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, FALSE)))
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 vFAIL4(m,a1,a2,a3) STMT_START { \
534 SAVEFREESV(RExC_rx_sv); \
535 Simple_vFAIL4(m, a1, a2, a3); \
538 /* m is not necessarily a "literal string", in this macro */
539 #define reg_warn_non_literal_string(loc, m) STMT_START { \
540 const IV offset = loc - RExC_precomp; \
541 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
542 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
545 #define ckWARNreg(loc,m) STMT_START { \
546 const IV offset = loc - RExC_precomp; \
547 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
548 (int)offset, RExC_precomp, RExC_precomp + offset); \
551 #define ckWARNregdep(loc,m) STMT_START { \
552 const IV offset = loc - RExC_precomp; \
553 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
555 (int)offset, RExC_precomp, RExC_precomp + offset); \
558 #define ckWARN2regdep(loc,m, a1) STMT_START { \
559 const IV offset = loc - RExC_precomp; \
560 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
562 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
565 #define ckWARN2reg(loc, m, a1) STMT_START { \
566 const IV offset = loc - RExC_precomp; \
567 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
568 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
571 #define vWARN3(loc, m, a1, a2) STMT_START { \
572 const IV offset = loc - RExC_precomp; \
573 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
574 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
577 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
578 const IV offset = loc - RExC_precomp; \
579 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
580 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
583 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
584 const IV offset = loc - RExC_precomp; \
585 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
586 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
589 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
590 const IV offset = loc - RExC_precomp; \
591 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
592 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
595 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
596 const IV offset = loc - RExC_precomp; \
597 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
598 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
602 /* Allow for side effects in s */
603 #define REGC(c,s) STMT_START { \
604 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
607 /* Macros for recording node offsets. 20001227 mjd@plover.com
608 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
609 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
610 * Element 0 holds the number n.
611 * Position is 1 indexed.
613 #ifndef RE_TRACK_PATTERN_OFFSETS
614 #define Set_Node_Offset_To_R(node,byte)
615 #define Set_Node_Offset(node,byte)
616 #define Set_Cur_Node_Offset
617 #define Set_Node_Length_To_R(node,len)
618 #define Set_Node_Length(node,len)
619 #define Set_Node_Cur_Length(node)
620 #define Node_Offset(n)
621 #define Node_Length(n)
622 #define Set_Node_Offset_Length(node,offset,len)
623 #define ProgLen(ri) ri->u.proglen
624 #define SetProgLen(ri,x) ri->u.proglen = x
626 #define ProgLen(ri) ri->u.offsets[0]
627 #define SetProgLen(ri,x) ri->u.offsets[0] = x
628 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
630 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
631 __LINE__, (int)(node), (int)(byte))); \
633 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
635 RExC_offsets[2*(node)-1] = (byte); \
640 #define Set_Node_Offset(node,byte) \
641 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
642 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
644 #define Set_Node_Length_To_R(node,len) STMT_START { \
646 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
647 __LINE__, (int)(node), (int)(len))); \
649 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
651 RExC_offsets[2*(node)] = (len); \
656 #define Set_Node_Length(node,len) \
657 Set_Node_Length_To_R((node)-RExC_emit_start, len)
658 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
659 #define Set_Node_Cur_Length(node) \
660 Set_Node_Length(node, RExC_parse - parse_start)
662 /* Get offsets and lengths */
663 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
664 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
666 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
667 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
668 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
672 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
673 #define EXPERIMENTAL_INPLACESCAN
674 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
676 #define DEBUG_STUDYDATA(str,data,depth) \
677 DEBUG_OPTIMISE_MORE_r(if(data){ \
678 PerlIO_printf(Perl_debug_log, \
679 "%*s" str "Pos:%"IVdf"/%"IVdf \
680 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
681 (int)(depth)*2, "", \
682 (IV)((data)->pos_min), \
683 (IV)((data)->pos_delta), \
684 (UV)((data)->flags), \
685 (IV)((data)->whilem_c), \
686 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
687 is_inf ? "INF " : "" \
689 if ((data)->last_found) \
690 PerlIO_printf(Perl_debug_log, \
691 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
692 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
693 SvPVX_const((data)->last_found), \
694 (IV)((data)->last_end), \
695 (IV)((data)->last_start_min), \
696 (IV)((data)->last_start_max), \
697 ((data)->longest && \
698 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
699 SvPVX_const((data)->longest_fixed), \
700 (IV)((data)->offset_fixed), \
701 ((data)->longest && \
702 (data)->longest==&((data)->longest_float)) ? "*" : "", \
703 SvPVX_const((data)->longest_float), \
704 (IV)((data)->offset_float_min), \
705 (IV)((data)->offset_float_max) \
707 PerlIO_printf(Perl_debug_log,"\n"); \
710 /* Mark that we cannot extend a found fixed substring at this point.
711 Update the longest found anchored substring and the longest found
712 floating substrings if needed. */
715 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
717 const STRLEN l = CHR_SVLEN(data->last_found);
718 const STRLEN old_l = CHR_SVLEN(*data->longest);
719 GET_RE_DEBUG_FLAGS_DECL;
721 PERL_ARGS_ASSERT_SCAN_COMMIT;
723 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
724 SvSetMagicSV(*data->longest, data->last_found);
725 if (*data->longest == data->longest_fixed) {
726 data->offset_fixed = l ? data->last_start_min : data->pos_min;
727 if (data->flags & SF_BEFORE_EOL)
729 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
731 data->flags &= ~SF_FIX_BEFORE_EOL;
732 data->minlen_fixed=minlenp;
733 data->lookbehind_fixed=0;
735 else { /* *data->longest == data->longest_float */
736 data->offset_float_min = l ? data->last_start_min : data->pos_min;
737 data->offset_float_max = (l
738 ? data->last_start_max
739 : data->pos_min + data->pos_delta);
740 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
741 data->offset_float_max = I32_MAX;
742 if (data->flags & SF_BEFORE_EOL)
744 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
746 data->flags &= ~SF_FL_BEFORE_EOL;
747 data->minlen_float=minlenp;
748 data->lookbehind_float=0;
751 SvCUR_set(data->last_found, 0);
753 SV * const sv = data->last_found;
754 if (SvUTF8(sv) && SvMAGICAL(sv)) {
755 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
761 data->flags &= ~SF_BEFORE_EOL;
762 DEBUG_STUDYDATA("commit: ",data,0);
765 /* These macros set, clear and test whether the synthetic start class ('ssc',
766 * given by the parameter) matches an empty string (EOS). This uses the
767 * 'next_off' field in the node, to save a bit in the flags field. The ssc
768 * stands alone, so there is never a next_off, so this field is otherwise
769 * unused. The EOS information is used only for compilation, but theoretically
770 * it could be passed on to the execution code. This could be used to store
771 * more than one bit of information, but only this one is currently used. */
772 #define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
773 #define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
774 #define TEST_SSC_EOS(node) cBOOL((node)->next_off)
776 /* Can match anything (initialization) */
778 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
780 PERL_ARGS_ASSERT_CL_ANYTHING;
782 ANYOF_BITMAP_SETALL(cl);
783 cl->flags = ANYOF_UNICODE_ALL;
786 /* If any portion of the regex is to operate under locale rules,
787 * initialization includes it. The reason this isn't done for all regexes
788 * is that the optimizer was written under the assumption that locale was
789 * all-or-nothing. Given the complexity and lack of documentation in the
790 * optimizer, and that there are inadequate test cases for locale, so many
791 * parts of it may not work properly, it is safest to avoid locale unless
793 if (RExC_contains_locale) {
794 ANYOF_CLASS_SETALL(cl); /* /l uses class */
795 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
798 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
802 /* Can match anything (initialization) */
804 S_cl_is_anything(const struct regnode_charclass_class *cl)
808 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
810 for (value = 0; value < ANYOF_MAX; value += 2)
811 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
813 if (!(cl->flags & ANYOF_UNICODE_ALL))
815 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
820 /* Can match anything (initialization) */
822 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
824 PERL_ARGS_ASSERT_CL_INIT;
826 Zero(cl, 1, struct regnode_charclass_class);
828 cl_anything(pRExC_state, cl);
829 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
832 /* These two functions currently do the exact same thing */
833 #define cl_init_zero S_cl_init
835 /* 'AND' a given class with another one. Can create false positives. 'cl'
836 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
837 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
839 S_cl_and(struct regnode_charclass_class *cl,
840 const struct regnode_charclass_class *and_with)
842 PERL_ARGS_ASSERT_CL_AND;
844 assert(PL_regkind[and_with->type] == ANYOF);
846 /* I (khw) am not sure all these restrictions are necessary XXX */
847 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
848 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
849 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
850 && !(and_with->flags & ANYOF_LOC_FOLD)
851 && !(cl->flags & ANYOF_LOC_FOLD)) {
854 if (and_with->flags & ANYOF_INVERT)
855 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
856 cl->bitmap[i] &= ~and_with->bitmap[i];
858 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
859 cl->bitmap[i] &= and_with->bitmap[i];
860 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
862 if (and_with->flags & ANYOF_INVERT) {
864 /* Here, the and'ed node is inverted. Get the AND of the flags that
865 * aren't affected by the inversion. Those that are affected are
866 * handled individually below */
867 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
868 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
869 cl->flags |= affected_flags;
871 /* We currently don't know how to deal with things that aren't in the
872 * bitmap, but we know that the intersection is no greater than what
873 * is already in cl, so let there be false positives that get sorted
874 * out after the synthetic start class succeeds, and the node is
875 * matched for real. */
877 /* The inversion of these two flags indicate that the resulting
878 * intersection doesn't have them */
879 if (and_with->flags & ANYOF_UNICODE_ALL) {
880 cl->flags &= ~ANYOF_UNICODE_ALL;
882 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
883 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
886 else { /* and'd node is not inverted */
887 U8 outside_bitmap_but_not_utf8; /* Temp variable */
889 if (! ANYOF_NONBITMAP(and_with)) {
891 /* Here 'and_with' doesn't match anything outside the bitmap
892 * (except possibly ANYOF_UNICODE_ALL), which means the
893 * intersection can't either, except for ANYOF_UNICODE_ALL, in
894 * which case we don't know what the intersection is, but it's no
895 * greater than what cl already has, so can just leave it alone,
896 * with possible false positives */
897 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
898 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
899 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
902 else if (! ANYOF_NONBITMAP(cl)) {
904 /* Here, 'and_with' does match something outside the bitmap, and cl
905 * doesn't have a list of things to match outside the bitmap. If
906 * cl can match all code points above 255, the intersection will
907 * be those above-255 code points that 'and_with' matches. If cl
908 * can't match all Unicode code points, it means that it can't
909 * match anything outside the bitmap (since the 'if' that got us
910 * into this block tested for that), so we leave the bitmap empty.
912 if (cl->flags & ANYOF_UNICODE_ALL) {
913 ARG_SET(cl, ARG(and_with));
915 /* and_with's ARG may match things that don't require UTF8.
916 * And now cl's will too, in spite of this being an 'and'. See
917 * the comments below about the kludge */
918 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
922 /* Here, both 'and_with' and cl match something outside the
923 * bitmap. Currently we do not do the intersection, so just match
924 * whatever cl had at the beginning. */
928 /* Take the intersection of the two sets of flags. However, the
929 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
930 * kludge around the fact that this flag is not treated like the others
931 * which are initialized in cl_anything(). The way the optimizer works
932 * is that the synthetic start class (SSC) is initialized to match
933 * anything, and then the first time a real node is encountered, its
934 * values are AND'd with the SSC's with the result being the values of
935 * the real node. However, there are paths through the optimizer where
936 * the AND never gets called, so those initialized bits are set
937 * inappropriately, which is not usually a big deal, as they just cause
938 * false positives in the SSC, which will just mean a probably
939 * imperceptible slow down in execution. However this bit has a
940 * higher false positive consequence in that it can cause utf8.pm,
941 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
942 * bigger slowdown and also causes significant extra memory to be used.
943 * In order to prevent this, the code now takes a different tack. The
944 * bit isn't set unless some part of the regular expression needs it,
945 * but once set it won't get cleared. This means that these extra
946 * modules won't get loaded unless there was some path through the
947 * pattern that would have required them anyway, and so any false
948 * positives that occur by not ANDing them out when they could be
949 * aren't as severe as they would be if we treated this bit like all
951 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
952 & ANYOF_NONBITMAP_NON_UTF8;
953 cl->flags &= and_with->flags;
954 cl->flags |= outside_bitmap_but_not_utf8;
958 /* 'OR' a given class with another one. Can create false positives. 'cl'
959 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
960 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
962 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
964 PERL_ARGS_ASSERT_CL_OR;
966 if (or_with->flags & ANYOF_INVERT) {
968 /* Here, the or'd node is to be inverted. This means we take the
969 * complement of everything not in the bitmap, but currently we don't
970 * know what that is, so give up and match anything */
971 if (ANYOF_NONBITMAP(or_with)) {
972 cl_anything(pRExC_state, cl);
975 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
976 * <= (B1 | !B2) | (CL1 | !CL2)
977 * which is wasteful if CL2 is small, but we ignore CL2:
978 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
979 * XXXX Can we handle case-fold? Unclear:
980 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
981 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
983 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
984 && !(or_with->flags & ANYOF_LOC_FOLD)
985 && !(cl->flags & ANYOF_LOC_FOLD) ) {
988 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
989 cl->bitmap[i] |= ~or_with->bitmap[i];
990 } /* XXXX: logic is complicated otherwise */
992 cl_anything(pRExC_state, cl);
995 /* And, we can just take the union of the flags that aren't affected
996 * by the inversion */
997 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
999 /* For the remaining flags:
1000 ANYOF_UNICODE_ALL and inverted means to not match anything above
1001 255, which means that the union with cl should just be
1002 what cl has in it, so can ignore this flag
1003 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
1004 is 127-255 to match them, but then invert that, so the
1005 union with cl should just be what cl has in it, so can
1008 } else { /* 'or_with' is not inverted */
1009 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1010 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1011 && (!(or_with->flags & ANYOF_LOC_FOLD)
1012 || (cl->flags & ANYOF_LOC_FOLD)) ) {
1015 /* OR char bitmap and class bitmap separately */
1016 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1017 cl->bitmap[i] |= or_with->bitmap[i];
1018 ANYOF_CLASS_OR(or_with, cl);
1020 else { /* XXXX: logic is complicated, leave it along for a moment. */
1021 cl_anything(pRExC_state, cl);
1024 if (ANYOF_NONBITMAP(or_with)) {
1026 /* Use the added node's outside-the-bit-map match if there isn't a
1027 * conflict. If there is a conflict (both nodes match something
1028 * outside the bitmap, but what they match outside is not the same
1029 * pointer, and hence not easily compared until XXX we extend
1030 * inversion lists this far), give up and allow the start class to
1031 * match everything outside the bitmap. If that stuff is all above
1032 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1033 if (! ANYOF_NONBITMAP(cl)) {
1034 ARG_SET(cl, ARG(or_with));
1036 else if (ARG(cl) != ARG(or_with)) {
1038 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1039 cl_anything(pRExC_state, cl);
1042 cl->flags |= ANYOF_UNICODE_ALL;
1047 /* Take the union */
1048 cl->flags |= or_with->flags;
1052 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1053 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1054 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1055 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1060 dump_trie(trie,widecharmap,revcharmap)
1061 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1062 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1064 These routines dump out a trie in a somewhat readable format.
1065 The _interim_ variants are used for debugging the interim
1066 tables that are used to generate the final compressed
1067 representation which is what dump_trie expects.
1069 Part of the reason for their existence is to provide a form
1070 of documentation as to how the different representations function.
1075 Dumps the final compressed table form of the trie to Perl_debug_log.
1076 Used for debugging make_trie().
1080 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1081 AV *revcharmap, U32 depth)
1084 SV *sv=sv_newmortal();
1085 int colwidth= widecharmap ? 6 : 4;
1087 GET_RE_DEBUG_FLAGS_DECL;
1089 PERL_ARGS_ASSERT_DUMP_TRIE;
1091 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1092 (int)depth * 2 + 2,"",
1093 "Match","Base","Ofs" );
1095 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1096 SV ** const tmp = av_fetch( revcharmap, state, 0);
1098 PerlIO_printf( Perl_debug_log, "%*s",
1100 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1101 PL_colors[0], PL_colors[1],
1102 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1103 PERL_PV_ESCAPE_FIRSTCHAR
1108 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1109 (int)depth * 2 + 2,"");
1111 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1112 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1113 PerlIO_printf( Perl_debug_log, "\n");
1115 for( state = 1 ; state < trie->statecount ; state++ ) {
1116 const U32 base = trie->states[ state ].trans.base;
1118 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1120 if ( trie->states[ state ].wordnum ) {
1121 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1123 PerlIO_printf( Perl_debug_log, "%6s", "" );
1126 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1131 while( ( base + ofs < trie->uniquecharcount ) ||
1132 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1133 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1136 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1138 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1139 if ( ( base + ofs >= trie->uniquecharcount ) &&
1140 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1141 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1143 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1145 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1147 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1151 PerlIO_printf( Perl_debug_log, "]");
1154 PerlIO_printf( Perl_debug_log, "\n" );
1156 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1157 for (word=1; word <= trie->wordcount; word++) {
1158 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1159 (int)word, (int)(trie->wordinfo[word].prev),
1160 (int)(trie->wordinfo[word].len));
1162 PerlIO_printf(Perl_debug_log, "\n" );
1165 Dumps a fully constructed but uncompressed trie in list form.
1166 List tries normally only are used for construction when the number of
1167 possible chars (trie->uniquecharcount) is very high.
1168 Used for debugging make_trie().
1171 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1172 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1176 SV *sv=sv_newmortal();
1177 int colwidth= widecharmap ? 6 : 4;
1178 GET_RE_DEBUG_FLAGS_DECL;
1180 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1182 /* print out the table precompression. */
1183 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1184 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1185 "------:-----+-----------------\n" );
1187 for( state=1 ; state < next_alloc ; state ++ ) {
1190 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1191 (int)depth * 2 + 2,"", (UV)state );
1192 if ( ! trie->states[ state ].wordnum ) {
1193 PerlIO_printf( Perl_debug_log, "%5s| ","");
1195 PerlIO_printf( Perl_debug_log, "W%4x| ",
1196 trie->states[ state ].wordnum
1199 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1200 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1202 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1204 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1205 PL_colors[0], PL_colors[1],
1206 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1207 PERL_PV_ESCAPE_FIRSTCHAR
1209 TRIE_LIST_ITEM(state,charid).forid,
1210 (UV)TRIE_LIST_ITEM(state,charid).newstate
1213 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1214 (int)((depth * 2) + 14), "");
1217 PerlIO_printf( Perl_debug_log, "\n");
1222 Dumps a fully constructed but uncompressed trie in table form.
1223 This is the normal DFA style state transition table, with a few
1224 twists to facilitate compression later.
1225 Used for debugging make_trie().
1228 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1229 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1234 SV *sv=sv_newmortal();
1235 int colwidth= widecharmap ? 6 : 4;
1236 GET_RE_DEBUG_FLAGS_DECL;
1238 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1241 print out the table precompression so that we can do a visual check
1242 that they are identical.
1245 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1247 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1248 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1250 PerlIO_printf( Perl_debug_log, "%*s",
1252 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1253 PL_colors[0], PL_colors[1],
1254 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1255 PERL_PV_ESCAPE_FIRSTCHAR
1261 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1263 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1264 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1267 PerlIO_printf( Perl_debug_log, "\n" );
1269 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1271 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1272 (int)depth * 2 + 2,"",
1273 (UV)TRIE_NODENUM( state ) );
1275 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1276 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1278 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1280 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1282 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1283 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1285 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1286 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1294 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1295 startbranch: the first branch in the whole branch sequence
1296 first : start branch of sequence of branch-exact nodes.
1297 May be the same as startbranch
1298 last : Thing following the last branch.
1299 May be the same as tail.
1300 tail : item following the branch sequence
1301 count : words in the sequence
1302 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1303 depth : indent depth
1305 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1307 A trie is an N'ary tree where the branches are determined by digital
1308 decomposition of the key. IE, at the root node you look up the 1st character and
1309 follow that branch repeat until you find the end of the branches. Nodes can be
1310 marked as "accepting" meaning they represent a complete word. Eg:
1314 would convert into the following structure. Numbers represent states, letters
1315 following numbers represent valid transitions on the letter from that state, if
1316 the number is in square brackets it represents an accepting state, otherwise it
1317 will be in parenthesis.
1319 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1323 (1) +-i->(6)-+-s->[7]
1325 +-s->(3)-+-h->(4)-+-e->[5]
1327 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1329 This shows that when matching against the string 'hers' we will begin at state 1
1330 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1331 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1332 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1333 single traverse. We store a mapping from accepting to state to which word was
1334 matched, and then when we have multiple possibilities we try to complete the
1335 rest of the regex in the order in which they occured in the alternation.
1337 The only prior NFA like behaviour that would be changed by the TRIE support is
1338 the silent ignoring of duplicate alternations which are of the form:
1340 / (DUPE|DUPE) X? (?{ ... }) Y /x
1342 Thus EVAL blocks following a trie may be called a different number of times with
1343 and without the optimisation. With the optimisations dupes will be silently
1344 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1345 the following demonstrates:
1347 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1349 which prints out 'word' three times, but
1351 'words'=~/(word|word|word)(?{ print $1 })S/
1353 which doesnt print it out at all. This is due to other optimisations kicking in.
1355 Example of what happens on a structural level:
1357 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1359 1: CURLYM[1] {1,32767}(18)
1370 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1371 and should turn into:
1373 1: CURLYM[1] {1,32767}(18)
1375 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1383 Cases where tail != last would be like /(?foo|bar)baz/:
1393 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1394 and would end up looking like:
1397 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1404 d = uvuni_to_utf8_flags(d, uv, 0);
1406 is the recommended Unicode-aware way of saying
1411 #define TRIE_STORE_REVCHAR(val) \
1414 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1415 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1416 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1417 SvCUR_set(zlopp, kapow - flrbbbbb); \
1420 av_push(revcharmap, zlopp); \
1422 char ooooff = (char)val; \
1423 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1427 #define TRIE_READ_CHAR STMT_START { \
1430 /* if it is UTF then it is either already folded, or does not need folding */ \
1431 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1433 else if (folder == PL_fold_latin1) { \
1434 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1435 if ( foldlen > 0 ) { \
1436 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1442 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1443 skiplen = UNISKIP(uvc); \
1444 foldlen -= skiplen; \
1445 scan = foldbuf + skiplen; \
1448 /* raw data, will be folded later if needed */ \
1456 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1457 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1458 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1459 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1461 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1462 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1463 TRIE_LIST_CUR( state )++; \
1466 #define TRIE_LIST_NEW(state) STMT_START { \
1467 Newxz( trie->states[ state ].trans.list, \
1468 4, reg_trie_trans_le ); \
1469 TRIE_LIST_CUR( state ) = 1; \
1470 TRIE_LIST_LEN( state ) = 4; \
1473 #define TRIE_HANDLE_WORD(state) STMT_START { \
1474 U16 dupe= trie->states[ state ].wordnum; \
1475 regnode * const noper_next = regnext( noper ); \
1478 /* store the word for dumping */ \
1480 if (OP(noper) != NOTHING) \
1481 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1483 tmp = newSVpvn_utf8( "", 0, UTF ); \
1484 av_push( trie_words, tmp ); \
1488 trie->wordinfo[curword].prev = 0; \
1489 trie->wordinfo[curword].len = wordlen; \
1490 trie->wordinfo[curword].accept = state; \
1492 if ( noper_next < tail ) { \
1494 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1495 trie->jump[curword] = (U16)(noper_next - convert); \
1497 jumper = noper_next; \
1499 nextbranch= regnext(cur); \
1503 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1504 /* chain, so that when the bits of chain are later */\
1505 /* linked together, the dups appear in the chain */\
1506 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1507 trie->wordinfo[dupe].prev = curword; \
1509 /* we haven't inserted this word yet. */ \
1510 trie->states[ state ].wordnum = curword; \
1515 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1516 ( ( base + charid >= ucharcount \
1517 && base + charid < ubound \
1518 && state == trie->trans[ base - ucharcount + charid ].check \
1519 && trie->trans[ base - ucharcount + charid ].next ) \
1520 ? trie->trans[ base - ucharcount + charid ].next \
1521 : ( state==1 ? special : 0 ) \
1525 #define MADE_JUMP_TRIE 2
1526 #define MADE_EXACT_TRIE 4
1529 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1532 /* first pass, loop through and scan words */
1533 reg_trie_data *trie;
1534 HV *widecharmap = NULL;
1535 AV *revcharmap = newAV();
1537 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1542 regnode *jumper = NULL;
1543 regnode *nextbranch = NULL;
1544 regnode *convert = NULL;
1545 U32 *prev_states; /* temp array mapping each state to previous one */
1546 /* we just use folder as a flag in utf8 */
1547 const U8 * folder = NULL;
1550 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1551 AV *trie_words = NULL;
1552 /* along with revcharmap, this only used during construction but both are
1553 * useful during debugging so we store them in the struct when debugging.
1556 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1557 STRLEN trie_charcount=0;
1559 SV *re_trie_maxbuff;
1560 GET_RE_DEBUG_FLAGS_DECL;
1562 PERL_ARGS_ASSERT_MAKE_TRIE;
1564 PERL_UNUSED_ARG(depth);
1571 case EXACTFU_TRICKYFOLD:
1572 case EXACTFU: folder = PL_fold_latin1; break;
1573 case EXACTF: folder = PL_fold; break;
1574 case EXACTFL: folder = PL_fold_locale; break;
1575 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1578 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1580 trie->startstate = 1;
1581 trie->wordcount = word_count;
1582 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1583 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1585 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1586 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1587 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1590 trie_words = newAV();
1593 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1594 if (!SvIOK(re_trie_maxbuff)) {
1595 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1597 DEBUG_TRIE_COMPILE_r({
1598 PerlIO_printf( Perl_debug_log,
1599 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1600 (int)depth * 2 + 2, "",
1601 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1602 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1606 /* Find the node we are going to overwrite */
1607 if ( first == startbranch && OP( last ) != BRANCH ) {
1608 /* whole branch chain */
1611 /* branch sub-chain */
1612 convert = NEXTOPER( first );
1615 /* -- First loop and Setup --
1617 We first traverse the branches and scan each word to determine if it
1618 contains widechars, and how many unique chars there are, this is
1619 important as we have to build a table with at least as many columns as we
1622 We use an array of integers to represent the character codes 0..255
1623 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1624 native representation of the character value as the key and IV's for the
1627 *TODO* If we keep track of how many times each character is used we can
1628 remap the columns so that the table compression later on is more
1629 efficient in terms of memory by ensuring the most common value is in the
1630 middle and the least common are on the outside. IMO this would be better
1631 than a most to least common mapping as theres a decent chance the most
1632 common letter will share a node with the least common, meaning the node
1633 will not be compressible. With a middle is most common approach the worst
1634 case is when we have the least common nodes twice.
1638 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1639 regnode *noper = NEXTOPER( cur );
1640 const U8 *uc = (U8*)STRING( noper );
1641 const U8 *e = uc + STR_LEN( noper );
1643 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1645 const U8 *scan = (U8*)NULL;
1646 U32 wordlen = 0; /* required init */
1648 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1650 if (OP(noper) == NOTHING) {
1651 regnode *noper_next= regnext(noper);
1652 if (noper_next != tail && OP(noper_next) == flags) {
1654 uc= (U8*)STRING(noper);
1655 e= uc + STR_LEN(noper);
1656 trie->minlen= STR_LEN(noper);
1663 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1664 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1665 regardless of encoding */
1666 if (OP( noper ) == EXACTFU_SS) {
1667 /* false positives are ok, so just set this */
1668 TRIE_BITMAP_SET(trie,0xDF);
1671 for ( ; uc < e ; uc += len ) {
1672 TRIE_CHARCOUNT(trie)++;
1677 U8 folded= folder[ (U8) uvc ];
1678 if ( !trie->charmap[ folded ] ) {
1679 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1680 TRIE_STORE_REVCHAR( folded );
1683 if ( !trie->charmap[ uvc ] ) {
1684 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1685 TRIE_STORE_REVCHAR( uvc );
1688 /* store the codepoint in the bitmap, and its folded
1690 TRIE_BITMAP_SET(trie, uvc);
1692 /* store the folded codepoint */
1693 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1696 /* store first byte of utf8 representation of
1697 variant codepoints */
1698 if (! UNI_IS_INVARIANT(uvc)) {
1699 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1702 set_bit = 0; /* We've done our bit :-) */
1707 widecharmap = newHV();
1709 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1712 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1714 if ( !SvTRUE( *svpp ) ) {
1715 sv_setiv( *svpp, ++trie->uniquecharcount );
1716 TRIE_STORE_REVCHAR(uvc);
1720 if( cur == first ) {
1721 trie->minlen = chars;
1722 trie->maxlen = chars;
1723 } else if (chars < trie->minlen) {
1724 trie->minlen = chars;
1725 } else if (chars > trie->maxlen) {
1726 trie->maxlen = chars;
1728 if (OP( noper ) == EXACTFU_SS) {
1729 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1730 if (trie->minlen > 1)
1733 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1734 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1735 * - We assume that any such sequence might match a 2 byte string */
1736 if (trie->minlen > 2 )
1740 } /* end first pass */
1741 DEBUG_TRIE_COMPILE_r(
1742 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1743 (int)depth * 2 + 2,"",
1744 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1745 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1746 (int)trie->minlen, (int)trie->maxlen )
1750 We now know what we are dealing with in terms of unique chars and
1751 string sizes so we can calculate how much memory a naive
1752 representation using a flat table will take. If it's over a reasonable
1753 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1754 conservative but potentially much slower representation using an array
1757 At the end we convert both representations into the same compressed
1758 form that will be used in regexec.c for matching with. The latter
1759 is a form that cannot be used to construct with but has memory
1760 properties similar to the list form and access properties similar
1761 to the table form making it both suitable for fast searches and
1762 small enough that its feasable to store for the duration of a program.
1764 See the comment in the code where the compressed table is produced
1765 inplace from the flat tabe representation for an explanation of how
1766 the compression works.
1771 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1774 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1776 Second Pass -- Array Of Lists Representation
1778 Each state will be represented by a list of charid:state records
1779 (reg_trie_trans_le) the first such element holds the CUR and LEN
1780 points of the allocated array. (See defines above).
1782 We build the initial structure using the lists, and then convert
1783 it into the compressed table form which allows faster lookups
1784 (but cant be modified once converted).
1787 STRLEN transcount = 1;
1789 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1790 "%*sCompiling trie using list compiler\n",
1791 (int)depth * 2 + 2, ""));
1793 trie->states = (reg_trie_state *)
1794 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1795 sizeof(reg_trie_state) );
1799 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1801 regnode *noper = NEXTOPER( cur );
1802 U8 *uc = (U8*)STRING( noper );
1803 const U8 *e = uc + STR_LEN( noper );
1804 U32 state = 1; /* required init */
1805 U16 charid = 0; /* sanity init */
1806 U8 *scan = (U8*)NULL; /* sanity init */
1807 STRLEN foldlen = 0; /* required init */
1808 U32 wordlen = 0; /* required init */
1809 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1812 if (OP(noper) == NOTHING) {
1813 regnode *noper_next= regnext(noper);
1814 if (noper_next != tail && OP(noper_next) == flags) {
1816 uc= (U8*)STRING(noper);
1817 e= uc + STR_LEN(noper);
1821 if (OP(noper) != NOTHING) {
1822 for ( ; uc < e ; uc += len ) {
1827 charid = trie->charmap[ uvc ];
1829 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1833 charid=(U16)SvIV( *svpp );
1836 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1843 if ( !trie->states[ state ].trans.list ) {
1844 TRIE_LIST_NEW( state );
1846 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1847 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1848 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1853 newstate = next_alloc++;
1854 prev_states[newstate] = state;
1855 TRIE_LIST_PUSH( state, charid, newstate );
1860 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1864 TRIE_HANDLE_WORD(state);
1866 } /* end second pass */
1868 /* next alloc is the NEXT state to be allocated */
1869 trie->statecount = next_alloc;
1870 trie->states = (reg_trie_state *)
1871 PerlMemShared_realloc( trie->states,
1873 * sizeof(reg_trie_state) );
1875 /* and now dump it out before we compress it */
1876 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1877 revcharmap, next_alloc,
1881 trie->trans = (reg_trie_trans *)
1882 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1889 for( state=1 ; state < next_alloc ; state ++ ) {
1893 DEBUG_TRIE_COMPILE_MORE_r(
1894 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1898 if (trie->states[state].trans.list) {
1899 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1903 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1904 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1905 if ( forid < minid ) {
1907 } else if ( forid > maxid ) {
1911 if ( transcount < tp + maxid - minid + 1) {
1913 trie->trans = (reg_trie_trans *)
1914 PerlMemShared_realloc( trie->trans,
1916 * sizeof(reg_trie_trans) );
1917 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1919 base = trie->uniquecharcount + tp - minid;
1920 if ( maxid == minid ) {
1922 for ( ; zp < tp ; zp++ ) {
1923 if ( ! trie->trans[ zp ].next ) {
1924 base = trie->uniquecharcount + zp - minid;
1925 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1926 trie->trans[ zp ].check = state;
1932 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1933 trie->trans[ tp ].check = state;
1938 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1939 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1940 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1941 trie->trans[ tid ].check = state;
1943 tp += ( maxid - minid + 1 );
1945 Safefree(trie->states[ state ].trans.list);
1948 DEBUG_TRIE_COMPILE_MORE_r(
1949 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1952 trie->states[ state ].trans.base=base;
1954 trie->lasttrans = tp + 1;
1958 Second Pass -- Flat Table Representation.
1960 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1961 We know that we will need Charcount+1 trans at most to store the data
1962 (one row per char at worst case) So we preallocate both structures
1963 assuming worst case.
1965 We then construct the trie using only the .next slots of the entry
1968 We use the .check field of the first entry of the node temporarily to
1969 make compression both faster and easier by keeping track of how many non
1970 zero fields are in the node.
1972 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1975 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1976 number representing the first entry of the node, and state as a
1977 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1978 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1979 are 2 entrys per node. eg:
1987 The table is internally in the right hand, idx form. However as we also
1988 have to deal with the states array which is indexed by nodenum we have to
1989 use TRIE_NODENUM() to convert.
1992 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1993 "%*sCompiling trie using table compiler\n",
1994 (int)depth * 2 + 2, ""));
1996 trie->trans = (reg_trie_trans *)
1997 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1998 * trie->uniquecharcount + 1,
1999 sizeof(reg_trie_trans) );
2000 trie->states = (reg_trie_state *)
2001 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2002 sizeof(reg_trie_state) );
2003 next_alloc = trie->uniquecharcount + 1;
2006 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2008 regnode *noper = NEXTOPER( cur );
2009 const U8 *uc = (U8*)STRING( noper );
2010 const U8 *e = uc + STR_LEN( noper );
2012 U32 state = 1; /* required init */
2014 U16 charid = 0; /* sanity init */
2015 U32 accept_state = 0; /* sanity init */
2016 U8 *scan = (U8*)NULL; /* sanity init */
2018 STRLEN foldlen = 0; /* required init */
2019 U32 wordlen = 0; /* required init */
2021 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2023 if (OP(noper) == NOTHING) {
2024 regnode *noper_next= regnext(noper);
2025 if (noper_next != tail && OP(noper_next) == flags) {
2027 uc= (U8*)STRING(noper);
2028 e= uc + STR_LEN(noper);
2032 if ( OP(noper) != NOTHING ) {
2033 for ( ; uc < e ; uc += len ) {
2038 charid = trie->charmap[ uvc ];
2040 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2041 charid = svpp ? (U16)SvIV(*svpp) : 0;
2045 if ( !trie->trans[ state + charid ].next ) {
2046 trie->trans[ state + charid ].next = next_alloc;
2047 trie->trans[ state ].check++;
2048 prev_states[TRIE_NODENUM(next_alloc)]
2049 = TRIE_NODENUM(state);
2050 next_alloc += trie->uniquecharcount;
2052 state = trie->trans[ state + charid ].next;
2054 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2056 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2059 accept_state = TRIE_NODENUM( state );
2060 TRIE_HANDLE_WORD(accept_state);
2062 } /* end second pass */
2064 /* and now dump it out before we compress it */
2065 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2067 next_alloc, depth+1));
2071 * Inplace compress the table.*
2073 For sparse data sets the table constructed by the trie algorithm will
2074 be mostly 0/FAIL transitions or to put it another way mostly empty.
2075 (Note that leaf nodes will not contain any transitions.)
2077 This algorithm compresses the tables by eliminating most such
2078 transitions, at the cost of a modest bit of extra work during lookup:
2080 - Each states[] entry contains a .base field which indicates the
2081 index in the state[] array wheres its transition data is stored.
2083 - If .base is 0 there are no valid transitions from that node.
2085 - If .base is nonzero then charid is added to it to find an entry in
2088 -If trans[states[state].base+charid].check!=state then the
2089 transition is taken to be a 0/Fail transition. Thus if there are fail
2090 transitions at the front of the node then the .base offset will point
2091 somewhere inside the previous nodes data (or maybe even into a node
2092 even earlier), but the .check field determines if the transition is
2096 The following process inplace converts the table to the compressed
2097 table: We first do not compress the root node 1,and mark all its
2098 .check pointers as 1 and set its .base pointer as 1 as well. This
2099 allows us to do a DFA construction from the compressed table later,
2100 and ensures that any .base pointers we calculate later are greater
2103 - We set 'pos' to indicate the first entry of the second node.
2105 - We then iterate over the columns of the node, finding the first and
2106 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2107 and set the .check pointers accordingly, and advance pos
2108 appropriately and repreat for the next node. Note that when we copy
2109 the next pointers we have to convert them from the original
2110 NODEIDX form to NODENUM form as the former is not valid post
2113 - If a node has no transitions used we mark its base as 0 and do not
2114 advance the pos pointer.
2116 - If a node only has one transition we use a second pointer into the
2117 structure to fill in allocated fail transitions from other states.
2118 This pointer is independent of the main pointer and scans forward
2119 looking for null transitions that are allocated to a state. When it
2120 finds one it writes the single transition into the "hole". If the
2121 pointer doesnt find one the single transition is appended as normal.
2123 - Once compressed we can Renew/realloc the structures to release the
2126 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2127 specifically Fig 3.47 and the associated pseudocode.
2131 const U32 laststate = TRIE_NODENUM( next_alloc );
2134 trie->statecount = laststate;
2136 for ( state = 1 ; state < laststate ; state++ ) {
2138 const U32 stateidx = TRIE_NODEIDX( state );
2139 const U32 o_used = trie->trans[ stateidx ].check;
2140 U32 used = trie->trans[ stateidx ].check;
2141 trie->trans[ stateidx ].check = 0;
2143 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2144 if ( flag || trie->trans[ stateidx + charid ].next ) {
2145 if ( trie->trans[ stateidx + charid ].next ) {
2147 for ( ; zp < pos ; zp++ ) {
2148 if ( ! trie->trans[ zp ].next ) {
2152 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2153 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2154 trie->trans[ zp ].check = state;
2155 if ( ++zp > pos ) pos = zp;
2162 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2164 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2165 trie->trans[ pos ].check = state;
2170 trie->lasttrans = pos + 1;
2171 trie->states = (reg_trie_state *)
2172 PerlMemShared_realloc( trie->states, laststate
2173 * sizeof(reg_trie_state) );
2174 DEBUG_TRIE_COMPILE_MORE_r(
2175 PerlIO_printf( Perl_debug_log,
2176 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2177 (int)depth * 2 + 2,"",
2178 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2181 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2184 } /* end table compress */
2186 DEBUG_TRIE_COMPILE_MORE_r(
2187 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2188 (int)depth * 2 + 2, "",
2189 (UV)trie->statecount,
2190 (UV)trie->lasttrans)
2192 /* resize the trans array to remove unused space */
2193 trie->trans = (reg_trie_trans *)
2194 PerlMemShared_realloc( trie->trans, trie->lasttrans
2195 * sizeof(reg_trie_trans) );
2197 { /* Modify the program and insert the new TRIE node */
2198 U8 nodetype =(U8)(flags & 0xFF);
2202 regnode *optimize = NULL;
2203 #ifdef RE_TRACK_PATTERN_OFFSETS
2206 U32 mjd_nodelen = 0;
2207 #endif /* RE_TRACK_PATTERN_OFFSETS */
2208 #endif /* DEBUGGING */
2210 This means we convert either the first branch or the first Exact,
2211 depending on whether the thing following (in 'last') is a branch
2212 or not and whther first is the startbranch (ie is it a sub part of
2213 the alternation or is it the whole thing.)
2214 Assuming its a sub part we convert the EXACT otherwise we convert
2215 the whole branch sequence, including the first.
2217 /* Find the node we are going to overwrite */
2218 if ( first != startbranch || OP( last ) == BRANCH ) {
2219 /* branch sub-chain */
2220 NEXT_OFF( first ) = (U16)(last - first);
2221 #ifdef RE_TRACK_PATTERN_OFFSETS
2223 mjd_offset= Node_Offset((convert));
2224 mjd_nodelen= Node_Length((convert));
2227 /* whole branch chain */
2229 #ifdef RE_TRACK_PATTERN_OFFSETS
2232 const regnode *nop = NEXTOPER( convert );
2233 mjd_offset= Node_Offset((nop));
2234 mjd_nodelen= Node_Length((nop));
2238 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2239 (int)depth * 2 + 2, "",
2240 (UV)mjd_offset, (UV)mjd_nodelen)
2243 /* But first we check to see if there is a common prefix we can
2244 split out as an EXACT and put in front of the TRIE node. */
2245 trie->startstate= 1;
2246 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2248 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2252 const U32 base = trie->states[ state ].trans.base;
2254 if ( trie->states[state].wordnum )
2257 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2258 if ( ( base + ofs >= trie->uniquecharcount ) &&
2259 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2260 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2262 if ( ++count > 1 ) {
2263 SV **tmp = av_fetch( revcharmap, ofs, 0);
2264 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2265 if ( state == 1 ) break;
2267 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2269 PerlIO_printf(Perl_debug_log,
2270 "%*sNew Start State=%"UVuf" Class: [",
2271 (int)depth * 2 + 2, "",
2274 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2275 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2277 TRIE_BITMAP_SET(trie,*ch);
2279 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2281 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2285 TRIE_BITMAP_SET(trie,*ch);
2287 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2288 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2294 SV **tmp = av_fetch( revcharmap, idx, 0);
2296 char *ch = SvPV( *tmp, len );
2298 SV *sv=sv_newmortal();
2299 PerlIO_printf( Perl_debug_log,
2300 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2301 (int)depth * 2 + 2, "",
2303 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2304 PL_colors[0], PL_colors[1],
2305 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2306 PERL_PV_ESCAPE_FIRSTCHAR
2311 OP( convert ) = nodetype;
2312 str=STRING(convert);
2315 STR_LEN(convert) += len;
2321 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2326 trie->prefixlen = (state-1);
2328 regnode *n = convert+NODE_SZ_STR(convert);
2329 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2330 trie->startstate = state;
2331 trie->minlen -= (state - 1);
2332 trie->maxlen -= (state - 1);
2334 /* At least the UNICOS C compiler choked on this
2335 * being argument to DEBUG_r(), so let's just have
2338 #ifdef PERL_EXT_RE_BUILD
2344 regnode *fix = convert;
2345 U32 word = trie->wordcount;
2347 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2348 while( ++fix < n ) {
2349 Set_Node_Offset_Length(fix, 0, 0);
2352 SV ** const tmp = av_fetch( trie_words, word, 0 );
2354 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2355 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2357 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2365 NEXT_OFF(convert) = (U16)(tail - convert);
2366 DEBUG_r(optimize= n);
2372 if ( trie->maxlen ) {
2373 NEXT_OFF( convert ) = (U16)(tail - convert);
2374 ARG_SET( convert, data_slot );
2375 /* Store the offset to the first unabsorbed branch in
2376 jump[0], which is otherwise unused by the jump logic.
2377 We use this when dumping a trie and during optimisation. */
2379 trie->jump[0] = (U16)(nextbranch - convert);
2381 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2382 * and there is a bitmap
2383 * and the first "jump target" node we found leaves enough room
2384 * then convert the TRIE node into a TRIEC node, with the bitmap
2385 * embedded inline in the opcode - this is hypothetically faster.
2387 if ( !trie->states[trie->startstate].wordnum
2389 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2391 OP( convert ) = TRIEC;
2392 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2393 PerlMemShared_free(trie->bitmap);
2396 OP( convert ) = TRIE;
2398 /* store the type in the flags */
2399 convert->flags = nodetype;
2403 + regarglen[ OP( convert ) ];
2405 /* XXX We really should free up the resource in trie now,
2406 as we won't use them - (which resources?) dmq */
2408 /* needed for dumping*/
2409 DEBUG_r(if (optimize) {
2410 regnode *opt = convert;
2412 while ( ++opt < optimize) {
2413 Set_Node_Offset_Length(opt,0,0);
2416 Try to clean up some of the debris left after the
2419 while( optimize < jumper ) {
2420 mjd_nodelen += Node_Length((optimize));
2421 OP( optimize ) = OPTIMIZED;
2422 Set_Node_Offset_Length(optimize,0,0);
2425 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2427 } /* end node insert */
2429 /* Finish populating the prev field of the wordinfo array. Walk back
2430 * from each accept state until we find another accept state, and if
2431 * so, point the first word's .prev field at the second word. If the
2432 * second already has a .prev field set, stop now. This will be the
2433 * case either if we've already processed that word's accept state,
2434 * or that state had multiple words, and the overspill words were
2435 * already linked up earlier.
2442 for (word=1; word <= trie->wordcount; word++) {
2444 if (trie->wordinfo[word].prev)
2446 state = trie->wordinfo[word].accept;
2448 state = prev_states[state];
2451 prev = trie->states[state].wordnum;
2455 trie->wordinfo[word].prev = prev;
2457 Safefree(prev_states);
2461 /* and now dump out the compressed format */
2462 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2464 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2466 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2467 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2469 SvREFCNT_dec_NN(revcharmap);
2473 : trie->startstate>1
2479 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2481 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2483 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2484 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2487 We find the fail state for each state in the trie, this state is the longest proper
2488 suffix of the current state's 'word' that is also a proper prefix of another word in our
2489 trie. State 1 represents the word '' and is thus the default fail state. This allows
2490 the DFA not to have to restart after its tried and failed a word at a given point, it
2491 simply continues as though it had been matching the other word in the first place.
2493 'abcdgu'=~/abcdefg|cdgu/
2494 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2495 fail, which would bring us to the state representing 'd' in the second word where we would
2496 try 'g' and succeed, proceeding to match 'cdgu'.
2498 /* add a fail transition */
2499 const U32 trie_offset = ARG(source);
2500 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2502 const U32 ucharcount = trie->uniquecharcount;
2503 const U32 numstates = trie->statecount;
2504 const U32 ubound = trie->lasttrans + ucharcount;
2508 U32 base = trie->states[ 1 ].trans.base;
2511 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2512 GET_RE_DEBUG_FLAGS_DECL;
2514 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2516 PERL_UNUSED_ARG(depth);
2520 ARG_SET( stclass, data_slot );
2521 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2522 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2523 aho->trie=trie_offset;
2524 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2525 Copy( trie->states, aho->states, numstates, reg_trie_state );
2526 Newxz( q, numstates, U32);
2527 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2530 /* initialize fail[0..1] to be 1 so that we always have
2531 a valid final fail state */
2532 fail[ 0 ] = fail[ 1 ] = 1;
2534 for ( charid = 0; charid < ucharcount ; charid++ ) {
2535 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2537 q[ q_write ] = newstate;
2538 /* set to point at the root */
2539 fail[ q[ q_write++ ] ]=1;
2542 while ( q_read < q_write) {
2543 const U32 cur = q[ q_read++ % numstates ];
2544 base = trie->states[ cur ].trans.base;
2546 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2547 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2549 U32 fail_state = cur;
2552 fail_state = fail[ fail_state ];
2553 fail_base = aho->states[ fail_state ].trans.base;
2554 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2556 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2557 fail[ ch_state ] = fail_state;
2558 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2560 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2562 q[ q_write++ % numstates] = ch_state;
2566 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2567 when we fail in state 1, this allows us to use the
2568 charclass scan to find a valid start char. This is based on the principle
2569 that theres a good chance the string being searched contains lots of stuff
2570 that cant be a start char.
2572 fail[ 0 ] = fail[ 1 ] = 0;
2573 DEBUG_TRIE_COMPILE_r({
2574 PerlIO_printf(Perl_debug_log,
2575 "%*sStclass Failtable (%"UVuf" states): 0",
2576 (int)(depth * 2), "", (UV)numstates
2578 for( q_read=1; q_read<numstates; q_read++ ) {
2579 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2581 PerlIO_printf(Perl_debug_log, "\n");
2584 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2589 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2590 * These need to be revisited when a newer toolchain becomes available.
2592 #if defined(__sparc64__) && defined(__GNUC__)
2593 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2594 # undef SPARC64_GCC_WORKAROUND
2595 # define SPARC64_GCC_WORKAROUND 1
2599 #define DEBUG_PEEP(str,scan,depth) \
2600 DEBUG_OPTIMISE_r({if (scan){ \
2601 SV * const mysv=sv_newmortal(); \
2602 regnode *Next = regnext(scan); \
2603 regprop(RExC_rx, mysv, scan); \
2604 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2605 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2606 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2610 /* The below joins as many adjacent EXACTish nodes as possible into a single
2611 * one. The regop may be changed if the node(s) contain certain sequences that
2612 * require special handling. The joining is only done if:
2613 * 1) there is room in the current conglomerated node to entirely contain the
2615 * 2) they are the exact same node type
2617 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2618 * these get optimized out
2620 * If a node is to match under /i (folded), the number of characters it matches
2621 * can be different than its character length if it contains a multi-character
2622 * fold. *min_subtract is set to the total delta of the input nodes.
2624 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2625 * and contains LATIN SMALL LETTER SHARP S
2627 * This is as good a place as any to discuss the design of handling these
2628 * multi-character fold sequences. It's been wrong in Perl for a very long
2629 * time. There are three code points in Unicode whose multi-character folds
2630 * were long ago discovered to mess things up. The previous designs for
2631 * dealing with these involved assigning a special node for them. This
2632 * approach doesn't work, as evidenced by this example:
2633 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2634 * Both these fold to "sss", but if the pattern is parsed to create a node that
2635 * would match just the \xDF, it won't be able to handle the case where a
2636 * successful match would have to cross the node's boundary. The new approach
2637 * that hopefully generally solves the problem generates an EXACTFU_SS node
2640 * It turns out that there are problems with all multi-character folds, and not
2641 * just these three. Now the code is general, for all such cases, but the
2642 * three still have some special handling. The approach taken is:
2643 * 1) This routine examines each EXACTFish node that could contain multi-
2644 * character fold sequences. It returns in *min_subtract how much to
2645 * subtract from the the actual length of the string to get a real minimum
2646 * match length; it is 0 if there are no multi-char folds. This delta is
2647 * used by the caller to adjust the min length of the match, and the delta
2648 * between min and max, so that the optimizer doesn't reject these
2649 * possibilities based on size constraints.
2650 * 2) Certain of these sequences require special handling by the trie code,
2651 * so, if found, this code changes the joined node type to special ops:
2652 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2653 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2654 * is used for an EXACTFU node that contains at least one "ss" sequence in
2655 * it. For non-UTF-8 patterns and strings, this is the only case where
2656 * there is a possible fold length change. That means that a regular
2657 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2658 * with length changes, and so can be processed faster. regexec.c takes
2659 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2660 * pre-folded by regcomp.c. This saves effort in regex matching.
2661 * However, the pre-folding isn't done for non-UTF8 patterns because the
2662 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2663 * down by forcing the pattern into UTF8 unless necessary. Also what
2664 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2665 * possibilities for the non-UTF8 patterns are quite simple, except for
2666 * the sharp s. All the ones that don't involve a UTF-8 target string are
2667 * members of a fold-pair, and arrays are set up for all of them so that
2668 * the other member of the pair can be found quickly. Code elsewhere in
2669 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2670 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2671 * described in the next item.
2672 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2673 * 'ss' or not is not knowable at compile time. It will match iff the
2674 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2675 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2676 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
2677 * described in item 3). An assumption that the optimizer part of
2678 * regexec.c (probably unwittingly) makes is that a character in the
2679 * pattern corresponds to at most a single character in the target string.
2680 * (And I do mean character, and not byte here, unlike other parts of the
2681 * documentation that have never been updated to account for multibyte
2682 * Unicode.) This assumption is wrong only in this case, as all other
2683 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2684 * virtue of having this file pre-fold UTF-8 patterns. I'm
2685 * reluctant to try to change this assumption, so instead the code punts.
2686 * This routine examines EXACTF nodes for the sharp s, and returns a
2687 * boolean indicating whether or not the node is an EXACTF node that
2688 * contains a sharp s. When it is true, the caller sets a flag that later
2689 * causes the optimizer in this file to not set values for the floating
2690 * and fixed string lengths, and thus avoids the optimizer code in
2691 * regexec.c that makes the invalid assumption. Thus, there is no
2692 * optimization based on string lengths for EXACTF nodes that contain the
2693 * sharp s. This only happens for /id rules (which means the pattern
2697 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2698 if (PL_regkind[OP(scan)] == EXACT) \
2699 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2702 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) {
2703 /* Merge several consecutive EXACTish nodes into one. */
2704 regnode *n = regnext(scan);
2706 regnode *next = scan + NODE_SZ_STR(scan);
2710 regnode *stop = scan;
2711 GET_RE_DEBUG_FLAGS_DECL;
2713 PERL_UNUSED_ARG(depth);
2716 PERL_ARGS_ASSERT_JOIN_EXACT;
2717 #ifndef EXPERIMENTAL_INPLACESCAN
2718 PERL_UNUSED_ARG(flags);
2719 PERL_UNUSED_ARG(val);
2721 DEBUG_PEEP("join",scan,depth);
2723 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2724 * EXACT ones that are mergeable to the current one. */
2726 && (PL_regkind[OP(n)] == NOTHING
2727 || (stringok && OP(n) == OP(scan)))
2729 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2732 if (OP(n) == TAIL || n > next)
2734 if (PL_regkind[OP(n)] == NOTHING) {
2735 DEBUG_PEEP("skip:",n,depth);
2736 NEXT_OFF(scan) += NEXT_OFF(n);
2737 next = n + NODE_STEP_REGNODE;
2744 else if (stringok) {
2745 const unsigned int oldl = STR_LEN(scan);
2746 regnode * const nnext = regnext(n);
2748 /* XXX I (khw) kind of doubt that this works on platforms where
2749 * U8_MAX is above 255 because of lots of other assumptions */
2750 /* Don't join if the sum can't fit into a single node */
2751 if (oldl + STR_LEN(n) > U8_MAX)
2754 DEBUG_PEEP("merg",n,depth);
2757 NEXT_OFF(scan) += NEXT_OFF(n);
2758 STR_LEN(scan) += STR_LEN(n);
2759 next = n + NODE_SZ_STR(n);
2760 /* Now we can overwrite *n : */
2761 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2769 #ifdef EXPERIMENTAL_INPLACESCAN
2770 if (flags && !NEXT_OFF(n)) {
2771 DEBUG_PEEP("atch", val, depth);
2772 if (reg_off_by_arg[OP(n)]) {
2773 ARG_SET(n, val - n);
2776 NEXT_OFF(n) = val - n;
2784 *has_exactf_sharp_s = FALSE;
2786 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2787 * can now analyze for sequences of problematic code points. (Prior to
2788 * this final joining, sequences could have been split over boundaries, and
2789 * hence missed). The sequences only happen in folding, hence for any
2790 * non-EXACT EXACTish node */
2791 if (OP(scan) != EXACT) {
2792 const U8 * const s0 = (U8*) STRING(scan);
2794 const U8 * const s_end = s0 + STR_LEN(scan);
2796 /* One pass is made over the node's string looking for all the
2797 * possibilities. to avoid some tests in the loop, there are two main
2798 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2802 /* Examine the string for a multi-character fold sequence. UTF-8
2803 * patterns have all characters pre-folded by the time this code is
2805 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2806 length sequence we are looking for is 2 */
2809 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2810 if (! len) { /* Not a multi-char fold: get next char */
2815 /* Nodes with 'ss' require special handling, except for EXACTFL
2816 * and EXACTFA for which there is no multi-char fold to this */
2817 if (len == 2 && *s == 's' && *(s+1) == 's'
2818 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2821 OP(scan) = EXACTFU_SS;
2824 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2825 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2826 COMBINING_DIAERESIS_UTF8
2827 COMBINING_ACUTE_ACCENT_UTF8,
2829 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2830 COMBINING_DIAERESIS_UTF8
2831 COMBINING_ACUTE_ACCENT_UTF8,
2836 /* These two folds require special handling by trie's, so
2837 * change the node type to indicate this. If EXACTFA and
2838 * EXACTFL were ever to be handled by trie's, this would
2839 * have to be changed. If this node has already been
2840 * changed to EXACTFU_SS in this loop, leave it as is. (I
2841 * (khw) think it doesn't matter in regexec.c for UTF
2842 * patterns, but no need to change it */
2843 if (OP(scan) == EXACTFU) {
2844 OP(scan) = EXACTFU_TRICKYFOLD;
2848 else { /* Here is a generic multi-char fold. */
2849 const U8* multi_end = s + len;
2851 /* Count how many characters in it. In the case of /l and
2852 * /aa, no folds which contain ASCII code points are
2853 * allowed, so check for those, and skip if found. (In
2854 * EXACTFL, no folds are allowed to any Latin1 code point,
2855 * not just ASCII. But there aren't any of these
2856 * currently, nor ever likely, so don't take the time to
2857 * test for them. The code that generates the
2858 * is_MULTI_foo() macros croaks should one actually get put
2859 * into Unicode .) */
2860 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2861 count = utf8_length(s, multi_end);
2865 while (s < multi_end) {
2868 goto next_iteration;
2878 /* The delta is how long the sequence is minus 1 (1 is how long
2879 * the character that folds to the sequence is) */
2880 *min_subtract += count - 1;
2884 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2886 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2887 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2888 * nodes can't have multi-char folds to this range (and there are
2889 * no existing ones in the upper latin1 range). In the EXACTF
2890 * case we look also for the sharp s, which can be in the final
2891 * position. Otherwise we can stop looking 1 byte earlier because
2892 * have to find at least two characters for a multi-fold */
2893 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2895 /* The below is perhaps overboard, but this allows us to save a
2896 * test each time through the loop at the expense of a mask. This
2897 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2898 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2899 * are 64. This uses an exclusive 'or' to find that bit and then
2900 * inverts it to form a mask, with just a single 0, in the bit
2901 * position where 'S' and 's' differ. */
2902 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2903 const U8 s_masked = 's' & S_or_s_mask;
2906 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2907 if (! len) { /* Not a multi-char fold. */
2908 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2910 *has_exactf_sharp_s = TRUE;
2917 && ((*s & S_or_s_mask) == s_masked)
2918 && ((*(s+1) & S_or_s_mask) == s_masked))
2921 /* EXACTF nodes need to know that the minimum length
2922 * changed so that a sharp s in the string can match this
2923 * ss in the pattern, but they remain EXACTF nodes, as they
2924 * won't match this unless the target string is is UTF-8,
2925 * which we don't know until runtime */
2926 if (OP(scan) != EXACTF) {
2927 OP(scan) = EXACTFU_SS;
2931 *min_subtract += len - 1;
2938 /* Allow dumping but overwriting the collection of skipped
2939 * ops and/or strings with fake optimized ops */
2940 n = scan + NODE_SZ_STR(scan);
2948 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2952 /* REx optimizer. Converts nodes into quicker variants "in place".
2953 Finds fixed substrings. */
2955 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2956 to the position after last scanned or to NULL. */
2958 #define INIT_AND_WITHP \
2959 assert(!and_withp); \
2960 Newx(and_withp,1,struct regnode_charclass_class); \
2961 SAVEFREEPV(and_withp)
2963 /* this is a chain of data about sub patterns we are processing that
2964 need to be handled separately/specially in study_chunk. Its so
2965 we can simulate recursion without losing state. */
2967 typedef struct scan_frame {
2968 regnode *last; /* last node to process in this frame */
2969 regnode *next; /* next node to process when last is reached */
2970 struct scan_frame *prev; /*previous frame*/
2971 I32 stop; /* what stopparen do we use */
2975 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2978 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2979 I32 *minlenp, I32 *deltap,
2984 struct regnode_charclass_class *and_withp,
2985 U32 flags, U32 depth)
2986 /* scanp: Start here (read-write). */
2987 /* deltap: Write maxlen-minlen here. */
2988 /* last: Stop before this one. */
2989 /* data: string data about the pattern */
2990 /* stopparen: treat close N as END */
2991 /* recursed: which subroutines have we recursed into */
2992 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
2995 I32 min = 0; /* There must be at least this number of characters to match */
2997 regnode *scan = *scanp, *next;
2999 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3000 int is_inf_internal = 0; /* The studied chunk is infinite */
3001 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3002 scan_data_t data_fake;
3003 SV *re_trie_maxbuff = NULL;
3004 regnode *first_non_open = scan;
3005 I32 stopmin = I32_MAX;
3006 scan_frame *frame = NULL;
3007 GET_RE_DEBUG_FLAGS_DECL;
3009 PERL_ARGS_ASSERT_STUDY_CHUNK;
3012 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3016 while (first_non_open && OP(first_non_open) == OPEN)
3017 first_non_open=regnext(first_non_open);
3022 while ( scan && OP(scan) != END && scan < last ){
3023 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3024 node length to get a real minimum (because
3025 the folded version may be shorter) */
3026 bool has_exactf_sharp_s = FALSE;
3027 /* Peephole optimizer: */
3028 DEBUG_STUDYDATA("Peep:", data,depth);
3029 DEBUG_PEEP("Peep",scan,depth);
3031 /* Its not clear to khw or hv why this is done here, and not in the
3032 * clauses that deal with EXACT nodes. khw's guess is that it's
3033 * because of a previous design */
3034 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3036 /* Follow the next-chain of the current node and optimize
3037 away all the NOTHINGs from it. */
3038 if (OP(scan) != CURLYX) {
3039 const int max = (reg_off_by_arg[OP(scan)]
3041 /* I32 may be smaller than U16 on CRAYs! */
3042 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3043 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3047 /* Skip NOTHING and LONGJMP. */
3048 while ((n = regnext(n))
3049 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3050 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3051 && off + noff < max)
3053 if (reg_off_by_arg[OP(scan)])
3056 NEXT_OFF(scan) = off;
3061 /* The principal pseudo-switch. Cannot be a switch, since we
3062 look into several different things. */
3063 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3064 || OP(scan) == IFTHEN) {
3065 next = regnext(scan);
3067 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3069 if (OP(next) == code || code == IFTHEN) {
3070 /* NOTE - There is similar code to this block below for handling
3071 TRIE nodes on a re-study. If you change stuff here check there
3073 I32 max1 = 0, min1 = I32_MAX, num = 0;
3074 struct regnode_charclass_class accum;
3075 regnode * const startbranch=scan;
3077 if (flags & SCF_DO_SUBSTR)
3078 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3079 if (flags & SCF_DO_STCLASS)
3080 cl_init_zero(pRExC_state, &accum);
3082 while (OP(scan) == code) {
3083 I32 deltanext, minnext, f = 0, fake;
3084 struct regnode_charclass_class this_class;
3087 data_fake.flags = 0;
3089 data_fake.whilem_c = data->whilem_c;
3090 data_fake.last_closep = data->last_closep;
3093 data_fake.last_closep = &fake;
3095 data_fake.pos_delta = delta;
3096 next = regnext(scan);
3097 scan = NEXTOPER(scan);
3099 scan = NEXTOPER(scan);
3100 if (flags & SCF_DO_STCLASS) {
3101 cl_init(pRExC_state, &this_class);
3102 data_fake.start_class = &this_class;
3103 f = SCF_DO_STCLASS_AND;
3105 if (flags & SCF_WHILEM_VISITED_POS)
3106 f |= SCF_WHILEM_VISITED_POS;
3108 /* we suppose the run is continuous, last=next...*/
3109 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3111 stopparen, recursed, NULL, f,depth+1);
3114 if (max1 < minnext + deltanext)
3115 max1 = minnext + deltanext;
3116 if (deltanext == I32_MAX)
3117 is_inf = is_inf_internal = 1;
3119 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3121 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3122 if ( stopmin > minnext)
3123 stopmin = min + min1;
3124 flags &= ~SCF_DO_SUBSTR;
3126 data->flags |= SCF_SEEN_ACCEPT;
3129 if (data_fake.flags & SF_HAS_EVAL)
3130 data->flags |= SF_HAS_EVAL;
3131 data->whilem_c = data_fake.whilem_c;
3133 if (flags & SCF_DO_STCLASS)
3134 cl_or(pRExC_state, &accum, &this_class);
3136 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3138 if (flags & SCF_DO_SUBSTR) {
3139 data->pos_min += min1;
3140 data->pos_delta += max1 - min1;
3141 if (max1 != min1 || is_inf)
3142 data->longest = &(data->longest_float);
3145 delta += max1 - min1;
3146 if (flags & SCF_DO_STCLASS_OR) {
3147 cl_or(pRExC_state, data->start_class, &accum);
3149 cl_and(data->start_class, and_withp);
3150 flags &= ~SCF_DO_STCLASS;
3153 else if (flags & SCF_DO_STCLASS_AND) {
3155 cl_and(data->start_class, &accum);
3156 flags &= ~SCF_DO_STCLASS;
3159 /* Switch to OR mode: cache the old value of
3160 * data->start_class */
3162 StructCopy(data->start_class, and_withp,
3163 struct regnode_charclass_class);
3164 flags &= ~SCF_DO_STCLASS_AND;
3165 StructCopy(&accum, data->start_class,
3166 struct regnode_charclass_class);
3167 flags |= SCF_DO_STCLASS_OR;
3168 SET_SSC_EOS(data->start_class);
3172 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3175 Assuming this was/is a branch we are dealing with: 'scan' now
3176 points at the item that follows the branch sequence, whatever
3177 it is. We now start at the beginning of the sequence and look
3184 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3186 If we can find such a subsequence we need to turn the first
3187 element into a trie and then add the subsequent branch exact
3188 strings to the trie.
3192 1. patterns where the whole set of branches can be converted.
3194 2. patterns where only a subset can be converted.
3196 In case 1 we can replace the whole set with a single regop
3197 for the trie. In case 2 we need to keep the start and end
3200 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3201 becomes BRANCH TRIE; BRANCH X;
3203 There is an additional case, that being where there is a
3204 common prefix, which gets split out into an EXACT like node
3205 preceding the TRIE node.
3207 If x(1..n)==tail then we can do a simple trie, if not we make
3208 a "jump" trie, such that when we match the appropriate word
3209 we "jump" to the appropriate tail node. Essentially we turn
3210 a nested if into a case structure of sorts.
3215 if (!re_trie_maxbuff) {
3216 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3217 if (!SvIOK(re_trie_maxbuff))
3218 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3220 if ( SvIV(re_trie_maxbuff)>=0 ) {
3222 regnode *first = (regnode *)NULL;
3223 regnode *last = (regnode *)NULL;
3224 regnode *tail = scan;
3229 SV * const mysv = sv_newmortal(); /* for dumping */
3231 /* var tail is used because there may be a TAIL
3232 regop in the way. Ie, the exacts will point to the
3233 thing following the TAIL, but the last branch will
3234 point at the TAIL. So we advance tail. If we
3235 have nested (?:) we may have to move through several
3239 while ( OP( tail ) == TAIL ) {
3240 /* this is the TAIL generated by (?:) */
3241 tail = regnext( tail );
3245 DEBUG_TRIE_COMPILE_r({
3246 regprop(RExC_rx, mysv, tail );
3247 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3248 (int)depth * 2 + 2, "",
3249 "Looking for TRIE'able sequences. Tail node is: ",
3250 SvPV_nolen_const( mysv )
3256 Step through the branches
3257 cur represents each branch,
3258 noper is the first thing to be matched as part of that branch
3259 noper_next is the regnext() of that node.
3261 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3262 via a "jump trie" but we also support building with NOJUMPTRIE,
3263 which restricts the trie logic to structures like /FOO|BAR/.
3265 If noper is a trieable nodetype then the branch is a possible optimization
3266 target. If we are building under NOJUMPTRIE then we require that noper_next
3267 is the same as scan (our current position in the regex program).
3269 Once we have two or more consecutive such branches we can create a
3270 trie of the EXACT's contents and stitch it in place into the program.
3272 If the sequence represents all of the branches in the alternation we
3273 replace the entire thing with a single TRIE node.
3275 Otherwise when it is a subsequence we need to stitch it in place and
3276 replace only the relevant branches. This means the first branch has
3277 to remain as it is used by the alternation logic, and its next pointer,
3278 and needs to be repointed at the item on the branch chain following
3279 the last branch we have optimized away.
3281 This could be either a BRANCH, in which case the subsequence is internal,
3282 or it could be the item following the branch sequence in which case the
3283 subsequence is at the end (which does not necessarily mean the first node
3284 is the start of the alternation).
3286 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3289 ----------------+-----------
3293 EXACTFU_SS | EXACTFU
3294 EXACTFU_TRICKYFOLD | EXACTFU
3299 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3300 ( EXACT == (X) ) ? EXACT : \
3301 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3304 /* dont use tail as the end marker for this traverse */
3305 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3306 regnode * const noper = NEXTOPER( cur );
3307 U8 noper_type = OP( noper );
3308 U8 noper_trietype = TRIE_TYPE( noper_type );
3309 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3310 regnode * const noper_next = regnext( noper );
3311 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3312 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3315 DEBUG_TRIE_COMPILE_r({
3316 regprop(RExC_rx, mysv, cur);
3317 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3318 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3320 regprop(RExC_rx, mysv, noper);
3321 PerlIO_printf( Perl_debug_log, " -> %s",
3322 SvPV_nolen_const(mysv));
3325 regprop(RExC_rx, mysv, noper_next );
3326 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3327 SvPV_nolen_const(mysv));
3329 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3330 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3331 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3335 /* Is noper a trieable nodetype that can be merged with the
3336 * current trie (if there is one)? */
3340 ( noper_trietype == NOTHING)
3341 || ( trietype == NOTHING )
3342 || ( trietype == noper_trietype )
3345 && noper_next == tail
3349 /* Handle mergable triable node
3350 * Either we are the first node in a new trieable sequence,
3351 * in which case we do some bookkeeping, otherwise we update
3352 * the end pointer. */
3355 if ( noper_trietype == NOTHING ) {
3356 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3357 regnode * const noper_next = regnext( noper );
3358 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3359 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3362 if ( noper_next_trietype ) {
3363 trietype = noper_next_trietype;
3364 } else if (noper_next_type) {
3365 /* a NOTHING regop is 1 regop wide. We need at least two
3366 * for a trie so we can't merge this in */
3370 trietype = noper_trietype;
3373 if ( trietype == NOTHING )
3374 trietype = noper_trietype;
3379 } /* end handle mergable triable node */
3381 /* handle unmergable node -
3382 * noper may either be a triable node which can not be tried
3383 * together with the current trie, or a non triable node */
3385 /* If last is set and trietype is not NOTHING then we have found
3386 * at least two triable branch sequences in a row of a similar
3387 * trietype so we can turn them into a trie. If/when we
3388 * allow NOTHING to start a trie sequence this condition will be
3389 * required, and it isn't expensive so we leave it in for now. */
3390 if ( trietype && trietype != NOTHING )
3391 make_trie( pRExC_state,
3392 startbranch, first, cur, tail, count,
3393 trietype, depth+1 );
3394 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3398 && noper_next == tail
3401 /* noper is triable, so we can start a new trie sequence */
3404 trietype = noper_trietype;
3406 /* if we already saw a first but the current node is not triable then we have
3407 * to reset the first information. */
3412 } /* end handle unmergable node */
3413 } /* loop over branches */
3414 DEBUG_TRIE_COMPILE_r({
3415 regprop(RExC_rx, mysv, cur);
3416 PerlIO_printf( Perl_debug_log,
3417 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3418 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3421 if ( last && trietype ) {
3422 if ( trietype != NOTHING ) {
3423 /* the last branch of the sequence was part of a trie,
3424 * so we have to construct it here outside of the loop
3426 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3427 #ifdef TRIE_STUDY_OPT
3428 if ( ((made == MADE_EXACT_TRIE &&
3429 startbranch == first)
3430 || ( first_non_open == first )) &&
3432 flags |= SCF_TRIE_RESTUDY;
3433 if ( startbranch == first
3436 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3441 /* at this point we know whatever we have is a NOTHING sequence/branch
3442 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3444 if ( startbranch == first ) {
3446 /* the entire thing is a NOTHING sequence, something like this:
3447 * (?:|) So we can turn it into a plain NOTHING op. */
3448 DEBUG_TRIE_COMPILE_r({
3449 regprop(RExC_rx, mysv, cur);
3450 PerlIO_printf( Perl_debug_log,
3451 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3452 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3455 OP(startbranch)= NOTHING;
3456 NEXT_OFF(startbranch)= tail - startbranch;
3457 for ( opt= startbranch + 1; opt < tail ; opt++ )
3461 } /* end if ( last) */
3462 } /* TRIE_MAXBUF is non zero */
3467 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3468 scan = NEXTOPER(NEXTOPER(scan));
3469 } else /* single branch is optimized. */
3470 scan = NEXTOPER(scan);
3472 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3473 scan_frame *newframe = NULL;
3478 if (OP(scan) != SUSPEND) {
3479 /* set the pointer */
3480 if (OP(scan) == GOSUB) {
3482 RExC_recurse[ARG2L(scan)] = scan;
3483 start = RExC_open_parens[paren-1];
3484 end = RExC_close_parens[paren-1];
3487 start = RExC_rxi->program + 1;
3491 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3492 SAVEFREEPV(recursed);
3494 if (!PAREN_TEST(recursed,paren+1)) {
3495 PAREN_SET(recursed,paren+1);
3496 Newx(newframe,1,scan_frame);
3498 if (flags & SCF_DO_SUBSTR) {
3499 SCAN_COMMIT(pRExC_state,data,minlenp);
3500 data->longest = &(data->longest_float);
3502 is_inf = is_inf_internal = 1;
3503 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3504 cl_anything(pRExC_state, data->start_class);
3505 flags &= ~SCF_DO_STCLASS;
3508 Newx(newframe,1,scan_frame);
3511 end = regnext(scan);
3516 SAVEFREEPV(newframe);
3517 newframe->next = regnext(scan);
3518 newframe->last = last;
3519 newframe->stop = stopparen;
3520 newframe->prev = frame;
3530 else if (OP(scan) == EXACT) {
3531 I32 l = STR_LEN(scan);
3534 const U8 * const s = (U8*)STRING(scan);
3535 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3536 l = utf8_length(s, s + l);
3538 uc = *((U8*)STRING(scan));
3541 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3542 /* The code below prefers earlier match for fixed
3543 offset, later match for variable offset. */
3544 if (data->last_end == -1) { /* Update the start info. */
3545 data->last_start_min = data->pos_min;
3546 data->last_start_max = is_inf
3547 ? I32_MAX : data->pos_min + data->pos_delta;
3549 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3551 SvUTF8_on(data->last_found);
3553 SV * const sv = data->last_found;
3554 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3555 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3556 if (mg && mg->mg_len >= 0)
3557 mg->mg_len += utf8_length((U8*)STRING(scan),
3558 (U8*)STRING(scan)+STR_LEN(scan));
3560 data->last_end = data->pos_min + l;
3561 data->pos_min += l; /* As in the first entry. */
3562 data->flags &= ~SF_BEFORE_EOL;
3564 if (flags & SCF_DO_STCLASS_AND) {
3565 /* Check whether it is compatible with what we know already! */
3569 /* If compatible, we or it in below. It is compatible if is
3570 * in the bitmp and either 1) its bit or its fold is set, or 2)
3571 * it's for a locale. Even if there isn't unicode semantics
3572 * here, at runtime there may be because of matching against a
3573 * utf8 string, so accept a possible false positive for
3574 * latin1-range folds */
3576 (!(data->start_class->flags & ANYOF_LOCALE)
3577 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3578 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3579 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3584 ANYOF_CLASS_ZERO(data->start_class);
3585 ANYOF_BITMAP_ZERO(data->start_class);
3587 ANYOF_BITMAP_SET(data->start_class, uc);
3588 else if (uc >= 0x100) {
3591 /* Some Unicode code points fold to the Latin1 range; as
3592 * XXX temporary code, instead of figuring out if this is
3593 * one, just assume it is and set all the start class bits
3594 * that could be some such above 255 code point's fold
3595 * which will generate fals positives. As the code
3596 * elsewhere that does compute the fold settles down, it
3597 * can be extracted out and re-used here */
3598 for (i = 0; i < 256; i++){
3599 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3600 ANYOF_BITMAP_SET(data->start_class, i);
3604 CLEAR_SSC_EOS(data->start_class);
3606 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3608 else if (flags & SCF_DO_STCLASS_OR) {
3609 /* false positive possible if the class is case-folded */
3611 ANYOF_BITMAP_SET(data->start_class, uc);
3613 data->start_class->flags |= ANYOF_UNICODE_ALL;
3614 CLEAR_SSC_EOS(data->start_class);
3615 cl_and(data->start_class, and_withp);
3617 flags &= ~SCF_DO_STCLASS;
3619 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3620 I32 l = STR_LEN(scan);
3621 UV uc = *((U8*)STRING(scan));
3623 /* Search for fixed substrings supports EXACT only. */
3624 if (flags & SCF_DO_SUBSTR) {
3626 SCAN_COMMIT(pRExC_state, data, minlenp);
3629 const U8 * const s = (U8 *)STRING(scan);
3630 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3631 l = utf8_length(s, s + l);
3633 if (has_exactf_sharp_s) {
3634 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3636 min += l - min_subtract;
3638 delta += min_subtract;
3639 if (flags & SCF_DO_SUBSTR) {
3640 data->pos_min += l - min_subtract;
3641 if (data->pos_min < 0) {
3644 data->pos_delta += min_subtract;
3646 data->longest = &(data->longest_float);
3649 if (flags & SCF_DO_STCLASS_AND) {
3650 /* Check whether it is compatible with what we know already! */
3653 (!(data->start_class->flags & ANYOF_LOCALE)
3654 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3655 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3659 ANYOF_CLASS_ZERO(data->start_class);
3660 ANYOF_BITMAP_ZERO(data->start_class);
3662 ANYOF_BITMAP_SET(data->start_class, uc);
3663 CLEAR_SSC_EOS(data->start_class);
3664 if (OP(scan) == EXACTFL) {
3665 /* XXX This set is probably no longer necessary, and
3666 * probably wrong as LOCALE now is on in the initial
3668 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3672 /* Also set the other member of the fold pair. In case
3673 * that unicode semantics is called for at runtime, use
3674 * the full latin1 fold. (Can't do this for locale,
3675 * because not known until runtime) */
3676 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3678 /* All other (EXACTFL handled above) folds except under
3679 * /iaa that include s, S, and sharp_s also may include
3681 if (OP(scan) != EXACTFA) {
3682 if (uc == 's' || uc == 'S') {
3683 ANYOF_BITMAP_SET(data->start_class,
3684 LATIN_SMALL_LETTER_SHARP_S);
3686 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3687 ANYOF_BITMAP_SET(data->start_class, 's');
3688 ANYOF_BITMAP_SET(data->start_class, 'S');
3693 else if (uc >= 0x100) {
3695 for (i = 0; i < 256; i++){
3696 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3697 ANYOF_BITMAP_SET(data->start_class, i);
3702 else if (flags & SCF_DO_STCLASS_OR) {
3703 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3704 /* false positive possible if the class is case-folded.
3705 Assume that the locale settings are the same... */
3707 ANYOF_BITMAP_SET(data->start_class, uc);
3708 if (OP(scan) != EXACTFL) {
3710 /* And set the other member of the fold pair, but
3711 * can't do that in locale because not known until
3713 ANYOF_BITMAP_SET(data->start_class,
3714 PL_fold_latin1[uc]);
3716 /* All folds except under /iaa that include s, S,
3717 * and sharp_s also may include the others */
3718 if (OP(scan) != EXACTFA) {
3719 if (uc == 's' || uc == 'S') {
3720 ANYOF_BITMAP_SET(data->start_class,
3721 LATIN_SMALL_LETTER_SHARP_S);
3723 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3724 ANYOF_BITMAP_SET(data->start_class, 's');
3725 ANYOF_BITMAP_SET(data->start_class, 'S');
3730 CLEAR_SSC_EOS(data->start_class);
3732 cl_and(data->start_class, and_withp);
3734 flags &= ~SCF_DO_STCLASS;
3736 else if (REGNODE_VARIES(OP(scan))) {
3737 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3738 I32 f = flags, pos_before = 0;
3739 regnode * const oscan = scan;
3740 struct regnode_charclass_class this_class;
3741 struct regnode_charclass_class *oclass = NULL;
3742 I32 next_is_eval = 0;
3744 switch (PL_regkind[OP(scan)]) {
3745 case WHILEM: /* End of (?:...)* . */
3746 scan = NEXTOPER(scan);
3749 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3750 next = NEXTOPER(scan);
3751 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3753 maxcount = REG_INFTY;
3754 next = regnext(scan);
3755 scan = NEXTOPER(scan);
3759 if (flags & SCF_DO_SUBSTR)
3764 if (flags & SCF_DO_STCLASS) {
3766 maxcount = REG_INFTY;
3767 next = regnext(scan);
3768 scan = NEXTOPER(scan);
3771 is_inf = is_inf_internal = 1;
3772 scan = regnext(scan);
3773 if (flags & SCF_DO_SUBSTR) {
3774 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3775 data->longest = &(data->longest_float);
3777 goto optimize_curly_tail;
3779 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3780 && (scan->flags == stopparen))
3785 mincount = ARG1(scan);
3786 maxcount = ARG2(scan);
3788 next = regnext(scan);
3789 if (OP(scan) == CURLYX) {
3790 I32 lp = (data ? *(data->last_closep) : 0);
3791 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3793 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3794 next_is_eval = (OP(scan) == EVAL);
3796 if (flags & SCF_DO_SUBSTR) {
3797 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3798 pos_before = data->pos_min;
3802 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3804 data->flags |= SF_IS_INF;
3806 if (flags & SCF_DO_STCLASS) {
3807 cl_init(pRExC_state, &this_class);
3808 oclass = data->start_class;
3809 data->start_class = &this_class;
3810 f |= SCF_DO_STCLASS_AND;
3811 f &= ~SCF_DO_STCLASS_OR;
3813 /* Exclude from super-linear cache processing any {n,m}
3814 regops for which the combination of input pos and regex
3815 pos is not enough information to determine if a match
3818 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3819 regex pos at the \s*, the prospects for a match depend not
3820 only on the input position but also on how many (bar\s*)
3821 repeats into the {4,8} we are. */
3822 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3823 f &= ~SCF_WHILEM_VISITED_POS;
3825 /* This will finish on WHILEM, setting scan, or on NULL: */
3826 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3827 last, data, stopparen, recursed, NULL,
3829 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3831 if (flags & SCF_DO_STCLASS)
3832 data->start_class = oclass;
3833 if (mincount == 0 || minnext == 0) {
3834 if (flags & SCF_DO_STCLASS_OR) {
3835 cl_or(pRExC_state, data->start_class, &this_class);
3837 else if (flags & SCF_DO_STCLASS_AND) {
3838 /* Switch to OR mode: cache the old value of
https://perl5.git.perl.org / perl5.git / blob