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; ®dummy = don't = compiling */
130 I32 naughty; /* How bad is this pattern? */
131 I32 sawback; /* Did we see \1, ...? */
133 I32 size; /* Code size. */
134 I32 npar; /* Capture buffer count, (OPEN). */
135 I32 cpar; /* Capture buffer count, (CLOSE). */
136 I32 nestroot; /* root parens we are in - used by accept */
139 regnode **open_parens; /* pointers to open parens */
140 regnode **close_parens; /* pointers to close parens */
141 regnode *opend; /* END node in program */
142 I32 utf8; /* whether the pattern is utf8 or not */
143 I32 orig_utf8; /* whether the pattern was originally in utf8 */
144 /* XXX use this for future optimisation of case
145 * where pattern must be upgraded to utf8. */
146 I32 uni_semantics; /* If a d charset modifier should use unicode
147 rules, even if the pattern is not in
149 HV *paren_names; /* Paren names */
151 regnode **recurse; /* Recurse regops */
152 I32 recurse_count; /* Number of recurse regops */
155 I32 override_recoding;
156 I32 in_multi_char_class;
157 struct reg_code_block *code_blocks; /* positions of literal (?{})
159 int num_code_blocks; /* size of code_blocks[] */
160 int code_index; /* next code_blocks[] slot */
162 char *starttry; /* -Dr: where regtry was called. */
163 #define RExC_starttry (pRExC_state->starttry)
165 SV *runtime_code_qr; /* qr with the runtime code blocks */
167 const char *lastparse;
169 AV *paren_name_list; /* idx -> name */
170 #define RExC_lastparse (pRExC_state->lastparse)
171 #define RExC_lastnum (pRExC_state->lastnum)
172 #define RExC_paren_name_list (pRExC_state->paren_name_list)
176 #define RExC_flags (pRExC_state->flags)
177 #define RExC_pm_flags (pRExC_state->pm_flags)
178 #define RExC_precomp (pRExC_state->precomp)
179 #define RExC_rx_sv (pRExC_state->rx_sv)
180 #define RExC_rx (pRExC_state->rx)
181 #define RExC_rxi (pRExC_state->rxi)
182 #define RExC_start (pRExC_state->start)
183 #define RExC_end (pRExC_state->end)
184 #define RExC_parse (pRExC_state->parse)
185 #define RExC_whilem_seen (pRExC_state->whilem_seen)
186 #ifdef RE_TRACK_PATTERN_OFFSETS
187 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
189 #define RExC_emit (pRExC_state->emit)
190 #define RExC_emit_start (pRExC_state->emit_start)
191 #define RExC_emit_bound (pRExC_state->emit_bound)
192 #define RExC_naughty (pRExC_state->naughty)
193 #define RExC_sawback (pRExC_state->sawback)
194 #define RExC_seen (pRExC_state->seen)
195 #define RExC_size (pRExC_state->size)
196 #define RExC_npar (pRExC_state->npar)
197 #define RExC_nestroot (pRExC_state->nestroot)
198 #define RExC_extralen (pRExC_state->extralen)
199 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
200 #define RExC_utf8 (pRExC_state->utf8)
201 #define RExC_uni_semantics (pRExC_state->uni_semantics)
202 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
203 #define RExC_open_parens (pRExC_state->open_parens)
204 #define RExC_close_parens (pRExC_state->close_parens)
205 #define RExC_opend (pRExC_state->opend)
206 #define RExC_paren_names (pRExC_state->paren_names)
207 #define RExC_recurse (pRExC_state->recurse)
208 #define RExC_recurse_count (pRExC_state->recurse_count)
209 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
210 #define RExC_contains_locale (pRExC_state->contains_locale)
211 #define RExC_override_recoding (pRExC_state->override_recoding)
212 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
215 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
216 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
217 ((*s) == '{' && regcurly(s)))
220 #undef SPSTART /* dratted cpp namespace... */
223 * Flags to be passed up and down.
225 #define WORST 0 /* Worst case. */
226 #define HASWIDTH 0x01 /* Known to match non-null strings. */
228 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
229 * character. (There needs to be a case: in the switch statement in regexec.c
230 * for any node marked SIMPLE.) Note that this is not the same thing as
233 #define SPSTART 0x04 /* Starts with * or + */
234 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
235 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
237 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
239 /* whether trie related optimizations are enabled */
240 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
241 #define TRIE_STUDY_OPT
242 #define FULL_TRIE_STUDY
248 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
249 #define PBITVAL(paren) (1 << ((paren) & 7))
250 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
251 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
252 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
254 /* If not already in utf8, do a longjmp back to the beginning */
255 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
256 #define REQUIRE_UTF8 STMT_START { \
257 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
260 /* This converts the named class defined in regcomp.h to its equivalent class
261 * number defined in handy.h. */
262 #define namedclass_to_classnum(class) ((int) ((class) / 2))
263 #define classnum_to_namedclass(classnum) ((classnum) * 2)
265 /* About scan_data_t.
267 During optimisation we recurse through the regexp program performing
268 various inplace (keyhole style) optimisations. In addition study_chunk
269 and scan_commit populate this data structure with information about
270 what strings MUST appear in the pattern. We look for the longest
271 string that must appear at a fixed location, and we look for the
272 longest string that may appear at a floating location. So for instance
277 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
278 strings (because they follow a .* construct). study_chunk will identify
279 both FOO and BAR as being the longest fixed and floating strings respectively.
281 The strings can be composites, for instance
285 will result in a composite fixed substring 'foo'.
287 For each string some basic information is maintained:
289 - offset or min_offset
290 This is the position the string must appear at, or not before.
291 It also implicitly (when combined with minlenp) tells us how many
292 characters must match before the string we are searching for.
293 Likewise when combined with minlenp and the length of the string it
294 tells us how many characters must appear after the string we have
298 Only used for floating strings. This is the rightmost point that
299 the string can appear at. If set to I32 max it indicates that the
300 string can occur infinitely far to the right.
303 A pointer to the minimum number of characters of the pattern that the
304 string was found inside. This is important as in the case of positive
305 lookahead or positive lookbehind we can have multiple patterns
310 The minimum length of the pattern overall is 3, the minimum length
311 of the lookahead part is 3, but the minimum length of the part that
312 will actually match is 1. So 'FOO's minimum length is 3, but the
313 minimum length for the F is 1. This is important as the minimum length
314 is used to determine offsets in front of and behind the string being
315 looked for. Since strings can be composites this is the length of the
316 pattern at the time it was committed with a scan_commit. Note that
317 the length is calculated by study_chunk, so that the minimum lengths
318 are not known until the full pattern has been compiled, thus the
319 pointer to the value.
323 In the case of lookbehind the string being searched for can be
324 offset past the start point of the final matching string.
325 If this value was just blithely removed from the min_offset it would
326 invalidate some of the calculations for how many chars must match
327 before or after (as they are derived from min_offset and minlen and
328 the length of the string being searched for).
329 When the final pattern is compiled and the data is moved from the
330 scan_data_t structure into the regexp structure the information
331 about lookbehind is factored in, with the information that would
332 have been lost precalculated in the end_shift field for the
335 The fields pos_min and pos_delta are used to store the minimum offset
336 and the delta to the maximum offset at the current point in the pattern.
340 typedef struct scan_data_t {
341 /*I32 len_min; unused */
342 /*I32 len_delta; unused */
346 I32 last_end; /* min value, <0 unless valid. */
349 SV **longest; /* Either &l_fixed, or &l_float. */
350 SV *longest_fixed; /* longest fixed string found in pattern */
351 I32 offset_fixed; /* offset where it starts */
352 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
353 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
354 SV *longest_float; /* longest floating string found in pattern */
355 I32 offset_float_min; /* earliest point in string it can appear */
356 I32 offset_float_max; /* latest point in string it can appear */
357 I32 *minlen_float; /* pointer to the minlen relevant to the string */
358 I32 lookbehind_float; /* is the position of the string modified by LB */
362 struct regnode_charclass_class *start_class;
366 * Forward declarations for pregcomp()'s friends.
369 static const scan_data_t zero_scan_data =
370 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
372 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
373 #define SF_BEFORE_SEOL 0x0001
374 #define SF_BEFORE_MEOL 0x0002
375 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
376 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
379 # define SF_FIX_SHIFT_EOL (0+2)
380 # define SF_FL_SHIFT_EOL (0+4)
382 # define SF_FIX_SHIFT_EOL (+2)
383 # define SF_FL_SHIFT_EOL (+4)
386 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
387 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
389 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
390 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
391 #define SF_IS_INF 0x0040
392 #define SF_HAS_PAR 0x0080
393 #define SF_IN_PAR 0x0100
394 #define SF_HAS_EVAL 0x0200
395 #define SCF_DO_SUBSTR 0x0400
396 #define SCF_DO_STCLASS_AND 0x0800
397 #define SCF_DO_STCLASS_OR 0x1000
398 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
399 #define SCF_WHILEM_VISITED_POS 0x2000
401 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
402 #define SCF_SEEN_ACCEPT 0x8000
404 #define UTF cBOOL(RExC_utf8)
406 /* The enums for all these are ordered so things work out correctly */
407 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
408 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
409 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
410 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
411 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
412 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
413 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
415 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
417 #define OOB_NAMEDCLASS -1
419 /* There is no code point that is out-of-bounds, so this is problematic. But
420 * its only current use is to initialize a variable that is always set before
422 #define OOB_UNICODE 0xDEADBEEF
424 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
425 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
428 /* length of regex to show in messages that don't mark a position within */
429 #define RegexLengthToShowInErrorMessages 127
432 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
433 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
434 * op/pragma/warn/regcomp.
436 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
437 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
439 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
442 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
443 * arg. Show regex, up to a maximum length. If it's too long, chop and add
446 #define _FAIL(code) STMT_START { \
447 const char *ellipses = ""; \
448 IV len = RExC_end - RExC_precomp; \
451 SAVEFREESV(RExC_rx_sv); \
452 if (len > RegexLengthToShowInErrorMessages) { \
453 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
454 len = RegexLengthToShowInErrorMessages - 10; \
460 #define FAIL(msg) _FAIL( \
461 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
462 msg, (int)len, RExC_precomp, ellipses))
464 #define FAIL2(msg,arg) _FAIL( \
465 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
466 arg, (int)len, RExC_precomp, ellipses))
469 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
471 #define Simple_vFAIL(m) STMT_START { \
472 const IV offset = RExC_parse - RExC_precomp; \
473 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
474 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
478 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
480 #define vFAIL(m) STMT_START { \
482 SAVEFREESV(RExC_rx_sv); \
487 * Like Simple_vFAIL(), but accepts two arguments.
489 #define Simple_vFAIL2(m,a1) STMT_START { \
490 const IV offset = RExC_parse - RExC_precomp; \
491 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
492 (int)offset, RExC_precomp, RExC_precomp + offset); \
496 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
498 #define vFAIL2(m,a1) STMT_START { \
500 SAVEFREESV(RExC_rx_sv); \
501 Simple_vFAIL2(m, a1); \
506 * Like Simple_vFAIL(), but accepts three arguments.
508 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
509 const IV offset = RExC_parse - RExC_precomp; \
510 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
511 (int)offset, RExC_precomp, RExC_precomp + offset); \
515 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
517 #define vFAIL3(m,a1,a2) STMT_START { \
519 SAVEFREESV(RExC_rx_sv); \
520 Simple_vFAIL3(m, a1, a2); \
524 * Like Simple_vFAIL(), but accepts four arguments.
526 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
527 const IV offset = RExC_parse - RExC_precomp; \
528 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
529 (int)offset, RExC_precomp, RExC_precomp + offset); \
532 #define ckWARNreg(loc,m) STMT_START { \
533 const IV offset = loc - RExC_precomp; \
534 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
535 (int)offset, RExC_precomp, RExC_precomp + offset); \
538 #define ckWARNregdep(loc,m) STMT_START { \
539 const IV offset = loc - RExC_precomp; \
540 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
542 (int)offset, RExC_precomp, RExC_precomp + offset); \
545 #define ckWARN2regdep(loc,m, a1) STMT_START { \
546 const IV offset = loc - RExC_precomp; \
547 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
549 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
552 #define ckWARN2reg(loc, m, a1) STMT_START { \
553 const IV offset = loc - RExC_precomp; \
554 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
555 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
558 #define vWARN3(loc, m, a1, a2) STMT_START { \
559 const IV offset = loc - RExC_precomp; \
560 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
561 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
564 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
565 const IV offset = loc - RExC_precomp; \
566 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
567 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
570 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
571 const IV offset = loc - RExC_precomp; \
572 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
573 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
576 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
577 const IV offset = loc - RExC_precomp; \
578 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
579 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
582 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
583 const IV offset = loc - RExC_precomp; \
584 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
585 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
589 /* Allow for side effects in s */
590 #define REGC(c,s) STMT_START { \
591 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
594 /* Macros for recording node offsets. 20001227 mjd@plover.com
595 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
596 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
597 * Element 0 holds the number n.
598 * Position is 1 indexed.
600 #ifndef RE_TRACK_PATTERN_OFFSETS
601 #define Set_Node_Offset_To_R(node,byte)
602 #define Set_Node_Offset(node,byte)
603 #define Set_Cur_Node_Offset
604 #define Set_Node_Length_To_R(node,len)
605 #define Set_Node_Length(node,len)
606 #define Set_Node_Cur_Length(node)
607 #define Node_Offset(n)
608 #define Node_Length(n)
609 #define Set_Node_Offset_Length(node,offset,len)
610 #define ProgLen(ri) ri->u.proglen
611 #define SetProgLen(ri,x) ri->u.proglen = x
613 #define ProgLen(ri) ri->u.offsets[0]
614 #define SetProgLen(ri,x) ri->u.offsets[0] = x
615 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
617 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
618 __LINE__, (int)(node), (int)(byte))); \
620 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
622 RExC_offsets[2*(node)-1] = (byte); \
627 #define Set_Node_Offset(node,byte) \
628 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
629 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
631 #define Set_Node_Length_To_R(node,len) STMT_START { \
633 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
634 __LINE__, (int)(node), (int)(len))); \
636 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
638 RExC_offsets[2*(node)] = (len); \
643 #define Set_Node_Length(node,len) \
644 Set_Node_Length_To_R((node)-RExC_emit_start, len)
645 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
646 #define Set_Node_Cur_Length(node) \
647 Set_Node_Length(node, RExC_parse - parse_start)
649 /* Get offsets and lengths */
650 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
651 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
653 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
654 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
655 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
659 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
660 #define EXPERIMENTAL_INPLACESCAN
661 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
663 #define DEBUG_STUDYDATA(str,data,depth) \
664 DEBUG_OPTIMISE_MORE_r(if(data){ \
665 PerlIO_printf(Perl_debug_log, \
666 "%*s" str "Pos:%"IVdf"/%"IVdf \
667 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
668 (int)(depth)*2, "", \
669 (IV)((data)->pos_min), \
670 (IV)((data)->pos_delta), \
671 (UV)((data)->flags), \
672 (IV)((data)->whilem_c), \
673 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
674 is_inf ? "INF " : "" \
676 if ((data)->last_found) \
677 PerlIO_printf(Perl_debug_log, \
678 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
679 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
680 SvPVX_const((data)->last_found), \
681 (IV)((data)->last_end), \
682 (IV)((data)->last_start_min), \
683 (IV)((data)->last_start_max), \
684 ((data)->longest && \
685 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
686 SvPVX_const((data)->longest_fixed), \
687 (IV)((data)->offset_fixed), \
688 ((data)->longest && \
689 (data)->longest==&((data)->longest_float)) ? "*" : "", \
690 SvPVX_const((data)->longest_float), \
691 (IV)((data)->offset_float_min), \
692 (IV)((data)->offset_float_max) \
694 PerlIO_printf(Perl_debug_log,"\n"); \
697 /* Mark that we cannot extend a found fixed substring at this point.
698 Update the longest found anchored substring and the longest found
699 floating substrings if needed. */
702 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
704 const STRLEN l = CHR_SVLEN(data->last_found);
705 const STRLEN old_l = CHR_SVLEN(*data->longest);
706 GET_RE_DEBUG_FLAGS_DECL;
708 PERL_ARGS_ASSERT_SCAN_COMMIT;
710 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
711 SvSetMagicSV(*data->longest, data->last_found);
712 if (*data->longest == data->longest_fixed) {
713 data->offset_fixed = l ? data->last_start_min : data->pos_min;
714 if (data->flags & SF_BEFORE_EOL)
716 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
718 data->flags &= ~SF_FIX_BEFORE_EOL;
719 data->minlen_fixed=minlenp;
720 data->lookbehind_fixed=0;
722 else { /* *data->longest == data->longest_float */
723 data->offset_float_min = l ? data->last_start_min : data->pos_min;
724 data->offset_float_max = (l
725 ? data->last_start_max
726 : data->pos_min + data->pos_delta);
727 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
728 data->offset_float_max = I32_MAX;
729 if (data->flags & SF_BEFORE_EOL)
731 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
733 data->flags &= ~SF_FL_BEFORE_EOL;
734 data->minlen_float=minlenp;
735 data->lookbehind_float=0;
738 SvCUR_set(data->last_found, 0);
740 SV * const sv = data->last_found;
741 if (SvUTF8(sv) && SvMAGICAL(sv)) {
742 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
748 data->flags &= ~SF_BEFORE_EOL;
749 DEBUG_STUDYDATA("commit: ",data,0);
752 /* These macros set, clear and test whether the synthetic start class ('ssc',
753 * given by the parameter) matches an empty string (EOS). This uses the
754 * 'next_off' field in the node, to save a bit in the flags field. The ssc
755 * stands alone, so there is never a next_off, so this field is otherwise
756 * unused. The EOS information is used only for compilation, but theoretically
757 * it could be passed on to the execution code. This could be used to store
758 * more than one bit of information, but only this one is currently used. */
759 #define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
760 #define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
761 #define TEST_SSC_EOS(node) cBOOL((node)->next_off)
763 /* Can match anything (initialization) */
765 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
767 PERL_ARGS_ASSERT_CL_ANYTHING;
769 ANYOF_BITMAP_SETALL(cl);
770 cl->flags = ANYOF_UNICODE_ALL;
773 /* If any portion of the regex is to operate under locale rules,
774 * initialization includes it. The reason this isn't done for all regexes
775 * is that the optimizer was written under the assumption that locale was
776 * all-or-nothing. Given the complexity and lack of documentation in the
777 * optimizer, and that there are inadequate test cases for locale, so many
778 * parts of it may not work properly, it is safest to avoid locale unless
780 if (RExC_contains_locale) {
781 ANYOF_CLASS_SETALL(cl); /* /l uses class */
782 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
785 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
789 /* Can match anything (initialization) */
791 S_cl_is_anything(const struct regnode_charclass_class *cl)
795 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
797 for (value = 0; value <= ANYOF_MAX; value += 2)
798 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
800 if (!(cl->flags & ANYOF_UNICODE_ALL))
802 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
807 /* Can match anything (initialization) */
809 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
811 PERL_ARGS_ASSERT_CL_INIT;
813 Zero(cl, 1, struct regnode_charclass_class);
815 cl_anything(pRExC_state, cl);
816 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
819 /* These two functions currently do the exact same thing */
820 #define cl_init_zero S_cl_init
822 /* 'AND' a given class with another one. Can create false positives. 'cl'
823 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
824 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
826 S_cl_and(struct regnode_charclass_class *cl,
827 const struct regnode_charclass_class *and_with)
829 PERL_ARGS_ASSERT_CL_AND;
831 assert(PL_regkind[and_with->type] == ANYOF);
833 /* I (khw) am not sure all these restrictions are necessary XXX */
834 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
835 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
836 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
837 && !(and_with->flags & ANYOF_LOC_FOLD)
838 && !(cl->flags & ANYOF_LOC_FOLD)) {
841 if (and_with->flags & ANYOF_INVERT)
842 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
843 cl->bitmap[i] &= ~and_with->bitmap[i];
845 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
846 cl->bitmap[i] &= and_with->bitmap[i];
847 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
849 if (and_with->flags & ANYOF_INVERT) {
851 /* Here, the and'ed node is inverted. Get the AND of the flags that
852 * aren't affected by the inversion. Those that are affected are
853 * handled individually below */
854 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
855 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
856 cl->flags |= affected_flags;
858 /* We currently don't know how to deal with things that aren't in the
859 * bitmap, but we know that the intersection is no greater than what
860 * is already in cl, so let there be false positives that get sorted
861 * out after the synthetic start class succeeds, and the node is
862 * matched for real. */
864 /* The inversion of these two flags indicate that the resulting
865 * intersection doesn't have them */
866 if (and_with->flags & ANYOF_UNICODE_ALL) {
867 cl->flags &= ~ANYOF_UNICODE_ALL;
869 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
870 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
873 else { /* and'd node is not inverted */
874 U8 outside_bitmap_but_not_utf8; /* Temp variable */
876 if (! ANYOF_NONBITMAP(and_with)) {
878 /* Here 'and_with' doesn't match anything outside the bitmap
879 * (except possibly ANYOF_UNICODE_ALL), which means the
880 * intersection can't either, except for ANYOF_UNICODE_ALL, in
881 * which case we don't know what the intersection is, but it's no
882 * greater than what cl already has, so can just leave it alone,
883 * with possible false positives */
884 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
885 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
886 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
889 else if (! ANYOF_NONBITMAP(cl)) {
891 /* Here, 'and_with' does match something outside the bitmap, and cl
892 * doesn't have a list of things to match outside the bitmap. If
893 * cl can match all code points above 255, the intersection will
894 * be those above-255 code points that 'and_with' matches. If cl
895 * can't match all Unicode code points, it means that it can't
896 * match anything outside the bitmap (since the 'if' that got us
897 * into this block tested for that), so we leave the bitmap empty.
899 if (cl->flags & ANYOF_UNICODE_ALL) {
900 ARG_SET(cl, ARG(and_with));
902 /* and_with's ARG may match things that don't require UTF8.
903 * And now cl's will too, in spite of this being an 'and'. See
904 * the comments below about the kludge */
905 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
909 /* Here, both 'and_with' and cl match something outside the
910 * bitmap. Currently we do not do the intersection, so just match
911 * whatever cl had at the beginning. */
915 /* Take the intersection of the two sets of flags. However, the
916 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
917 * kludge around the fact that this flag is not treated like the others
918 * which are initialized in cl_anything(). The way the optimizer works
919 * is that the synthetic start class (SSC) is initialized to match
920 * anything, and then the first time a real node is encountered, its
921 * values are AND'd with the SSC's with the result being the values of
922 * the real node. However, there are paths through the optimizer where
923 * the AND never gets called, so those initialized bits are set
924 * inappropriately, which is not usually a big deal, as they just cause
925 * false positives in the SSC, which will just mean a probably
926 * imperceptible slow down in execution. However this bit has a
927 * higher false positive consequence in that it can cause utf8.pm,
928 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
929 * bigger slowdown and also causes significant extra memory to be used.
930 * In order to prevent this, the code now takes a different tack. The
931 * bit isn't set unless some part of the regular expression needs it,
932 * but once set it won't get cleared. This means that these extra
933 * modules won't get loaded unless there was some path through the
934 * pattern that would have required them anyway, and so any false
935 * positives that occur by not ANDing them out when they could be
936 * aren't as severe as they would be if we treated this bit like all
938 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
939 & ANYOF_NONBITMAP_NON_UTF8;
940 cl->flags &= and_with->flags;
941 cl->flags |= outside_bitmap_but_not_utf8;
945 /* 'OR' a given class with another one. Can create false positives. 'cl'
946 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
947 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
949 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
951 PERL_ARGS_ASSERT_CL_OR;
953 if (or_with->flags & ANYOF_INVERT) {
955 /* Here, the or'd node is to be inverted. This means we take the
956 * complement of everything not in the bitmap, but currently we don't
957 * know what that is, so give up and match anything */
958 if (ANYOF_NONBITMAP(or_with)) {
959 cl_anything(pRExC_state, cl);
962 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
963 * <= (B1 | !B2) | (CL1 | !CL2)
964 * which is wasteful if CL2 is small, but we ignore CL2:
965 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
966 * XXXX Can we handle case-fold? Unclear:
967 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
968 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
970 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
971 && !(or_with->flags & ANYOF_LOC_FOLD)
972 && !(cl->flags & ANYOF_LOC_FOLD) ) {
975 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
976 cl->bitmap[i] |= ~or_with->bitmap[i];
977 } /* XXXX: logic is complicated otherwise */
979 cl_anything(pRExC_state, cl);
982 /* And, we can just take the union of the flags that aren't affected
983 * by the inversion */
984 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
986 /* For the remaining flags:
987 ANYOF_UNICODE_ALL and inverted means to not match anything above
988 255, which means that the union with cl should just be
989 what cl has in it, so can ignore this flag
990 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
991 is 127-255 to match them, but then invert that, so the
992 union with cl should just be what cl has in it, so can
995 } else { /* 'or_with' is not inverted */
996 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
997 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
998 && (!(or_with->flags & ANYOF_LOC_FOLD)
999 || (cl->flags & ANYOF_LOC_FOLD)) ) {
1002 /* OR char bitmap and class bitmap separately */
1003 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1004 cl->bitmap[i] |= or_with->bitmap[i];
1005 ANYOF_CLASS_OR(or_with, cl);
1007 else { /* XXXX: logic is complicated, leave it along for a moment. */
1008 cl_anything(pRExC_state, cl);
1011 if (ANYOF_NONBITMAP(or_with)) {
1013 /* Use the added node's outside-the-bit-map match if there isn't a
1014 * conflict. If there is a conflict (both nodes match something
1015 * outside the bitmap, but what they match outside is not the same
1016 * pointer, and hence not easily compared until XXX we extend
1017 * inversion lists this far), give up and allow the start class to
1018 * match everything outside the bitmap. If that stuff is all above
1019 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1020 if (! ANYOF_NONBITMAP(cl)) {
1021 ARG_SET(cl, ARG(or_with));
1023 else if (ARG(cl) != ARG(or_with)) {
1025 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1026 cl_anything(pRExC_state, cl);
1029 cl->flags |= ANYOF_UNICODE_ALL;
1034 /* Take the union */
1035 cl->flags |= or_with->flags;
1039 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1040 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1041 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1042 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1047 dump_trie(trie,widecharmap,revcharmap)
1048 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1049 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1051 These routines dump out a trie in a somewhat readable format.
1052 The _interim_ variants are used for debugging the interim
1053 tables that are used to generate the final compressed
1054 representation which is what dump_trie expects.
1056 Part of the reason for their existence is to provide a form
1057 of documentation as to how the different representations function.
1062 Dumps the final compressed table form of the trie to Perl_debug_log.
1063 Used for debugging make_trie().
1067 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1068 AV *revcharmap, U32 depth)
1071 SV *sv=sv_newmortal();
1072 int colwidth= widecharmap ? 6 : 4;
1074 GET_RE_DEBUG_FLAGS_DECL;
1076 PERL_ARGS_ASSERT_DUMP_TRIE;
1078 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1079 (int)depth * 2 + 2,"",
1080 "Match","Base","Ofs" );
1082 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1083 SV ** const tmp = av_fetch( revcharmap, state, 0);
1085 PerlIO_printf( Perl_debug_log, "%*s",
1087 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1088 PL_colors[0], PL_colors[1],
1089 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1090 PERL_PV_ESCAPE_FIRSTCHAR
1095 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1096 (int)depth * 2 + 2,"");
1098 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1099 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1100 PerlIO_printf( Perl_debug_log, "\n");
1102 for( state = 1 ; state < trie->statecount ; state++ ) {
1103 const U32 base = trie->states[ state ].trans.base;
1105 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1107 if ( trie->states[ state ].wordnum ) {
1108 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1110 PerlIO_printf( Perl_debug_log, "%6s", "" );
1113 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1118 while( ( base + ofs < trie->uniquecharcount ) ||
1119 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1120 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1123 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1125 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1126 if ( ( base + ofs >= trie->uniquecharcount ) &&
1127 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1128 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1130 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1132 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1134 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1138 PerlIO_printf( Perl_debug_log, "]");
1141 PerlIO_printf( Perl_debug_log, "\n" );
1143 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1144 for (word=1; word <= trie->wordcount; word++) {
1145 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1146 (int)word, (int)(trie->wordinfo[word].prev),
1147 (int)(trie->wordinfo[word].len));
1149 PerlIO_printf(Perl_debug_log, "\n" );
1152 Dumps a fully constructed but uncompressed trie in list form.
1153 List tries normally only are used for construction when the number of
1154 possible chars (trie->uniquecharcount) is very high.
1155 Used for debugging make_trie().
1158 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1159 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1163 SV *sv=sv_newmortal();
1164 int colwidth= widecharmap ? 6 : 4;
1165 GET_RE_DEBUG_FLAGS_DECL;
1167 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1169 /* print out the table precompression. */
1170 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1171 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1172 "------:-----+-----------------\n" );
1174 for( state=1 ; state < next_alloc ; state ++ ) {
1177 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1178 (int)depth * 2 + 2,"", (UV)state );
1179 if ( ! trie->states[ state ].wordnum ) {
1180 PerlIO_printf( Perl_debug_log, "%5s| ","");
1182 PerlIO_printf( Perl_debug_log, "W%4x| ",
1183 trie->states[ state ].wordnum
1186 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1187 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1189 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1191 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1192 PL_colors[0], PL_colors[1],
1193 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1194 PERL_PV_ESCAPE_FIRSTCHAR
1196 TRIE_LIST_ITEM(state,charid).forid,
1197 (UV)TRIE_LIST_ITEM(state,charid).newstate
1200 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1201 (int)((depth * 2) + 14), "");
1204 PerlIO_printf( Perl_debug_log, "\n");
1209 Dumps a fully constructed but uncompressed trie in table form.
1210 This is the normal DFA style state transition table, with a few
1211 twists to facilitate compression later.
1212 Used for debugging make_trie().
1215 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1216 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1221 SV *sv=sv_newmortal();
1222 int colwidth= widecharmap ? 6 : 4;
1223 GET_RE_DEBUG_FLAGS_DECL;
1225 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1228 print out the table precompression so that we can do a visual check
1229 that they are identical.
1232 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1234 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1235 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1237 PerlIO_printf( Perl_debug_log, "%*s",
1239 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1240 PL_colors[0], PL_colors[1],
1241 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1242 PERL_PV_ESCAPE_FIRSTCHAR
1248 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1250 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1251 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1254 PerlIO_printf( Perl_debug_log, "\n" );
1256 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1258 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1259 (int)depth * 2 + 2,"",
1260 (UV)TRIE_NODENUM( state ) );
1262 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1263 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1265 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1267 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1269 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1270 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1272 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1273 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1281 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1282 startbranch: the first branch in the whole branch sequence
1283 first : start branch of sequence of branch-exact nodes.
1284 May be the same as startbranch
1285 last : Thing following the last branch.
1286 May be the same as tail.
1287 tail : item following the branch sequence
1288 count : words in the sequence
1289 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1290 depth : indent depth
1292 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1294 A trie is an N'ary tree where the branches are determined by digital
1295 decomposition of the key. IE, at the root node you look up the 1st character and
1296 follow that branch repeat until you find the end of the branches. Nodes can be
1297 marked as "accepting" meaning they represent a complete word. Eg:
1301 would convert into the following structure. Numbers represent states, letters
1302 following numbers represent valid transitions on the letter from that state, if
1303 the number is in square brackets it represents an accepting state, otherwise it
1304 will be in parenthesis.
1306 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1310 (1) +-i->(6)-+-s->[7]
1312 +-s->(3)-+-h->(4)-+-e->[5]
1314 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1316 This shows that when matching against the string 'hers' we will begin at state 1
1317 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1318 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1319 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1320 single traverse. We store a mapping from accepting to state to which word was
1321 matched, and then when we have multiple possibilities we try to complete the
1322 rest of the regex in the order in which they occured in the alternation.
1324 The only prior NFA like behaviour that would be changed by the TRIE support is
1325 the silent ignoring of duplicate alternations which are of the form:
1327 / (DUPE|DUPE) X? (?{ ... }) Y /x
1329 Thus EVAL blocks following a trie may be called a different number of times with
1330 and without the optimisation. With the optimisations dupes will be silently
1331 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1332 the following demonstrates:
1334 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1336 which prints out 'word' three times, but
1338 'words'=~/(word|word|word)(?{ print $1 })S/
1340 which doesnt print it out at all. This is due to other optimisations kicking in.
1342 Example of what happens on a structural level:
1344 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1346 1: CURLYM[1] {1,32767}(18)
1357 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1358 and should turn into:
1360 1: CURLYM[1] {1,32767}(18)
1362 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1370 Cases where tail != last would be like /(?foo|bar)baz/:
1380 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1381 and would end up looking like:
1384 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1391 d = uvuni_to_utf8_flags(d, uv, 0);
1393 is the recommended Unicode-aware way of saying
1398 #define TRIE_STORE_REVCHAR(val) \
1401 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1402 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1403 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1404 SvCUR_set(zlopp, kapow - flrbbbbb); \
1407 av_push(revcharmap, zlopp); \
1409 char ooooff = (char)val; \
1410 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1414 #define TRIE_READ_CHAR STMT_START { \
1417 /* if it is UTF then it is either already folded, or does not need folding */ \
1418 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1420 else if (folder == PL_fold_latin1) { \
1421 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1422 if ( foldlen > 0 ) { \
1423 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1429 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1430 skiplen = UNISKIP(uvc); \
1431 foldlen -= skiplen; \
1432 scan = foldbuf + skiplen; \
1435 /* raw data, will be folded later if needed */ \
1443 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1444 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1445 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1446 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1448 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1449 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1450 TRIE_LIST_CUR( state )++; \
1453 #define TRIE_LIST_NEW(state) STMT_START { \
1454 Newxz( trie->states[ state ].trans.list, \
1455 4, reg_trie_trans_le ); \
1456 TRIE_LIST_CUR( state ) = 1; \
1457 TRIE_LIST_LEN( state ) = 4; \
1460 #define TRIE_HANDLE_WORD(state) STMT_START { \
1461 U16 dupe= trie->states[ state ].wordnum; \
1462 regnode * const noper_next = regnext( noper ); \
1465 /* store the word for dumping */ \
1467 if (OP(noper) != NOTHING) \
1468 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1470 tmp = newSVpvn_utf8( "", 0, UTF ); \
1471 av_push( trie_words, tmp ); \
1475 trie->wordinfo[curword].prev = 0; \
1476 trie->wordinfo[curword].len = wordlen; \
1477 trie->wordinfo[curword].accept = state; \
1479 if ( noper_next < tail ) { \
1481 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1482 trie->jump[curword] = (U16)(noper_next - convert); \
1484 jumper = noper_next; \
1486 nextbranch= regnext(cur); \
1490 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1491 /* chain, so that when the bits of chain are later */\
1492 /* linked together, the dups appear in the chain */\
1493 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1494 trie->wordinfo[dupe].prev = curword; \
1496 /* we haven't inserted this word yet. */ \
1497 trie->states[ state ].wordnum = curword; \
1502 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1503 ( ( base + charid >= ucharcount \
1504 && base + charid < ubound \
1505 && state == trie->trans[ base - ucharcount + charid ].check \
1506 && trie->trans[ base - ucharcount + charid ].next ) \
1507 ? trie->trans[ base - ucharcount + charid ].next \
1508 : ( state==1 ? special : 0 ) \
1512 #define MADE_JUMP_TRIE 2
1513 #define MADE_EXACT_TRIE 4
1516 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1519 /* first pass, loop through and scan words */
1520 reg_trie_data *trie;
1521 HV *widecharmap = NULL;
1522 AV *revcharmap = newAV();
1524 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1529 regnode *jumper = NULL;
1530 regnode *nextbranch = NULL;
1531 regnode *convert = NULL;
1532 U32 *prev_states; /* temp array mapping each state to previous one */
1533 /* we just use folder as a flag in utf8 */
1534 const U8 * folder = NULL;
1537 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1538 AV *trie_words = NULL;
1539 /* along with revcharmap, this only used during construction but both are
1540 * useful during debugging so we store them in the struct when debugging.
1543 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1544 STRLEN trie_charcount=0;
1546 SV *re_trie_maxbuff;
1547 GET_RE_DEBUG_FLAGS_DECL;
1549 PERL_ARGS_ASSERT_MAKE_TRIE;
1551 PERL_UNUSED_ARG(depth);
1558 case EXACTFU_TRICKYFOLD:
1559 case EXACTFU: folder = PL_fold_latin1; break;
1560 case EXACTF: folder = PL_fold; break;
1561 case EXACTFL: folder = PL_fold_locale; break;
1562 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1565 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1567 trie->startstate = 1;
1568 trie->wordcount = word_count;
1569 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1570 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1572 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1573 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1574 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1577 trie_words = newAV();
1580 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1581 if (!SvIOK(re_trie_maxbuff)) {
1582 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1584 DEBUG_TRIE_COMPILE_r({
1585 PerlIO_printf( Perl_debug_log,
1586 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1587 (int)depth * 2 + 2, "",
1588 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1589 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1593 /* Find the node we are going to overwrite */
1594 if ( first == startbranch && OP( last ) != BRANCH ) {
1595 /* whole branch chain */
1598 /* branch sub-chain */
1599 convert = NEXTOPER( first );
1602 /* -- First loop and Setup --
1604 We first traverse the branches and scan each word to determine if it
1605 contains widechars, and how many unique chars there are, this is
1606 important as we have to build a table with at least as many columns as we
1609 We use an array of integers to represent the character codes 0..255
1610 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1611 native representation of the character value as the key and IV's for the
1614 *TODO* If we keep track of how many times each character is used we can
1615 remap the columns so that the table compression later on is more
1616 efficient in terms of memory by ensuring the most common value is in the
1617 middle and the least common are on the outside. IMO this would be better
1618 than a most to least common mapping as theres a decent chance the most
1619 common letter will share a node with the least common, meaning the node
1620 will not be compressible. With a middle is most common approach the worst
1621 case is when we have the least common nodes twice.
1625 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1626 regnode *noper = NEXTOPER( cur );
1627 const U8 *uc = (U8*)STRING( noper );
1628 const U8 *e = uc + STR_LEN( noper );
1630 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1632 const U8 *scan = (U8*)NULL;
1633 U32 wordlen = 0; /* required init */
1635 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1637 if (OP(noper) == NOTHING) {
1638 regnode *noper_next= regnext(noper);
1639 if (noper_next != tail && OP(noper_next) == flags) {
1641 uc= (U8*)STRING(noper);
1642 e= uc + STR_LEN(noper);
1643 trie->minlen= STR_LEN(noper);
1650 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1651 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1652 regardless of encoding */
1653 if (OP( noper ) == EXACTFU_SS) {
1654 /* false positives are ok, so just set this */
1655 TRIE_BITMAP_SET(trie,0xDF);
1658 for ( ; uc < e ; uc += len ) {
1659 TRIE_CHARCOUNT(trie)++;
1664 U8 folded= folder[ (U8) uvc ];
1665 if ( !trie->charmap[ folded ] ) {
1666 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1667 TRIE_STORE_REVCHAR( folded );
1670 if ( !trie->charmap[ uvc ] ) {
1671 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1672 TRIE_STORE_REVCHAR( uvc );
1675 /* store the codepoint in the bitmap, and its folded
1677 TRIE_BITMAP_SET(trie, uvc);
1679 /* store the folded codepoint */
1680 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1683 /* store first byte of utf8 representation of
1684 variant codepoints */
1685 if (! UNI_IS_INVARIANT(uvc)) {
1686 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1689 set_bit = 0; /* We've done our bit :-) */
1694 widecharmap = newHV();
1696 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1699 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1701 if ( !SvTRUE( *svpp ) ) {
1702 sv_setiv( *svpp, ++trie->uniquecharcount );
1703 TRIE_STORE_REVCHAR(uvc);
1707 if( cur == first ) {
1708 trie->minlen = chars;
1709 trie->maxlen = chars;
1710 } else if (chars < trie->minlen) {
1711 trie->minlen = chars;
1712 } else if (chars > trie->maxlen) {
1713 trie->maxlen = chars;
1715 if (OP( noper ) == EXACTFU_SS) {
1716 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1717 if (trie->minlen > 1)
1720 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1721 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1722 * - We assume that any such sequence might match a 2 byte string */
1723 if (trie->minlen > 2 )
1727 } /* end first pass */
1728 DEBUG_TRIE_COMPILE_r(
1729 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1730 (int)depth * 2 + 2,"",
1731 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1732 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1733 (int)trie->minlen, (int)trie->maxlen )
1737 We now know what we are dealing with in terms of unique chars and
1738 string sizes so we can calculate how much memory a naive
1739 representation using a flat table will take. If it's over a reasonable
1740 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1741 conservative but potentially much slower representation using an array
1744 At the end we convert both representations into the same compressed
1745 form that will be used in regexec.c for matching with. The latter
1746 is a form that cannot be used to construct with but has memory
1747 properties similar to the list form and access properties similar
1748 to the table form making it both suitable for fast searches and
1749 small enough that its feasable to store for the duration of a program.
1751 See the comment in the code where the compressed table is produced
1752 inplace from the flat tabe representation for an explanation of how
1753 the compression works.
1758 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1761 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1763 Second Pass -- Array Of Lists Representation
1765 Each state will be represented by a list of charid:state records
1766 (reg_trie_trans_le) the first such element holds the CUR and LEN
1767 points of the allocated array. (See defines above).
1769 We build the initial structure using the lists, and then convert
1770 it into the compressed table form which allows faster lookups
1771 (but cant be modified once converted).
1774 STRLEN transcount = 1;
1776 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1777 "%*sCompiling trie using list compiler\n",
1778 (int)depth * 2 + 2, ""));
1780 trie->states = (reg_trie_state *)
1781 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1782 sizeof(reg_trie_state) );
1786 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1788 regnode *noper = NEXTOPER( cur );
1789 U8 *uc = (U8*)STRING( noper );
1790 const U8 *e = uc + STR_LEN( noper );
1791 U32 state = 1; /* required init */
1792 U16 charid = 0; /* sanity init */
1793 U8 *scan = (U8*)NULL; /* sanity init */
1794 STRLEN foldlen = 0; /* required init */
1795 U32 wordlen = 0; /* required init */
1796 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1799 if (OP(noper) == NOTHING) {
1800 regnode *noper_next= regnext(noper);
1801 if (noper_next != tail && OP(noper_next) == flags) {
1803 uc= (U8*)STRING(noper);
1804 e= uc + STR_LEN(noper);
1808 if (OP(noper) != NOTHING) {
1809 for ( ; uc < e ; uc += len ) {
1814 charid = trie->charmap[ uvc ];
1816 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1820 charid=(U16)SvIV( *svpp );
1823 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1830 if ( !trie->states[ state ].trans.list ) {
1831 TRIE_LIST_NEW( state );
1833 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1834 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1835 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1840 newstate = next_alloc++;
1841 prev_states[newstate] = state;
1842 TRIE_LIST_PUSH( state, charid, newstate );
1847 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1851 TRIE_HANDLE_WORD(state);
1853 } /* end second pass */
1855 /* next alloc is the NEXT state to be allocated */
1856 trie->statecount = next_alloc;
1857 trie->states = (reg_trie_state *)
1858 PerlMemShared_realloc( trie->states,
1860 * sizeof(reg_trie_state) );
1862 /* and now dump it out before we compress it */
1863 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1864 revcharmap, next_alloc,
1868 trie->trans = (reg_trie_trans *)
1869 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1876 for( state=1 ; state < next_alloc ; state ++ ) {
1880 DEBUG_TRIE_COMPILE_MORE_r(
1881 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1885 if (trie->states[state].trans.list) {
1886 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1890 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1891 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1892 if ( forid < minid ) {
1894 } else if ( forid > maxid ) {
1898 if ( transcount < tp + maxid - minid + 1) {
1900 trie->trans = (reg_trie_trans *)
1901 PerlMemShared_realloc( trie->trans,
1903 * sizeof(reg_trie_trans) );
1904 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1906 base = trie->uniquecharcount + tp - minid;
1907 if ( maxid == minid ) {
1909 for ( ; zp < tp ; zp++ ) {
1910 if ( ! trie->trans[ zp ].next ) {
1911 base = trie->uniquecharcount + zp - minid;
1912 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1913 trie->trans[ zp ].check = state;
1919 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1920 trie->trans[ tp ].check = state;
1925 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1926 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1927 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1928 trie->trans[ tid ].check = state;
1930 tp += ( maxid - minid + 1 );
1932 Safefree(trie->states[ state ].trans.list);
1935 DEBUG_TRIE_COMPILE_MORE_r(
1936 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1939 trie->states[ state ].trans.base=base;
1941 trie->lasttrans = tp + 1;
1945 Second Pass -- Flat Table Representation.
1947 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1948 We know that we will need Charcount+1 trans at most to store the data
1949 (one row per char at worst case) So we preallocate both structures
1950 assuming worst case.
1952 We then construct the trie using only the .next slots of the entry
1955 We use the .check field of the first entry of the node temporarily to
1956 make compression both faster and easier by keeping track of how many non
1957 zero fields are in the node.
1959 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1962 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1963 number representing the first entry of the node, and state as a
1964 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1965 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1966 are 2 entrys per node. eg:
1974 The table is internally in the right hand, idx form. However as we also
1975 have to deal with the states array which is indexed by nodenum we have to
1976 use TRIE_NODENUM() to convert.
1979 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1980 "%*sCompiling trie using table compiler\n",
1981 (int)depth * 2 + 2, ""));
1983 trie->trans = (reg_trie_trans *)
1984 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1985 * trie->uniquecharcount + 1,
1986 sizeof(reg_trie_trans) );
1987 trie->states = (reg_trie_state *)
1988 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1989 sizeof(reg_trie_state) );
1990 next_alloc = trie->uniquecharcount + 1;
1993 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1995 regnode *noper = NEXTOPER( cur );
1996 const U8 *uc = (U8*)STRING( noper );
1997 const U8 *e = uc + STR_LEN( noper );
1999 U32 state = 1; /* required init */
2001 U16 charid = 0; /* sanity init */
2002 U32 accept_state = 0; /* sanity init */
2003 U8 *scan = (U8*)NULL; /* sanity init */
2005 STRLEN foldlen = 0; /* required init */
2006 U32 wordlen = 0; /* required init */
2008 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2010 if (OP(noper) == NOTHING) {
2011 regnode *noper_next= regnext(noper);
2012 if (noper_next != tail && OP(noper_next) == flags) {
2014 uc= (U8*)STRING(noper);
2015 e= uc + STR_LEN(noper);
2019 if ( OP(noper) != NOTHING ) {
2020 for ( ; uc < e ; uc += len ) {
2025 charid = trie->charmap[ uvc ];
2027 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2028 charid = svpp ? (U16)SvIV(*svpp) : 0;
2032 if ( !trie->trans[ state + charid ].next ) {
2033 trie->trans[ state + charid ].next = next_alloc;
2034 trie->trans[ state ].check++;
2035 prev_states[TRIE_NODENUM(next_alloc)]
2036 = TRIE_NODENUM(state);
2037 next_alloc += trie->uniquecharcount;
2039 state = trie->trans[ state + charid ].next;
2041 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2043 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2046 accept_state = TRIE_NODENUM( state );
2047 TRIE_HANDLE_WORD(accept_state);
2049 } /* end second pass */
2051 /* and now dump it out before we compress it */
2052 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2054 next_alloc, depth+1));
2058 * Inplace compress the table.*
2060 For sparse data sets the table constructed by the trie algorithm will
2061 be mostly 0/FAIL transitions or to put it another way mostly empty.
2062 (Note that leaf nodes will not contain any transitions.)
2064 This algorithm compresses the tables by eliminating most such
2065 transitions, at the cost of a modest bit of extra work during lookup:
2067 - Each states[] entry contains a .base field which indicates the
2068 index in the state[] array wheres its transition data is stored.
2070 - If .base is 0 there are no valid transitions from that node.
2072 - If .base is nonzero then charid is added to it to find an entry in
2075 -If trans[states[state].base+charid].check!=state then the
2076 transition is taken to be a 0/Fail transition. Thus if there are fail
2077 transitions at the front of the node then the .base offset will point
2078 somewhere inside the previous nodes data (or maybe even into a node
2079 even earlier), but the .check field determines if the transition is
2083 The following process inplace converts the table to the compressed
2084 table: We first do not compress the root node 1,and mark all its
2085 .check pointers as 1 and set its .base pointer as 1 as well. This
2086 allows us to do a DFA construction from the compressed table later,
2087 and ensures that any .base pointers we calculate later are greater
2090 - We set 'pos' to indicate the first entry of the second node.
2092 - We then iterate over the columns of the node, finding the first and
2093 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2094 and set the .check pointers accordingly, and advance pos
2095 appropriately and repreat for the next node. Note that when we copy
2096 the next pointers we have to convert them from the original
2097 NODEIDX form to NODENUM form as the former is not valid post
2100 - If a node has no transitions used we mark its base as 0 and do not
2101 advance the pos pointer.
2103 - If a node only has one transition we use a second pointer into the
2104 structure to fill in allocated fail transitions from other states.
2105 This pointer is independent of the main pointer and scans forward
2106 looking for null transitions that are allocated to a state. When it
2107 finds one it writes the single transition into the "hole". If the
2108 pointer doesnt find one the single transition is appended as normal.
2110 - Once compressed we can Renew/realloc the structures to release the
2113 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2114 specifically Fig 3.47 and the associated pseudocode.
2118 const U32 laststate = TRIE_NODENUM( next_alloc );
2121 trie->statecount = laststate;
2123 for ( state = 1 ; state < laststate ; state++ ) {
2125 const U32 stateidx = TRIE_NODEIDX( state );
2126 const U32 o_used = trie->trans[ stateidx ].check;
2127 U32 used = trie->trans[ stateidx ].check;
2128 trie->trans[ stateidx ].check = 0;
2130 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2131 if ( flag || trie->trans[ stateidx + charid ].next ) {
2132 if ( trie->trans[ stateidx + charid ].next ) {
2134 for ( ; zp < pos ; zp++ ) {
2135 if ( ! trie->trans[ zp ].next ) {
2139 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2140 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2141 trie->trans[ zp ].check = state;
2142 if ( ++zp > pos ) pos = zp;
2149 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2151 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2152 trie->trans[ pos ].check = state;
2157 trie->lasttrans = pos + 1;
2158 trie->states = (reg_trie_state *)
2159 PerlMemShared_realloc( trie->states, laststate
2160 * sizeof(reg_trie_state) );
2161 DEBUG_TRIE_COMPILE_MORE_r(
2162 PerlIO_printf( Perl_debug_log,
2163 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2164 (int)depth * 2 + 2,"",
2165 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2168 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2171 } /* end table compress */
2173 DEBUG_TRIE_COMPILE_MORE_r(
2174 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2175 (int)depth * 2 + 2, "",
2176 (UV)trie->statecount,
2177 (UV)trie->lasttrans)
2179 /* resize the trans array to remove unused space */
2180 trie->trans = (reg_trie_trans *)
2181 PerlMemShared_realloc( trie->trans, trie->lasttrans
2182 * sizeof(reg_trie_trans) );
2184 { /* Modify the program and insert the new TRIE node */
2185 U8 nodetype =(U8)(flags & 0xFF);
2189 regnode *optimize = NULL;
2190 #ifdef RE_TRACK_PATTERN_OFFSETS
2193 U32 mjd_nodelen = 0;
2194 #endif /* RE_TRACK_PATTERN_OFFSETS */
2195 #endif /* DEBUGGING */
2197 This means we convert either the first branch or the first Exact,
2198 depending on whether the thing following (in 'last') is a branch
2199 or not and whther first is the startbranch (ie is it a sub part of
2200 the alternation or is it the whole thing.)
2201 Assuming its a sub part we convert the EXACT otherwise we convert
2202 the whole branch sequence, including the first.
2204 /* Find the node we are going to overwrite */
2205 if ( first != startbranch || OP( last ) == BRANCH ) {
2206 /* branch sub-chain */
2207 NEXT_OFF( first ) = (U16)(last - first);
2208 #ifdef RE_TRACK_PATTERN_OFFSETS
2210 mjd_offset= Node_Offset((convert));
2211 mjd_nodelen= Node_Length((convert));
2214 /* whole branch chain */
2216 #ifdef RE_TRACK_PATTERN_OFFSETS
2219 const regnode *nop = NEXTOPER( convert );
2220 mjd_offset= Node_Offset((nop));
2221 mjd_nodelen= Node_Length((nop));
2225 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2226 (int)depth * 2 + 2, "",
2227 (UV)mjd_offset, (UV)mjd_nodelen)
2230 /* But first we check to see if there is a common prefix we can
2231 split out as an EXACT and put in front of the TRIE node. */
2232 trie->startstate= 1;
2233 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2235 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2239 const U32 base = trie->states[ state ].trans.base;
2241 if ( trie->states[state].wordnum )
2244 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2245 if ( ( base + ofs >= trie->uniquecharcount ) &&
2246 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2247 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2249 if ( ++count > 1 ) {
2250 SV **tmp = av_fetch( revcharmap, ofs, 0);
2251 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2252 if ( state == 1 ) break;
2254 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2256 PerlIO_printf(Perl_debug_log,
2257 "%*sNew Start State=%"UVuf" Class: [",
2258 (int)depth * 2 + 2, "",
2261 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2262 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2264 TRIE_BITMAP_SET(trie,*ch);
2266 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2268 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2272 TRIE_BITMAP_SET(trie,*ch);
2274 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2275 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2281 SV **tmp = av_fetch( revcharmap, idx, 0);
2283 char *ch = SvPV( *tmp, len );
2285 SV *sv=sv_newmortal();
2286 PerlIO_printf( Perl_debug_log,
2287 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2288 (int)depth * 2 + 2, "",
2290 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2291 PL_colors[0], PL_colors[1],
2292 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2293 PERL_PV_ESCAPE_FIRSTCHAR
2298 OP( convert ) = nodetype;
2299 str=STRING(convert);
2302 STR_LEN(convert) += len;
2308 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2313 trie->prefixlen = (state-1);
2315 regnode *n = convert+NODE_SZ_STR(convert);
2316 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2317 trie->startstate = state;
2318 trie->minlen -= (state - 1);
2319 trie->maxlen -= (state - 1);
2321 /* At least the UNICOS C compiler choked on this
2322 * being argument to DEBUG_r(), so let's just have
2325 #ifdef PERL_EXT_RE_BUILD
2331 regnode *fix = convert;
2332 U32 word = trie->wordcount;
2334 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2335 while( ++fix < n ) {
2336 Set_Node_Offset_Length(fix, 0, 0);
2339 SV ** const tmp = av_fetch( trie_words, word, 0 );
2341 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2342 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2344 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2352 NEXT_OFF(convert) = (U16)(tail - convert);
2353 DEBUG_r(optimize= n);
2359 if ( trie->maxlen ) {
2360 NEXT_OFF( convert ) = (U16)(tail - convert);
2361 ARG_SET( convert, data_slot );
2362 /* Store the offset to the first unabsorbed branch in
2363 jump[0], which is otherwise unused by the jump logic.
2364 We use this when dumping a trie and during optimisation. */
2366 trie->jump[0] = (U16)(nextbranch - convert);
2368 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2369 * and there is a bitmap
2370 * and the first "jump target" node we found leaves enough room
2371 * then convert the TRIE node into a TRIEC node, with the bitmap
2372 * embedded inline in the opcode - this is hypothetically faster.
2374 if ( !trie->states[trie->startstate].wordnum
2376 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2378 OP( convert ) = TRIEC;
2379 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2380 PerlMemShared_free(trie->bitmap);
2383 OP( convert ) = TRIE;
2385 /* store the type in the flags */
2386 convert->flags = nodetype;
2390 + regarglen[ OP( convert ) ];
2392 /* XXX We really should free up the resource in trie now,
2393 as we won't use them - (which resources?) dmq */
2395 /* needed for dumping*/
2396 DEBUG_r(if (optimize) {
2397 regnode *opt = convert;
2399 while ( ++opt < optimize) {
2400 Set_Node_Offset_Length(opt,0,0);
2403 Try to clean up some of the debris left after the
2406 while( optimize < jumper ) {
2407 mjd_nodelen += Node_Length((optimize));
2408 OP( optimize ) = OPTIMIZED;
2409 Set_Node_Offset_Length(optimize,0,0);
2412 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2414 } /* end node insert */
2416 /* Finish populating the prev field of the wordinfo array. Walk back
2417 * from each accept state until we find another accept state, and if
2418 * so, point the first word's .prev field at the second word. If the
2419 * second already has a .prev field set, stop now. This will be the
2420 * case either if we've already processed that word's accept state,
2421 * or that state had multiple words, and the overspill words were
2422 * already linked up earlier.
2429 for (word=1; word <= trie->wordcount; word++) {
2431 if (trie->wordinfo[word].prev)
2433 state = trie->wordinfo[word].accept;
2435 state = prev_states[state];
2438 prev = trie->states[state].wordnum;
2442 trie->wordinfo[word].prev = prev;
2444 Safefree(prev_states);
2448 /* and now dump out the compressed format */
2449 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2451 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2453 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2454 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2456 SvREFCNT_dec_NN(revcharmap);
2460 : trie->startstate>1
2466 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2468 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2470 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2471 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2474 We find the fail state for each state in the trie, this state is the longest proper
2475 suffix of the current state's 'word' that is also a proper prefix of another word in our
2476 trie. State 1 represents the word '' and is thus the default fail state. This allows
2477 the DFA not to have to restart after its tried and failed a word at a given point, it
2478 simply continues as though it had been matching the other word in the first place.
2480 'abcdgu'=~/abcdefg|cdgu/
2481 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2482 fail, which would bring us to the state representing 'd' in the second word where we would
2483 try 'g' and succeed, proceeding to match 'cdgu'.
2485 /* add a fail transition */
2486 const U32 trie_offset = ARG(source);
2487 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2489 const U32 ucharcount = trie->uniquecharcount;
2490 const U32 numstates = trie->statecount;
2491 const U32 ubound = trie->lasttrans + ucharcount;
2495 U32 base = trie->states[ 1 ].trans.base;
2498 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2499 GET_RE_DEBUG_FLAGS_DECL;
2501 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2503 PERL_UNUSED_ARG(depth);
2507 ARG_SET( stclass, data_slot );
2508 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2509 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2510 aho->trie=trie_offset;
2511 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2512 Copy( trie->states, aho->states, numstates, reg_trie_state );
2513 Newxz( q, numstates, U32);
2514 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2517 /* initialize fail[0..1] to be 1 so that we always have
2518 a valid final fail state */
2519 fail[ 0 ] = fail[ 1 ] = 1;
2521 for ( charid = 0; charid < ucharcount ; charid++ ) {
2522 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2524 q[ q_write ] = newstate;
2525 /* set to point at the root */
2526 fail[ q[ q_write++ ] ]=1;
2529 while ( q_read < q_write) {
2530 const U32 cur = q[ q_read++ % numstates ];
2531 base = trie->states[ cur ].trans.base;
2533 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2534 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2536 U32 fail_state = cur;
2539 fail_state = fail[ fail_state ];
2540 fail_base = aho->states[ fail_state ].trans.base;
2541 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2543 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2544 fail[ ch_state ] = fail_state;
2545 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2547 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2549 q[ q_write++ % numstates] = ch_state;
2553 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2554 when we fail in state 1, this allows us to use the
2555 charclass scan to find a valid start char. This is based on the principle
2556 that theres a good chance the string being searched contains lots of stuff
2557 that cant be a start char.
2559 fail[ 0 ] = fail[ 1 ] = 0;
2560 DEBUG_TRIE_COMPILE_r({
2561 PerlIO_printf(Perl_debug_log,
2562 "%*sStclass Failtable (%"UVuf" states): 0",
2563 (int)(depth * 2), "", (UV)numstates
2565 for( q_read=1; q_read<numstates; q_read++ ) {
2566 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2568 PerlIO_printf(Perl_debug_log, "\n");
2571 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2576 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2577 * These need to be revisited when a newer toolchain becomes available.
2579 #if defined(__sparc64__) && defined(__GNUC__)
2580 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2581 # undef SPARC64_GCC_WORKAROUND
2582 # define SPARC64_GCC_WORKAROUND 1
2586 #define DEBUG_PEEP(str,scan,depth) \
2587 DEBUG_OPTIMISE_r({if (scan){ \
2588 SV * const mysv=sv_newmortal(); \
2589 regnode *Next = regnext(scan); \
2590 regprop(RExC_rx, mysv, scan); \
2591 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2592 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2593 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2597 /* The below joins as many adjacent EXACTish nodes as possible into a single
2598 * one. The regop may be changed if the node(s) contain certain sequences that
2599 * require special handling. The joining is only done if:
2600 * 1) there is room in the current conglomerated node to entirely contain the
2602 * 2) they are the exact same node type
2604 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2605 * these get optimized out
2607 * If a node is to match under /i (folded), the number of characters it matches
2608 * can be different than its character length if it contains a multi-character
2609 * fold. *min_subtract is set to the total delta of the input nodes.
2611 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2612 * and contains LATIN SMALL LETTER SHARP S
2614 * This is as good a place as any to discuss the design of handling these
2615 * multi-character fold sequences. It's been wrong in Perl for a very long
2616 * time. There are three code points in Unicode whose multi-character folds
2617 * were long ago discovered to mess things up. The previous designs for
2618 * dealing with these involved assigning a special node for them. This
2619 * approach doesn't work, as evidenced by this example:
2620 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2621 * Both these fold to "sss", but if the pattern is parsed to create a node that
2622 * would match just the \xDF, it won't be able to handle the case where a
2623 * successful match would have to cross the node's boundary. The new approach
2624 * that hopefully generally solves the problem generates an EXACTFU_SS node
2627 * It turns out that there are problems with all multi-character folds, and not
2628 * just these three. Now the code is general, for all such cases, but the
2629 * three still have some special handling. The approach taken is:
2630 * 1) This routine examines each EXACTFish node that could contain multi-
2631 * character fold sequences. It returns in *min_subtract how much to
2632 * subtract from the the actual length of the string to get a real minimum
2633 * match length; it is 0 if there are no multi-char folds. This delta is
2634 * used by the caller to adjust the min length of the match, and the delta
2635 * between min and max, so that the optimizer doesn't reject these
2636 * possibilities based on size constraints.
2637 * 2) Certain of these sequences require special handling by the trie code,
2638 * so, if found, this code changes the joined node type to special ops:
2639 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2640 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2641 * is used for an EXACTFU node that contains at least one "ss" sequence in
2642 * it. For non-UTF-8 patterns and strings, this is the only case where
2643 * there is a possible fold length change. That means that a regular
2644 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2645 * with length changes, and so can be processed faster. regexec.c takes
2646 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2647 * pre-folded by regcomp.c. This saves effort in regex matching.
2648 * However, the pre-folding isn't done for non-UTF8 patterns because the
2649 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2650 * down by forcing the pattern into UTF8 unless necessary. Also what
2651 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2652 * possibilities for the non-UTF8 patterns are quite simple, except for
2653 * the sharp s. All the ones that don't involve a UTF-8 target string are
2654 * members of a fold-pair, and arrays are set up for all of them so that
2655 * the other member of the pair can be found quickly. Code elsewhere in
2656 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2657 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2658 * described in the next item.
2659 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2660 * 'ss' or not is not knowable at compile time. It will match iff the
2661 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2662 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2663 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
2664 * described in item 3). An assumption that the optimizer part of
2665 * regexec.c (probably unwittingly) makes is that a character in the
2666 * pattern corresponds to at most a single character in the target string.
2667 * (And I do mean character, and not byte here, unlike other parts of the
2668 * documentation that have never been updated to account for multibyte
2669 * Unicode.) This assumption is wrong only in this case, as all other
2670 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2671 * virtue of having this file pre-fold UTF-8 patterns. I'm
2672 * reluctant to try to change this assumption, so instead the code punts.
2673 * This routine examines EXACTF nodes for the sharp s, and returns a
2674 * boolean indicating whether or not the node is an EXACTF node that
2675 * contains a sharp s. When it is true, the caller sets a flag that later
2676 * causes the optimizer in this file to not set values for the floating
2677 * and fixed string lengths, and thus avoids the optimizer code in
2678 * regexec.c that makes the invalid assumption. Thus, there is no
2679 * optimization based on string lengths for EXACTF nodes that contain the
2680 * sharp s. This only happens for /id rules (which means the pattern
2684 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2685 if (PL_regkind[OP(scan)] == EXACT) \
2686 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2689 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) {
2690 /* Merge several consecutive EXACTish nodes into one. */
2691 regnode *n = regnext(scan);
2693 regnode *next = scan + NODE_SZ_STR(scan);
2697 regnode *stop = scan;
2698 GET_RE_DEBUG_FLAGS_DECL;
2700 PERL_UNUSED_ARG(depth);
2703 PERL_ARGS_ASSERT_JOIN_EXACT;
2704 #ifndef EXPERIMENTAL_INPLACESCAN
2705 PERL_UNUSED_ARG(flags);
2706 PERL_UNUSED_ARG(val);
2708 DEBUG_PEEP("join",scan,depth);
2710 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2711 * EXACT ones that are mergeable to the current one. */
2713 && (PL_regkind[OP(n)] == NOTHING
2714 || (stringok && OP(n) == OP(scan)))
2716 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2719 if (OP(n) == TAIL || n > next)
2721 if (PL_regkind[OP(n)] == NOTHING) {
2722 DEBUG_PEEP("skip:",n,depth);
2723 NEXT_OFF(scan) += NEXT_OFF(n);
2724 next = n + NODE_STEP_REGNODE;
2731 else if (stringok) {
2732 const unsigned int oldl = STR_LEN(scan);
2733 regnode * const nnext = regnext(n);
2735 /* XXX I (khw) kind of doubt that this works on platforms where
2736 * U8_MAX is above 255 because of lots of other assumptions */
2737 /* Don't join if the sum can't fit into a single node */
2738 if (oldl + STR_LEN(n) > U8_MAX)
2741 DEBUG_PEEP("merg",n,depth);
2744 NEXT_OFF(scan) += NEXT_OFF(n);
2745 STR_LEN(scan) += STR_LEN(n);
2746 next = n + NODE_SZ_STR(n);
2747 /* Now we can overwrite *n : */
2748 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2756 #ifdef EXPERIMENTAL_INPLACESCAN
2757 if (flags && !NEXT_OFF(n)) {
2758 DEBUG_PEEP("atch", val, depth);
2759 if (reg_off_by_arg[OP(n)]) {
2760 ARG_SET(n, val - n);
2763 NEXT_OFF(n) = val - n;
2771 *has_exactf_sharp_s = FALSE;
2773 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2774 * can now analyze for sequences of problematic code points. (Prior to
2775 * this final joining, sequences could have been split over boundaries, and
2776 * hence missed). The sequences only happen in folding, hence for any
2777 * non-EXACT EXACTish node */
2778 if (OP(scan) != EXACT) {
2779 const U8 * const s0 = (U8*) STRING(scan);
2781 const U8 * const s_end = s0 + STR_LEN(scan);
2783 /* One pass is made over the node's string looking for all the
2784 * possibilities. to avoid some tests in the loop, there are two main
2785 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2789 /* Examine the string for a multi-character fold sequence. UTF-8
2790 * patterns have all characters pre-folded by the time this code is
2792 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2793 length sequence we are looking for is 2 */
2796 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2797 if (! len) { /* Not a multi-char fold: get next char */
2802 /* Nodes with 'ss' require special handling, except for EXACTFL
2803 * and EXACTFA for which there is no multi-char fold to this */
2804 if (len == 2 && *s == 's' && *(s+1) == 's'
2805 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2808 OP(scan) = EXACTFU_SS;
2811 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2812 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2813 COMBINING_DIAERESIS_UTF8
2814 COMBINING_ACUTE_ACCENT_UTF8,
2816 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2817 COMBINING_DIAERESIS_UTF8
2818 COMBINING_ACUTE_ACCENT_UTF8,
2823 /* These two folds require special handling by trie's, so
2824 * change the node type to indicate this. If EXACTFA and
2825 * EXACTFL were ever to be handled by trie's, this would
2826 * have to be changed. If this node has already been
2827 * changed to EXACTFU_SS in this loop, leave it as is. (I
2828 * (khw) think it doesn't matter in regexec.c for UTF
2829 * patterns, but no need to change it */
2830 if (OP(scan) == EXACTFU) {
2831 OP(scan) = EXACTFU_TRICKYFOLD;
2835 else { /* Here is a generic multi-char fold. */
2836 const U8* multi_end = s + len;
2838 /* Count how many characters in it. In the case of /l and
2839 * /aa, no folds which contain ASCII code points are
2840 * allowed, so check for those, and skip if found. (In
2841 * EXACTFL, no folds are allowed to any Latin1 code point,
2842 * not just ASCII. But there aren't any of these
2843 * currently, nor ever likely, so don't take the time to
2844 * test for them. The code that generates the
2845 * is_MULTI_foo() macros croaks should one actually get put
2846 * into Unicode .) */
2847 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2848 count = utf8_length(s, multi_end);
2852 while (s < multi_end) {
2855 goto next_iteration;
2865 /* The delta is how long the sequence is minus 1 (1 is how long
2866 * the character that folds to the sequence is) */
2867 *min_subtract += count - 1;
2871 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2873 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2874 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2875 * nodes can't have multi-char folds to this range (and there are
2876 * no existing ones in the upper latin1 range). In the EXACTF
2877 * case we look also for the sharp s, which can be in the final
2878 * position. Otherwise we can stop looking 1 byte earlier because
2879 * have to find at least two characters for a multi-fold */
2880 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2882 /* The below is perhaps overboard, but this allows us to save a
2883 * test each time through the loop at the expense of a mask. This
2884 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2885 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2886 * are 64. This uses an exclusive 'or' to find that bit and then
2887 * inverts it to form a mask, with just a single 0, in the bit
2888 * position where 'S' and 's' differ. */
2889 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2890 const U8 s_masked = 's' & S_or_s_mask;
2893 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2894 if (! len) { /* Not a multi-char fold. */
2895 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2897 *has_exactf_sharp_s = TRUE;
2904 && ((*s & S_or_s_mask) == s_masked)
2905 && ((*(s+1) & S_or_s_mask) == s_masked))
2908 /* EXACTF nodes need to know that the minimum length
2909 * changed so that a sharp s in the string can match this
2910 * ss in the pattern, but they remain EXACTF nodes, as they
2911 * won't match this unless the target string is is UTF-8,
2912 * which we don't know until runtime */
2913 if (OP(scan) != EXACTF) {
2914 OP(scan) = EXACTFU_SS;
2918 *min_subtract += len - 1;
2925 /* Allow dumping but overwriting the collection of skipped
2926 * ops and/or strings with fake optimized ops */
2927 n = scan + NODE_SZ_STR(scan);
2935 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2939 /* REx optimizer. Converts nodes into quicker variants "in place".
2940 Finds fixed substrings. */
2942 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2943 to the position after last scanned or to NULL. */
2945 #define INIT_AND_WITHP \
2946 assert(!and_withp); \
2947 Newx(and_withp,1,struct regnode_charclass_class); \
2948 SAVEFREEPV(and_withp)
2950 /* this is a chain of data about sub patterns we are processing that
2951 need to be handled separately/specially in study_chunk. Its so
2952 we can simulate recursion without losing state. */
2954 typedef struct scan_frame {
2955 regnode *last; /* last node to process in this frame */
2956 regnode *next; /* next node to process when last is reached */
2957 struct scan_frame *prev; /*previous frame*/
2958 I32 stop; /* what stopparen do we use */
2962 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2965 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2966 I32 *minlenp, I32 *deltap,
2971 struct regnode_charclass_class *and_withp,
2972 U32 flags, U32 depth)
2973 /* scanp: Start here (read-write). */
2974 /* deltap: Write maxlen-minlen here. */
2975 /* last: Stop before this one. */
2976 /* data: string data about the pattern */
2977 /* stopparen: treat close N as END */
2978 /* recursed: which subroutines have we recursed into */
2979 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
2982 I32 min = 0; /* There must be at least this number of characters to match */
2984 regnode *scan = *scanp, *next;
2986 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2987 int is_inf_internal = 0; /* The studied chunk is infinite */
2988 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2989 scan_data_t data_fake;
2990 SV *re_trie_maxbuff = NULL;
2991 regnode *first_non_open = scan;
2992 I32 stopmin = I32_MAX;
2993 scan_frame *frame = NULL;
2994 GET_RE_DEBUG_FLAGS_DECL;
2996 PERL_ARGS_ASSERT_STUDY_CHUNK;
2999 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3003 while (first_non_open && OP(first_non_open) == OPEN)
3004 first_non_open=regnext(first_non_open);
3009 while ( scan && OP(scan) != END && scan < last ){
3010 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3011 node length to get a real minimum (because
3012 the folded version may be shorter) */
3013 bool has_exactf_sharp_s = FALSE;
3014 /* Peephole optimizer: */
3015 DEBUG_STUDYDATA("Peep:", data,depth);
3016 DEBUG_PEEP("Peep",scan,depth);
3018 /* Its not clear to khw or hv why this is done here, and not in the
3019 * clauses that deal with EXACT nodes. khw's guess is that it's
3020 * because of a previous design */
3021 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3023 /* Follow the next-chain of the current node and optimize
3024 away all the NOTHINGs from it. */
3025 if (OP(scan) != CURLYX) {
3026 const int max = (reg_off_by_arg[OP(scan)]
3028 /* I32 may be smaller than U16 on CRAYs! */
3029 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3030 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3034 /* Skip NOTHING and LONGJMP. */
3035 while ((n = regnext(n))
3036 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3037 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3038 && off + noff < max)
3040 if (reg_off_by_arg[OP(scan)])
3043 NEXT_OFF(scan) = off;
3048 /* The principal pseudo-switch. Cannot be a switch, since we
3049 look into several different things. */
3050 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3051 || OP(scan) == IFTHEN) {
3052 next = regnext(scan);
3054 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3056 if (OP(next) == code || code == IFTHEN) {
3057 /* NOTE - There is similar code to this block below for handling
3058 TRIE nodes on a re-study. If you change stuff here check there
3060 I32 max1 = 0, min1 = I32_MAX, num = 0;
3061 struct regnode_charclass_class accum;
3062 regnode * const startbranch=scan;
3064 if (flags & SCF_DO_SUBSTR)
3065 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3066 if (flags & SCF_DO_STCLASS)
3067 cl_init_zero(pRExC_state, &accum);
3069 while (OP(scan) == code) {
3070 I32 deltanext, minnext, f = 0, fake;
3071 struct regnode_charclass_class this_class;
3074 data_fake.flags = 0;
3076 data_fake.whilem_c = data->whilem_c;
3077 data_fake.last_closep = data->last_closep;
3080 data_fake.last_closep = &fake;
3082 data_fake.pos_delta = delta;
3083 next = regnext(scan);
3084 scan = NEXTOPER(scan);
3086 scan = NEXTOPER(scan);
3087 if (flags & SCF_DO_STCLASS) {
3088 cl_init(pRExC_state, &this_class);
3089 data_fake.start_class = &this_class;
3090 f = SCF_DO_STCLASS_AND;
3092 if (flags & SCF_WHILEM_VISITED_POS)
3093 f |= SCF_WHILEM_VISITED_POS;
3095 /* we suppose the run is continuous, last=next...*/
3096 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3098 stopparen, recursed, NULL, f,depth+1);
3101 if (max1 < minnext + deltanext)
3102 max1 = minnext + deltanext;
3103 if (deltanext == I32_MAX)
3104 is_inf = is_inf_internal = 1;
3106 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3108 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3109 if ( stopmin > minnext)
3110 stopmin = min + min1;
3111 flags &= ~SCF_DO_SUBSTR;
3113 data->flags |= SCF_SEEN_ACCEPT;
3116 if (data_fake.flags & SF_HAS_EVAL)
3117 data->flags |= SF_HAS_EVAL;
3118 data->whilem_c = data_fake.whilem_c;
3120 if (flags & SCF_DO_STCLASS)
3121 cl_or(pRExC_state, &accum, &this_class);
3123 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3125 if (flags & SCF_DO_SUBSTR) {
3126 data->pos_min += min1;
3127 data->pos_delta += max1 - min1;
3128 if (max1 != min1 || is_inf)
3129 data->longest = &(data->longest_float);
3132 delta += max1 - min1;
3133 if (flags & SCF_DO_STCLASS_OR) {
3134 cl_or(pRExC_state, data->start_class, &accum);
3136 cl_and(data->start_class, and_withp);
3137 flags &= ~SCF_DO_STCLASS;
3140 else if (flags & SCF_DO_STCLASS_AND) {
3142 cl_and(data->start_class, &accum);
3143 flags &= ~SCF_DO_STCLASS;
3146 /* Switch to OR mode: cache the old value of
3147 * data->start_class */
3149 StructCopy(data->start_class, and_withp,
3150 struct regnode_charclass_class);
3151 flags &= ~SCF_DO_STCLASS_AND;
3152 StructCopy(&accum, data->start_class,
3153 struct regnode_charclass_class);
3154 flags |= SCF_DO_STCLASS_OR;
3155 SET_SSC_EOS(data->start_class);
3159 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3162 Assuming this was/is a branch we are dealing with: 'scan' now
3163 points at the item that follows the branch sequence, whatever
3164 it is. We now start at the beginning of the sequence and look
3171 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3173 If we can find such a subsequence we need to turn the first
3174 element into a trie and then add the subsequent branch exact
3175 strings to the trie.
3179 1. patterns where the whole set of branches can be converted.
3181 2. patterns where only a subset can be converted.
3183 In case 1 we can replace the whole set with a single regop
3184 for the trie. In case 2 we need to keep the start and end
3187 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3188 becomes BRANCH TRIE; BRANCH X;
3190 There is an additional case, that being where there is a
3191 common prefix, which gets split out into an EXACT like node
3192 preceding the TRIE node.
3194 If x(1..n)==tail then we can do a simple trie, if not we make
3195 a "jump" trie, such that when we match the appropriate word
3196 we "jump" to the appropriate tail node. Essentially we turn
3197 a nested if into a case structure of sorts.
3202 if (!re_trie_maxbuff) {
3203 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3204 if (!SvIOK(re_trie_maxbuff))
3205 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3207 if ( SvIV(re_trie_maxbuff)>=0 ) {
3209 regnode *first = (regnode *)NULL;
3210 regnode *last = (regnode *)NULL;
3211 regnode *tail = scan;
3216 SV * const mysv = sv_newmortal(); /* for dumping */
3218 /* var tail is used because there may be a TAIL
3219 regop in the way. Ie, the exacts will point to the
3220 thing following the TAIL, but the last branch will
3221 point at the TAIL. So we advance tail. If we
3222 have nested (?:) we may have to move through several
3226 while ( OP( tail ) == TAIL ) {
3227 /* this is the TAIL generated by (?:) */
3228 tail = regnext( tail );
3232 DEBUG_TRIE_COMPILE_r({
3233 regprop(RExC_rx, mysv, tail );
3234 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3235 (int)depth * 2 + 2, "",
3236 "Looking for TRIE'able sequences. Tail node is: ",
3237 SvPV_nolen_const( mysv )
3243 Step through the branches
3244 cur represents each branch,
3245 noper is the first thing to be matched as part of that branch
3246 noper_next is the regnext() of that node.
3248 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3249 via a "jump trie" but we also support building with NOJUMPTRIE,
3250 which restricts the trie logic to structures like /FOO|BAR/.
3252 If noper is a trieable nodetype then the branch is a possible optimization
3253 target. If we are building under NOJUMPTRIE then we require that noper_next
3254 is the same as scan (our current position in the regex program).
3256 Once we have two or more consecutive such branches we can create a
3257 trie of the EXACT's contents and stitch it in place into the program.
3259 If the sequence represents all of the branches in the alternation we
3260 replace the entire thing with a single TRIE node.
3262 Otherwise when it is a subsequence we need to stitch it in place and
3263 replace only the relevant branches. This means the first branch has
3264 to remain as it is used by the alternation logic, and its next pointer,
3265 and needs to be repointed at the item on the branch chain following
3266 the last branch we have optimized away.
3268 This could be either a BRANCH, in which case the subsequence is internal,
3269 or it could be the item following the branch sequence in which case the
3270 subsequence is at the end (which does not necessarily mean the first node
3271 is the start of the alternation).
3273 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3276 ----------------+-----------
3280 EXACTFU_SS | EXACTFU
3281 EXACTFU_TRICKYFOLD | EXACTFU
3286 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3287 ( EXACT == (X) ) ? EXACT : \
3288 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3291 /* dont use tail as the end marker for this traverse */
3292 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3293 regnode * const noper = NEXTOPER( cur );
3294 U8 noper_type = OP( noper );
3295 U8 noper_trietype = TRIE_TYPE( noper_type );
3296 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3297 regnode * const noper_next = regnext( noper );
3298 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3299 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3302 DEBUG_TRIE_COMPILE_r({
3303 regprop(RExC_rx, mysv, cur);
3304 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3305 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3307 regprop(RExC_rx, mysv, noper);
3308 PerlIO_printf( Perl_debug_log, " -> %s",
3309 SvPV_nolen_const(mysv));
3312 regprop(RExC_rx, mysv, noper_next );
3313 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3314 SvPV_nolen_const(mysv));
3316 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3317 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3318 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3322 /* Is noper a trieable nodetype that can be merged with the
3323 * current trie (if there is one)? */
3327 ( noper_trietype == NOTHING)
3328 || ( trietype == NOTHING )
3329 || ( trietype == noper_trietype )
3332 && noper_next == tail
3336 /* Handle mergable triable node
3337 * Either we are the first node in a new trieable sequence,
3338 * in which case we do some bookkeeping, otherwise we update
3339 * the end pointer. */
3342 if ( noper_trietype == NOTHING ) {
3343 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3344 regnode * const noper_next = regnext( noper );
3345 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3346 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3349 if ( noper_next_trietype ) {
3350 trietype = noper_next_trietype;
3351 } else if (noper_next_type) {
3352 /* a NOTHING regop is 1 regop wide. We need at least two
3353 * for a trie so we can't merge this in */
3357 trietype = noper_trietype;
3360 if ( trietype == NOTHING )
3361 trietype = noper_trietype;
3366 } /* end handle mergable triable node */
3368 /* handle unmergable node -
3369 * noper may either be a triable node which can not be tried
3370 * together with the current trie, or a non triable node */
3372 /* If last is set and trietype is not NOTHING then we have found
3373 * at least two triable branch sequences in a row of a similar
3374 * trietype so we can turn them into a trie. If/when we
3375 * allow NOTHING to start a trie sequence this condition will be
3376 * required, and it isn't expensive so we leave it in for now. */
3377 if ( trietype && trietype != NOTHING )
3378 make_trie( pRExC_state,
3379 startbranch, first, cur, tail, count,
3380 trietype, depth+1 );
3381 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3385 && noper_next == tail
3388 /* noper is triable, so we can start a new trie sequence */
3391 trietype = noper_trietype;
3393 /* if we already saw a first but the current node is not triable then we have
3394 * to reset the first information. */
3399 } /* end handle unmergable node */
3400 } /* loop over branches */
3401 DEBUG_TRIE_COMPILE_r({
3402 regprop(RExC_rx, mysv, cur);
3403 PerlIO_printf( Perl_debug_log,
3404 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3405 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3408 if ( last && trietype ) {
3409 if ( trietype != NOTHING ) {
3410 /* the last branch of the sequence was part of a trie,
3411 * so we have to construct it here outside of the loop
3413 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3414 #ifdef TRIE_STUDY_OPT
3415 if ( ((made == MADE_EXACT_TRIE &&
3416 startbranch == first)
3417 || ( first_non_open == first )) &&
3419 flags |= SCF_TRIE_RESTUDY;
3420 if ( startbranch == first
3423 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3428 /* at this point we know whatever we have is a NOTHING sequence/branch
3429 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3431 if ( startbranch == first ) {
3433 /* the entire thing is a NOTHING sequence, something like this:
3434 * (?:|) So we can turn it into a plain NOTHING op. */
3435 DEBUG_TRIE_COMPILE_r({
3436 regprop(RExC_rx, mysv, cur);
3437 PerlIO_printf( Perl_debug_log,
3438 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3439 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3442 OP(startbranch)= NOTHING;
3443 NEXT_OFF(startbranch)= tail - startbranch;
3444 for ( opt= startbranch + 1; opt < tail ; opt++ )
3448 } /* end if ( last) */
3449 } /* TRIE_MAXBUF is non zero */
3454 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3455 scan = NEXTOPER(NEXTOPER(scan));
3456 } else /* single branch is optimized. */
3457 scan = NEXTOPER(scan);
3459 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3460 scan_frame *newframe = NULL;
3465 if (OP(scan) != SUSPEND) {
3466 /* set the pointer */
3467 if (OP(scan) == GOSUB) {
3469 RExC_recurse[ARG2L(scan)] = scan;
3470 start = RExC_open_parens[paren-1];
3471 end = RExC_close_parens[paren-1];
3474 start = RExC_rxi->program + 1;
3478 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3479 SAVEFREEPV(recursed);
3481 if (!PAREN_TEST(recursed,paren+1)) {
3482 PAREN_SET(recursed,paren+1);
3483 Newx(newframe,1,scan_frame);
3485 if (flags & SCF_DO_SUBSTR) {
3486 SCAN_COMMIT(pRExC_state,data,minlenp);
3487 data->longest = &(data->longest_float);
3489 is_inf = is_inf_internal = 1;
3490 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3491 cl_anything(pRExC_state, data->start_class);
3492 flags &= ~SCF_DO_STCLASS;
3495 Newx(newframe,1,scan_frame);
3498 end = regnext(scan);
3503 SAVEFREEPV(newframe);
3504 newframe->next = regnext(scan);
3505 newframe->last = last;
3506 newframe->stop = stopparen;
3507 newframe->prev = frame;
3517 else if (OP(scan) == EXACT) {
3518 I32 l = STR_LEN(scan);
3521 const U8 * const s = (U8*)STRING(scan);
3522 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3523 l = utf8_length(s, s + l);
3525 uc = *((U8*)STRING(scan));
3528 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3529 /* The code below prefers earlier match for fixed
3530 offset, later match for variable offset. */
3531 if (data->last_end == -1) { /* Update the start info. */
3532 data->last_start_min = data->pos_min;
3533 data->last_start_max = is_inf
3534 ? I32_MAX : data->pos_min + data->pos_delta;
3536 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3538 SvUTF8_on(data->last_found);
3540 SV * const sv = data->last_found;
3541 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3542 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3543 if (mg && mg->mg_len >= 0)
3544 mg->mg_len += utf8_length((U8*)STRING(scan),
3545 (U8*)STRING(scan)+STR_LEN(scan));
3547 data->last_end = data->pos_min + l;
3548 data->pos_min += l; /* As in the first entry. */
3549 data->flags &= ~SF_BEFORE_EOL;
3551 if (flags & SCF_DO_STCLASS_AND) {
3552 /* Check whether it is compatible with what we know already! */
3556 /* If compatible, we or it in below. It is compatible if is
3557 * in the bitmp and either 1) its bit or its fold is set, or 2)
3558 * it's for a locale. Even if there isn't unicode semantics
3559 * here, at runtime there may be because of matching against a
3560 * utf8 string, so accept a possible false positive for
3561 * latin1-range folds */
3563 (!(data->start_class->flags & ANYOF_LOCALE)
3564 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3565 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3566 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3571 ANYOF_CLASS_ZERO(data->start_class);
3572 ANYOF_BITMAP_ZERO(data->start_class);
3574 ANYOF_BITMAP_SET(data->start_class, uc);
3575 else if (uc >= 0x100) {
3578 /* Some Unicode code points fold to the Latin1 range; as
3579 * XXX temporary code, instead of figuring out if this is
3580 * one, just assume it is and set all the start class bits
3581 * that could be some such above 255 code point's fold
3582 * which will generate fals positives. As the code
3583 * elsewhere that does compute the fold settles down, it
3584 * can be extracted out and re-used here */
3585 for (i = 0; i < 256; i++){
3586 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3587 ANYOF_BITMAP_SET(data->start_class, i);
3591 CLEAR_SSC_EOS(data->start_class);
3593 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3595 else if (flags & SCF_DO_STCLASS_OR) {
3596 /* false positive possible if the class is case-folded */
3598 ANYOF_BITMAP_SET(data->start_class, uc);
3600 data->start_class->flags |= ANYOF_UNICODE_ALL;
3601 CLEAR_SSC_EOS(data->start_class);
3602 cl_and(data->start_class, and_withp);
3604 flags &= ~SCF_DO_STCLASS;
3606 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3607 I32 l = STR_LEN(scan);
3608 UV uc = *((U8*)STRING(scan));
3610 /* Search for fixed substrings supports EXACT only. */
3611 if (flags & SCF_DO_SUBSTR) {
3613 SCAN_COMMIT(pRExC_state, data, minlenp);
3616 const U8 * const s = (U8 *)STRING(scan);
3617 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3618 l = utf8_length(s, s + l);
3620 if (has_exactf_sharp_s) {
3621 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3623 min += l - min_subtract;
3625 delta += min_subtract;
3626 if (flags & SCF_DO_SUBSTR) {
3627 data->pos_min += l - min_subtract;
3628 if (data->pos_min < 0) {
3631 data->pos_delta += min_subtract;
3633 data->longest = &(data->longest_float);
3636 if (flags & SCF_DO_STCLASS_AND) {
3637 /* Check whether it is compatible with what we know already! */
3640 (!(data->start_class->flags & ANYOF_LOCALE)
3641 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3642 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3646 ANYOF_CLASS_ZERO(data->start_class);
3647 ANYOF_BITMAP_ZERO(data->start_class);
3649 ANYOF_BITMAP_SET(data->start_class, uc);
3650 CLEAR_SSC_EOS(data->start_class);
3651 if (OP(scan) == EXACTFL) {
3652 /* XXX This set is probably no longer necessary, and
3653 * probably wrong as LOCALE now is on in the initial
3655 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3659 /* Also set the other member of the fold pair. In case
3660 * that unicode semantics is called for at runtime, use
3661 * the full latin1 fold. (Can't do this for locale,
3662 * because not known until runtime) */
3663 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3665 /* All other (EXACTFL handled above) folds except under
3666 * /iaa that include s, S, and sharp_s also may include
3668 if (OP(scan) != EXACTFA) {
3669 if (uc == 's' || uc == 'S') {
3670 ANYOF_BITMAP_SET(data->start_class,
3671 LATIN_SMALL_LETTER_SHARP_S);
3673 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3674 ANYOF_BITMAP_SET(data->start_class, 's');
3675 ANYOF_BITMAP_SET(data->start_class, 'S');
3680 else if (uc >= 0x100) {
3682 for (i = 0; i < 256; i++){
3683 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3684 ANYOF_BITMAP_SET(data->start_class, i);
3689 else if (flags & SCF_DO_STCLASS_OR) {
3690 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3691 /* false positive possible if the class is case-folded.
3692 Assume that the locale settings are the same... */
3694 ANYOF_BITMAP_SET(data->start_class, uc);
3695 if (OP(scan) != EXACTFL) {
3697 /* And set the other member of the fold pair, but
3698 * can't do that in locale because not known until
3700 ANYOF_BITMAP_SET(data->start_class,
3701 PL_fold_latin1[uc]);
3703 /* All folds except under /iaa that include s, S,
3704 * and sharp_s also may include the others */
3705 if (OP(scan) != EXACTFA) {
3706 if (uc == 's' || uc == 'S') {
3707 ANYOF_BITMAP_SET(data->start_class,
3708 LATIN_SMALL_LETTER_SHARP_S);
3710 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3711 ANYOF_BITMAP_SET(data->start_class, 's');
3712 ANYOF_BITMAP_SET(data->start_class, 'S');
3717 CLEAR_SSC_EOS(data->start_class);
3719 cl_and(data->start_class, and_withp);
3721 flags &= ~SCF_DO_STCLASS;
3723 else if (REGNODE_VARIES(OP(scan))) {
3724 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3725 I32 f = flags, pos_before = 0;
3726 regnode * const oscan = scan;
3727 struct regnode_charclass_class this_class;
3728 struct regnode_charclass_class *oclass = NULL;
3729 I32 next_is_eval = 0;
3731 switch (PL_regkind[OP(scan)]) {
3732 case WHILEM: /* End of (?:...)* . */
3733 scan = NEXTOPER(scan);
3736 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3737 next = NEXTOPER(scan);
3738 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3740 maxcount = REG_INFTY;
3741 next = regnext(scan);
3742 scan = NEXTOPER(scan);
3746 if (flags & SCF_DO_SUBSTR)
3751 if (flags & SCF_DO_STCLASS) {
3753 maxcount = REG_INFTY;
3754 next = regnext(scan);
3755 scan = NEXTOPER(scan);
3758 is_inf = is_inf_internal = 1;
3759 scan = regnext(scan);
3760 if (flags & SCF_DO_SUBSTR) {
3761 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3762 data->longest = &(data->longest_float);
3764 goto optimize_curly_tail;
3766 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3767 && (scan->flags == stopparen))
3772 mincount = ARG1(scan);
3773 maxcount = ARG2(scan);
3775 next = regnext(scan);
3776 if (OP(scan) == CURLYX) {
3777 I32 lp = (data ? *(data->last_closep) : 0);
3778 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3780 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3781 next_is_eval = (OP(scan) == EVAL);
3783 if (flags & SCF_DO_SUBSTR) {
3784 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3785 pos_before = data->pos_min;
3789 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3791 data->flags |= SF_IS_INF;
3793 if (flags & SCF_DO_STCLASS) {
3794 cl_init(pRExC_state, &this_class);
3795 oclass = data->start_class;
3796 data->start_class = &this_class;
3797 f |= SCF_DO_STCLASS_AND;
3798 f &= ~SCF_DO_STCLASS_OR;
3800 /* Exclude from super-linear cache processing any {n,m}
3801 regops for which the combination of input pos and regex
3802 pos is not enough information to determine if a match
3805 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3806 regex pos at the \s*, the prospects for a match depend not
3807 only on the input position but also on how many (bar\s*)
3808 repeats into the {4,8} we are. */
3809 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3810 f &= ~SCF_WHILEM_VISITED_POS;
3812 /* This will finish on WHILEM, setting scan, or on NULL: */
3813 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3814 last, data, stopparen, recursed, NULL,
3816 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3818 if (flags & SCF_DO_STCLASS)
3819 data->start_class = oclass;
3820 if (mincount == 0 || minnext == 0) {
3821 if (flags & SCF_DO_STCLASS_OR) {
3822 cl_or(pRExC_state, data->start_class, &this_class);
3824 else if (flags & SCF_DO_STCLASS_AND) {
3825 /* Switch to OR mode: cache the old value of
3826 * data->start_class */
https://perl5.git.perl.org / perl5.git / blob