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 vWARN_dep(loc, m) STMT_START { \
552 const IV offset = loc - RExC_precomp; \
553 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
554 (int)offset, RExC_precomp, RExC_precomp + offset); \
557 #define ckWARNdep(loc,m) STMT_START { \
558 const IV offset = loc - RExC_precomp; \
559 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
561 (int)offset, RExC_precomp, RExC_precomp + offset); \
564 #define ckWARNregdep(loc,m) STMT_START { \
565 const IV offset = loc - RExC_precomp; \
566 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
568 (int)offset, RExC_precomp, RExC_precomp + offset); \
571 #define ckWARN2regdep(loc,m, a1) STMT_START { \
572 const IV offset = loc - RExC_precomp; \
573 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
575 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
578 #define ckWARN2reg(loc, m, a1) STMT_START { \
579 const IV offset = loc - RExC_precomp; \
580 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
581 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
584 #define vWARN3(loc, m, a1, a2) STMT_START { \
585 const IV offset = loc - RExC_precomp; \
586 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
587 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
590 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
591 const IV offset = loc - RExC_precomp; \
592 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
593 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
596 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
597 const IV offset = loc - RExC_precomp; \
598 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
599 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
602 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
603 const IV offset = loc - RExC_precomp; \
604 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
605 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
608 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
609 const IV offset = loc - RExC_precomp; \
610 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
611 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
615 /* Allow for side effects in s */
616 #define REGC(c,s) STMT_START { \
617 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
620 /* Macros for recording node offsets. 20001227 mjd@plover.com
621 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
622 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
623 * Element 0 holds the number n.
624 * Position is 1 indexed.
626 #ifndef RE_TRACK_PATTERN_OFFSETS
627 #define Set_Node_Offset_To_R(node,byte)
628 #define Set_Node_Offset(node,byte)
629 #define Set_Cur_Node_Offset
630 #define Set_Node_Length_To_R(node,len)
631 #define Set_Node_Length(node,len)
632 #define Set_Node_Cur_Length(node)
633 #define Node_Offset(n)
634 #define Node_Length(n)
635 #define Set_Node_Offset_Length(node,offset,len)
636 #define ProgLen(ri) ri->u.proglen
637 #define SetProgLen(ri,x) ri->u.proglen = x
639 #define ProgLen(ri) ri->u.offsets[0]
640 #define SetProgLen(ri,x) ri->u.offsets[0] = x
641 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
643 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
644 __LINE__, (int)(node), (int)(byte))); \
646 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
648 RExC_offsets[2*(node)-1] = (byte); \
653 #define Set_Node_Offset(node,byte) \
654 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
655 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
657 #define Set_Node_Length_To_R(node,len) STMT_START { \
659 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
660 __LINE__, (int)(node), (int)(len))); \
662 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
664 RExC_offsets[2*(node)] = (len); \
669 #define Set_Node_Length(node,len) \
670 Set_Node_Length_To_R((node)-RExC_emit_start, len)
671 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
672 #define Set_Node_Cur_Length(node) \
673 Set_Node_Length(node, RExC_parse - parse_start)
675 /* Get offsets and lengths */
676 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
677 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
679 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
680 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
681 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
685 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
686 #define EXPERIMENTAL_INPLACESCAN
687 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
689 #define DEBUG_STUDYDATA(str,data,depth) \
690 DEBUG_OPTIMISE_MORE_r(if(data){ \
691 PerlIO_printf(Perl_debug_log, \
692 "%*s" str "Pos:%"IVdf"/%"IVdf \
693 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
694 (int)(depth)*2, "", \
695 (IV)((data)->pos_min), \
696 (IV)((data)->pos_delta), \
697 (UV)((data)->flags), \
698 (IV)((data)->whilem_c), \
699 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
700 is_inf ? "INF " : "" \
702 if ((data)->last_found) \
703 PerlIO_printf(Perl_debug_log, \
704 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
705 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
706 SvPVX_const((data)->last_found), \
707 (IV)((data)->last_end), \
708 (IV)((data)->last_start_min), \
709 (IV)((data)->last_start_max), \
710 ((data)->longest && \
711 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
712 SvPVX_const((data)->longest_fixed), \
713 (IV)((data)->offset_fixed), \
714 ((data)->longest && \
715 (data)->longest==&((data)->longest_float)) ? "*" : "", \
716 SvPVX_const((data)->longest_float), \
717 (IV)((data)->offset_float_min), \
718 (IV)((data)->offset_float_max) \
720 PerlIO_printf(Perl_debug_log,"\n"); \
723 /* Mark that we cannot extend a found fixed substring at this point.
724 Update the longest found anchored substring and the longest found
725 floating substrings if needed. */
728 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
730 const STRLEN l = CHR_SVLEN(data->last_found);
731 const STRLEN old_l = CHR_SVLEN(*data->longest);
732 GET_RE_DEBUG_FLAGS_DECL;
734 PERL_ARGS_ASSERT_SCAN_COMMIT;
736 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
737 SvSetMagicSV(*data->longest, data->last_found);
738 if (*data->longest == data->longest_fixed) {
739 data->offset_fixed = l ? data->last_start_min : data->pos_min;
740 if (data->flags & SF_BEFORE_EOL)
742 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
744 data->flags &= ~SF_FIX_BEFORE_EOL;
745 data->minlen_fixed=minlenp;
746 data->lookbehind_fixed=0;
748 else { /* *data->longest == data->longest_float */
749 data->offset_float_min = l ? data->last_start_min : data->pos_min;
750 data->offset_float_max = (l
751 ? data->last_start_max
752 : (data->pos_delta == I32_MAX ? I32_MAX : data->pos_min + data->pos_delta));
753 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
754 data->offset_float_max = I32_MAX;
755 if (data->flags & SF_BEFORE_EOL)
757 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
759 data->flags &= ~SF_FL_BEFORE_EOL;
760 data->minlen_float=minlenp;
761 data->lookbehind_float=0;
764 SvCUR_set(data->last_found, 0);
766 SV * const sv = data->last_found;
767 if (SvUTF8(sv) && SvMAGICAL(sv)) {
768 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
774 data->flags &= ~SF_BEFORE_EOL;
775 DEBUG_STUDYDATA("commit: ",data,0);
778 /* These macros set, clear and test whether the synthetic start class ('ssc',
779 * given by the parameter) matches an empty string (EOS). This uses the
780 * 'next_off' field in the node, to save a bit in the flags field. The ssc
781 * stands alone, so there is never a next_off, so this field is otherwise
782 * unused. The EOS information is used only for compilation, but theoretically
783 * it could be passed on to the execution code. This could be used to store
784 * more than one bit of information, but only this one is currently used. */
785 #define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
786 #define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
787 #define TEST_SSC_EOS(node) cBOOL((node)->next_off)
789 /* Can match anything (initialization) */
791 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
793 PERL_ARGS_ASSERT_CL_ANYTHING;
795 ANYOF_BITMAP_SETALL(cl);
796 cl->flags = ANYOF_UNICODE_ALL;
799 /* If any portion of the regex is to operate under locale rules,
800 * initialization includes it. The reason this isn't done for all regexes
801 * is that the optimizer was written under the assumption that locale was
802 * all-or-nothing. Given the complexity and lack of documentation in the
803 * optimizer, and that there are inadequate test cases for locale, so many
804 * parts of it may not work properly, it is safest to avoid locale unless
806 if (RExC_contains_locale) {
807 ANYOF_CLASS_SETALL(cl); /* /l uses class */
808 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
811 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
815 /* Can match anything (initialization) */
817 S_cl_is_anything(const struct regnode_charclass_class *cl)
821 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
823 for (value = 0; value < ANYOF_MAX; value += 2)
824 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
826 if (!(cl->flags & ANYOF_UNICODE_ALL))
828 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
833 /* Can match anything (initialization) */
835 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
837 PERL_ARGS_ASSERT_CL_INIT;
839 Zero(cl, 1, struct regnode_charclass_class);
841 cl_anything(pRExC_state, cl);
842 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
845 /* These two functions currently do the exact same thing */
846 #define cl_init_zero S_cl_init
848 /* 'AND' a given class with another one. Can create false positives. 'cl'
849 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
850 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
852 S_cl_and(struct regnode_charclass_class *cl,
853 const struct regnode_charclass_class *and_with)
855 PERL_ARGS_ASSERT_CL_AND;
857 assert(PL_regkind[and_with->type] == ANYOF);
859 /* I (khw) am not sure all these restrictions are necessary XXX */
860 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
861 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
862 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
863 && !(and_with->flags & ANYOF_LOC_FOLD)
864 && !(cl->flags & ANYOF_LOC_FOLD)) {
867 if (and_with->flags & ANYOF_INVERT)
868 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
869 cl->bitmap[i] &= ~and_with->bitmap[i];
871 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
872 cl->bitmap[i] &= and_with->bitmap[i];
873 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
875 if (and_with->flags & ANYOF_INVERT) {
877 /* Here, the and'ed node is inverted. Get the AND of the flags that
878 * aren't affected by the inversion. Those that are affected are
879 * handled individually below */
880 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
881 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
882 cl->flags |= affected_flags;
884 /* We currently don't know how to deal with things that aren't in the
885 * bitmap, but we know that the intersection is no greater than what
886 * is already in cl, so let there be false positives that get sorted
887 * out after the synthetic start class succeeds, and the node is
888 * matched for real. */
890 /* The inversion of these two flags indicate that the resulting
891 * intersection doesn't have them */
892 if (and_with->flags & ANYOF_UNICODE_ALL) {
893 cl->flags &= ~ANYOF_UNICODE_ALL;
895 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
896 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
899 else { /* and'd node is not inverted */
900 U8 outside_bitmap_but_not_utf8; /* Temp variable */
902 if (! ANYOF_NONBITMAP(and_with)) {
904 /* Here 'and_with' doesn't match anything outside the bitmap
905 * (except possibly ANYOF_UNICODE_ALL), which means the
906 * intersection can't either, except for ANYOF_UNICODE_ALL, in
907 * which case we don't know what the intersection is, but it's no
908 * greater than what cl already has, so can just leave it alone,
909 * with possible false positives */
910 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
911 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
912 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
915 else if (! ANYOF_NONBITMAP(cl)) {
917 /* Here, 'and_with' does match something outside the bitmap, and cl
918 * doesn't have a list of things to match outside the bitmap. If
919 * cl can match all code points above 255, the intersection will
920 * be those above-255 code points that 'and_with' matches. If cl
921 * can't match all Unicode code points, it means that it can't
922 * match anything outside the bitmap (since the 'if' that got us
923 * into this block tested for that), so we leave the bitmap empty.
925 if (cl->flags & ANYOF_UNICODE_ALL) {
926 ARG_SET(cl, ARG(and_with));
928 /* and_with's ARG may match things that don't require UTF8.
929 * And now cl's will too, in spite of this being an 'and'. See
930 * the comments below about the kludge */
931 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
935 /* Here, both 'and_with' and cl match something outside the
936 * bitmap. Currently we do not do the intersection, so just match
937 * whatever cl had at the beginning. */
941 /* Take the intersection of the two sets of flags. However, the
942 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
943 * kludge around the fact that this flag is not treated like the others
944 * which are initialized in cl_anything(). The way the optimizer works
945 * is that the synthetic start class (SSC) is initialized to match
946 * anything, and then the first time a real node is encountered, its
947 * values are AND'd with the SSC's with the result being the values of
948 * the real node. However, there are paths through the optimizer where
949 * the AND never gets called, so those initialized bits are set
950 * inappropriately, which is not usually a big deal, as they just cause
951 * false positives in the SSC, which will just mean a probably
952 * imperceptible slow down in execution. However this bit has a
953 * higher false positive consequence in that it can cause utf8.pm,
954 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
955 * bigger slowdown and also causes significant extra memory to be used.
956 * In order to prevent this, the code now takes a different tack. The
957 * bit isn't set unless some part of the regular expression needs it,
958 * but once set it won't get cleared. This means that these extra
959 * modules won't get loaded unless there was some path through the
960 * pattern that would have required them anyway, and so any false
961 * positives that occur by not ANDing them out when they could be
962 * aren't as severe as they would be if we treated this bit like all
964 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
965 & ANYOF_NONBITMAP_NON_UTF8;
966 cl->flags &= and_with->flags;
967 cl->flags |= outside_bitmap_but_not_utf8;
971 /* 'OR' a given class with another one. Can create false positives. 'cl'
972 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
973 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
975 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
977 PERL_ARGS_ASSERT_CL_OR;
979 if (or_with->flags & ANYOF_INVERT) {
981 /* Here, the or'd node is to be inverted. This means we take the
982 * complement of everything not in the bitmap, but currently we don't
983 * know what that is, so give up and match anything */
984 if (ANYOF_NONBITMAP(or_with)) {
985 cl_anything(pRExC_state, cl);
988 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
989 * <= (B1 | !B2) | (CL1 | !CL2)
990 * which is wasteful if CL2 is small, but we ignore CL2:
991 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
992 * XXXX Can we handle case-fold? Unclear:
993 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
994 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
996 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
997 && !(or_with->flags & ANYOF_LOC_FOLD)
998 && !(cl->flags & ANYOF_LOC_FOLD) ) {
1001 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1002 cl->bitmap[i] |= ~or_with->bitmap[i];
1003 } /* XXXX: logic is complicated otherwise */
1005 cl_anything(pRExC_state, cl);
1008 /* And, we can just take the union of the flags that aren't affected
1009 * by the inversion */
1010 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1012 /* For the remaining flags:
1013 ANYOF_UNICODE_ALL and inverted means to not match anything above
1014 255, which means that the union with cl should just be
1015 what cl has in it, so can ignore this flag
1016 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
1017 is 127-255 to match them, but then invert that, so the
1018 union with cl should just be what cl has in it, so can
1021 } else { /* 'or_with' is not inverted */
1022 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1023 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1024 && (!(or_with->flags & ANYOF_LOC_FOLD)
1025 || (cl->flags & ANYOF_LOC_FOLD)) ) {
1028 /* OR char bitmap and class bitmap separately */
1029 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1030 cl->bitmap[i] |= or_with->bitmap[i];
1031 if (or_with->flags & ANYOF_CLASS) {
1032 ANYOF_CLASS_OR(or_with, cl);
1035 else { /* XXXX: logic is complicated, leave it along for a moment. */
1036 cl_anything(pRExC_state, cl);
1039 if (ANYOF_NONBITMAP(or_with)) {
1041 /* Use the added node's outside-the-bit-map match if there isn't a
1042 * conflict. If there is a conflict (both nodes match something
1043 * outside the bitmap, but what they match outside is not the same
1044 * pointer, and hence not easily compared until XXX we extend
1045 * inversion lists this far), give up and allow the start class to
1046 * match everything outside the bitmap. If that stuff is all above
1047 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1048 if (! ANYOF_NONBITMAP(cl)) {
1049 ARG_SET(cl, ARG(or_with));
1051 else if (ARG(cl) != ARG(or_with)) {
1053 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1054 cl_anything(pRExC_state, cl);
1057 cl->flags |= ANYOF_UNICODE_ALL;
1062 /* Take the union */
1063 cl->flags |= or_with->flags;
1067 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1068 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1069 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1070 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1075 dump_trie(trie,widecharmap,revcharmap)
1076 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1077 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1079 These routines dump out a trie in a somewhat readable format.
1080 The _interim_ variants are used for debugging the interim
1081 tables that are used to generate the final compressed
1082 representation which is what dump_trie expects.
1084 Part of the reason for their existence is to provide a form
1085 of documentation as to how the different representations function.
1090 Dumps the final compressed table form of the trie to Perl_debug_log.
1091 Used for debugging make_trie().
1095 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1096 AV *revcharmap, U32 depth)
1099 SV *sv=sv_newmortal();
1100 int colwidth= widecharmap ? 6 : 4;
1102 GET_RE_DEBUG_FLAGS_DECL;
1104 PERL_ARGS_ASSERT_DUMP_TRIE;
1106 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1107 (int)depth * 2 + 2,"",
1108 "Match","Base","Ofs" );
1110 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1111 SV ** const tmp = av_fetch( revcharmap, state, 0);
1113 PerlIO_printf( Perl_debug_log, "%*s",
1115 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1116 PL_colors[0], PL_colors[1],
1117 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1118 PERL_PV_ESCAPE_FIRSTCHAR
1123 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1124 (int)depth * 2 + 2,"");
1126 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1127 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1128 PerlIO_printf( Perl_debug_log, "\n");
1130 for( state = 1 ; state < trie->statecount ; state++ ) {
1131 const U32 base = trie->states[ state ].trans.base;
1133 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1135 if ( trie->states[ state ].wordnum ) {
1136 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1138 PerlIO_printf( Perl_debug_log, "%6s", "" );
1141 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1146 while( ( base + ofs < trie->uniquecharcount ) ||
1147 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1148 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1151 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1153 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1154 if ( ( base + ofs >= trie->uniquecharcount ) &&
1155 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1156 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1158 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1160 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1162 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1166 PerlIO_printf( Perl_debug_log, "]");
1169 PerlIO_printf( Perl_debug_log, "\n" );
1171 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1172 for (word=1; word <= trie->wordcount; word++) {
1173 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1174 (int)word, (int)(trie->wordinfo[word].prev),
1175 (int)(trie->wordinfo[word].len));
1177 PerlIO_printf(Perl_debug_log, "\n" );
1180 Dumps a fully constructed but uncompressed trie in list form.
1181 List tries normally only are used for construction when the number of
1182 possible chars (trie->uniquecharcount) is very high.
1183 Used for debugging make_trie().
1186 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1187 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1191 SV *sv=sv_newmortal();
1192 int colwidth= widecharmap ? 6 : 4;
1193 GET_RE_DEBUG_FLAGS_DECL;
1195 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1197 /* print out the table precompression. */
1198 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1199 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1200 "------:-----+-----------------\n" );
1202 for( state=1 ; state < next_alloc ; state ++ ) {
1205 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1206 (int)depth * 2 + 2,"", (UV)state );
1207 if ( ! trie->states[ state ].wordnum ) {
1208 PerlIO_printf( Perl_debug_log, "%5s| ","");
1210 PerlIO_printf( Perl_debug_log, "W%4x| ",
1211 trie->states[ state ].wordnum
1214 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1215 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1217 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1219 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1220 PL_colors[0], PL_colors[1],
1221 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1222 PERL_PV_ESCAPE_FIRSTCHAR
1224 TRIE_LIST_ITEM(state,charid).forid,
1225 (UV)TRIE_LIST_ITEM(state,charid).newstate
1228 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1229 (int)((depth * 2) + 14), "");
1232 PerlIO_printf( Perl_debug_log, "\n");
1237 Dumps a fully constructed but uncompressed trie in table form.
1238 This is the normal DFA style state transition table, with a few
1239 twists to facilitate compression later.
1240 Used for debugging make_trie().
1243 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1244 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1249 SV *sv=sv_newmortal();
1250 int colwidth= widecharmap ? 6 : 4;
1251 GET_RE_DEBUG_FLAGS_DECL;
1253 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1256 print out the table precompression so that we can do a visual check
1257 that they are identical.
1260 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1262 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1263 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1265 PerlIO_printf( Perl_debug_log, "%*s",
1267 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1268 PL_colors[0], PL_colors[1],
1269 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1270 PERL_PV_ESCAPE_FIRSTCHAR
1276 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1278 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1279 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1282 PerlIO_printf( Perl_debug_log, "\n" );
1284 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1286 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1287 (int)depth * 2 + 2,"",
1288 (UV)TRIE_NODENUM( state ) );
1290 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1291 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1293 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1295 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1297 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1298 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1300 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1301 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1309 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1310 startbranch: the first branch in the whole branch sequence
1311 first : start branch of sequence of branch-exact nodes.
1312 May be the same as startbranch
1313 last : Thing following the last branch.
1314 May be the same as tail.
1315 tail : item following the branch sequence
1316 count : words in the sequence
1317 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1318 depth : indent depth
1320 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1322 A trie is an N'ary tree where the branches are determined by digital
1323 decomposition of the key. IE, at the root node you look up the 1st character and
1324 follow that branch repeat until you find the end of the branches. Nodes can be
1325 marked as "accepting" meaning they represent a complete word. Eg:
1329 would convert into the following structure. Numbers represent states, letters
1330 following numbers represent valid transitions on the letter from that state, if
1331 the number is in square brackets it represents an accepting state, otherwise it
1332 will be in parenthesis.
1334 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1338 (1) +-i->(6)-+-s->[7]
1340 +-s->(3)-+-h->(4)-+-e->[5]
1342 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1344 This shows that when matching against the string 'hers' we will begin at state 1
1345 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1346 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1347 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1348 single traverse. We store a mapping from accepting to state to which word was
1349 matched, and then when we have multiple possibilities we try to complete the
1350 rest of the regex in the order in which they occured in the alternation.
1352 The only prior NFA like behaviour that would be changed by the TRIE support is
1353 the silent ignoring of duplicate alternations which are of the form:
1355 / (DUPE|DUPE) X? (?{ ... }) Y /x
1357 Thus EVAL blocks following a trie may be called a different number of times with
1358 and without the optimisation. With the optimisations dupes will be silently
1359 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1360 the following demonstrates:
1362 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1364 which prints out 'word' three times, but
1366 'words'=~/(word|word|word)(?{ print $1 })S/
1368 which doesnt print it out at all. This is due to other optimisations kicking in.
1370 Example of what happens on a structural level:
1372 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1374 1: CURLYM[1] {1,32767}(18)
1385 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1386 and should turn into:
1388 1: CURLYM[1] {1,32767}(18)
1390 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1398 Cases where tail != last would be like /(?foo|bar)baz/:
1408 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1409 and would end up looking like:
1412 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1419 d = uvuni_to_utf8_flags(d, uv, 0);
1421 is the recommended Unicode-aware way of saying
1426 #define TRIE_STORE_REVCHAR(val) \
1429 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1430 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1431 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1432 SvCUR_set(zlopp, kapow - flrbbbbb); \
1435 av_push(revcharmap, zlopp); \
1437 char ooooff = (char)val; \
1438 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1442 #define TRIE_READ_CHAR STMT_START { \
1445 /* if it is UTF then it is either already folded, or does not need folding */ \
1446 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1448 else if (folder == PL_fold_latin1) { \
1449 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1450 if ( foldlen > 0 ) { \
1451 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1457 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1458 skiplen = UNISKIP(uvc); \
1459 foldlen -= skiplen; \
1460 scan = foldbuf + skiplen; \
1463 /* raw data, will be folded later if needed */ \
1471 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1472 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1473 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1474 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1476 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1477 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1478 TRIE_LIST_CUR( state )++; \
1481 #define TRIE_LIST_NEW(state) STMT_START { \
1482 Newxz( trie->states[ state ].trans.list, \
1483 4, reg_trie_trans_le ); \
1484 TRIE_LIST_CUR( state ) = 1; \
1485 TRIE_LIST_LEN( state ) = 4; \
1488 #define TRIE_HANDLE_WORD(state) STMT_START { \
1489 U16 dupe= trie->states[ state ].wordnum; \
1490 regnode * const noper_next = regnext( noper ); \
1493 /* store the word for dumping */ \
1495 if (OP(noper) != NOTHING) \
1496 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1498 tmp = newSVpvn_utf8( "", 0, UTF ); \
1499 av_push( trie_words, tmp ); \
1503 trie->wordinfo[curword].prev = 0; \
1504 trie->wordinfo[curword].len = wordlen; \
1505 trie->wordinfo[curword].accept = state; \
1507 if ( noper_next < tail ) { \
1509 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1510 trie->jump[curword] = (U16)(noper_next - convert); \
1512 jumper = noper_next; \
1514 nextbranch= regnext(cur); \
1518 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1519 /* chain, so that when the bits of chain are later */\
1520 /* linked together, the dups appear in the chain */\
1521 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1522 trie->wordinfo[dupe].prev = curword; \
1524 /* we haven't inserted this word yet. */ \
1525 trie->states[ state ].wordnum = curword; \
1530 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1531 ( ( base + charid >= ucharcount \
1532 && base + charid < ubound \
1533 && state == trie->trans[ base - ucharcount + charid ].check \
1534 && trie->trans[ base - ucharcount + charid ].next ) \
1535 ? trie->trans[ base - ucharcount + charid ].next \
1536 : ( state==1 ? special : 0 ) \
1540 #define MADE_JUMP_TRIE 2
1541 #define MADE_EXACT_TRIE 4
1544 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1547 /* first pass, loop through and scan words */
1548 reg_trie_data *trie;
1549 HV *widecharmap = NULL;
1550 AV *revcharmap = newAV();
1552 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1557 regnode *jumper = NULL;
1558 regnode *nextbranch = NULL;
1559 regnode *convert = NULL;
1560 U32 *prev_states; /* temp array mapping each state to previous one */
1561 /* we just use folder as a flag in utf8 */
1562 const U8 * folder = NULL;
1565 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1566 AV *trie_words = NULL;
1567 /* along with revcharmap, this only used during construction but both are
1568 * useful during debugging so we store them in the struct when debugging.
1571 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1572 STRLEN trie_charcount=0;
1574 SV *re_trie_maxbuff;
1575 GET_RE_DEBUG_FLAGS_DECL;
1577 PERL_ARGS_ASSERT_MAKE_TRIE;
1579 PERL_UNUSED_ARG(depth);
1586 case EXACTFU_TRICKYFOLD:
1587 case EXACTFU: folder = PL_fold_latin1; break;
1588 case EXACTF: folder = PL_fold; break;
1589 case EXACTFL: folder = PL_fold_locale; break;
1590 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1593 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1595 trie->startstate = 1;
1596 trie->wordcount = word_count;
1597 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1598 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1600 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1601 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1602 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1605 trie_words = newAV();
1608 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1609 if (!SvIOK(re_trie_maxbuff)) {
1610 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1612 DEBUG_TRIE_COMPILE_r({
1613 PerlIO_printf( Perl_debug_log,
1614 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1615 (int)depth * 2 + 2, "",
1616 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1617 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1621 /* Find the node we are going to overwrite */
1622 if ( first == startbranch && OP( last ) != BRANCH ) {
1623 /* whole branch chain */
1626 /* branch sub-chain */
1627 convert = NEXTOPER( first );
1630 /* -- First loop and Setup --
1632 We first traverse the branches and scan each word to determine if it
1633 contains widechars, and how many unique chars there are, this is
1634 important as we have to build a table with at least as many columns as we
1637 We use an array of integers to represent the character codes 0..255
1638 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1639 native representation of the character value as the key and IV's for the
1642 *TODO* If we keep track of how many times each character is used we can
1643 remap the columns so that the table compression later on is more
1644 efficient in terms of memory by ensuring the most common value is in the
1645 middle and the least common are on the outside. IMO this would be better
1646 than a most to least common mapping as theres a decent chance the most
1647 common letter will share a node with the least common, meaning the node
1648 will not be compressible. With a middle is most common approach the worst
1649 case is when we have the least common nodes twice.
1653 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1654 regnode *noper = NEXTOPER( cur );
1655 const U8 *uc = (U8*)STRING( noper );
1656 const U8 *e = uc + STR_LEN( noper );
1658 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1660 const U8 *scan = (U8*)NULL;
1661 U32 wordlen = 0; /* required init */
1663 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1665 if (OP(noper) == NOTHING) {
1666 regnode *noper_next= regnext(noper);
1667 if (noper_next != tail && OP(noper_next) == flags) {
1669 uc= (U8*)STRING(noper);
1670 e= uc + STR_LEN(noper);
1671 trie->minlen= STR_LEN(noper);
1678 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1679 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1680 regardless of encoding */
1681 if (OP( noper ) == EXACTFU_SS) {
1682 /* false positives are ok, so just set this */
1683 TRIE_BITMAP_SET(trie,0xDF);
1686 for ( ; uc < e ; uc += len ) {
1687 TRIE_CHARCOUNT(trie)++;
1692 U8 folded= folder[ (U8) uvc ];
1693 if ( !trie->charmap[ folded ] ) {
1694 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1695 TRIE_STORE_REVCHAR( folded );
1698 if ( !trie->charmap[ uvc ] ) {
1699 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1700 TRIE_STORE_REVCHAR( uvc );
1703 /* store the codepoint in the bitmap, and its folded
1705 TRIE_BITMAP_SET(trie, uvc);
1707 /* store the folded codepoint */
1708 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1711 /* store first byte of utf8 representation of
1712 variant codepoints */
1713 if (! UNI_IS_INVARIANT(uvc)) {
1714 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1717 set_bit = 0; /* We've done our bit :-) */
1722 widecharmap = newHV();
1724 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1727 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1729 if ( !SvTRUE( *svpp ) ) {
1730 sv_setiv( *svpp, ++trie->uniquecharcount );
1731 TRIE_STORE_REVCHAR(uvc);
1735 if( cur == first ) {
1736 trie->minlen = chars;
1737 trie->maxlen = chars;
1738 } else if (chars < trie->minlen) {
1739 trie->minlen = chars;
1740 } else if (chars > trie->maxlen) {
1741 trie->maxlen = chars;
1743 if (OP( noper ) == EXACTFU_SS) {
1744 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1745 if (trie->minlen > 1)
1748 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1749 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1750 * - We assume that any such sequence might match a 2 byte string */
1751 if (trie->minlen > 2 )
1755 } /* end first pass */
1756 DEBUG_TRIE_COMPILE_r(
1757 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1758 (int)depth * 2 + 2,"",
1759 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1760 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1761 (int)trie->minlen, (int)trie->maxlen )
1765 We now know what we are dealing with in terms of unique chars and
1766 string sizes so we can calculate how much memory a naive
1767 representation using a flat table will take. If it's over a reasonable
1768 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1769 conservative but potentially much slower representation using an array
1772 At the end we convert both representations into the same compressed
1773 form that will be used in regexec.c for matching with. The latter
1774 is a form that cannot be used to construct with but has memory
1775 properties similar to the list form and access properties similar
1776 to the table form making it both suitable for fast searches and
1777 small enough that its feasable to store for the duration of a program.
1779 See the comment in the code where the compressed table is produced
1780 inplace from the flat tabe representation for an explanation of how
1781 the compression works.
1786 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1789 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1791 Second Pass -- Array Of Lists Representation
1793 Each state will be represented by a list of charid:state records
1794 (reg_trie_trans_le) the first such element holds the CUR and LEN
1795 points of the allocated array. (See defines above).
1797 We build the initial structure using the lists, and then convert
1798 it into the compressed table form which allows faster lookups
1799 (but cant be modified once converted).
1802 STRLEN transcount = 1;
1804 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1805 "%*sCompiling trie using list compiler\n",
1806 (int)depth * 2 + 2, ""));
1808 trie->states = (reg_trie_state *)
1809 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1810 sizeof(reg_trie_state) );
1814 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1816 regnode *noper = NEXTOPER( cur );
1817 U8 *uc = (U8*)STRING( noper );
1818 const U8 *e = uc + STR_LEN( noper );
1819 U32 state = 1; /* required init */
1820 U16 charid = 0; /* sanity init */
1821 U8 *scan = (U8*)NULL; /* sanity init */
1822 STRLEN foldlen = 0; /* required init */
1823 U32 wordlen = 0; /* required init */
1824 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1827 if (OP(noper) == NOTHING) {
1828 regnode *noper_next= regnext(noper);
1829 if (noper_next != tail && OP(noper_next) == flags) {
1831 uc= (U8*)STRING(noper);
1832 e= uc + STR_LEN(noper);
1836 if (OP(noper) != NOTHING) {
1837 for ( ; uc < e ; uc += len ) {
1842 charid = trie->charmap[ uvc ];
1844 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1848 charid=(U16)SvIV( *svpp );
1851 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1858 if ( !trie->states[ state ].trans.list ) {
1859 TRIE_LIST_NEW( state );
1861 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1862 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1863 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1868 newstate = next_alloc++;
1869 prev_states[newstate] = state;
1870 TRIE_LIST_PUSH( state, charid, newstate );
1875 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1879 TRIE_HANDLE_WORD(state);
1881 } /* end second pass */
1883 /* next alloc is the NEXT state to be allocated */
1884 trie->statecount = next_alloc;
1885 trie->states = (reg_trie_state *)
1886 PerlMemShared_realloc( trie->states,
1888 * sizeof(reg_trie_state) );
1890 /* and now dump it out before we compress it */
1891 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1892 revcharmap, next_alloc,
1896 trie->trans = (reg_trie_trans *)
1897 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1904 for( state=1 ; state < next_alloc ; state ++ ) {
1908 DEBUG_TRIE_COMPILE_MORE_r(
1909 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1913 if (trie->states[state].trans.list) {
1914 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1918 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1919 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1920 if ( forid < minid ) {
1922 } else if ( forid > maxid ) {
1926 if ( transcount < tp + maxid - minid + 1) {
1928 trie->trans = (reg_trie_trans *)
1929 PerlMemShared_realloc( trie->trans,
1931 * sizeof(reg_trie_trans) );
1932 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1934 base = trie->uniquecharcount + tp - minid;
1935 if ( maxid == minid ) {
1937 for ( ; zp < tp ; zp++ ) {
1938 if ( ! trie->trans[ zp ].next ) {
1939 base = trie->uniquecharcount + zp - minid;
1940 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1941 trie->trans[ zp ].check = state;
1947 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1948 trie->trans[ tp ].check = state;
1953 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1954 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1955 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1956 trie->trans[ tid ].check = state;
1958 tp += ( maxid - minid + 1 );
1960 Safefree(trie->states[ state ].trans.list);
1963 DEBUG_TRIE_COMPILE_MORE_r(
1964 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1967 trie->states[ state ].trans.base=base;
1969 trie->lasttrans = tp + 1;
1973 Second Pass -- Flat Table Representation.
1975 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1976 We know that we will need Charcount+1 trans at most to store the data
1977 (one row per char at worst case) So we preallocate both structures
1978 assuming worst case.
1980 We then construct the trie using only the .next slots of the entry
1983 We use the .check field of the first entry of the node temporarily to
1984 make compression both faster and easier by keeping track of how many non
1985 zero fields are in the node.
1987 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1990 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1991 number representing the first entry of the node, and state as a
1992 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1993 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1994 are 2 entrys per node. eg:
2002 The table is internally in the right hand, idx form. However as we also
2003 have to deal with the states array which is indexed by nodenum we have to
2004 use TRIE_NODENUM() to convert.
2007 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2008 "%*sCompiling trie using table compiler\n",
2009 (int)depth * 2 + 2, ""));
2011 trie->trans = (reg_trie_trans *)
2012 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2013 * trie->uniquecharcount + 1,
2014 sizeof(reg_trie_trans) );
2015 trie->states = (reg_trie_state *)
2016 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2017 sizeof(reg_trie_state) );
2018 next_alloc = trie->uniquecharcount + 1;
2021 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2023 regnode *noper = NEXTOPER( cur );
2024 const U8 *uc = (U8*)STRING( noper );
2025 const U8 *e = uc + STR_LEN( noper );
2027 U32 state = 1; /* required init */
2029 U16 charid = 0; /* sanity init */
2030 U32 accept_state = 0; /* sanity init */
2031 U8 *scan = (U8*)NULL; /* sanity init */
2033 STRLEN foldlen = 0; /* required init */
2034 U32 wordlen = 0; /* required init */
2036 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2038 if (OP(noper) == NOTHING) {
2039 regnode *noper_next= regnext(noper);
2040 if (noper_next != tail && OP(noper_next) == flags) {
2042 uc= (U8*)STRING(noper);
2043 e= uc + STR_LEN(noper);
2047 if ( OP(noper) != NOTHING ) {
2048 for ( ; uc < e ; uc += len ) {
2053 charid = trie->charmap[ uvc ];
2055 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2056 charid = svpp ? (U16)SvIV(*svpp) : 0;
2060 if ( !trie->trans[ state + charid ].next ) {
2061 trie->trans[ state + charid ].next = next_alloc;
2062 trie->trans[ state ].check++;
2063 prev_states[TRIE_NODENUM(next_alloc)]
2064 = TRIE_NODENUM(state);
2065 next_alloc += trie->uniquecharcount;
2067 state = trie->trans[ state + charid ].next;
2069 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2071 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2074 accept_state = TRIE_NODENUM( state );
2075 TRIE_HANDLE_WORD(accept_state);
2077 } /* end second pass */
2079 /* and now dump it out before we compress it */
2080 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2082 next_alloc, depth+1));
2086 * Inplace compress the table.*
2088 For sparse data sets the table constructed by the trie algorithm will
2089 be mostly 0/FAIL transitions or to put it another way mostly empty.
2090 (Note that leaf nodes will not contain any transitions.)
2092 This algorithm compresses the tables by eliminating most such
2093 transitions, at the cost of a modest bit of extra work during lookup:
2095 - Each states[] entry contains a .base field which indicates the
2096 index in the state[] array wheres its transition data is stored.
2098 - If .base is 0 there are no valid transitions from that node.
2100 - If .base is nonzero then charid is added to it to find an entry in
2103 -If trans[states[state].base+charid].check!=state then the
2104 transition is taken to be a 0/Fail transition. Thus if there are fail
2105 transitions at the front of the node then the .base offset will point
2106 somewhere inside the previous nodes data (or maybe even into a node
2107 even earlier), but the .check field determines if the transition is
2111 The following process inplace converts the table to the compressed
2112 table: We first do not compress the root node 1,and mark all its
2113 .check pointers as 1 and set its .base pointer as 1 as well. This
2114 allows us to do a DFA construction from the compressed table later,
2115 and ensures that any .base pointers we calculate later are greater
2118 - We set 'pos' to indicate the first entry of the second node.
2120 - We then iterate over the columns of the node, finding the first and
2121 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2122 and set the .check pointers accordingly, and advance pos
2123 appropriately and repreat for the next node. Note that when we copy
2124 the next pointers we have to convert them from the original
2125 NODEIDX form to NODENUM form as the former is not valid post
2128 - If a node has no transitions used we mark its base as 0 and do not
2129 advance the pos pointer.
2131 - If a node only has one transition we use a second pointer into the
2132 structure to fill in allocated fail transitions from other states.
2133 This pointer is independent of the main pointer and scans forward
2134 looking for null transitions that are allocated to a state. When it
2135 finds one it writes the single transition into the "hole". If the
2136 pointer doesnt find one the single transition is appended as normal.
2138 - Once compressed we can Renew/realloc the structures to release the
2141 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2142 specifically Fig 3.47 and the associated pseudocode.
2146 const U32 laststate = TRIE_NODENUM( next_alloc );
2149 trie->statecount = laststate;
2151 for ( state = 1 ; state < laststate ; state++ ) {
2153 const U32 stateidx = TRIE_NODEIDX( state );
2154 const U32 o_used = trie->trans[ stateidx ].check;
2155 U32 used = trie->trans[ stateidx ].check;
2156 trie->trans[ stateidx ].check = 0;
2158 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2159 if ( flag || trie->trans[ stateidx + charid ].next ) {
2160 if ( trie->trans[ stateidx + charid ].next ) {
2162 for ( ; zp < pos ; zp++ ) {
2163 if ( ! trie->trans[ zp ].next ) {
2167 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2168 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2169 trie->trans[ zp ].check = state;
2170 if ( ++zp > pos ) pos = zp;
2177 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2179 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2180 trie->trans[ pos ].check = state;
2185 trie->lasttrans = pos + 1;
2186 trie->states = (reg_trie_state *)
2187 PerlMemShared_realloc( trie->states, laststate
2188 * sizeof(reg_trie_state) );
2189 DEBUG_TRIE_COMPILE_MORE_r(
2190 PerlIO_printf( Perl_debug_log,
2191 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2192 (int)depth * 2 + 2,"",
2193 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2196 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2199 } /* end table compress */
2201 DEBUG_TRIE_COMPILE_MORE_r(
2202 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2203 (int)depth * 2 + 2, "",
2204 (UV)trie->statecount,
2205 (UV)trie->lasttrans)
2207 /* resize the trans array to remove unused space */
2208 trie->trans = (reg_trie_trans *)
2209 PerlMemShared_realloc( trie->trans, trie->lasttrans
2210 * sizeof(reg_trie_trans) );
2212 { /* Modify the program and insert the new TRIE node */
2213 U8 nodetype =(U8)(flags & 0xFF);
2217 regnode *optimize = NULL;
2218 #ifdef RE_TRACK_PATTERN_OFFSETS
2221 U32 mjd_nodelen = 0;
2222 #endif /* RE_TRACK_PATTERN_OFFSETS */
2223 #endif /* DEBUGGING */
2225 This means we convert either the first branch or the first Exact,
2226 depending on whether the thing following (in 'last') is a branch
2227 or not and whther first is the startbranch (ie is it a sub part of
2228 the alternation or is it the whole thing.)
2229 Assuming its a sub part we convert the EXACT otherwise we convert
2230 the whole branch sequence, including the first.
2232 /* Find the node we are going to overwrite */
2233 if ( first != startbranch || OP( last ) == BRANCH ) {
2234 /* branch sub-chain */
2235 NEXT_OFF( first ) = (U16)(last - first);
2236 #ifdef RE_TRACK_PATTERN_OFFSETS
2238 mjd_offset= Node_Offset((convert));
2239 mjd_nodelen= Node_Length((convert));
2242 /* whole branch chain */
2244 #ifdef RE_TRACK_PATTERN_OFFSETS
2247 const regnode *nop = NEXTOPER( convert );
2248 mjd_offset= Node_Offset((nop));
2249 mjd_nodelen= Node_Length((nop));
2253 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2254 (int)depth * 2 + 2, "",
2255 (UV)mjd_offset, (UV)mjd_nodelen)
2258 /* But first we check to see if there is a common prefix we can
2259 split out as an EXACT and put in front of the TRIE node. */
2260 trie->startstate= 1;
2261 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2263 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2267 const U32 base = trie->states[ state ].trans.base;
2269 if ( trie->states[state].wordnum )
2272 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2273 if ( ( base + ofs >= trie->uniquecharcount ) &&
2274 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2275 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2277 if ( ++count > 1 ) {
2278 SV **tmp = av_fetch( revcharmap, ofs, 0);
2279 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2280 if ( state == 1 ) break;
2282 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2284 PerlIO_printf(Perl_debug_log,
2285 "%*sNew Start State=%"UVuf" Class: [",
2286 (int)depth * 2 + 2, "",
2289 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2290 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2292 TRIE_BITMAP_SET(trie,*ch);
2294 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2296 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2300 TRIE_BITMAP_SET(trie,*ch);
2302 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2303 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2309 SV **tmp = av_fetch( revcharmap, idx, 0);
2311 char *ch = SvPV( *tmp, len );
2313 SV *sv=sv_newmortal();
2314 PerlIO_printf( Perl_debug_log,
2315 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2316 (int)depth * 2 + 2, "",
2318 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2319 PL_colors[0], PL_colors[1],
2320 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2321 PERL_PV_ESCAPE_FIRSTCHAR
2326 OP( convert ) = nodetype;
2327 str=STRING(convert);
2330 STR_LEN(convert) += len;
2336 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2341 trie->prefixlen = (state-1);
2343 regnode *n = convert+NODE_SZ_STR(convert);
2344 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2345 trie->startstate = state;
2346 trie->minlen -= (state - 1);
2347 trie->maxlen -= (state - 1);
2349 /* At least the UNICOS C compiler choked on this
2350 * being argument to DEBUG_r(), so let's just have
2353 #ifdef PERL_EXT_RE_BUILD
2359 regnode *fix = convert;
2360 U32 word = trie->wordcount;
2362 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2363 while( ++fix < n ) {
2364 Set_Node_Offset_Length(fix, 0, 0);
2367 SV ** const tmp = av_fetch( trie_words, word, 0 );
2369 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2370 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2372 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2380 NEXT_OFF(convert) = (U16)(tail - convert);
2381 DEBUG_r(optimize= n);
2387 if ( trie->maxlen ) {
2388 NEXT_OFF( convert ) = (U16)(tail - convert);
2389 ARG_SET( convert, data_slot );
2390 /* Store the offset to the first unabsorbed branch in
2391 jump[0], which is otherwise unused by the jump logic.
2392 We use this when dumping a trie and during optimisation. */
2394 trie->jump[0] = (U16)(nextbranch - convert);
2396 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2397 * and there is a bitmap
2398 * and the first "jump target" node we found leaves enough room
2399 * then convert the TRIE node into a TRIEC node, with the bitmap
2400 * embedded inline in the opcode - this is hypothetically faster.
2402 if ( !trie->states[trie->startstate].wordnum
2404 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2406 OP( convert ) = TRIEC;
2407 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2408 PerlMemShared_free(trie->bitmap);
2411 OP( convert ) = TRIE;
2413 /* store the type in the flags */
2414 convert->flags = nodetype;
2418 + regarglen[ OP( convert ) ];
2420 /* XXX We really should free up the resource in trie now,
2421 as we won't use them - (which resources?) dmq */
2423 /* needed for dumping*/
2424 DEBUG_r(if (optimize) {
2425 regnode *opt = convert;
2427 while ( ++opt < optimize) {
2428 Set_Node_Offset_Length(opt,0,0);
2431 Try to clean up some of the debris left after the
2434 while( optimize < jumper ) {
2435 mjd_nodelen += Node_Length((optimize));
2436 OP( optimize ) = OPTIMIZED;
2437 Set_Node_Offset_Length(optimize,0,0);
2440 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2442 } /* end node insert */
2444 /* Finish populating the prev field of the wordinfo array. Walk back
2445 * from each accept state until we find another accept state, and if
2446 * so, point the first word's .prev field at the second word. If the
2447 * second already has a .prev field set, stop now. This will be the
2448 * case either if we've already processed that word's accept state,
2449 * or that state had multiple words, and the overspill words were
2450 * already linked up earlier.
2457 for (word=1; word <= trie->wordcount; word++) {
2459 if (trie->wordinfo[word].prev)
2461 state = trie->wordinfo[word].accept;
2463 state = prev_states[state];
2466 prev = trie->states[state].wordnum;
2470 trie->wordinfo[word].prev = prev;
2472 Safefree(prev_states);
2476 /* and now dump out the compressed format */
2477 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2479 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2481 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2482 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2484 SvREFCNT_dec_NN(revcharmap);
2488 : trie->startstate>1
2494 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2496 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2498 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2499 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2502 We find the fail state for each state in the trie, this state is the longest proper
2503 suffix of the current state's 'word' that is also a proper prefix of another word in our
2504 trie. State 1 represents the word '' and is thus the default fail state. This allows
2505 the DFA not to have to restart after its tried and failed a word at a given point, it
2506 simply continues as though it had been matching the other word in the first place.
2508 'abcdgu'=~/abcdefg|cdgu/
2509 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2510 fail, which would bring us to the state representing 'd' in the second word where we would
2511 try 'g' and succeed, proceeding to match 'cdgu'.
2513 /* add a fail transition */
2514 const U32 trie_offset = ARG(source);
2515 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2517 const U32 ucharcount = trie->uniquecharcount;
2518 const U32 numstates = trie->statecount;
2519 const U32 ubound = trie->lasttrans + ucharcount;
2523 U32 base = trie->states[ 1 ].trans.base;
2526 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2527 GET_RE_DEBUG_FLAGS_DECL;
2529 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2531 PERL_UNUSED_ARG(depth);
2535 ARG_SET( stclass, data_slot );
2536 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2537 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2538 aho->trie=trie_offset;
2539 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2540 Copy( trie->states, aho->states, numstates, reg_trie_state );
2541 Newxz( q, numstates, U32);
2542 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2545 /* initialize fail[0..1] to be 1 so that we always have
2546 a valid final fail state */
2547 fail[ 0 ] = fail[ 1 ] = 1;
2549 for ( charid = 0; charid < ucharcount ; charid++ ) {
2550 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2552 q[ q_write ] = newstate;
2553 /* set to point at the root */
2554 fail[ q[ q_write++ ] ]=1;
2557 while ( q_read < q_write) {
2558 const U32 cur = q[ q_read++ % numstates ];
2559 base = trie->states[ cur ].trans.base;
2561 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2562 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2564 U32 fail_state = cur;
2567 fail_state = fail[ fail_state ];
2568 fail_base = aho->states[ fail_state ].trans.base;
2569 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2571 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2572 fail[ ch_state ] = fail_state;
2573 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2575 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2577 q[ q_write++ % numstates] = ch_state;
2581 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2582 when we fail in state 1, this allows us to use the
2583 charclass scan to find a valid start char. This is based on the principle
2584 that theres a good chance the string being searched contains lots of stuff
2585 that cant be a start char.
2587 fail[ 0 ] = fail[ 1 ] = 0;
2588 DEBUG_TRIE_COMPILE_r({
2589 PerlIO_printf(Perl_debug_log,
2590 "%*sStclass Failtable (%"UVuf" states): 0",
2591 (int)(depth * 2), "", (UV)numstates
2593 for( q_read=1; q_read<numstates; q_read++ ) {
2594 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2596 PerlIO_printf(Perl_debug_log, "\n");
2599 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2604 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2605 * These need to be revisited when a newer toolchain becomes available.
2607 #if defined(__sparc64__) && defined(__GNUC__)
2608 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2609 # undef SPARC64_GCC_WORKAROUND
2610 # define SPARC64_GCC_WORKAROUND 1
2614 #define DEBUG_PEEP(str,scan,depth) \
2615 DEBUG_OPTIMISE_r({if (scan){ \
2616 SV * const mysv=sv_newmortal(); \
2617 regnode *Next = regnext(scan); \
2618 regprop(RExC_rx, mysv, scan); \
2619 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2620 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2621 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2625 /* The below joins as many adjacent EXACTish nodes as possible into a single
2626 * one. The regop may be changed if the node(s) contain certain sequences that
2627 * require special handling. The joining is only done if:
2628 * 1) there is room in the current conglomerated node to entirely contain the
2630 * 2) they are the exact same node type
2632 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2633 * these get optimized out
2635 * If a node is to match under /i (folded), the number of characters it matches
2636 * can be different than its character length if it contains a multi-character
2637 * fold. *min_subtract is set to the total delta of the input nodes.
2639 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2640 * and contains LATIN SMALL LETTER SHARP S
2642 * This is as good a place as any to discuss the design of handling these
2643 * multi-character fold sequences. It's been wrong in Perl for a very long
2644 * time. There are three code points in Unicode whose multi-character folds
2645 * were long ago discovered to mess things up. The previous designs for
2646 * dealing with these involved assigning a special node for them. This
2647 * approach doesn't work, as evidenced by this example:
2648 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2649 * Both these fold to "sss", but if the pattern is parsed to create a node that
2650 * would match just the \xDF, it won't be able to handle the case where a
2651 * successful match would have to cross the node's boundary. The new approach
2652 * that hopefully generally solves the problem generates an EXACTFU_SS node
2655 * It turns out that there are problems with all multi-character folds, and not
2656 * just these three. Now the code is general, for all such cases, but the
2657 * three still have some special handling. The approach taken is:
2658 * 1) This routine examines each EXACTFish node that could contain multi-
2659 * character fold sequences. It returns in *min_subtract how much to
2660 * subtract from the the actual length of the string to get a real minimum
2661 * match length; it is 0 if there are no multi-char folds. This delta is
2662 * used by the caller to adjust the min length of the match, and the delta
2663 * between min and max, so that the optimizer doesn't reject these
2664 * possibilities based on size constraints.
2665 * 2) Certain of these sequences require special handling by the trie code,
2666 * so, if found, this code changes the joined node type to special ops:
2667 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2668 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2669 * is used for an EXACTFU node that contains at least one "ss" sequence in
2670 * it. For non-UTF-8 patterns and strings, this is the only case where
2671 * there is a possible fold length change. That means that a regular
2672 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2673 * with length changes, and so can be processed faster. regexec.c takes
2674 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2675 * pre-folded by regcomp.c. This saves effort in regex matching.
2676 * However, the pre-folding isn't done for non-UTF8 patterns because the
2677 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2678 * down by forcing the pattern into UTF8 unless necessary. Also what
2679 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2680 * possibilities for the non-UTF8 patterns are quite simple, except for
2681 * the sharp s. All the ones that don't involve a UTF-8 target string are
2682 * members of a fold-pair, and arrays are set up for all of them so that
2683 * the other member of the pair can be found quickly. Code elsewhere in
2684 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2685 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2686 * described in the next item.
2687 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2688 * 'ss' or not is not knowable at compile time. It will match iff the
2689 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2690 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2691 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
2692 * described in item 3). An assumption that the optimizer part of
2693 * regexec.c (probably unwittingly) makes is that a character in the
2694 * pattern corresponds to at most a single character in the target string.
2695 * (And I do mean character, and not byte here, unlike other parts of the
2696 * documentation that have never been updated to account for multibyte
2697 * Unicode.) This assumption is wrong only in this case, as all other
2698 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2699 * virtue of having this file pre-fold UTF-8 patterns. I'm
2700 * reluctant to try to change this assumption, so instead the code punts.
2701 * This routine examines EXACTF nodes for the sharp s, and returns a
2702 * boolean indicating whether or not the node is an EXACTF node that
2703 * contains a sharp s. When it is true, the caller sets a flag that later
2704 * causes the optimizer in this file to not set values for the floating
2705 * and fixed string lengths, and thus avoids the optimizer code in
2706 * regexec.c that makes the invalid assumption. Thus, there is no
2707 * optimization based on string lengths for EXACTF nodes that contain the
2708 * sharp s. This only happens for /id rules (which means the pattern
2712 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2713 if (PL_regkind[OP(scan)] == EXACT) \
2714 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2717 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) {
2718 /* Merge several consecutive EXACTish nodes into one. */
2719 regnode *n = regnext(scan);
2721 regnode *next = scan + NODE_SZ_STR(scan);
2725 regnode *stop = scan;
2726 GET_RE_DEBUG_FLAGS_DECL;
2728 PERL_UNUSED_ARG(depth);
2731 PERL_ARGS_ASSERT_JOIN_EXACT;
2732 #ifndef EXPERIMENTAL_INPLACESCAN
2733 PERL_UNUSED_ARG(flags);
2734 PERL_UNUSED_ARG(val);
2736 DEBUG_PEEP("join",scan,depth);
2738 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2739 * EXACT ones that are mergeable to the current one. */
2741 && (PL_regkind[OP(n)] == NOTHING
2742 || (stringok && OP(n) == OP(scan)))
2744 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2747 if (OP(n) == TAIL || n > next)
2749 if (PL_regkind[OP(n)] == NOTHING) {
2750 DEBUG_PEEP("skip:",n,depth);
2751 NEXT_OFF(scan) += NEXT_OFF(n);
2752 next = n + NODE_STEP_REGNODE;
2759 else if (stringok) {
2760 const unsigned int oldl = STR_LEN(scan);
2761 regnode * const nnext = regnext(n);
2763 /* XXX I (khw) kind of doubt that this works on platforms where
2764 * U8_MAX is above 255 because of lots of other assumptions */
2765 /* Don't join if the sum can't fit into a single node */
2766 if (oldl + STR_LEN(n) > U8_MAX)
2769 DEBUG_PEEP("merg",n,depth);
2772 NEXT_OFF(scan) += NEXT_OFF(n);
2773 STR_LEN(scan) += STR_LEN(n);
2774 next = n + NODE_SZ_STR(n);
2775 /* Now we can overwrite *n : */
2776 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2784 #ifdef EXPERIMENTAL_INPLACESCAN
2785 if (flags && !NEXT_OFF(n)) {
2786 DEBUG_PEEP("atch", val, depth);
2787 if (reg_off_by_arg[OP(n)]) {
2788 ARG_SET(n, val - n);
2791 NEXT_OFF(n) = val - n;
2799 *has_exactf_sharp_s = FALSE;
2801 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2802 * can now analyze for sequences of problematic code points. (Prior to
2803 * this final joining, sequences could have been split over boundaries, and
2804 * hence missed). The sequences only happen in folding, hence for any
2805 * non-EXACT EXACTish node */
2806 if (OP(scan) != EXACT) {
2807 const U8 * const s0 = (U8*) STRING(scan);
2809 const U8 * const s_end = s0 + STR_LEN(scan);
2811 /* One pass is made over the node's string looking for all the
2812 * possibilities. to avoid some tests in the loop, there are two main
2813 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2817 /* Examine the string for a multi-character fold sequence. UTF-8
2818 * patterns have all characters pre-folded by the time this code is
2820 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2821 length sequence we are looking for is 2 */
2824 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2825 if (! len) { /* Not a multi-char fold: get next char */
2830 /* Nodes with 'ss' require special handling, except for EXACTFL
2831 * and EXACTFA for which there is no multi-char fold to this */
2832 if (len == 2 && *s == 's' && *(s+1) == 's'
2833 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2836 OP(scan) = EXACTFU_SS;
2839 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2840 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2841 COMBINING_DIAERESIS_UTF8
2842 COMBINING_ACUTE_ACCENT_UTF8,
2844 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2845 COMBINING_DIAERESIS_UTF8
2846 COMBINING_ACUTE_ACCENT_UTF8,
2851 /* These two folds require special handling by trie's, so
2852 * change the node type to indicate this. If EXACTFA and
2853 * EXACTFL were ever to be handled by trie's, this would
2854 * have to be changed. If this node has already been
2855 * changed to EXACTFU_SS in this loop, leave it as is. (I
2856 * (khw) think it doesn't matter in regexec.c for UTF
2857 * patterns, but no need to change it */
2858 if (OP(scan) == EXACTFU) {
2859 OP(scan) = EXACTFU_TRICKYFOLD;
2863 else { /* Here is a generic multi-char fold. */
2864 const U8* multi_end = s + len;
2866 /* Count how many characters in it. In the case of /l and
2867 * /aa, no folds which contain ASCII code points are
2868 * allowed, so check for those, and skip if found. (In
2869 * EXACTFL, no folds are allowed to any Latin1 code point,
2870 * not just ASCII. But there aren't any of these
2871 * currently, nor ever likely, so don't take the time to
2872 * test for them. The code that generates the
2873 * is_MULTI_foo() macros croaks should one actually get put
2874 * into Unicode .) */
2875 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2876 count = utf8_length(s, multi_end);
2880 while (s < multi_end) {
2883 goto next_iteration;
2893 /* The delta is how long the sequence is minus 1 (1 is how long
2894 * the character that folds to the sequence is) */
2895 *min_subtract += count - 1;
2899 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2901 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2902 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2903 * nodes can't have multi-char folds to this range (and there are
2904 * no existing ones in the upper latin1 range). In the EXACTF
2905 * case we look also for the sharp s, which can be in the final
2906 * position. Otherwise we can stop looking 1 byte earlier because
2907 * have to find at least two characters for a multi-fold */
2908 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2910 /* The below is perhaps overboard, but this allows us to save a
2911 * test each time through the loop at the expense of a mask. This
2912 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2913 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2914 * are 64. This uses an exclusive 'or' to find that bit and then
2915 * inverts it to form a mask, with just a single 0, in the bit
2916 * position where 'S' and 's' differ. */
2917 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2918 const U8 s_masked = 's' & S_or_s_mask;
2921 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2922 if (! len) { /* Not a multi-char fold. */
2923 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2925 *has_exactf_sharp_s = TRUE;
2932 && ((*s & S_or_s_mask) == s_masked)
2933 && ((*(s+1) & S_or_s_mask) == s_masked))
2936 /* EXACTF nodes need to know that the minimum length
2937 * changed so that a sharp s in the string can match this
2938 * ss in the pattern, but they remain EXACTF nodes, as they
2939 * won't match this unless the target string is is UTF-8,
2940 * which we don't know until runtime */
2941 if (OP(scan) != EXACTF) {
2942 OP(scan) = EXACTFU_SS;
2946 *min_subtract += len - 1;
2953 /* Allow dumping but overwriting the collection of skipped
2954 * ops and/or strings with fake optimized ops */
2955 n = scan + NODE_SZ_STR(scan);
2963 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2967 /* REx optimizer. Converts nodes into quicker variants "in place".
2968 Finds fixed substrings. */
2970 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2971 to the position after last scanned or to NULL. */
2973 #define INIT_AND_WITHP \
2974 assert(!and_withp); \
2975 Newx(and_withp,1,struct regnode_charclass_class); \
2976 SAVEFREEPV(and_withp)
2978 /* this is a chain of data about sub patterns we are processing that
2979 need to be handled separately/specially in study_chunk. Its so
2980 we can simulate recursion without losing state. */
2982 typedef struct scan_frame {
2983 regnode *last; /* last node to process in this frame */
2984 regnode *next; /* next node to process when last is reached */
2985 struct scan_frame *prev; /*previous frame*/
2986 I32 stop; /* what stopparen do we use */
2990 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2993 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2994 I32 *minlenp, I32 *deltap,
2999 struct regnode_charclass_class *and_withp,
3000 U32 flags, U32 depth)
3001 /* scanp: Start here (read-write). */
3002 /* deltap: Write maxlen-minlen here. */
3003 /* last: Stop before this one. */
3004 /* data: string data about the pattern */
3005 /* stopparen: treat close N as END */
3006 /* recursed: which subroutines have we recursed into */
3007 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3010 I32 min = 0; /* There must be at least this number of characters to match */
3012 regnode *scan = *scanp, *next;
3014 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3015 int is_inf_internal = 0; /* The studied chunk is infinite */
3016 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3017 scan_data_t data_fake;
3018 SV *re_trie_maxbuff = NULL;
3019 regnode *first_non_open = scan;
3020 I32 stopmin = I32_MAX;
3021 scan_frame *frame = NULL;
3022 GET_RE_DEBUG_FLAGS_DECL;
3024 PERL_ARGS_ASSERT_STUDY_CHUNK;
3027 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3031 while (first_non_open && OP(first_non_open) == OPEN)
3032 first_non_open=regnext(first_non_open);
3037 while ( scan && OP(scan) != END && scan < last ){
3038 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3039 node length to get a real minimum (because
3040 the folded version may be shorter) */
3041 bool has_exactf_sharp_s = FALSE;
3042 /* Peephole optimizer: */
3043 DEBUG_STUDYDATA("Peep:", data,depth);
3044 DEBUG_PEEP("Peep",scan,depth);
3046 /* Its not clear to khw or hv why this is done here, and not in the
3047 * clauses that deal with EXACT nodes. khw's guess is that it's
3048 * because of a previous design */
3049 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3051 /* Follow the next-chain of the current node and optimize
3052 away all the NOTHINGs from it. */
3053 if (OP(scan) != CURLYX) {
3054 const int max = (reg_off_by_arg[OP(scan)]
3056 /* I32 may be smaller than U16 on CRAYs! */
3057 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3058 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3062 /* Skip NOTHING and LONGJMP. */
3063 while ((n = regnext(n))
3064 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3065 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3066 && off + noff < max)
3068 if (reg_off_by_arg[OP(scan)])
3071 NEXT_OFF(scan) = off;
3076 /* The principal pseudo-switch. Cannot be a switch, since we
3077 look into several different things. */
3078 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3079 || OP(scan) == IFTHEN) {
3080 next = regnext(scan);
3082 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3084 if (OP(next) == code || code == IFTHEN) {
3085 /* NOTE - There is similar code to this block below for handling
3086 TRIE nodes on a re-study. If you change stuff here check there
3088 I32 max1 = 0, min1 = I32_MAX, num = 0;
3089 struct regnode_charclass_class accum;
3090 regnode * const startbranch=scan;
3092 if (flags & SCF_DO_SUBSTR)
3093 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3094 if (flags & SCF_DO_STCLASS)
3095 cl_init_zero(pRExC_state, &accum);
3097 while (OP(scan) == code) {
3098 I32 deltanext, minnext, f = 0, fake;
3099 struct regnode_charclass_class this_class;
3102 data_fake.flags = 0;
3104 data_fake.whilem_c = data->whilem_c;
3105 data_fake.last_closep = data->last_closep;
3108 data_fake.last_closep = &fake;
3110 data_fake.pos_delta = delta;
3111 next = regnext(scan);
3112 scan = NEXTOPER(scan);
3114 scan = NEXTOPER(scan);
3115 if (flags & SCF_DO_STCLASS) {
3116 cl_init(pRExC_state, &this_class);
3117 data_fake.start_class = &this_class;
3118 f = SCF_DO_STCLASS_AND;
3120 if (flags & SCF_WHILEM_VISITED_POS)
3121 f |= SCF_WHILEM_VISITED_POS;
3123 /* we suppose the run is continuous, last=next...*/
3124 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3126 stopparen, recursed, NULL, f,depth+1);
3129 if (deltanext == I32_MAX) {
3130 is_inf = is_inf_internal = 1;
3132 } else if (max1 < minnext + deltanext)
3133 max1 = minnext + deltanext;
3135 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3137 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3138 if ( stopmin > minnext)
3139 stopmin = min + min1;
3140 flags &= ~SCF_DO_SUBSTR;
3142 data->flags |= SCF_SEEN_ACCEPT;
3145 if (data_fake.flags & SF_HAS_EVAL)
3146 data->flags |= SF_HAS_EVAL;
3147 data->whilem_c = data_fake.whilem_c;
3149 if (flags & SCF_DO_STCLASS)
3150 cl_or(pRExC_state, &accum, &this_class);
3152 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3154 if (flags & SCF_DO_SUBSTR) {
3155 data->pos_min += min1;
3156 if (data->pos_delta >= I32_MAX - (max1 - min1))
3157 data->pos_delta = I32_MAX;
3159 data->pos_delta += max1 - min1;
3160 if (max1 != min1 || is_inf)
3161 data->longest = &(data->longest_float);
3164 if (delta == I32_MAX || I32_MAX - delta - (max1 - min1) < 0)
3167 delta += max1 - min1;
3168 if (flags & SCF_DO_STCLASS_OR) {
3169 cl_or(pRExC_state, data->start_class, &accum);
3171 cl_and(data->start_class, and_withp);
3172 flags &= ~SCF_DO_STCLASS;
3175 else if (flags & SCF_DO_STCLASS_AND) {
3177 cl_and(data->start_class, &accum);
3178 flags &= ~SCF_DO_STCLASS;
3181 /* Switch to OR mode: cache the old value of
3182 * data->start_class */
3184 StructCopy(data->start_class, and_withp,
3185 struct regnode_charclass_class);
3186 flags &= ~SCF_DO_STCLASS_AND;
3187 StructCopy(&accum, data->start_class,
3188 struct regnode_charclass_class);
3189 flags |= SCF_DO_STCLASS_OR;
3190 SET_SSC_EOS(data->start_class);
3194 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3197 Assuming this was/is a branch we are dealing with: 'scan' now
3198 points at the item that follows the branch sequence, whatever
3199 it is. We now start at the beginning of the sequence and look
3206 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3208 If we can find such a subsequence we need to turn the first
3209 element into a trie and then add the subsequent branch exact
3210 strings to the trie.
3214 1. patterns where the whole set of branches can be converted.
3216 2. patterns where only a subset can be converted.
3218 In case 1 we can replace the whole set with a single regop
3219 for the trie. In case 2 we need to keep the start and end
3222 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3223 becomes BRANCH TRIE; BRANCH X;
3225 There is an additional case, that being where there is a
3226 common prefix, which gets split out into an EXACT like node
3227 preceding the TRIE node.
3229 If x(1..n)==tail then we can do a simple trie, if not we make
3230 a "jump" trie, such that when we match the appropriate word
3231 we "jump" to the appropriate tail node. Essentially we turn
3232 a nested if into a case structure of sorts.
3237 if (!re_trie_maxbuff) {
3238 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3239 if (!SvIOK(re_trie_maxbuff))
3240 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3242 if ( SvIV(re_trie_maxbuff)>=0 ) {
3244 regnode *first = (regnode *)NULL;
3245 regnode *last = (regnode *)NULL;
3246 regnode *tail = scan;
3251 SV * const mysv = sv_newmortal(); /* for dumping */
3253 /* var tail is used because there may be a TAIL
3254 regop in the way. Ie, the exacts will point to the
3255 thing following the TAIL, but the last branch will
3256 point at the TAIL. So we advance tail. If we
3257 have nested (?:) we may have to move through several
3261 while ( OP( tail ) == TAIL ) {
3262 /* this is the TAIL generated by (?:) */
3263 tail = regnext( tail );
3267 DEBUG_TRIE_COMPILE_r({
3268 regprop(RExC_rx, mysv, tail );
3269 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3270 (int)depth * 2 + 2, "",
3271 "Looking for TRIE'able sequences. Tail node is: ",
3272 SvPV_nolen_const( mysv )
3278 Step through the branches
3279 cur represents each branch,
3280 noper is the first thing to be matched as part of that branch
3281 noper_next is the regnext() of that node.
3283 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3284 via a "jump trie" but we also support building with NOJUMPTRIE,
3285 which restricts the trie logic to structures like /FOO|BAR/.
3287 If noper is a trieable nodetype then the branch is a possible optimization
3288 target. If we are building under NOJUMPTRIE then we require that noper_next
3289 is the same as scan (our current position in the regex program).
3291 Once we have two or more consecutive such branches we can create a
3292 trie of the EXACT's contents and stitch it in place into the program.
3294 If the sequence represents all of the branches in the alternation we
3295 replace the entire thing with a single TRIE node.
3297 Otherwise when it is a subsequence we need to stitch it in place and
3298 replace only the relevant branches. This means the first branch has
3299 to remain as it is used by the alternation logic, and its next pointer,
3300 and needs to be repointed at the item on the branch chain following
3301 the last branch we have optimized away.
3303 This could be either a BRANCH, in which case the subsequence is internal,
3304 or it could be the item following the branch sequence in which case the
3305 subsequence is at the end (which does not necessarily mean the first node
3306 is the start of the alternation).
3308 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3311 ----------------+-----------
3315 EXACTFU_SS | EXACTFU
3316 EXACTFU_TRICKYFOLD | EXACTFU
3321 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3322 ( EXACT == (X) ) ? EXACT : \
3323 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3326 /* dont use tail as the end marker for this traverse */
3327 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3328 regnode * const noper = NEXTOPER( cur );
3329 U8 noper_type = OP( noper );
3330 U8 noper_trietype = TRIE_TYPE( noper_type );
3331 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3332 regnode * const noper_next = regnext( noper );
3333 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3334 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3337 DEBUG_TRIE_COMPILE_r({
3338 regprop(RExC_rx, mysv, cur);
3339 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3340 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3342 regprop(RExC_rx, mysv, noper);
3343 PerlIO_printf( Perl_debug_log, " -> %s",
3344 SvPV_nolen_const(mysv));
3347 regprop(RExC_rx, mysv, noper_next );
3348 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3349 SvPV_nolen_const(mysv));
3351 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3352 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3353 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3357 /* Is noper a trieable nodetype that can be merged with the
3358 * current trie (if there is one)? */
3362 ( noper_trietype == NOTHING)
3363 || ( trietype == NOTHING )
3364 || ( trietype == noper_trietype )
3367 && noper_next == tail
3371 /* Handle mergable triable node
3372 * Either we are the first node in a new trieable sequence,
3373 * in which case we do some bookkeeping, otherwise we update
3374 * the end pointer. */
3377 if ( noper_trietype == NOTHING ) {
3378 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3379 regnode * const noper_next = regnext( noper );
3380 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3381 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3384 if ( noper_next_trietype ) {
3385 trietype = noper_next_trietype;
3386 } else if (noper_next_type) {
3387 /* a NOTHING regop is 1 regop wide. We need at least two
3388 * for a trie so we can't merge this in */
3392 trietype = noper_trietype;
3395 if ( trietype == NOTHING )
3396 trietype = noper_trietype;
3401 } /* end handle mergable triable node */
3403 /* handle unmergable node -
3404 * noper may either be a triable node which can not be tried
3405 * together with the current trie, or a non triable node */
3407 /* If last is set and trietype is not NOTHING then we have found
3408 * at least two triable branch sequences in a row of a similar
3409 * trietype so we can turn them into a trie. If/when we
3410 * allow NOTHING to start a trie sequence this condition will be
3411 * required, and it isn't expensive so we leave it in for now. */
3412 if ( trietype && trietype != NOTHING )
3413 make_trie( pRExC_state,
3414 startbranch, first, cur, tail, count,
3415 trietype, depth+1 );
3416 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3420 && noper_next == tail
3423 /* noper is triable, so we can start a new trie sequence */
3426 trietype = noper_trietype;
3428 /* if we already saw a first but the current node is not triable then we have
3429 * to reset the first information. */
3434 } /* end handle unmergable node */
3435 } /* loop over branches */
3436 DEBUG_TRIE_COMPILE_r({
3437 regprop(RExC_rx, mysv, cur);
3438 PerlIO_printf( Perl_debug_log,
3439 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3440 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3443 if ( last && trietype ) {
3444 if ( trietype != NOTHING ) {
3445 /* the last branch of the sequence was part of a trie,
3446 * so we have to construct it here outside of the loop
3448 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3449 #ifdef TRIE_STUDY_OPT
3450 if ( ((made == MADE_EXACT_TRIE &&
3451 startbranch == first)
3452 || ( first_non_open == first )) &&
3454 flags |= SCF_TRIE_RESTUDY;
3455 if ( startbranch == first
3458 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3463 /* at this point we know whatever we have is a NOTHING sequence/branch
3464 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3466 if ( startbranch == first ) {
3468 /* the entire thing is a NOTHING sequence, something like this:
3469 * (?:|) So we can turn it into a plain NOTHING op. */
3470 DEBUG_TRIE_COMPILE_r({
3471 regprop(RExC_rx, mysv, cur);
3472 PerlIO_printf( Perl_debug_log,
3473 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3474 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3477 OP(startbranch)= NOTHING;
3478 NEXT_OFF(startbranch)= tail - startbranch;
3479 for ( opt= startbranch + 1; opt < tail ; opt++ )
3483 } /* end if ( last) */
3484 } /* TRIE_MAXBUF is non zero */
3489 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3490 scan = NEXTOPER(NEXTOPER(scan));
3491 } else /* single branch is optimized. */
3492 scan = NEXTOPER(scan);
3494 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3495 scan_frame *newframe = NULL;
3500 if (OP(scan) != SUSPEND) {
3501 /* set the pointer */
3502 if (OP(scan) == GOSUB) {
3504 RExC_recurse[ARG2L(scan)] = scan;
3505 start = RExC_open_parens[paren-1];
3506 end = RExC_close_parens[paren-1];
3509 start = RExC_rxi->program + 1;
3513 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3514 SAVEFREEPV(recursed);
3516 if (!PAREN_TEST(recursed,paren+1)) {
3517 PAREN_SET(recursed,paren+1);
3518 Newx(newframe,1,scan_frame);
3520 if (flags & SCF_DO_SUBSTR) {
3521 SCAN_COMMIT(pRExC_state,data,minlenp);
3522 data->longest = &(data->longest_float);
3524 is_inf = is_inf_internal = 1;
3525 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3526 cl_anything(pRExC_state, data->start_class);
3527 flags &= ~SCF_DO_STCLASS;
3530 Newx(newframe,1,scan_frame);
3533 end = regnext(scan);
3538 SAVEFREEPV(newframe);
3539 newframe->next = regnext(scan);
3540 newframe->last = last;
3541 newframe->stop = stopparen;
3542 newframe->prev = frame;
3552 else if (OP(scan) == EXACT) {
3553 I32 l = STR_LEN(scan);
3556 const U8 * const s = (U8*)STRING(scan);
3557 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3558 l = utf8_length(s, s + l);
3560 uc = *((U8*)STRING(scan));
3563 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3564 /* The code below prefers earlier match for fixed
3565 offset, later match for variable offset. */
3566 if (data->last_end == -1) { /* Update the start info. */
3567 data->last_start_min = data->pos_min;
3568 data->last_start_max = is_inf
3569 ? I32_MAX : data->pos_min + data->pos_delta;
3571 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3573 SvUTF8_on(data->last_found);
3575 SV * const sv = data->last_found;
3576 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3577 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3578 if (mg && mg->mg_len >= 0)
3579 mg->mg_len += utf8_length((U8*)STRING(scan),
3580 (U8*)STRING(scan)+STR_LEN(scan));
3582 data->last_end = data->pos_min + l;
3583 data->pos_min += l; /* As in the first entry. */
3584 data->flags &= ~SF_BEFORE_EOL;
3586 if (flags & SCF_DO_STCLASS_AND) {
3587 /* Check whether it is compatible with what we know already! */
3591 /* If compatible, we or it in below. It is compatible if is
3592 * in the bitmp and either 1) its bit or its fold is set, or 2)
3593 * it's for a locale. Even if there isn't unicode semantics
3594 * here, at runtime there may be because of matching against a
3595 * utf8 string, so accept a possible false positive for
3596 * latin1-range folds */
3598 (!(data->start_class->flags & ANYOF_LOCALE)
3599 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3600 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3601 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3606 ANYOF_CLASS_ZERO(data->start_class);
3607 ANYOF_BITMAP_ZERO(data->start_class);
3609 ANYOF_BITMAP_SET(data->start_class, uc);
3610 else if (uc >= 0x100) {
3613 /* Some Unicode code points fold to the Latin1 range; as
3614 * XXX temporary code, instead of figuring out if this is
3615 * one, just assume it is and set all the start class bits
3616 * that could be some such above 255 code point's fold
3617 * which will generate fals positives. As the code
3618 * elsewhere that does compute the fold settles down, it
3619 * can be extracted out and re-used here */
3620 for (i = 0; i < 256; i++){
3621 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3622 ANYOF_BITMAP_SET(data->start_class, i);
3626 CLEAR_SSC_EOS(data->start_class);
3628 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3630 else if (flags & SCF_DO_STCLASS_OR) {
3631 /* false positive possible if the class is case-folded */
3633 ANYOF_BITMAP_SET(data->start_class, uc);
3635 data->start_class->flags |= ANYOF_UNICODE_ALL;
3636 CLEAR_SSC_EOS(data->start_class);
3637 cl_and(data->start_class, and_withp);
3639 flags &= ~SCF_DO_STCLASS;
3641 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3642 I32 l = STR_LEN(scan);
3643 UV uc = *((U8*)STRING(scan));
3645 /* Search for fixed substrings supports EXACT only. */
3646 if (flags & SCF_DO_SUBSTR) {
3648 SCAN_COMMIT(pRExC_state, data, minlenp);
3651 const U8 * const s = (U8 *)STRING(scan);
3652 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3653 l = utf8_length(s, s + l);
3655 if (has_exactf_sharp_s) {
3656 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3658 min += l - min_subtract;
3660 delta += min_subtract;
3661 if (flags & SCF_DO_SUBSTR) {
3662 data->pos_min += l - min_subtract;
3663 if (data->pos_min < 0) {
3666 data->pos_delta += min_subtract;
3668 data->longest = &(data->longest_float);
3671 if (flags & SCF_DO_STCLASS_AND) {
3672 /* Check whether it is compatible with what we know already! */
3675 (!(data->start_class->flags & ANYOF_LOCALE)
3676 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3677 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3681 ANYOF_CLASS_ZERO(data->start_class);
3682 ANYOF_BITMAP_ZERO(data->start_class);
3684 ANYOF_BITMAP_SET(data->start_class, uc);
3685 CLEAR_SSC_EOS(data->start_class);
3686 if (OP(scan) == EXACTFL) {
3687 /* XXX This set is probably no longer necessary, and
3688 * probably wrong as LOCALE now is on in the initial
3690 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3694 /* Also set the other member of the fold pair. In case
3695 * that unicode semantics is called for at runtime, use
3696 * the full latin1 fold. (Can't do this for locale,
3697 * because not known until runtime) */
3698 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3700 /* All other (EXACTFL handled above) folds except under
3701 * /iaa that include s, S, and sharp_s also may include
3703 if (OP(scan) != EXACTFA) {
3704 if (uc == 's' || uc == 'S') {
3705 ANYOF_BITMAP_SET(data->start_class,
3706 LATIN_SMALL_LETTER_SHARP_S);
3708 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3709 ANYOF_BITMAP_SET(data->start_class, 's');
3710 ANYOF_BITMAP_SET(data->start_class, 'S');
3715 else if (uc >= 0x100) {
3717 for (i = 0; i < 256; i++){
3718 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3719 ANYOF_BITMAP_SET(data->start_class, i);
3724 else if (flags & SCF_DO_STCLASS_OR) {
3725 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3726 /* false positive possible if the class is case-folded.
3727 Assume that the locale settings are the same... */
3729 ANYOF_BITMAP_SET(data->start_class, uc);
3730 if (OP(scan) != EXACTFL) {
3732 /* And set the other member of the fold pair, but
3733 * can't do that in locale because not known until
3735 ANYOF_BITMAP_SET(data->start_class,
3736 PL_fold_latin1[uc]);
3738 /* All folds except under /iaa that include s, S,
3739 * and sharp_s also may include the others */
3740 if (OP(scan) != EXACTFA) {
3741 if (uc == 's' || uc == 'S') {
3742 ANYOF_BITMAP_SET(data->start_class,
3743 LATIN_SMALL_LETTER_SHARP_S);
3745 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3746 ANYOF_BITMAP_SET(data->start_class, 's');
3747 ANYOF_BITMAP_SET(data->start_class, 'S');
3752 CLEAR_SSC_EOS(data->start_class);
3754 cl_and(data->start_class, and_withp);
3756 flags &= ~SCF_DO_STCLASS;
3758 else if (REGNODE_VARIES(OP(scan))) {
3759 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3760 I32 f = flags, pos_before = 0;
3761 regnode * const oscan = scan;
3762 struct regnode_charclass_class this_class;
3763 struct regnode_charclass_class *oclass = NULL;
3764 I32 next_is_eval = 0;
3766 switch (PL_regkind[OP(scan)]) {
3767 case WHILEM: /* End of (?:...)* . */
3768 scan = NEXTOPER(scan);
3771 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3772 next = NEXTOPER(scan);
3773 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3775 maxcount = REG_INFTY;
3776 next = regnext(scan);
3777 scan = NEXTOPER(scan);
3781 if (flags & SCF_DO_SUBSTR)
3786 if (flags & SCF_DO_STCLASS) {
3788 maxcount = REG_INFTY;
3789 next = regnext(scan);
3790 scan = NEXTOPER(scan);
3793 is_inf = is_inf_internal = 1;
3794 scan = regnext(scan);
3795 if (flags & SCF_DO_SUBSTR) {
3796 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3797 data->longest = &(data->longest_float);
3799 goto optimize_curly_tail;
3801 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3802 && (scan->flags == stopparen))
3807 mincount = ARG1(scan);
3808 maxcount = ARG2(scan);
3810 next = regnext(scan);
3811 if (OP(scan) == CURLYX) {
3812 I32 lp = (data ? *(data->last_closep) : 0);
3813 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3815 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3816 next_is_eval = (OP(scan) == EVAL);
3818 if (flags & SCF_DO_SUBSTR) {
3819 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3820 pos_before = data->pos_min;
3824 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3826 data->flags |= SF_IS_INF;
3828 if (flags & SCF_DO_STCLASS) {
3829 cl_init(pRExC_state, &this_class);
3830 oclass = data->start_class;
3831 data->start_class = &this_class;
3832 f |= SCF_DO_STCLASS_AND;
3833 f &= ~SCF_DO_STCLASS_OR;
3835 /* Exclude from super-linear cache processing any {n,m}
3836 regops for which the combination of input pos and regex
3837 pos is not enough information to determine if a match
https://perl5.git.perl.org / perl5.git / blob