/is a branch we are dealing with: 'scan' now
3229 points at the item that follows the branch sequence, whatever
3230 it is. We now start at the beginning of the sequence and look
3237 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3239 If we can find such a subsequence we need to turn the first
3240 element into a trie and then add the subsequent branch exact
3241 strings to the trie.
3245 1. patterns where the whole set of branches can be converted.
3247 2. patterns where only a subset can be converted.
3249 In case 1 we can replace the whole set with a single regop
3250 for the trie. In case 2 we need to keep the start and end
3253 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3254 becomes BRANCH TRIE; BRANCH X;
3256 There is an additional case, that being where there is a
3257 common prefix, which gets split out into an EXACT like node
3258 preceding the TRIE node.
3260 If x(1..n)==tail then we can do a simple trie, if not we make
3261 a "jump" trie, such that when we match the appropriate word
3262 we "jump" to the appropriate tail node. Essentially we turn
3263 a nested if into a case structure of sorts.
3268 if (!re_trie_maxbuff) {
3269 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3270 if (!SvIOK(re_trie_maxbuff))
3271 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3273 if ( SvIV(re_trie_maxbuff)>=0 ) {
3275 regnode *first = (regnode *)NULL;
3276 regnode *last = (regnode *)NULL;
3277 regnode *tail = scan;
3282 SV * const mysv = sv_newmortal(); /* for dumping */
3284 /* var tail is used because there may be a TAIL
3285 regop in the way. Ie, the exacts will point to the
3286 thing following the TAIL, but the last branch will
3287 point at the TAIL. So we advance tail. If we
3288 have nested (?:) we may have to move through several
3292 while ( OP( tail ) == TAIL ) {
3293 /* this is the TAIL generated by (?:) */
3294 tail = regnext( tail );
3298 DEBUG_TRIE_COMPILE_r({
3299 regprop(RExC_rx, mysv, tail );
3300 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3301 (int)depth * 2 + 2, "",
3302 "Looking for TRIE'able sequences. Tail node is: ",
3303 SvPV_nolen_const( mysv )
3309 Step through the branches
3310 cur represents each branch,
3311 noper is the first thing to be matched as part of that branch
3312 noper_next is the regnext() of that node.
3314 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3315 via a "jump trie" but we also support building with NOJUMPTRIE,
3316 which restricts the trie logic to structures like /FOO|BAR/.
3318 If noper is a trieable nodetype then the branch is a possible optimization
3319 target. If we are building under NOJUMPTRIE then we require that noper_next
3320 is the same as scan (our current position in the regex program).
3322 Once we have two or more consecutive such branches we can create a
3323 trie of the EXACT's contents and stitch it in place into the program.
3325 If the sequence represents all of the branches in the alternation we
3326 replace the entire thing with a single TRIE node.
3328 Otherwise when it is a subsequence we need to stitch it in place and
3329 replace only the relevant branches. This means the first branch has
3330 to remain as it is used by the alternation logic, and its next pointer,
3331 and needs to be repointed at the item on the branch chain following
3332 the last branch we have optimized away.
3334 This could be either a BRANCH, in which case the subsequence is internal,
3335 or it could be the item following the branch sequence in which case the
3336 subsequence is at the end (which does not necessarily mean the first node
3337 is the start of the alternation).
3339 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3342 ----------------+-----------
3346 EXACTFU_SS | EXACTFU
3347 EXACTFU_TRICKYFOLD | EXACTFU
3352 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3353 ( EXACT == (X) ) ? EXACT : \
3354 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3357 /* dont use tail as the end marker for this traverse */
3358 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3359 regnode * const noper = NEXTOPER( cur );
3360 U8 noper_type = OP( noper );
3361 U8 noper_trietype = TRIE_TYPE( noper_type );
3362 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3363 regnode * const noper_next = regnext( noper );
3364 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3365 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3368 DEBUG_TRIE_COMPILE_r({
3369 regprop(RExC_rx, mysv, cur);
3370 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3371 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3373 regprop(RExC_rx, mysv, noper);
3374 PerlIO_printf( Perl_debug_log, " -> %s",
3375 SvPV_nolen_const(mysv));
3378 regprop(RExC_rx, mysv, noper_next );
3379 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3380 SvPV_nolen_const(mysv));
3382 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3383 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3384 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3388 /* Is noper a trieable nodetype that can be merged with the
3389 * current trie (if there is one)? */
3393 ( noper_trietype == NOTHING)
3394 || ( trietype == NOTHING )
3395 || ( trietype == noper_trietype )
3398 && noper_next == tail
3402 /* Handle mergable triable node
3403 * Either we are the first node in a new trieable sequence,
3404 * in which case we do some bookkeeping, otherwise we update
3405 * the end pointer. */
3408 if ( noper_trietype == NOTHING ) {
3409 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3410 regnode * const noper_next = regnext( noper );
3411 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3412 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3415 if ( noper_next_trietype ) {
3416 trietype = noper_next_trietype;
3417 } else if (noper_next_type) {
3418 /* a NOTHING regop is 1 regop wide. We need at least two
3419 * for a trie so we can't merge this in */
3423 trietype = noper_trietype;
3426 if ( trietype == NOTHING )
3427 trietype = noper_trietype;
3432 } /* end handle mergable triable node */
3434 /* handle unmergable node -
3435 * noper may either be a triable node which can not be tried
3436 * together with the current trie, or a non triable node */
3438 /* If last is set and trietype is not NOTHING then we have found
3439 * at least two triable branch sequences in a row of a similar
3440 * trietype so we can turn them into a trie. If/when we
3441 * allow NOTHING to start a trie sequence this condition will be
3442 * required, and it isn't expensive so we leave it in for now. */
3443 if ( trietype && trietype != NOTHING )
3444 make_trie( pRExC_state,
3445 startbranch, first, cur, tail, count,
3446 trietype, depth+1 );
3447 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3451 && noper_next == tail
3454 /* noper is triable, so we can start a new trie sequence */
3457 trietype = noper_trietype;
3459 /* if we already saw a first but the current node is not triable then we have
3460 * to reset the first information. */
3465 } /* end handle unmergable node */
3466 } /* loop over branches */
3467 DEBUG_TRIE_COMPILE_r({
3468 regprop(RExC_rx, mysv, cur);
3469 PerlIO_printf( Perl_debug_log,
3470 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3471 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3474 if ( last && trietype ) {
3475 if ( trietype != NOTHING ) {
3476 /* the last branch of the sequence was part of a trie,
3477 * so we have to construct it here outside of the loop
3479 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3480 #ifdef TRIE_STUDY_OPT
3481 if ( ((made == MADE_EXACT_TRIE &&
3482 startbranch == first)
3483 || ( first_non_open == first )) &&
3485 flags |= SCF_TRIE_RESTUDY;
3486 if ( startbranch == first
3489 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3494 /* at this point we know whatever we have is a NOTHING sequence/branch
3495 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3497 if ( startbranch == first ) {
3499 /* the entire thing is a NOTHING sequence, something like this:
3500 * (?:|) So we can turn it into a plain NOTHING op. */
3501 DEBUG_TRIE_COMPILE_r({
3502 regprop(RExC_rx, mysv, cur);
3503 PerlIO_printf( Perl_debug_log,
3504 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3505 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3508 OP(startbranch)= NOTHING;
3509 NEXT_OFF(startbranch)= tail - startbranch;
3510 for ( opt= startbranch + 1; opt < tail ; opt++ )
3514 } /* end if ( last) */
3515 } /* TRIE_MAXBUF is non zero */
3520 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3521 scan = NEXTOPER(NEXTOPER(scan));
3522 } else /* single branch is optimized. */
3523 scan = NEXTOPER(scan);
3525 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3526 scan_frame *newframe = NULL;
3531 if (OP(scan) != SUSPEND) {
3532 /* set the pointer */
3533 if (OP(scan) == GOSUB) {
3535 RExC_recurse[ARG2L(scan)] = scan;
3536 start = RExC_open_parens[paren-1];
3537 end = RExC_close_parens[paren-1];
3540 start = RExC_rxi->program + 1;
3544 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3545 SAVEFREEPV(recursed);
3547 if (!PAREN_TEST(recursed,paren+1)) {
3548 PAREN_SET(recursed,paren+1);
3549 Newx(newframe,1,scan_frame);
3551 if (flags & SCF_DO_SUBSTR) {
3552 SCAN_COMMIT(pRExC_state,data,minlenp);
3553 data->longest = &(data->longest_float);
3555 is_inf = is_inf_internal = 1;
3556 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3557 cl_anything(pRExC_state, data->start_class);
3558 flags &= ~SCF_DO_STCLASS;
3561 Newx(newframe,1,scan_frame);
3564 end = regnext(scan);
3569 SAVEFREEPV(newframe);
3570 newframe->next = regnext(scan);
3571 newframe->last = last;
3572 newframe->stop = stopparen;
3573 newframe->prev = frame;
3583 else if (OP(scan) == EXACT) {
3584 I32 l = STR_LEN(scan);
3587 const U8 * const s = (U8*)STRING(scan);
3588 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3589 l = utf8_length(s, s + l);
3591 uc = *((U8*)STRING(scan));
3594 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3595 /* The code below prefers earlier match for fixed
3596 offset, later match for variable offset. */
3597 if (data->last_end == -1) { /* Update the start info. */
3598 data->last_start_min = data->pos_min;
3599 data->last_start_max = is_inf
3600 ? I32_MAX : data->pos_min + data->pos_delta;
3602 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3604 SvUTF8_on(data->last_found);
3606 SV * const sv = data->last_found;
3607 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3608 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3609 if (mg && mg->mg_len >= 0)
3610 mg->mg_len += utf8_length((U8*)STRING(scan),
3611 (U8*)STRING(scan)+STR_LEN(scan));
3613 data->last_end = data->pos_min + l;
3614 data->pos_min += l; /* As in the first entry. */
3615 data->flags &= ~SF_BEFORE_EOL;
3617 if (flags & SCF_DO_STCLASS_AND) {
3618 /* Check whether it is compatible with what we know already! */
3622 /* If compatible, we or it in below. It is compatible if is
3623 * in the bitmp and either 1) its bit or its fold is set, or 2)
3624 * it's for a locale. Even if there isn't unicode semantics
3625 * here, at runtime there may be because of matching against a
3626 * utf8 string, so accept a possible false positive for
3627 * latin1-range folds */
3629 (!(data->start_class->flags & ANYOF_LOCALE)
3630 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3631 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3632 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3637 ANYOF_CLASS_ZERO(data->start_class);
3638 ANYOF_BITMAP_ZERO(data->start_class);
3640 ANYOF_BITMAP_SET(data->start_class, uc);
3641 else if (uc >= 0x100) {
3644 /* Some Unicode code points fold to the Latin1 range; as
3645 * XXX temporary code, instead of figuring out if this is
3646 * one, just assume it is and set all the start class bits
3647 * that could be some such above 255 code point's fold
3648 * which will generate fals positives. As the code
3649 * elsewhere that does compute the fold settles down, it
3650 * can be extracted out and re-used here */
3651 for (i = 0; i < 256; i++){
3652 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3653 ANYOF_BITMAP_SET(data->start_class, i);
3657 CLEAR_SSC_EOS(data->start_class);
3659 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3661 else if (flags & SCF_DO_STCLASS_OR) {
3662 /* false positive possible if the class is case-folded */
3664 ANYOF_BITMAP_SET(data->start_class, uc);
3666 data->start_class->flags |= ANYOF_UNICODE_ALL;
3667 CLEAR_SSC_EOS(data->start_class);
3668 cl_and(data->start_class, and_withp);
3670 flags &= ~SCF_DO_STCLASS;
3672 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3673 I32 l = STR_LEN(scan);
3674 UV uc = *((U8*)STRING(scan));
3676 /* Search for fixed substrings supports EXACT only. */
3677 if (flags & SCF_DO_SUBSTR) {
3679 SCAN_COMMIT(pRExC_state, data, minlenp);
3682 const U8 * const s = (U8 *)STRING(scan);
3683 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3684 l = utf8_length(s, s + l);
3686 if (has_exactf_sharp_s) {
3687 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3689 min += l - min_subtract;
3691 delta += min_subtract;
3692 if (flags & SCF_DO_SUBSTR) {
3693 data->pos_min += l - min_subtract;
3694 if (data->pos_min < 0) {
3697 data->pos_delta += min_subtract;
3699 data->longest = &(data->longest_float);
3702 if (flags & SCF_DO_STCLASS_AND) {
3703 /* Check whether it is compatible with what we know already! */
3706 (!(data->start_class->flags & ANYOF_LOCALE)
3707 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3708 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3712 ANYOF_CLASS_ZERO(data->start_class);
3713 ANYOF_BITMAP_ZERO(data->start_class);
3715 ANYOF_BITMAP_SET(data->start_class, uc);
3716 CLEAR_SSC_EOS(data->start_class);
3717 if (OP(scan) == EXACTFL) {
3718 /* XXX This set is probably no longer necessary, and
3719 * probably wrong as LOCALE now is on in the initial
3721 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3725 /* Also set the other member of the fold pair. In case
3726 * that unicode semantics is called for at runtime, use
3727 * the full latin1 fold. (Can't do this for locale,
3728 * because not known until runtime) */
3729 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3731 /* All other (EXACTFL handled above) folds except under
3732 * /iaa that include s, S, and sharp_s also may include
3734 if (OP(scan) != EXACTFA) {
3735 if (uc == 's' || uc == 'S') {
3736 ANYOF_BITMAP_SET(data->start_class,
3737 LATIN_SMALL_LETTER_SHARP_S);
3739 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3740 ANYOF_BITMAP_SET(data->start_class, 's');
3741 ANYOF_BITMAP_SET(data->start_class, 'S');
3746 else if (uc >= 0x100) {
3748 for (i = 0; i < 256; i++){
3749 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3750 ANYOF_BITMAP_SET(data->start_class, i);
3755 else if (flags & SCF_DO_STCLASS_OR) {
3756 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3757 /* false positive possible if the class is case-folded.
3758 Assume that the locale settings are the same... */
3760 ANYOF_BITMAP_SET(data->start_class, uc);
3761 if (OP(scan) != EXACTFL) {
3763 /* And set the other member of the fold pair, but
3764 * can't do that in locale because not known until
3766 ANYOF_BITMAP_SET(data->start_class,
3767 PL_fold_latin1[uc]);
3769 /* All folds except under /iaa that include s, S,
3770 * and sharp_s also may include the others */
3771 if (OP(scan) != EXACTFA) {
3772 if (uc == 's' || uc == 'S') {
3773 ANYOF_BITMAP_SET(data->start_class,
3774 LATIN_SMALL_LETTER_SHARP_S);
3776 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3777 ANYOF_BITMAP_SET(data->start_class, 's');
3778 ANYOF_BITMAP_SET(data->start_class, 'S');
3783 CLEAR_SSC_EOS(data->start_class);
3785 cl_and(data->start_class, and_withp);
3787 flags &= ~SCF_DO_STCLASS;
3789 else if (REGNODE_VARIES(OP(scan))) {
3790 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3791 I32 f = flags, pos_before = 0;
3792 regnode * const oscan = scan;
3793 struct regnode_charclass_class this_class;
3794 struct regnode_charclass_class *oclass = NULL;
3795 I32 next_is_eval = 0;
3797 switch (PL_regkind[OP(scan)]) {
3798 case WHILEM: /* End of (?:...)* . */
3799 scan = NEXTOPER(scan);
3802 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3803 next = NEXTOPER(scan);
3804 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3806 maxcount = REG_INFTY;
3807 next = regnext(scan);
3808 scan = NEXTOPER(scan);
3812 if (flags & SCF_DO_SUBSTR)
3817 if (flags & SCF_DO_STCLASS) {
3819 maxcount = REG_INFTY;
3820 next = regnext(scan);
3821 scan = NEXTOPER(scan);
3824 is_inf = is_inf_internal = 1;
3825 scan = regnext(scan);
3826 if (flags & SCF_DO_SUBSTR) {
3827 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3828 data->longest = &(data->longest_float);
3830 goto optimize_curly_tail;
3832 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3833 && (scan->flags == stopparen))
https://perl5.git.perl.org / perl5.git / blob