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"
100 #define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
101 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
102 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
109 # if defined(BUGGY_MSC6)
110 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
111 # pragma optimize("a",off)
112 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
113 # pragma optimize("w",on )
114 # endif /* BUGGY_MSC6 */
118 #define STATIC static
122 typedef struct RExC_state_t {
123 U32 flags; /* RXf_* are we folding, multilining? */
124 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
125 char *precomp; /* uncompiled string. */
126 REGEXP *rx_sv; /* The SV that is the regexp. */
127 regexp *rx; /* perl core regexp structure */
128 regexp_internal *rxi; /* internal data for regexp object pprivate field */
129 char *start; /* Start of input for compile */
130 char *end; /* End of input for compile */
131 char *parse; /* Input-scan pointer. */
132 I32 whilem_seen; /* number of WHILEM in this expr */
133 regnode *emit_start; /* Start of emitted-code area */
134 regnode *emit_bound; /* First regnode outside of the allocated space */
135 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
136 I32 naughty; /* How bad is this pattern? */
137 I32 sawback; /* Did we see \1, ...? */
139 I32 size; /* Code size. */
140 I32 npar; /* Capture buffer count, (OPEN). */
141 I32 cpar; /* Capture buffer count, (CLOSE). */
142 I32 nestroot; /* root parens we are in - used by accept */
145 regnode **open_parens; /* pointers to open parens */
146 regnode **close_parens; /* pointers to close parens */
147 regnode *opend; /* END node in program */
148 I32 utf8; /* whether the pattern is utf8 or not */
149 I32 orig_utf8; /* whether the pattern was originally in utf8 */
150 /* XXX use this for future optimisation of case
151 * where pattern must be upgraded to utf8. */
152 I32 uni_semantics; /* If a d charset modifier should use unicode
153 rules, even if the pattern is not in
155 HV *paren_names; /* Paren names */
157 regnode **recurse; /* Recurse regops */
158 I32 recurse_count; /* Number of recurse regops */
161 I32 override_recoding;
162 I32 in_multi_char_class;
163 struct reg_code_block *code_blocks; /* positions of literal (?{})
165 int num_code_blocks; /* size of code_blocks[] */
166 int code_index; /* next code_blocks[] slot */
168 char *starttry; /* -Dr: where regtry was called. */
169 #define RExC_starttry (pRExC_state->starttry)
171 SV *runtime_code_qr; /* qr with the runtime code blocks */
173 const char *lastparse;
175 AV *paren_name_list; /* idx -> name */
176 #define RExC_lastparse (pRExC_state->lastparse)
177 #define RExC_lastnum (pRExC_state->lastnum)
178 #define RExC_paren_name_list (pRExC_state->paren_name_list)
182 #define RExC_flags (pRExC_state->flags)
183 #define RExC_pm_flags (pRExC_state->pm_flags)
184 #define RExC_precomp (pRExC_state->precomp)
185 #define RExC_rx_sv (pRExC_state->rx_sv)
186 #define RExC_rx (pRExC_state->rx)
187 #define RExC_rxi (pRExC_state->rxi)
188 #define RExC_start (pRExC_state->start)
189 #define RExC_end (pRExC_state->end)
190 #define RExC_parse (pRExC_state->parse)
191 #define RExC_whilem_seen (pRExC_state->whilem_seen)
192 #ifdef RE_TRACK_PATTERN_OFFSETS
193 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
195 #define RExC_emit (pRExC_state->emit)
196 #define RExC_emit_start (pRExC_state->emit_start)
197 #define RExC_emit_bound (pRExC_state->emit_bound)
198 #define RExC_naughty (pRExC_state->naughty)
199 #define RExC_sawback (pRExC_state->sawback)
200 #define RExC_seen (pRExC_state->seen)
201 #define RExC_size (pRExC_state->size)
202 #define RExC_npar (pRExC_state->npar)
203 #define RExC_nestroot (pRExC_state->nestroot)
204 #define RExC_extralen (pRExC_state->extralen)
205 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
206 #define RExC_utf8 (pRExC_state->utf8)
207 #define RExC_uni_semantics (pRExC_state->uni_semantics)
208 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
209 #define RExC_open_parens (pRExC_state->open_parens)
210 #define RExC_close_parens (pRExC_state->close_parens)
211 #define RExC_opend (pRExC_state->opend)
212 #define RExC_paren_names (pRExC_state->paren_names)
213 #define RExC_recurse (pRExC_state->recurse)
214 #define RExC_recurse_count (pRExC_state->recurse_count)
215 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
216 #define RExC_contains_locale (pRExC_state->contains_locale)
217 #define RExC_override_recoding (pRExC_state->override_recoding)
218 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
221 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
222 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
223 ((*s) == '{' && regcurly(s)))
226 #undef SPSTART /* dratted cpp namespace... */
229 * Flags to be passed up and down.
231 #define WORST 0 /* Worst case. */
232 #define HASWIDTH 0x01 /* Known to match non-null strings. */
234 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
235 * character. (There needs to be a case: in the switch statement in regexec.c
236 * for any node marked SIMPLE.) Note that this is not the same thing as
239 #define SPSTART 0x04 /* Starts with * or + */
240 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
241 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
243 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
245 /* whether trie related optimizations are enabled */
246 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
247 #define TRIE_STUDY_OPT
248 #define FULL_TRIE_STUDY
254 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
255 #define PBITVAL(paren) (1 << ((paren) & 7))
256 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
257 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
258 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
260 /* If not already in utf8, do a longjmp back to the beginning */
261 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
262 #define REQUIRE_UTF8 STMT_START { \
263 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
266 /* About scan_data_t.
268 During optimisation we recurse through the regexp program performing
269 various inplace (keyhole style) optimisations. In addition study_chunk
270 and scan_commit populate this data structure with information about
271 what strings MUST appear in the pattern. We look for the longest
272 string that must appear at a fixed location, and we look for the
273 longest string that may appear at a floating location. So for instance
278 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
279 strings (because they follow a .* construct). study_chunk will identify
280 both FOO and BAR as being the longest fixed and floating strings respectively.
282 The strings can be composites, for instance
286 will result in a composite fixed substring 'foo'.
288 For each string some basic information is maintained:
290 - offset or min_offset
291 This is the position the string must appear at, or not before.
292 It also implicitly (when combined with minlenp) tells us how many
293 characters must match before the string we are searching for.
294 Likewise when combined with minlenp and the length of the string it
295 tells us how many characters must appear after the string we have
299 Only used for floating strings. This is the rightmost point that
300 the string can appear at. If set to I32 max it indicates that the
301 string can occur infinitely far to the right.
304 A pointer to the minimum number of characters of the pattern that the
305 string was found inside. This is important as in the case of positive
306 lookahead or positive lookbehind we can have multiple patterns
311 The minimum length of the pattern overall is 3, the minimum length
312 of the lookahead part is 3, but the minimum length of the part that
313 will actually match is 1. So 'FOO's minimum length is 3, but the
314 minimum length for the F is 1. This is important as the minimum length
315 is used to determine offsets in front of and behind the string being
316 looked for. Since strings can be composites this is the length of the
317 pattern at the time it was committed with a scan_commit. Note that
318 the length is calculated by study_chunk, so that the minimum lengths
319 are not known until the full pattern has been compiled, thus the
320 pointer to the value.
324 In the case of lookbehind the string being searched for can be
325 offset past the start point of the final matching string.
326 If this value was just blithely removed from the min_offset it would
327 invalidate some of the calculations for how many chars must match
328 before or after (as they are derived from min_offset and minlen and
329 the length of the string being searched for).
330 When the final pattern is compiled and the data is moved from the
331 scan_data_t structure into the regexp structure the information
332 about lookbehind is factored in, with the information that would
333 have been lost precalculated in the end_shift field for the
336 The fields pos_min and pos_delta are used to store the minimum offset
337 and the delta to the maximum offset at the current point in the pattern.
341 typedef struct scan_data_t {
342 /*I32 len_min; unused */
343 /*I32 len_delta; unused */
347 I32 last_end; /* min value, <0 unless valid. */
350 SV **longest; /* Either &l_fixed, or &l_float. */
351 SV *longest_fixed; /* longest fixed string found in pattern */
352 I32 offset_fixed; /* offset where it starts */
353 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
354 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
355 SV *longest_float; /* longest floating string found in pattern */
356 I32 offset_float_min; /* earliest point in string it can appear */
357 I32 offset_float_max; /* latest point in string it can appear */
358 I32 *minlen_float; /* pointer to the minlen relevant to the string */
359 I32 lookbehind_float; /* is the position of the string modified by LB */
363 struct regnode_charclass_class *start_class;
367 * Forward declarations for pregcomp()'s friends.
370 static const scan_data_t zero_scan_data =
371 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
373 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
374 #define SF_BEFORE_SEOL 0x0001
375 #define SF_BEFORE_MEOL 0x0002
376 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
377 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
380 # define SF_FIX_SHIFT_EOL (0+2)
381 # define SF_FL_SHIFT_EOL (0+4)
383 # define SF_FIX_SHIFT_EOL (+2)
384 # define SF_FL_SHIFT_EOL (+4)
387 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
388 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
390 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
391 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
392 #define SF_IS_INF 0x0040
393 #define SF_HAS_PAR 0x0080
394 #define SF_IN_PAR 0x0100
395 #define SF_HAS_EVAL 0x0200
396 #define SCF_DO_SUBSTR 0x0400
397 #define SCF_DO_STCLASS_AND 0x0800
398 #define SCF_DO_STCLASS_OR 0x1000
399 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
400 #define SCF_WHILEM_VISITED_POS 0x2000
402 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
403 #define SCF_SEEN_ACCEPT 0x8000
405 #define UTF cBOOL(RExC_utf8)
407 /* The enums for all these are ordered so things work out correctly */
408 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
409 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
410 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
411 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
412 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
413 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
414 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
416 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
418 #define OOB_NAMEDCLASS -1
420 /* There is no code point that is out-of-bounds, so this is problematic. But
421 * its only current use is to initialize a variable that is always set before
423 #define OOB_UNICODE 0xDEADBEEF
425 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
426 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
429 /* length of regex to show in messages that don't mark a position within */
430 #define RegexLengthToShowInErrorMessages 127
433 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
434 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
435 * op/pragma/warn/regcomp.
437 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
438 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
440 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
443 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
444 * arg. Show regex, up to a maximum length. If it's too long, chop and add
447 #define _FAIL(code) STMT_START { \
448 const char *ellipses = ""; \
449 IV len = RExC_end - RExC_precomp; \
452 SAVEFREESV(RExC_rx_sv); \
453 if (len > RegexLengthToShowInErrorMessages) { \
454 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
455 len = RegexLengthToShowInErrorMessages - 10; \
461 #define FAIL(msg) _FAIL( \
462 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
463 msg, (int)len, RExC_precomp, ellipses))
465 #define FAIL2(msg,arg) _FAIL( \
466 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
467 arg, (int)len, RExC_precomp, ellipses))
470 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
472 #define Simple_vFAIL(m) STMT_START { \
473 const IV offset = RExC_parse - RExC_precomp; \
474 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
475 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
479 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
481 #define vFAIL(m) STMT_START { \
483 SAVEFREESV(RExC_rx_sv); \
488 * Like Simple_vFAIL(), but accepts two arguments.
490 #define Simple_vFAIL2(m,a1) STMT_START { \
491 const IV offset = RExC_parse - RExC_precomp; \
492 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
493 (int)offset, RExC_precomp, RExC_precomp + offset); \
497 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
499 #define vFAIL2(m,a1) STMT_START { \
501 SAVEFREESV(RExC_rx_sv); \
502 Simple_vFAIL2(m, a1); \
507 * Like Simple_vFAIL(), but accepts three arguments.
509 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
510 const IV offset = RExC_parse - RExC_precomp; \
511 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
512 (int)offset, RExC_precomp, RExC_precomp + offset); \
516 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
518 #define vFAIL3(m,a1,a2) STMT_START { \
520 SAVEFREESV(RExC_rx_sv); \
521 Simple_vFAIL3(m, a1, a2); \
525 * Like Simple_vFAIL(), but accepts four arguments.
527 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
528 const IV offset = RExC_parse - RExC_precomp; \
529 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
530 (int)offset, RExC_precomp, RExC_precomp + offset); \
533 #define ckWARNreg(loc,m) STMT_START { \
534 const IV offset = loc - RExC_precomp; \
535 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
536 (int)offset, RExC_precomp, RExC_precomp + offset); \
539 #define ckWARNregdep(loc,m) STMT_START { \
540 const IV offset = loc - RExC_precomp; \
541 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
543 (int)offset, RExC_precomp, RExC_precomp + offset); \
546 #define ckWARN2regdep(loc,m, a1) STMT_START { \
547 const IV offset = loc - RExC_precomp; \
548 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
550 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
553 #define ckWARN2reg(loc, m, a1) STMT_START { \
554 const IV offset = loc - RExC_precomp; \
555 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
556 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
559 #define vWARN3(loc, m, a1, a2) STMT_START { \
560 const IV offset = loc - RExC_precomp; \
561 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
562 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
565 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
566 const IV offset = loc - RExC_precomp; \
567 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
568 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
571 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
572 const IV offset = loc - RExC_precomp; \
573 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
574 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
577 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
578 const IV offset = loc - RExC_precomp; \
579 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
580 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
583 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
584 const IV offset = loc - RExC_precomp; \
585 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
586 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
590 /* Allow for side effects in s */
591 #define REGC(c,s) STMT_START { \
592 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
595 /* Macros for recording node offsets. 20001227 mjd@plover.com
596 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
597 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
598 * Element 0 holds the number n.
599 * Position is 1 indexed.
601 #ifndef RE_TRACK_PATTERN_OFFSETS
602 #define Set_Node_Offset_To_R(node,byte)
603 #define Set_Node_Offset(node,byte)
604 #define Set_Cur_Node_Offset
605 #define Set_Node_Length_To_R(node,len)
606 #define Set_Node_Length(node,len)
607 #define Set_Node_Cur_Length(node)
608 #define Node_Offset(n)
609 #define Node_Length(n)
610 #define Set_Node_Offset_Length(node,offset,len)
611 #define ProgLen(ri) ri->u.proglen
612 #define SetProgLen(ri,x) ri->u.proglen = x
614 #define ProgLen(ri) ri->u.offsets[0]
615 #define SetProgLen(ri,x) ri->u.offsets[0] = x
616 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
618 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
619 __LINE__, (int)(node), (int)(byte))); \
621 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
623 RExC_offsets[2*(node)-1] = (byte); \
628 #define Set_Node_Offset(node,byte) \
629 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
630 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
632 #define Set_Node_Length_To_R(node,len) STMT_START { \
634 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
635 __LINE__, (int)(node), (int)(len))); \
637 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
639 RExC_offsets[2*(node)] = (len); \
644 #define Set_Node_Length(node,len) \
645 Set_Node_Length_To_R((node)-RExC_emit_start, len)
646 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
647 #define Set_Node_Cur_Length(node) \
648 Set_Node_Length(node, RExC_parse - parse_start)
650 /* Get offsets and lengths */
651 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
652 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
654 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
655 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
656 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
660 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
661 #define EXPERIMENTAL_INPLACESCAN
662 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
664 #define DEBUG_STUDYDATA(str,data,depth) \
665 DEBUG_OPTIMISE_MORE_r(if(data){ \
666 PerlIO_printf(Perl_debug_log, \
667 "%*s" str "Pos:%"IVdf"/%"IVdf \
668 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
669 (int)(depth)*2, "", \
670 (IV)((data)->pos_min), \
671 (IV)((data)->pos_delta), \
672 (UV)((data)->flags), \
673 (IV)((data)->whilem_c), \
674 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
675 is_inf ? "INF " : "" \
677 if ((data)->last_found) \
678 PerlIO_printf(Perl_debug_log, \
679 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
680 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
681 SvPVX_const((data)->last_found), \
682 (IV)((data)->last_end), \
683 (IV)((data)->last_start_min), \
684 (IV)((data)->last_start_max), \
685 ((data)->longest && \
686 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
687 SvPVX_const((data)->longest_fixed), \
688 (IV)((data)->offset_fixed), \
689 ((data)->longest && \
690 (data)->longest==&((data)->longest_float)) ? "*" : "", \
691 SvPVX_const((data)->longest_float), \
692 (IV)((data)->offset_float_min), \
693 (IV)((data)->offset_float_max) \
695 PerlIO_printf(Perl_debug_log,"\n"); \
698 /* Mark that we cannot extend a found fixed substring at this point.
699 Update the longest found anchored substring and the longest found
700 floating substrings if needed. */
703 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
705 const STRLEN l = CHR_SVLEN(data->last_found);
706 const STRLEN old_l = CHR_SVLEN(*data->longest);
707 GET_RE_DEBUG_FLAGS_DECL;
709 PERL_ARGS_ASSERT_SCAN_COMMIT;
711 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
712 SvSetMagicSV(*data->longest, data->last_found);
713 if (*data->longest == data->longest_fixed) {
714 data->offset_fixed = l ? data->last_start_min : data->pos_min;
715 if (data->flags & SF_BEFORE_EOL)
717 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
719 data->flags &= ~SF_FIX_BEFORE_EOL;
720 data->minlen_fixed=minlenp;
721 data->lookbehind_fixed=0;
723 else { /* *data->longest == data->longest_float */
724 data->offset_float_min = l ? data->last_start_min : data->pos_min;
725 data->offset_float_max = (l
726 ? data->last_start_max
727 : data->pos_min + data->pos_delta);
728 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
729 data->offset_float_max = I32_MAX;
730 if (data->flags & SF_BEFORE_EOL)
732 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
734 data->flags &= ~SF_FL_BEFORE_EOL;
735 data->minlen_float=minlenp;
736 data->lookbehind_float=0;
739 SvCUR_set(data->last_found, 0);
741 SV * const sv = data->last_found;
742 if (SvUTF8(sv) && SvMAGICAL(sv)) {
743 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
749 data->flags &= ~SF_BEFORE_EOL;
750 DEBUG_STUDYDATA("commit: ",data,0);
753 /* Can match anything (initialization) */
755 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
757 PERL_ARGS_ASSERT_CL_ANYTHING;
759 ANYOF_BITMAP_SETALL(cl);
760 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
761 |ANYOF_NON_UTF8_LATIN1_ALL;
763 /* If any portion of the regex is to operate under locale rules,
764 * initialization includes it. The reason this isn't done for all regexes
765 * is that the optimizer was written under the assumption that locale was
766 * all-or-nothing. Given the complexity and lack of documentation in the
767 * optimizer, and that there are inadequate test cases for locale, so many
768 * parts of it may not work properly, it is safest to avoid locale unless
770 if (RExC_contains_locale) {
771 ANYOF_CLASS_SETALL(cl); /* /l uses class */
772 cl->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
775 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
779 /* Can match anything (initialization) */
781 S_cl_is_anything(const struct regnode_charclass_class *cl)
785 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
787 for (value = 0; value <= ANYOF_MAX; value += 2)
788 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
790 if (!(cl->flags & ANYOF_UNICODE_ALL))
792 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
797 /* Can match anything (initialization) */
799 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
801 PERL_ARGS_ASSERT_CL_INIT;
803 Zero(cl, 1, struct regnode_charclass_class);
805 cl_anything(pRExC_state, cl);
806 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
809 /* These two functions currently do the exact same thing */
810 #define cl_init_zero S_cl_init
812 /* 'AND' a given class with another one. Can create false positives. 'cl'
813 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
814 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
816 S_cl_and(struct regnode_charclass_class *cl,
817 const struct regnode_charclass_class *and_with)
819 PERL_ARGS_ASSERT_CL_AND;
821 assert(and_with->type == ANYOF);
823 /* I (khw) am not sure all these restrictions are necessary XXX */
824 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
825 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
826 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
827 && !(and_with->flags & ANYOF_LOC_FOLD)
828 && !(cl->flags & ANYOF_LOC_FOLD)) {
831 if (and_with->flags & ANYOF_INVERT)
832 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
833 cl->bitmap[i] &= ~and_with->bitmap[i];
835 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
836 cl->bitmap[i] &= and_with->bitmap[i];
837 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
839 if (and_with->flags & ANYOF_INVERT) {
841 /* Here, the and'ed node is inverted. Get the AND of the flags that
842 * aren't affected by the inversion. Those that are affected are
843 * handled individually below */
844 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
845 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
846 cl->flags |= affected_flags;
848 /* We currently don't know how to deal with things that aren't in the
849 * bitmap, but we know that the intersection is no greater than what
850 * is already in cl, so let there be false positives that get sorted
851 * out after the synthetic start class succeeds, and the node is
852 * matched for real. */
854 /* The inversion of these two flags indicate that the resulting
855 * intersection doesn't have them */
856 if (and_with->flags & ANYOF_UNICODE_ALL) {
857 cl->flags &= ~ANYOF_UNICODE_ALL;
859 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
860 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
863 else { /* and'd node is not inverted */
864 U8 outside_bitmap_but_not_utf8; /* Temp variable */
866 if (! ANYOF_NONBITMAP(and_with)) {
868 /* Here 'and_with' doesn't match anything outside the bitmap
869 * (except possibly ANYOF_UNICODE_ALL), which means the
870 * intersection can't either, except for ANYOF_UNICODE_ALL, in
871 * which case we don't know what the intersection is, but it's no
872 * greater than what cl already has, so can just leave it alone,
873 * with possible false positives */
874 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
875 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
876 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
879 else if (! ANYOF_NONBITMAP(cl)) {
881 /* Here, 'and_with' does match something outside the bitmap, and cl
882 * doesn't have a list of things to match outside the bitmap. If
883 * cl can match all code points above 255, the intersection will
884 * be those above-255 code points that 'and_with' matches. If cl
885 * can't match all Unicode code points, it means that it can't
886 * match anything outside the bitmap (since the 'if' that got us
887 * into this block tested for that), so we leave the bitmap empty.
889 if (cl->flags & ANYOF_UNICODE_ALL) {
890 ARG_SET(cl, ARG(and_with));
892 /* and_with's ARG may match things that don't require UTF8.
893 * And now cl's will too, in spite of this being an 'and'. See
894 * the comments below about the kludge */
895 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
899 /* Here, both 'and_with' and cl match something outside the
900 * bitmap. Currently we do not do the intersection, so just match
901 * whatever cl had at the beginning. */
905 /* Take the intersection of the two sets of flags. However, the
906 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
907 * kludge around the fact that this flag is not treated like the others
908 * which are initialized in cl_anything(). The way the optimizer works
909 * is that the synthetic start class (SSC) is initialized to match
910 * anything, and then the first time a real node is encountered, its
911 * values are AND'd with the SSC's with the result being the values of
912 * the real node. However, there are paths through the optimizer where
913 * the AND never gets called, so those initialized bits are set
914 * inappropriately, which is not usually a big deal, as they just cause
915 * false positives in the SSC, which will just mean a probably
916 * imperceptible slow down in execution. However this bit has a
917 * higher false positive consequence in that it can cause utf8.pm,
918 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
919 * bigger slowdown and also causes significant extra memory to be used.
920 * In order to prevent this, the code now takes a different tack. The
921 * bit isn't set unless some part of the regular expression needs it,
922 * but once set it won't get cleared. This means that these extra
923 * modules won't get loaded unless there was some path through the
924 * pattern that would have required them anyway, and so any false
925 * positives that occur by not ANDing them out when they could be
926 * aren't as severe as they would be if we treated this bit like all
928 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
929 & ANYOF_NONBITMAP_NON_UTF8;
930 cl->flags &= and_with->flags;
931 cl->flags |= outside_bitmap_but_not_utf8;
935 /* 'OR' a given class with another one. Can create false positives. 'cl'
936 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
937 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
939 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
941 PERL_ARGS_ASSERT_CL_OR;
943 if (or_with->flags & ANYOF_INVERT) {
945 /* Here, the or'd node is to be inverted. This means we take the
946 * complement of everything not in the bitmap, but currently we don't
947 * know what that is, so give up and match anything */
948 if (ANYOF_NONBITMAP(or_with)) {
949 cl_anything(pRExC_state, cl);
952 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
953 * <= (B1 | !B2) | (CL1 | !CL2)
954 * which is wasteful if CL2 is small, but we ignore CL2:
955 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
956 * XXXX Can we handle case-fold? Unclear:
957 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
958 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
960 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
961 && !(or_with->flags & ANYOF_LOC_FOLD)
962 && !(cl->flags & ANYOF_LOC_FOLD) ) {
965 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
966 cl->bitmap[i] |= ~or_with->bitmap[i];
967 } /* XXXX: logic is complicated otherwise */
969 cl_anything(pRExC_state, cl);
972 /* And, we can just take the union of the flags that aren't affected
973 * by the inversion */
974 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
976 /* For the remaining flags:
977 ANYOF_UNICODE_ALL and inverted means to not match anything above
978 255, which means that the union with cl should just be
979 what cl has in it, so can ignore this flag
980 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
981 is 127-255 to match them, but then invert that, so the
982 union with cl should just be what cl has in it, so can
985 } else { /* 'or_with' is not inverted */
986 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
987 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
988 && (!(or_with->flags & ANYOF_LOC_FOLD)
989 || (cl->flags & ANYOF_LOC_FOLD)) ) {
992 /* OR char bitmap and class bitmap separately */
993 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
994 cl->bitmap[i] |= or_with->bitmap[i];
995 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
996 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
997 cl->classflags[i] |= or_with->classflags[i];
998 cl->flags |= ANYOF_CLASS;
1001 else { /* XXXX: logic is complicated, leave it along for a moment. */
1002 cl_anything(pRExC_state, cl);
1005 if (ANYOF_NONBITMAP(or_with)) {
1007 /* Use the added node's outside-the-bit-map match if there isn't a
1008 * conflict. If there is a conflict (both nodes match something
1009 * outside the bitmap, but what they match outside is not the same
1010 * pointer, and hence not easily compared until XXX we extend
1011 * inversion lists this far), give up and allow the start class to
1012 * match everything outside the bitmap. If that stuff is all above
1013 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1014 if (! ANYOF_NONBITMAP(cl)) {
1015 ARG_SET(cl, ARG(or_with));
1017 else if (ARG(cl) != ARG(or_with)) {
1019 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1020 cl_anything(pRExC_state, cl);
1023 cl->flags |= ANYOF_UNICODE_ALL;
1028 /* Take the union */
1029 cl->flags |= or_with->flags;
1033 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1034 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1035 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1036 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1041 dump_trie(trie,widecharmap,revcharmap)
1042 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1043 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1045 These routines dump out a trie in a somewhat readable format.
1046 The _interim_ variants are used for debugging the interim
1047 tables that are used to generate the final compressed
1048 representation which is what dump_trie expects.
1050 Part of the reason for their existence is to provide a form
1051 of documentation as to how the different representations function.
1056 Dumps the final compressed table form of the trie to Perl_debug_log.
1057 Used for debugging make_trie().
1061 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1062 AV *revcharmap, U32 depth)
1065 SV *sv=sv_newmortal();
1066 int colwidth= widecharmap ? 6 : 4;
1068 GET_RE_DEBUG_FLAGS_DECL;
1070 PERL_ARGS_ASSERT_DUMP_TRIE;
1072 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1073 (int)depth * 2 + 2,"",
1074 "Match","Base","Ofs" );
1076 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1077 SV ** const tmp = av_fetch( revcharmap, state, 0);
1079 PerlIO_printf( Perl_debug_log, "%*s",
1081 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1082 PL_colors[0], PL_colors[1],
1083 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1084 PERL_PV_ESCAPE_FIRSTCHAR
1089 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1090 (int)depth * 2 + 2,"");
1092 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1093 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1094 PerlIO_printf( Perl_debug_log, "\n");
1096 for( state = 1 ; state < trie->statecount ; state++ ) {
1097 const U32 base = trie->states[ state ].trans.base;
1099 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1101 if ( trie->states[ state ].wordnum ) {
1102 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1104 PerlIO_printf( Perl_debug_log, "%6s", "" );
1107 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1112 while( ( base + ofs < trie->uniquecharcount ) ||
1113 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1114 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1117 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1119 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1120 if ( ( base + ofs >= trie->uniquecharcount ) &&
1121 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1122 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1124 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1126 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1128 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1132 PerlIO_printf( Perl_debug_log, "]");
1135 PerlIO_printf( Perl_debug_log, "\n" );
1137 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1138 for (word=1; word <= trie->wordcount; word++) {
1139 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1140 (int)word, (int)(trie->wordinfo[word].prev),
1141 (int)(trie->wordinfo[word].len));
1143 PerlIO_printf(Perl_debug_log, "\n" );
1146 Dumps a fully constructed but uncompressed trie in list form.
1147 List tries normally only are used for construction when the number of
1148 possible chars (trie->uniquecharcount) is very high.
1149 Used for debugging make_trie().
1152 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1153 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1157 SV *sv=sv_newmortal();
1158 int colwidth= widecharmap ? 6 : 4;
1159 GET_RE_DEBUG_FLAGS_DECL;
1161 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1163 /* print out the table precompression. */
1164 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1165 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1166 "------:-----+-----------------\n" );
1168 for( state=1 ; state < next_alloc ; state ++ ) {
1171 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1172 (int)depth * 2 + 2,"", (UV)state );
1173 if ( ! trie->states[ state ].wordnum ) {
1174 PerlIO_printf( Perl_debug_log, "%5s| ","");
1176 PerlIO_printf( Perl_debug_log, "W%4x| ",
1177 trie->states[ state ].wordnum
1180 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1181 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1183 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1185 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1186 PL_colors[0], PL_colors[1],
1187 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1188 PERL_PV_ESCAPE_FIRSTCHAR
1190 TRIE_LIST_ITEM(state,charid).forid,
1191 (UV)TRIE_LIST_ITEM(state,charid).newstate
1194 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1195 (int)((depth * 2) + 14), "");
1198 PerlIO_printf( Perl_debug_log, "\n");
1203 Dumps a fully constructed but uncompressed trie in table form.
1204 This is the normal DFA style state transition table, with a few
1205 twists to facilitate compression later.
1206 Used for debugging make_trie().
1209 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1210 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1215 SV *sv=sv_newmortal();
1216 int colwidth= widecharmap ? 6 : 4;
1217 GET_RE_DEBUG_FLAGS_DECL;
1219 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1222 print out the table precompression so that we can do a visual check
1223 that they are identical.
1226 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1228 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1229 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1231 PerlIO_printf( Perl_debug_log, "%*s",
1233 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1234 PL_colors[0], PL_colors[1],
1235 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1236 PERL_PV_ESCAPE_FIRSTCHAR
1242 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1244 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1245 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1248 PerlIO_printf( Perl_debug_log, "\n" );
1250 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1252 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1253 (int)depth * 2 + 2,"",
1254 (UV)TRIE_NODENUM( state ) );
1256 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1257 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1259 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1261 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1263 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1264 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1266 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1267 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1275 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1276 startbranch: the first branch in the whole branch sequence
1277 first : start branch of sequence of branch-exact nodes.
1278 May be the same as startbranch
1279 last : Thing following the last branch.
1280 May be the same as tail.
1281 tail : item following the branch sequence
1282 count : words in the sequence
1283 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1284 depth : indent depth
1286 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1288 A trie is an N'ary tree where the branches are determined by digital
1289 decomposition of the key. IE, at the root node you look up the 1st character and
1290 follow that branch repeat until you find the end of the branches. Nodes can be
1291 marked as "accepting" meaning they represent a complete word. Eg:
1295 would convert into the following structure. Numbers represent states, letters
1296 following numbers represent valid transitions on the letter from that state, if
1297 the number is in square brackets it represents an accepting state, otherwise it
1298 will be in parenthesis.
1300 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1304 (1) +-i->(6)-+-s->[7]
1306 +-s->(3)-+-h->(4)-+-e->[5]
1308 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1310 This shows that when matching against the string 'hers' we will begin at state 1
1311 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1312 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1313 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1314 single traverse. We store a mapping from accepting to state to which word was
1315 matched, and then when we have multiple possibilities we try to complete the
1316 rest of the regex in the order in which they occured in the alternation.
1318 The only prior NFA like behaviour that would be changed by the TRIE support is
1319 the silent ignoring of duplicate alternations which are of the form:
1321 / (DUPE|DUPE) X? (?{ ... }) Y /x
1323 Thus EVAL blocks following a trie may be called a different number of times with
1324 and without the optimisation. With the optimisations dupes will be silently
1325 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1326 the following demonstrates:
1328 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1330 which prints out 'word' three times, but
1332 'words'=~/(word|word|word)(?{ print $1 })S/
1334 which doesnt print it out at all. This is due to other optimisations kicking in.
1336 Example of what happens on a structural level:
1338 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1340 1: CURLYM[1] {1,32767}(18)
1351 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1352 and should turn into:
1354 1: CURLYM[1] {1,32767}(18)
1356 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1364 Cases where tail != last would be like /(?foo|bar)baz/:
1374 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1375 and would end up looking like:
1378 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1385 d = uvuni_to_utf8_flags(d, uv, 0);
1387 is the recommended Unicode-aware way of saying
1392 #define TRIE_STORE_REVCHAR(val) \
1395 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1396 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1397 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1398 SvCUR_set(zlopp, kapow - flrbbbbb); \
1401 av_push(revcharmap, zlopp); \
1403 char ooooff = (char)val; \
1404 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1408 #define TRIE_READ_CHAR STMT_START { \
1411 /* if it is UTF then it is either already folded, or does not need folding */ \
1412 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1414 else if (folder == PL_fold_latin1) { \
1415 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1416 if ( foldlen > 0 ) { \
1417 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1423 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1424 skiplen = UNISKIP(uvc); \
1425 foldlen -= skiplen; \
1426 scan = foldbuf + skiplen; \
1429 /* raw data, will be folded later if needed */ \
1437 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1438 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1439 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1440 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1442 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1443 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1444 TRIE_LIST_CUR( state )++; \
1447 #define TRIE_LIST_NEW(state) STMT_START { \
1448 Newxz( trie->states[ state ].trans.list, \
1449 4, reg_trie_trans_le ); \
1450 TRIE_LIST_CUR( state ) = 1; \
1451 TRIE_LIST_LEN( state ) = 4; \
1454 #define TRIE_HANDLE_WORD(state) STMT_START { \
1455 U16 dupe= trie->states[ state ].wordnum; \
1456 regnode * const noper_next = regnext( noper ); \
1459 /* store the word for dumping */ \
1461 if (OP(noper) != NOTHING) \
1462 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1464 tmp = newSVpvn_utf8( "", 0, UTF ); \
1465 av_push( trie_words, tmp ); \
1469 trie->wordinfo[curword].prev = 0; \
1470 trie->wordinfo[curword].len = wordlen; \
1471 trie->wordinfo[curword].accept = state; \
1473 if ( noper_next < tail ) { \
1475 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1476 trie->jump[curword] = (U16)(noper_next - convert); \
1478 jumper = noper_next; \
1480 nextbranch= regnext(cur); \
1484 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1485 /* chain, so that when the bits of chain are later */\
1486 /* linked together, the dups appear in the chain */\
1487 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1488 trie->wordinfo[dupe].prev = curword; \
1490 /* we haven't inserted this word yet. */ \
1491 trie->states[ state ].wordnum = curword; \
1496 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1497 ( ( base + charid >= ucharcount \
1498 && base + charid < ubound \
1499 && state == trie->trans[ base - ucharcount + charid ].check \
1500 && trie->trans[ base - ucharcount + charid ].next ) \
1501 ? trie->trans[ base - ucharcount + charid ].next \
1502 : ( state==1 ? special : 0 ) \
1506 #define MADE_JUMP_TRIE 2
1507 #define MADE_EXACT_TRIE 4
1510 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1513 /* first pass, loop through and scan words */
1514 reg_trie_data *trie;
1515 HV *widecharmap = NULL;
1516 AV *revcharmap = newAV();
1518 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1523 regnode *jumper = NULL;
1524 regnode *nextbranch = NULL;
1525 regnode *convert = NULL;
1526 U32 *prev_states; /* temp array mapping each state to previous one */
1527 /* we just use folder as a flag in utf8 */
1528 const U8 * folder = NULL;
1531 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1532 AV *trie_words = NULL;
1533 /* along with revcharmap, this only used during construction but both are
1534 * useful during debugging so we store them in the struct when debugging.
1537 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1538 STRLEN trie_charcount=0;
1540 SV *re_trie_maxbuff;
1541 GET_RE_DEBUG_FLAGS_DECL;
1543 PERL_ARGS_ASSERT_MAKE_TRIE;
1545 PERL_UNUSED_ARG(depth);
1552 case EXACTFU_TRICKYFOLD:
1553 case EXACTFU: folder = PL_fold_latin1; break;
1554 case EXACTF: folder = PL_fold; break;
1555 case EXACTFL: folder = PL_fold_locale; break;
1556 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1559 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1561 trie->startstate = 1;
1562 trie->wordcount = word_count;
1563 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1564 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1566 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1567 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1568 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1571 trie_words = newAV();
1574 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1575 if (!SvIOK(re_trie_maxbuff)) {
1576 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1578 DEBUG_TRIE_COMPILE_r({
1579 PerlIO_printf( Perl_debug_log,
1580 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1581 (int)depth * 2 + 2, "",
1582 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1583 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1587 /* Find the node we are going to overwrite */
1588 if ( first == startbranch && OP( last ) != BRANCH ) {
1589 /* whole branch chain */
1592 /* branch sub-chain */
1593 convert = NEXTOPER( first );
1596 /* -- First loop and Setup --
1598 We first traverse the branches and scan each word to determine if it
1599 contains widechars, and how many unique chars there are, this is
1600 important as we have to build a table with at least as many columns as we
1603 We use an array of integers to represent the character codes 0..255
1604 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1605 native representation of the character value as the key and IV's for the
1608 *TODO* If we keep track of how many times each character is used we can
1609 remap the columns so that the table compression later on is more
1610 efficient in terms of memory by ensuring the most common value is in the
1611 middle and the least common are on the outside. IMO this would be better
1612 than a most to least common mapping as theres a decent chance the most
1613 common letter will share a node with the least common, meaning the node
1614 will not be compressible. With a middle is most common approach the worst
1615 case is when we have the least common nodes twice.
1619 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1620 regnode *noper = NEXTOPER( cur );
1621 const U8 *uc = (U8*)STRING( noper );
1622 const U8 *e = uc + STR_LEN( noper );
1624 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1626 const U8 *scan = (U8*)NULL;
1627 U32 wordlen = 0; /* required init */
1629 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1631 if (OP(noper) == NOTHING) {
1632 regnode *noper_next= regnext(noper);
1633 if (noper_next != tail && OP(noper_next) == flags) {
1635 uc= (U8*)STRING(noper);
1636 e= uc + STR_LEN(noper);
1637 trie->minlen= STR_LEN(noper);
1644 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1645 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1646 regardless of encoding */
1647 if (OP( noper ) == EXACTFU_SS) {
1648 /* false positives are ok, so just set this */
1649 TRIE_BITMAP_SET(trie,0xDF);
1652 for ( ; uc < e ; uc += len ) {
1653 TRIE_CHARCOUNT(trie)++;
1658 U8 folded= folder[ (U8) uvc ];
1659 if ( !trie->charmap[ folded ] ) {
1660 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1661 TRIE_STORE_REVCHAR( folded );
1664 if ( !trie->charmap[ uvc ] ) {
1665 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1666 TRIE_STORE_REVCHAR( uvc );
1669 /* store the codepoint in the bitmap, and its folded
1671 TRIE_BITMAP_SET(trie, uvc);
1673 /* store the folded codepoint */
1674 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1677 /* store first byte of utf8 representation of
1678 variant codepoints */
1679 if (! UNI_IS_INVARIANT(uvc)) {
1680 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1683 set_bit = 0; /* We've done our bit :-) */
1688 widecharmap = newHV();
1690 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1693 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1695 if ( !SvTRUE( *svpp ) ) {
1696 sv_setiv( *svpp, ++trie->uniquecharcount );
1697 TRIE_STORE_REVCHAR(uvc);
1701 if( cur == first ) {
1702 trie->minlen = chars;
1703 trie->maxlen = chars;
1704 } else if (chars < trie->minlen) {
1705 trie->minlen = chars;
1706 } else if (chars > trie->maxlen) {
1707 trie->maxlen = chars;
1709 if (OP( noper ) == EXACTFU_SS) {
1710 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1711 if (trie->minlen > 1)
1714 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1715 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1716 * - We assume that any such sequence might match a 2 byte string */
1717 if (trie->minlen > 2 )
1721 } /* end first pass */
1722 DEBUG_TRIE_COMPILE_r(
1723 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1724 (int)depth * 2 + 2,"",
1725 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1726 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1727 (int)trie->minlen, (int)trie->maxlen )
1731 We now know what we are dealing with in terms of unique chars and
1732 string sizes so we can calculate how much memory a naive
1733 representation using a flat table will take. If it's over a reasonable
1734 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1735 conservative but potentially much slower representation using an array
1738 At the end we convert both representations into the same compressed
1739 form that will be used in regexec.c for matching with. The latter
1740 is a form that cannot be used to construct with but has memory
1741 properties similar to the list form and access properties similar
1742 to the table form making it both suitable for fast searches and
1743 small enough that its feasable to store for the duration of a program.
1745 See the comment in the code where the compressed table is produced
1746 inplace from the flat tabe representation for an explanation of how
1747 the compression works.
1752 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1755 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1757 Second Pass -- Array Of Lists Representation
1759 Each state will be represented by a list of charid:state records
1760 (reg_trie_trans_le) the first such element holds the CUR and LEN
1761 points of the allocated array. (See defines above).
1763 We build the initial structure using the lists, and then convert
1764 it into the compressed table form which allows faster lookups
1765 (but cant be modified once converted).
1768 STRLEN transcount = 1;
1770 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1771 "%*sCompiling trie using list compiler\n",
1772 (int)depth * 2 + 2, ""));
1774 trie->states = (reg_trie_state *)
1775 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1776 sizeof(reg_trie_state) );
1780 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1782 regnode *noper = NEXTOPER( cur );
1783 U8 *uc = (U8*)STRING( noper );
1784 const U8 *e = uc + STR_LEN( noper );
1785 U32 state = 1; /* required init */
1786 U16 charid = 0; /* sanity init */
1787 U8 *scan = (U8*)NULL; /* sanity init */
1788 STRLEN foldlen = 0; /* required init */
1789 U32 wordlen = 0; /* required init */
1790 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1793 if (OP(noper) == NOTHING) {
1794 regnode *noper_next= regnext(noper);
1795 if (noper_next != tail && OP(noper_next) == flags) {
1797 uc= (U8*)STRING(noper);
1798 e= uc + STR_LEN(noper);
1802 if (OP(noper) != NOTHING) {
1803 for ( ; uc < e ; uc += len ) {
1808 charid = trie->charmap[ uvc ];
1810 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1814 charid=(U16)SvIV( *svpp );
1817 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1824 if ( !trie->states[ state ].trans.list ) {
1825 TRIE_LIST_NEW( state );
1827 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1828 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1829 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1834 newstate = next_alloc++;
1835 prev_states[newstate] = state;
1836 TRIE_LIST_PUSH( state, charid, newstate );
1841 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1845 TRIE_HANDLE_WORD(state);
1847 } /* end second pass */
1849 /* next alloc is the NEXT state to be allocated */
1850 trie->statecount = next_alloc;
1851 trie->states = (reg_trie_state *)
1852 PerlMemShared_realloc( trie->states,
1854 * sizeof(reg_trie_state) );
1856 /* and now dump it out before we compress it */
1857 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1858 revcharmap, next_alloc,
1862 trie->trans = (reg_trie_trans *)
1863 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1870 for( state=1 ; state < next_alloc ; state ++ ) {
1874 DEBUG_TRIE_COMPILE_MORE_r(
1875 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1879 if (trie->states[state].trans.list) {
1880 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1884 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1885 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1886 if ( forid < minid ) {
1888 } else if ( forid > maxid ) {
1892 if ( transcount < tp + maxid - minid + 1) {
1894 trie->trans = (reg_trie_trans *)
1895 PerlMemShared_realloc( trie->trans,
1897 * sizeof(reg_trie_trans) );
1898 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1900 base = trie->uniquecharcount + tp - minid;
1901 if ( maxid == minid ) {
1903 for ( ; zp < tp ; zp++ ) {
1904 if ( ! trie->trans[ zp ].next ) {
1905 base = trie->uniquecharcount + zp - minid;
1906 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1907 trie->trans[ zp ].check = state;
1913 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1914 trie->trans[ tp ].check = state;
1919 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1920 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1921 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1922 trie->trans[ tid ].check = state;
1924 tp += ( maxid - minid + 1 );
1926 Safefree(trie->states[ state ].trans.list);
1929 DEBUG_TRIE_COMPILE_MORE_r(
1930 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1933 trie->states[ state ].trans.base=base;
1935 trie->lasttrans = tp + 1;
1939 Second Pass -- Flat Table Representation.
1941 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1942 We know that we will need Charcount+1 trans at most to store the data
1943 (one row per char at worst case) So we preallocate both structures
1944 assuming worst case.
1946 We then construct the trie using only the .next slots of the entry
1949 We use the .check field of the first entry of the node temporarily to
1950 make compression both faster and easier by keeping track of how many non
1951 zero fields are in the node.
1953 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1956 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1957 number representing the first entry of the node, and state as a
1958 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1959 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1960 are 2 entrys per node. eg:
1968 The table is internally in the right hand, idx form. However as we also
1969 have to deal with the states array which is indexed by nodenum we have to
1970 use TRIE_NODENUM() to convert.
1973 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1974 "%*sCompiling trie using table compiler\n",
1975 (int)depth * 2 + 2, ""));
1977 trie->trans = (reg_trie_trans *)
1978 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1979 * trie->uniquecharcount + 1,
1980 sizeof(reg_trie_trans) );
1981 trie->states = (reg_trie_state *)
1982 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1983 sizeof(reg_trie_state) );
1984 next_alloc = trie->uniquecharcount + 1;
1987 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1989 regnode *noper = NEXTOPER( cur );
1990 const U8 *uc = (U8*)STRING( noper );
1991 const U8 *e = uc + STR_LEN( noper );
1993 U32 state = 1; /* required init */
1995 U16 charid = 0; /* sanity init */
1996 U32 accept_state = 0; /* sanity init */
1997 U8 *scan = (U8*)NULL; /* sanity init */
1999 STRLEN foldlen = 0; /* required init */
2000 U32 wordlen = 0; /* required init */
2002 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2004 if (OP(noper) == NOTHING) {
2005 regnode *noper_next= regnext(noper);
2006 if (noper_next != tail && OP(noper_next) == flags) {
2008 uc= (U8*)STRING(noper);
2009 e= uc + STR_LEN(noper);
2013 if ( OP(noper) != NOTHING ) {
2014 for ( ; uc < e ; uc += len ) {
2019 charid = trie->charmap[ uvc ];
2021 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2022 charid = svpp ? (U16)SvIV(*svpp) : 0;
2026 if ( !trie->trans[ state + charid ].next ) {
2027 trie->trans[ state + charid ].next = next_alloc;
2028 trie->trans[ state ].check++;
2029 prev_states[TRIE_NODENUM(next_alloc)]
2030 = TRIE_NODENUM(state);
2031 next_alloc += trie->uniquecharcount;
2033 state = trie->trans[ state + charid ].next;
2035 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2037 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2040 accept_state = TRIE_NODENUM( state );
2041 TRIE_HANDLE_WORD(accept_state);
2043 } /* end second pass */
2045 /* and now dump it out before we compress it */
2046 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2048 next_alloc, depth+1));
2052 * Inplace compress the table.*
2054 For sparse data sets the table constructed by the trie algorithm will
2055 be mostly 0/FAIL transitions or to put it another way mostly empty.
2056 (Note that leaf nodes will not contain any transitions.)
2058 This algorithm compresses the tables by eliminating most such
2059 transitions, at the cost of a modest bit of extra work during lookup:
2061 - Each states[] entry contains a .base field which indicates the
2062 index in the state[] array wheres its transition data is stored.
2064 - If .base is 0 there are no valid transitions from that node.
2066 - If .base is nonzero then charid is added to it to find an entry in
2069 -If trans[states[state].base+charid].check!=state then the
2070 transition is taken to be a 0/Fail transition. Thus if there are fail
2071 transitions at the front of the node then the .base offset will point
2072 somewhere inside the previous nodes data (or maybe even into a node
2073 even earlier), but the .check field determines if the transition is
2077 The following process inplace converts the table to the compressed
2078 table: We first do not compress the root node 1,and mark all its
2079 .check pointers as 1 and set its .base pointer as 1 as well. This
2080 allows us to do a DFA construction from the compressed table later,
2081 and ensures that any .base pointers we calculate later are greater
2084 - We set 'pos' to indicate the first entry of the second node.
2086 - We then iterate over the columns of the node, finding the first and
2087 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2088 and set the .check pointers accordingly, and advance pos
2089 appropriately and repreat for the next node. Note that when we copy
2090 the next pointers we have to convert them from the original
2091 NODEIDX form to NODENUM form as the former is not valid post
2094 - If a node has no transitions used we mark its base as 0 and do not
2095 advance the pos pointer.
2097 - If a node only has one transition we use a second pointer into the
2098 structure to fill in allocated fail transitions from other states.
2099 This pointer is independent of the main pointer and scans forward
2100 looking for null transitions that are allocated to a state. When it
2101 finds one it writes the single transition into the "hole". If the
2102 pointer doesnt find one the single transition is appended as normal.
2104 - Once compressed we can Renew/realloc the structures to release the
2107 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2108 specifically Fig 3.47 and the associated pseudocode.
2112 const U32 laststate = TRIE_NODENUM( next_alloc );
2115 trie->statecount = laststate;
2117 for ( state = 1 ; state < laststate ; state++ ) {
2119 const U32 stateidx = TRIE_NODEIDX( state );
2120 const U32 o_used = trie->trans[ stateidx ].check;
2121 U32 used = trie->trans[ stateidx ].check;
2122 trie->trans[ stateidx ].check = 0;
2124 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2125 if ( flag || trie->trans[ stateidx + charid ].next ) {
2126 if ( trie->trans[ stateidx + charid ].next ) {
2128 for ( ; zp < pos ; zp++ ) {
2129 if ( ! trie->trans[ zp ].next ) {
2133 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2134 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2135 trie->trans[ zp ].check = state;
2136 if ( ++zp > pos ) pos = zp;
2143 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2145 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2146 trie->trans[ pos ].check = state;
2151 trie->lasttrans = pos + 1;
2152 trie->states = (reg_trie_state *)
2153 PerlMemShared_realloc( trie->states, laststate
2154 * sizeof(reg_trie_state) );
2155 DEBUG_TRIE_COMPILE_MORE_r(
2156 PerlIO_printf( Perl_debug_log,
2157 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2158 (int)depth * 2 + 2,"",
2159 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2162 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2165 } /* end table compress */
2167 DEBUG_TRIE_COMPILE_MORE_r(
2168 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2169 (int)depth * 2 + 2, "",
2170 (UV)trie->statecount,
2171 (UV)trie->lasttrans)
2173 /* resize the trans array to remove unused space */
2174 trie->trans = (reg_trie_trans *)
2175 PerlMemShared_realloc( trie->trans, trie->lasttrans
2176 * sizeof(reg_trie_trans) );
2178 { /* Modify the program and insert the new TRIE node */
2179 U8 nodetype =(U8)(flags & 0xFF);
2183 regnode *optimize = NULL;
2184 #ifdef RE_TRACK_PATTERN_OFFSETS
2187 U32 mjd_nodelen = 0;
2188 #endif /* RE_TRACK_PATTERN_OFFSETS */
2189 #endif /* DEBUGGING */
2191 This means we convert either the first branch or the first Exact,
2192 depending on whether the thing following (in 'last') is a branch
2193 or not and whther first is the startbranch (ie is it a sub part of
2194 the alternation or is it the whole thing.)
2195 Assuming its a sub part we convert the EXACT otherwise we convert
2196 the whole branch sequence, including the first.
2198 /* Find the node we are going to overwrite */
2199 if ( first != startbranch || OP( last ) == BRANCH ) {
2200 /* branch sub-chain */
2201 NEXT_OFF( first ) = (U16)(last - first);
2202 #ifdef RE_TRACK_PATTERN_OFFSETS
2204 mjd_offset= Node_Offset((convert));
2205 mjd_nodelen= Node_Length((convert));
2208 /* whole branch chain */
2210 #ifdef RE_TRACK_PATTERN_OFFSETS
2213 const regnode *nop = NEXTOPER( convert );
2214 mjd_offset= Node_Offset((nop));
2215 mjd_nodelen= Node_Length((nop));
2219 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2220 (int)depth * 2 + 2, "",
2221 (UV)mjd_offset, (UV)mjd_nodelen)
2224 /* But first we check to see if there is a common prefix we can
2225 split out as an EXACT and put in front of the TRIE node. */
2226 trie->startstate= 1;
2227 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2229 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2233 const U32 base = trie->states[ state ].trans.base;
2235 if ( trie->states[state].wordnum )
2238 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2239 if ( ( base + ofs >= trie->uniquecharcount ) &&
2240 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2241 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2243 if ( ++count > 1 ) {
2244 SV **tmp = av_fetch( revcharmap, ofs, 0);
2245 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2246 if ( state == 1 ) break;
2248 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2250 PerlIO_printf(Perl_debug_log,
2251 "%*sNew Start State=%"UVuf" Class: [",
2252 (int)depth * 2 + 2, "",
2255 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2256 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2258 TRIE_BITMAP_SET(trie,*ch);
2260 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2262 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2266 TRIE_BITMAP_SET(trie,*ch);
2268 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2269 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2275 SV **tmp = av_fetch( revcharmap, idx, 0);
2277 char *ch = SvPV( *tmp, len );
2279 SV *sv=sv_newmortal();
2280 PerlIO_printf( Perl_debug_log,
2281 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2282 (int)depth * 2 + 2, "",
2284 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2285 PL_colors[0], PL_colors[1],
2286 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2287 PERL_PV_ESCAPE_FIRSTCHAR
2292 OP( convert ) = nodetype;
2293 str=STRING(convert);
2296 STR_LEN(convert) += len;
2302 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2307 trie->prefixlen = (state-1);
2309 regnode *n = convert+NODE_SZ_STR(convert);
2310 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2311 trie->startstate = state;
2312 trie->minlen -= (state - 1);
2313 trie->maxlen -= (state - 1);
2315 /* At least the UNICOS C compiler choked on this
2316 * being argument to DEBUG_r(), so let's just have
2319 #ifdef PERL_EXT_RE_BUILD
2325 regnode *fix = convert;
2326 U32 word = trie->wordcount;
2328 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2329 while( ++fix < n ) {
2330 Set_Node_Offset_Length(fix, 0, 0);
2333 SV ** const tmp = av_fetch( trie_words, word, 0 );
2335 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2336 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2338 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2346 NEXT_OFF(convert) = (U16)(tail - convert);
2347 DEBUG_r(optimize= n);
2353 if ( trie->maxlen ) {
2354 NEXT_OFF( convert ) = (U16)(tail - convert);
2355 ARG_SET( convert, data_slot );
2356 /* Store the offset to the first unabsorbed branch in
2357 jump[0], which is otherwise unused by the jump logic.
2358 We use this when dumping a trie and during optimisation. */
2360 trie->jump[0] = (U16)(nextbranch - convert);
2362 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2363 * and there is a bitmap
2364 * and the first "jump target" node we found leaves enough room
2365 * then convert the TRIE node into a TRIEC node, with the bitmap
2366 * embedded inline in the opcode - this is hypothetically faster.
2368 if ( !trie->states[trie->startstate].wordnum
2370 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2372 OP( convert ) = TRIEC;
2373 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2374 PerlMemShared_free(trie->bitmap);
2377 OP( convert ) = TRIE;
2379 /* store the type in the flags */
2380 convert->flags = nodetype;
2384 + regarglen[ OP( convert ) ];
2386 /* XXX We really should free up the resource in trie now,
2387 as we won't use them - (which resources?) dmq */
2389 /* needed for dumping*/
2390 DEBUG_r(if (optimize) {
2391 regnode *opt = convert;
2393 while ( ++opt < optimize) {
2394 Set_Node_Offset_Length(opt,0,0);
2397 Try to clean up some of the debris left after the
2400 while( optimize < jumper ) {
2401 mjd_nodelen += Node_Length((optimize));
2402 OP( optimize ) = OPTIMIZED;
2403 Set_Node_Offset_Length(optimize,0,0);
2406 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2408 } /* end node insert */
2410 /* Finish populating the prev field of the wordinfo array. Walk back
2411 * from each accept state until we find another accept state, and if
2412 * so, point the first word's .prev field at the second word. If the
2413 * second already has a .prev field set, stop now. This will be the
2414 * case either if we've already processed that word's accept state,
2415 * or that state had multiple words, and the overspill words were
2416 * already linked up earlier.
2423 for (word=1; word <= trie->wordcount; word++) {
2425 if (trie->wordinfo[word].prev)
2427 state = trie->wordinfo[word].accept;
2429 state = prev_states[state];
2432 prev = trie->states[state].wordnum;
2436 trie->wordinfo[word].prev = prev;
2438 Safefree(prev_states);
2442 /* and now dump out the compressed format */
2443 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2445 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2447 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2448 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2450 SvREFCNT_dec(revcharmap);
2454 : trie->startstate>1
2460 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2462 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2464 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2465 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2468 We find the fail state for each state in the trie, this state is the longest proper
2469 suffix of the current state's 'word' that is also a proper prefix of another word in our
2470 trie. State 1 represents the word '' and is thus the default fail state. This allows
2471 the DFA not to have to restart after its tried and failed a word at a given point, it
2472 simply continues as though it had been matching the other word in the first place.
2474 'abcdgu'=~/abcdefg|cdgu/
2475 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2476 fail, which would bring us to the state representing 'd' in the second word where we would
2477 try 'g' and succeed, proceeding to match 'cdgu'.
2479 /* add a fail transition */
2480 const U32 trie_offset = ARG(source);
2481 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2483 const U32 ucharcount = trie->uniquecharcount;
2484 const U32 numstates = trie->statecount;
2485 const U32 ubound = trie->lasttrans + ucharcount;
2489 U32 base = trie->states[ 1 ].trans.base;
2492 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2493 GET_RE_DEBUG_FLAGS_DECL;
2495 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2497 PERL_UNUSED_ARG(depth);
2501 ARG_SET( stclass, data_slot );
2502 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2503 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2504 aho->trie=trie_offset;
2505 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2506 Copy( trie->states, aho->states, numstates, reg_trie_state );
2507 Newxz( q, numstates, U32);
2508 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2511 /* initialize fail[0..1] to be 1 so that we always have
2512 a valid final fail state */
2513 fail[ 0 ] = fail[ 1 ] = 1;
2515 for ( charid = 0; charid < ucharcount ; charid++ ) {
2516 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2518 q[ q_write ] = newstate;
2519 /* set to point at the root */
2520 fail[ q[ q_write++ ] ]=1;
2523 while ( q_read < q_write) {
2524 const U32 cur = q[ q_read++ % numstates ];
2525 base = trie->states[ cur ].trans.base;
2527 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2528 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2530 U32 fail_state = cur;
2533 fail_state = fail[ fail_state ];
2534 fail_base = aho->states[ fail_state ].trans.base;
2535 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2537 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2538 fail[ ch_state ] = fail_state;
2539 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2541 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2543 q[ q_write++ % numstates] = ch_state;
2547 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2548 when we fail in state 1, this allows us to use the
2549 charclass scan to find a valid start char. This is based on the principle
2550 that theres a good chance the string being searched contains lots of stuff
2551 that cant be a start char.
2553 fail[ 0 ] = fail[ 1 ] = 0;
2554 DEBUG_TRIE_COMPILE_r({
2555 PerlIO_printf(Perl_debug_log,
2556 "%*sStclass Failtable (%"UVuf" states): 0",
2557 (int)(depth * 2), "", (UV)numstates
2559 for( q_read=1; q_read<numstates; q_read++ ) {
2560 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2562 PerlIO_printf(Perl_debug_log, "\n");
2565 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2570 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2571 * These need to be revisited when a newer toolchain becomes available.
2573 #if defined(__sparc64__) && defined(__GNUC__)
2574 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2575 # undef SPARC64_GCC_WORKAROUND
2576 # define SPARC64_GCC_WORKAROUND 1
2580 #define DEBUG_PEEP(str,scan,depth) \
2581 DEBUG_OPTIMISE_r({if (scan){ \
2582 SV * const mysv=sv_newmortal(); \
2583 regnode *Next = regnext(scan); \
2584 regprop(RExC_rx, mysv, scan); \
2585 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2586 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2587 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2591 /* The below joins as many adjacent EXACTish nodes as possible into a single
2592 * one. The regop may be changed if the node(s) contain certain sequences that
2593 * require special handling. The joining is only done if:
2594 * 1) there is room in the current conglomerated node to entirely contain the
2596 * 2) they are the exact same node type
2598 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2599 * these get optimized out
2601 * If a node is to match under /i (folded), the number of characters it matches
2602 * can be different than its character length if it contains a multi-character
2603 * fold. *min_subtract is set to the total delta of the input nodes.
2605 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2606 * and contains LATIN SMALL LETTER SHARP S
2608 * This is as good a place as any to discuss the design of handling these
2609 * multi-character fold sequences. It's been wrong in Perl for a very long
2610 * time. There are three code points in Unicode whose multi-character folds
2611 * were long ago discovered to mess things up. The previous designs for
2612 * dealing with these involved assigning a special node for them. This
2613 * approach doesn't work, as evidenced by this example:
2614 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2615 * Both these fold to "sss", but if the pattern is parsed to create a node that
2616 * would match just the \xDF, it won't be able to handle the case where a
2617 * successful match would have to cross the node's boundary. The new approach
2618 * that hopefully generally solves the problem generates an EXACTFU_SS node
2621 * It turns out that there are problems with all multi-character folds, and not
2622 * just these three. Now the code is general, for all such cases, but the
2623 * three still have some special handling. The approach taken is:
2624 * 1) This routine examines each EXACTFish node that could contain multi-
2625 * character fold sequences. It returns in *min_subtract how much to
2626 * subtract from the the actual length of the string to get a real minimum
2627 * match length; it is 0 if there are no multi-char folds. This delta is
2628 * used by the caller to adjust the min length of the match, and the delta
2629 * between min and max, so that the optimizer doesn't reject these
2630 * possibilities based on size constraints.
2631 * 2) Certain of these sequences require special handling by the trie code,
2632 * so, if found, this code changes the joined node type to special ops:
2633 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2634 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2635 * is used for an EXACTFU node that contains at least one "ss" sequence in
2636 * it. For non-UTF-8 patterns and strings, this is the only case where
2637 * there is a possible fold length change. That means that a regular
2638 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2639 * with length changes, and so can be processed faster. regexec.c takes
2640 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2641 * pre-folded by regcomp.c. This saves effort in regex matching.
2642 * However, the pre-folding isn't done for non-UTF8 patterns because the
2643 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2644 * down by forcing the pattern into UTF8 unless necessary. Also what
2645 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2646 * possibilities for the non-UTF8 patterns are quite simple, except for
2647 * the sharp s. All the ones that don't involve a UTF-8 target string are
2648 * members of a fold-pair, and arrays are set up for all of them so that
2649 * the other member of the pair can be found quickly. Code elsewhere in
2650 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2651 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2652 * described in the next item.
2653 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2654 * 'ss' or not is not knowable at compile time. It will match iff the
2655 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2656 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2657 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
2658 * described in item 3). An assumption that the optimizer part of
2659 * regexec.c (probably unwittingly) makes is that a character in the
2660 * pattern corresponds to at most a single character in the target string.
2661 * (And I do mean character, and not byte here, unlike other parts of the
2662 * documentation that have never been updated to account for multibyte
2663 * Unicode.) This assumption is wrong only in this case, as all other
2664 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2665 * virtue of having this file pre-fold UTF-8 patterns. I'm
2666 * reluctant to try to change this assumption, so instead the code punts.
2667 * This routine examines EXACTF nodes for the sharp s, and returns a
2668 * boolean indicating whether or not the node is an EXACTF node that
2669 * contains a sharp s. When it is true, the caller sets a flag that later
2670 * causes the optimizer in this file to not set values for the floating
2671 * and fixed string lengths, and thus avoids the optimizer code in
2672 * regexec.c that makes the invalid assumption. Thus, there is no
2673 * optimization based on string lengths for EXACTF nodes that contain the
2674 * sharp s. This only happens for /id rules (which means the pattern
2678 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2679 if (PL_regkind[OP(scan)] == EXACT) \
2680 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2683 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) {
2684 /* Merge several consecutive EXACTish nodes into one. */
2685 regnode *n = regnext(scan);
2687 regnode *next = scan + NODE_SZ_STR(scan);
2691 regnode *stop = scan;
2692 GET_RE_DEBUG_FLAGS_DECL;
2694 PERL_UNUSED_ARG(depth);
2697 PERL_ARGS_ASSERT_JOIN_EXACT;
2698 #ifndef EXPERIMENTAL_INPLACESCAN
2699 PERL_UNUSED_ARG(flags);
2700 PERL_UNUSED_ARG(val);
2702 DEBUG_PEEP("join",scan,depth);
2704 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2705 * EXACT ones that are mergeable to the current one. */
2707 && (PL_regkind[OP(n)] == NOTHING
2708 || (stringok && OP(n) == OP(scan)))
2710 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2713 if (OP(n) == TAIL || n > next)
2715 if (PL_regkind[OP(n)] == NOTHING) {
2716 DEBUG_PEEP("skip:",n,depth);
2717 NEXT_OFF(scan) += NEXT_OFF(n);
2718 next = n + NODE_STEP_REGNODE;
2725 else if (stringok) {
2726 const unsigned int oldl = STR_LEN(scan);
2727 regnode * const nnext = regnext(n);
2729 /* XXX I (khw) kind of doubt that this works on platforms where
2730 * U8_MAX is above 255 because of lots of other assumptions */
2731 if (oldl + STR_LEN(n) > U8_MAX)
2734 DEBUG_PEEP("merg",n,depth);
2737 NEXT_OFF(scan) += NEXT_OFF(n);
2738 STR_LEN(scan) += STR_LEN(n);
2739 next = n + NODE_SZ_STR(n);
2740 /* Now we can overwrite *n : */
2741 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2749 #ifdef EXPERIMENTAL_INPLACESCAN
2750 if (flags && !NEXT_OFF(n)) {
2751 DEBUG_PEEP("atch", val, depth);
2752 if (reg_off_by_arg[OP(n)]) {
2753 ARG_SET(n, val - n);
2756 NEXT_OFF(n) = val - n;
2764 *has_exactf_sharp_s = FALSE;
2766 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2767 * can now analyze for sequences of problematic code points. (Prior to
2768 * this final joining, sequences could have been split over boundaries, and
2769 * hence missed). The sequences only happen in folding, hence for any
2770 * non-EXACT EXACTish node */
2771 if (OP(scan) != EXACT) {
2772 const U8 * const s0 = (U8*) STRING(scan);
2774 const U8 * const s_end = s0 + STR_LEN(scan);
2776 /* One pass is made over the node's string looking for all the
2777 * possibilities. to avoid some tests in the loop, there are two main
2778 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2782 /* Examine the string for a multi-character fold sequence. UTF-8
2783 * patterns have all characters pre-folded by the time this code is
2785 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2786 length sequence we are looking for is 2 */
2789 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2790 if (! len) { /* Not a multi-char fold: get next char */
2795 /* Nodes with 'ss' require special handling, except for EXACTFL
2796 * and EXACTFA for which there is no multi-char fold to this */
2797 if (len == 2 && *s == 's' && *(s+1) == 's'
2798 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2801 OP(scan) = EXACTFU_SS;
2804 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2805 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2806 COMBINING_DIAERESIS_UTF8
2807 COMBINING_ACUTE_ACCENT_UTF8,
2809 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2810 COMBINING_DIAERESIS_UTF8
2811 COMBINING_ACUTE_ACCENT_UTF8,
2816 /* These two folds require special handling by trie's, so
2817 * change the node type to indicate this. If EXACTFA and
2818 * EXACTFL were ever to be handled by trie's, this would
2819 * have to be changed. If this node has already been
2820 * changed to EXACTFU_SS in this loop, leave it as is. (I
2821 * (khw) think it doesn't matter in regexec.c for UTF
2822 * patterns, but no need to change it */
2823 if (OP(scan) == EXACTFU) {
2824 OP(scan) = EXACTFU_TRICKYFOLD;
2828 else { /* Here is a generic multi-char fold. */
2829 const U8* multi_end = s + len;
2831 /* Count how many characters in it. In the case of /l and
2832 * /aa, no folds which contain ASCII code points are
2833 * allowed, so check for those, and skip if found. (In
2834 * EXACTFL, no folds are allowed to any Latin1 code point,
2835 * not just ASCII. But there aren't any of these
2836 * currently, nor ever likely, so don't take the time to
2837 * test for them. The code that generates the
2838 * is_MULTI_foo() macros croaks should one actually get put
2839 * into Unicode .) */
2840 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2841 count = utf8_length(s, multi_end);
2845 while (s < multi_end) {
2848 goto next_iteration;
2858 /* The delta is how long the sequence is minus 1 (1 is how long
2859 * the character that folds to the sequence is) */
2860 *min_subtract += count - 1;
2864 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2866 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2867 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2868 * nodes can't have multi-char folds to this range (and there are
2869 * no existing ones in the upper latin1 range). In the EXACTF
2870 * case we look also for the sharp s, which can be in the final
2871 * position. Otherwise we can stop looking 1 byte earlier because
2872 * have to find at least two characters for a multi-fold */
2873 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2875 /* The below is perhaps overboard, but this allows us to save a
2876 * test each time through the loop at the expense of a mask. This
2877 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2878 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2879 * are 64. This uses an exclusive 'or' to find that bit and then
2880 * inverts it to form a mask, with just a single 0, in the bit
2881 * position where 'S' and 's' differ. */
2882 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2883 const U8 s_masked = 's' & S_or_s_mask;
2886 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2887 if (! len) { /* Not a multi-char fold. */
2888 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2890 *has_exactf_sharp_s = TRUE;
2897 && ((*s & S_or_s_mask) == s_masked)
2898 && ((*(s+1) & S_or_s_mask) == s_masked))
2901 /* EXACTF nodes need to know that the minimum length
2902 * changed so that a sharp s in the string can match this
2903 * ss in the pattern, but they remain EXACTF nodes, as they
2904 * won't match this unless the target string is is UTF-8,
2905 * which we don't know until runtime */
2906 if (OP(scan) != EXACTF) {
2907 OP(scan) = EXACTFU_SS;
2911 *min_subtract += len - 1;
2918 /* Allow dumping but overwriting the collection of skipped
2919 * ops and/or strings with fake optimized ops */
2920 n = scan + NODE_SZ_STR(scan);
2928 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2932 /* REx optimizer. Converts nodes into quicker variants "in place".
2933 Finds fixed substrings. */
2935 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2936 to the position after last scanned or to NULL. */
2938 #define INIT_AND_WITHP \
2939 assert(!and_withp); \
2940 Newx(and_withp,1,struct regnode_charclass_class); \
2941 SAVEFREEPV(and_withp)
2943 /* this is a chain of data about sub patterns we are processing that
2944 need to be handled separately/specially in study_chunk. Its so
2945 we can simulate recursion without losing state. */
2947 typedef struct scan_frame {
2948 regnode *last; /* last node to process in this frame */
2949 regnode *next; /* next node to process when last is reached */
2950 struct scan_frame *prev; /*previous frame*/
2951 I32 stop; /* what stopparen do we use */
2955 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2957 #define CASE_SYNST_FNC(nAmE) \
2959 if (flags & SCF_DO_STCLASS_AND) { \
2960 for (value = 0; value < 256; value++) \
2961 if (!is_ ## nAmE ## _cp(value)) \
2962 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2965 for (value = 0; value < 256; value++) \
2966 if (is_ ## nAmE ## _cp(value)) \
2967 ANYOF_BITMAP_SET(data->start_class, value); \
2971 if (flags & SCF_DO_STCLASS_AND) { \
2972 for (value = 0; value < 256; value++) \
2973 if (is_ ## nAmE ## _cp(value)) \
2974 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2977 for (value = 0; value < 256; value++) \
2978 if (!is_ ## nAmE ## _cp(value)) \
2979 ANYOF_BITMAP_SET(data->start_class, value); \
2986 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2987 I32 *minlenp, I32 *deltap,
2992 struct regnode_charclass_class *and_withp,
2993 U32 flags, U32 depth)
2994 /* scanp: Start here (read-write). */
2995 /* deltap: Write maxlen-minlen here. */
2996 /* last: Stop before this one. */
2997 /* data: string data about the pattern */
2998 /* stopparen: treat close N as END */
2999 /* recursed: which subroutines have we recursed into */
3000 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3003 I32 min = 0; /* There must be at least this number of characters to match */
3005 regnode *scan = *scanp, *next;
3007 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3008 int is_inf_internal = 0; /* The studied chunk is infinite */
3009 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3010 scan_data_t data_fake;
3011 SV *re_trie_maxbuff = NULL;
3012 regnode *first_non_open = scan;
3013 I32 stopmin = I32_MAX;
3014 scan_frame *frame = NULL;
3015 GET_RE_DEBUG_FLAGS_DECL;
3017 PERL_ARGS_ASSERT_STUDY_CHUNK;
3020 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3024 while (first_non_open && OP(first_non_open) == OPEN)
3025 first_non_open=regnext(first_non_open);
3030 while ( scan && OP(scan) != END && scan < last ){
3031 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3032 node length to get a real minimum (because
3033 the folded version may be shorter) */
3034 bool has_exactf_sharp_s = FALSE;
3035 /* Peephole optimizer: */
3036 DEBUG_STUDYDATA("Peep:", data,depth);
3037 DEBUG_PEEP("Peep",scan,depth);
3039 /* Its not clear to khw or hv why this is done here, and not in the
3040 * clauses that deal with EXACT nodes. khw's guess is that it's
3041 * because of a previous design */
3042 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3044 /* Follow the next-chain of the current node and optimize
3045 away all the NOTHINGs from it. */
3046 if (OP(scan) != CURLYX) {
3047 const int max = (reg_off_by_arg[OP(scan)]
3049 /* I32 may be smaller than U16 on CRAYs! */
3050 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3051 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3055 /* Skip NOTHING and LONGJMP. */
3056 while ((n = regnext(n))
3057 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3058 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3059 && off + noff < max)
3061 if (reg_off_by_arg[OP(scan)])
3064 NEXT_OFF(scan) = off;
3069 /* The principal pseudo-switch. Cannot be a switch, since we
3070 look into several different things. */
3071 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3072 || OP(scan) == IFTHEN) {
3073 next = regnext(scan);
3075 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3077 if (OP(next) == code || code == IFTHEN) {
3078 /* NOTE - There is similar code to this block below for handling
3079 TRIE nodes on a re-study. If you change stuff here check there
3081 I32 max1 = 0, min1 = I32_MAX, num = 0;
3082 struct regnode_charclass_class accum;
3083 regnode * const startbranch=scan;
3085 if (flags & SCF_DO_SUBSTR)
3086 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3087 if (flags & SCF_DO_STCLASS)
3088 cl_init_zero(pRExC_state, &accum);
3090 while (OP(scan) == code) {
3091 I32 deltanext, minnext, f = 0, fake;
3092 struct regnode_charclass_class this_class;
3095 data_fake.flags = 0;
3097 data_fake.whilem_c = data->whilem_c;
3098 data_fake.last_closep = data->last_closep;
3101 data_fake.last_closep = &fake;
3103 data_fake.pos_delta = delta;
3104 next = regnext(scan);
3105 scan = NEXTOPER(scan);
3107 scan = NEXTOPER(scan);
3108 if (flags & SCF_DO_STCLASS) {
3109 cl_init(pRExC_state, &this_class);
3110 data_fake.start_class = &this_class;
3111 f = SCF_DO_STCLASS_AND;
3113 if (flags & SCF_WHILEM_VISITED_POS)
3114 f |= SCF_WHILEM_VISITED_POS;
3116 /* we suppose the run is continuous, last=next...*/
3117 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3119 stopparen, recursed, NULL, f,depth+1);
3122 if (max1 < minnext + deltanext)
3123 max1 = minnext + deltanext;
3124 if (deltanext == I32_MAX)
3125 is_inf = is_inf_internal = 1;
3127 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3129 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3130 if ( stopmin > minnext)
3131 stopmin = min + min1;
3132 flags &= ~SCF_DO_SUBSTR;
3134 data->flags |= SCF_SEEN_ACCEPT;
3137 if (data_fake.flags & SF_HAS_EVAL)
3138 data->flags |= SF_HAS_EVAL;
3139 data->whilem_c = data_fake.whilem_c;
3141 if (flags & SCF_DO_STCLASS)
3142 cl_or(pRExC_state, &accum, &this_class);
3144 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3146 if (flags & SCF_DO_SUBSTR) {
3147 data->pos_min += min1;
3148 data->pos_delta += max1 - min1;
3149 if (max1 != min1 || is_inf)
3150 data->longest = &(data->longest_float);
3153 delta += max1 - min1;
3154 if (flags & SCF_DO_STCLASS_OR) {
3155 cl_or(pRExC_state, data->start_class, &accum);
3157 cl_and(data->start_class, and_withp);
3158 flags &= ~SCF_DO_STCLASS;
3161 else if (flags & SCF_DO_STCLASS_AND) {
3163 cl_and(data->start_class, &accum);
3164 flags &= ~SCF_DO_STCLASS;
3167 /* Switch to OR mode: cache the old value of
3168 * data->start_class */
3170 StructCopy(data->start_class, and_withp,
3171 struct regnode_charclass_class);
3172 flags &= ~SCF_DO_STCLASS_AND;
3173 StructCopy(&accum, data->start_class,
3174 struct regnode_charclass_class);
3175 flags |= SCF_DO_STCLASS_OR;
3176 data->start_class->flags |= ANYOF_EOS;
3180 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3183 Assuming this was/is a branch we are dealing with: 'scan' now
3184 points at the item that follows the branch sequence, whatever
3185 it is. We now start at the beginning of the sequence and look
3192 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3194 If we can find such a subsequence we need to turn the first
3195 element into a trie and then add the subsequent branch exact
3196 strings to the trie.
3200 1. patterns where the whole set of branches can be converted.
3202 2. patterns where only a subset can be converted.
3204 In case 1 we can replace the whole set with a single regop
3205 for the trie. In case 2 we need to keep the start and end
3208 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3209 becomes BRANCH TRIE; BRANCH X;
3211 There is an additional case, that being where there is a
3212 common prefix, which gets split out into an EXACT like node
3213 preceding the TRIE node.
3215 If x(1..n)==tail then we can do a simple trie, if not we make
3216 a "jump" trie, such that when we match the appropriate word
3217 we "jump" to the appropriate tail node. Essentially we turn
3218 a nested if into a case structure of sorts.
3223 if (!re_trie_maxbuff) {
3224 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3225 if (!SvIOK(re_trie_maxbuff))
3226 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3228 if ( SvIV(re_trie_maxbuff)>=0 ) {
3230 regnode *first = (regnode *)NULL;
3231 regnode *last = (regnode *)NULL;
3232 regnode *tail = scan;
3237 SV * const mysv = sv_newmortal(); /* for dumping */
3239 /* var tail is used because there may be a TAIL
3240 regop in the way. Ie, the exacts will point to the
3241 thing following the TAIL, but the last branch will
3242 point at the TAIL. So we advance tail. If we
3243 have nested (?:) we may have to move through several
3247 while ( OP( tail ) == TAIL ) {
3248 /* this is the TAIL generated by (?:) */
3249 tail = regnext( tail );
3253 DEBUG_TRIE_COMPILE_r({
3254 regprop(RExC_rx, mysv, tail );
3255 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3256 (int)depth * 2 + 2, "",
3257 "Looking for TRIE'able sequences. Tail node is: ",
3258 SvPV_nolen_const( mysv )
3264 Step through the branches
3265 cur represents each branch,
3266 noper is the first thing to be matched as part of that branch
3267 noper_next is the regnext() of that node.
3269 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3270 via a "jump trie" but we also support building with NOJUMPTRIE,
3271 which restricts the trie logic to structures like /FOO|BAR/.
3273 If noper is a trieable nodetype then the branch is a possible optimization
3274 target. If we are building under NOJUMPTRIE then we require that noper_next
3275 is the same as scan (our current position in the regex program).
3277 Once we have two or more consecutive such branches we can create a
3278 trie of the EXACT's contents and stitch it in place into the program.
3280 If the sequence represents all of the branches in the alternation we
3281 replace the entire thing with a single TRIE node.
3283 Otherwise when it is a subsequence we need to stitch it in place and
3284 replace only the relevant branches. This means the first branch has
3285 to remain as it is used by the alternation logic, and its next pointer,
3286 and needs to be repointed at the item on the branch chain following
3287 the last branch we have optimized away.
3289 This could be either a BRANCH, in which case the subsequence is internal,
3290 or it could be the item following the branch sequence in which case the
3291 subsequence is at the end (which does not necessarily mean the first node
3292 is the start of the alternation).
3294 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3297 ----------------+-----------
3301 EXACTFU_SS | EXACTFU
3302 EXACTFU_TRICKYFOLD | EXACTFU
3307 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3308 ( EXACT == (X) ) ? EXACT : \
3309 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3312 /* dont use tail as the end marker for this traverse */
3313 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3314 regnode * const noper = NEXTOPER( cur );
3315 U8 noper_type = OP( noper );
3316 U8 noper_trietype = TRIE_TYPE( noper_type );
3317 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3318 regnode * const noper_next = regnext( noper );
3319 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3320 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3323 DEBUG_TRIE_COMPILE_r({
3324 regprop(RExC_rx, mysv, cur);
3325 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3326 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3328 regprop(RExC_rx, mysv, noper);
3329 PerlIO_printf( Perl_debug_log, " -> %s",
3330 SvPV_nolen_const(mysv));
3333 regprop(RExC_rx, mysv, noper_next );
3334 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3335 SvPV_nolen_const(mysv));
3337 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3338 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3339 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3343 /* Is noper a trieable nodetype that can be merged with the
3344 * current trie (if there is one)? */
3348 ( noper_trietype == NOTHING)
3349 || ( trietype == NOTHING )
3350 || ( trietype == noper_trietype )
3353 && noper_next == tail
3357 /* Handle mergable triable node
3358 * Either we are the first node in a new trieable sequence,
3359 * in which case we do some bookkeeping, otherwise we update
3360 * the end pointer. */
3363 if ( noper_trietype == NOTHING ) {
3364 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3365 regnode * const noper_next = regnext( noper );
3366 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3367 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3370 if ( noper_next_trietype ) {
3371 trietype = noper_next_trietype;
3372 } else if (noper_next_type) {
3373 /* a NOTHING regop is 1 regop wide. We need at least two
3374 * for a trie so we can't merge this in */
3378 trietype = noper_trietype;
3381 if ( trietype == NOTHING )
3382 trietype = noper_trietype;
3387 } /* end handle mergable triable node */
3389 /* handle unmergable node -
3390 * noper may either be a triable node which can not be tried
3391 * together with the current trie, or a non triable node */
3393 /* If last is set and trietype is not NOTHING then we have found
3394 * at least two triable branch sequences in a row of a similar
3395 * trietype so we can turn them into a trie. If/when we
3396 * allow NOTHING to start a trie sequence this condition will be
3397 * required, and it isn't expensive so we leave it in for now. */
3398 if ( trietype && trietype != NOTHING )
3399 make_trie( pRExC_state,
3400 startbranch, first, cur, tail, count,
3401 trietype, depth+1 );
3402 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3406 && noper_next == tail
3409 /* noper is triable, so we can start a new trie sequence */
3412 trietype = noper_trietype;
3414 /* if we already saw a first but the current node is not triable then we have
3415 * to reset the first information. */
3420 } /* end handle unmergable node */
3421 } /* loop over branches */
3422 DEBUG_TRIE_COMPILE_r({
3423 regprop(RExC_rx, mysv, cur);
3424 PerlIO_printf( Perl_debug_log,
3425 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3426 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3429 if ( last && trietype ) {
3430 if ( trietype != NOTHING ) {
3431 /* the last branch of the sequence was part of a trie,
3432 * so we have to construct it here outside of the loop
3434 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3435 #ifdef TRIE_STUDY_OPT
3436 if ( ((made == MADE_EXACT_TRIE &&
3437 startbranch == first)
3438 || ( first_non_open == first )) &&
3440 flags |= SCF_TRIE_RESTUDY;
3441 if ( startbranch == first
3444 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3449 /* at this point we know whatever we have is a NOTHING sequence/branch
3450 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3452 if ( startbranch == first ) {
3454 /* the entire thing is a NOTHING sequence, something like this:
3455 * (?:|) So we can turn it into a plain NOTHING op. */
3456 DEBUG_TRIE_COMPILE_r({
3457 regprop(RExC_rx, mysv, cur);
3458 PerlIO_printf( Perl_debug_log,
3459 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3460 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3463 OP(startbranch)= NOTHING;
3464 NEXT_OFF(startbranch)= tail - startbranch;
3465 for ( opt= startbranch + 1; opt < tail ; opt++ )
3469 } /* end if ( last) */
3470 } /* TRIE_MAXBUF is non zero */
3475 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3476 scan = NEXTOPER(NEXTOPER(scan));
3477 } else /* single branch is optimized. */
3478 scan = NEXTOPER(scan);
3480 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3481 scan_frame *newframe = NULL;
3486 if (OP(scan) != SUSPEND) {
3487 /* set the pointer */
3488 if (OP(scan) == GOSUB) {
3490 RExC_recurse[ARG2L(scan)] = scan;
3491 start = RExC_open_parens[paren-1];
3492 end = RExC_close_parens[paren-1];
3495 start = RExC_rxi->program + 1;
3499 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3500 SAVEFREEPV(recursed);
3502 if (!PAREN_TEST(recursed,paren+1)) {
3503 PAREN_SET(recursed,paren+1);
3504 Newx(newframe,1,scan_frame);
3506 if (flags & SCF_DO_SUBSTR) {
3507 SCAN_COMMIT(pRExC_state,data,minlenp);
3508 data->longest = &(data->longest_float);
3510 is_inf = is_inf_internal = 1;
3511 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3512 cl_anything(pRExC_state, data->start_class);
3513 flags &= ~SCF_DO_STCLASS;
3516 Newx(newframe,1,scan_frame);
3519 end = regnext(scan);
3524 SAVEFREEPV(newframe);
3525 newframe->next = regnext(scan);
3526 newframe->last = last;
3527 newframe->stop = stopparen;
3528 newframe->prev = frame;
3538 else if (OP(scan) == EXACT) {
3539 I32 l = STR_LEN(scan);
3542 const U8 * const s = (U8*)STRING(scan);
3543 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3544 l = utf8_length(s, s + l);
3546 uc = *((U8*)STRING(scan));
3549 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3550 /* The code below prefers earlier match for fixed
3551 offset, later match for variable offset. */
3552 if (data->last_end == -1) { /* Update the start info. */
3553 data->last_start_min = data->pos_min;
3554 data->last_start_max = is_inf
3555 ? I32_MAX : data->pos_min + data->pos_delta;
3557 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3559 SvUTF8_on(data->last_found);
3561 SV * const sv = data->last_found;
3562 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3563 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3564 if (mg && mg->mg_len >= 0)
3565 mg->mg_len += utf8_length((U8*)STRING(scan),
3566 (U8*)STRING(scan)+STR_LEN(scan));
3568 data->last_end = data->pos_min + l;
3569 data->pos_min += l; /* As in the first entry. */
3570 data->flags &= ~SF_BEFORE_EOL;
3572 if (flags & SCF_DO_STCLASS_AND) {
3573 /* Check whether it is compatible with what we know already! */
3577 /* If compatible, we or it in below. It is compatible if is
3578 * in the bitmp and either 1) its bit or its fold is set, or 2)
3579 * it's for a locale. Even if there isn't unicode semantics
3580 * here, at runtime there may be because of matching against a
3581 * utf8 string, so accept a possible false positive for
3582 * latin1-range folds */
3584 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3585 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3586 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3587 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3592 ANYOF_CLASS_ZERO(data->start_class);
3593 ANYOF_BITMAP_ZERO(data->start_class);
3595 ANYOF_BITMAP_SET(data->start_class, uc);
3596 else if (uc >= 0x100) {
3599 /* Some Unicode code points fold to the Latin1 range; as
3600 * XXX temporary code, instead of figuring out if this is
3601 * one, just assume it is and set all the start class bits
3602 * that could be some such above 255 code point's fold
3603 * which will generate fals positives. As the code
3604 * elsewhere that does compute the fold settles down, it
3605 * can be extracted out and re-used here */
3606 for (i = 0; i < 256; i++){
3607 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3608 ANYOF_BITMAP_SET(data->start_class, i);
3612 data->start_class->flags &= ~ANYOF_EOS;
3614 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3616 else if (flags & SCF_DO_STCLASS_OR) {
3617 /* false positive possible if the class is case-folded */
3619 ANYOF_BITMAP_SET(data->start_class, uc);
3621 data->start_class->flags |= ANYOF_UNICODE_ALL;
3622 data->start_class->flags &= ~ANYOF_EOS;
3623 cl_and(data->start_class, and_withp);
3625 flags &= ~SCF_DO_STCLASS;
3627 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3628 I32 l = STR_LEN(scan);
3629 UV uc = *((U8*)STRING(scan));
3631 /* Search for fixed substrings supports EXACT only. */
3632 if (flags & SCF_DO_SUBSTR) {
3634 SCAN_COMMIT(pRExC_state, data, minlenp);
3637 const U8 * const s = (U8 *)STRING(scan);
3638 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3639 l = utf8_length(s, s + l);
3641 if (has_exactf_sharp_s) {
3642 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3644 min += l - min_subtract;
3646 delta += min_subtract;
3647 if (flags & SCF_DO_SUBSTR) {
3648 data->pos_min += l - min_subtract;
3649 if (data->pos_min < 0) {
3652 data->pos_delta += min_subtract;
3654 data->longest = &(data->longest_float);
3657 if (flags & SCF_DO_STCLASS_AND) {
3658 /* Check whether it is compatible with what we know already! */
3661 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3662 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3663 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3667 ANYOF_CLASS_ZERO(data->start_class);
3668 ANYOF_BITMAP_ZERO(data->start_class);
3670 ANYOF_BITMAP_SET(data->start_class, uc);
3671 data->start_class->flags &= ~ANYOF_EOS;
3672 if (OP(scan) == EXACTFL) {
3673 /* XXX This set is probably no longer necessary, and
3674 * probably wrong as LOCALE now is on in the initial
3676 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3680 /* Also set the other member of the fold pair. In case
3681 * that unicode semantics is called for at runtime, use
3682 * the full latin1 fold. (Can't do this for locale,
3683 * because not known until runtime) */
3684 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3686 /* All other (EXACTFL handled above) folds except under
3687 * /iaa that include s, S, and sharp_s also may include
3689 if (OP(scan) != EXACTFA) {
3690 if (uc == 's' || uc == 'S') {
3691 ANYOF_BITMAP_SET(data->start_class,
3692 LATIN_SMALL_LETTER_SHARP_S);
3694 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3695 ANYOF_BITMAP_SET(data->start_class, 's');
3696 ANYOF_BITMAP_SET(data->start_class, 'S');
3701 else if (uc >= 0x100) {
3703 for (i = 0; i < 256; i++){
3704 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3705 ANYOF_BITMAP_SET(data->start_class, i);
3710 else if (flags & SCF_DO_STCLASS_OR) {
3711 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3712 /* false positive possible if the class is case-folded.
3713 Assume that the locale settings are the same... */
3715 ANYOF_BITMAP_SET(data->start_class, uc);
3716 if (OP(scan) != EXACTFL) {
3718 /* And set the other member of the fold pair, but
3719 * can't do that in locale because not known until
3721 ANYOF_BITMAP_SET(data->start_class,
3722 PL_fold_latin1[uc]);
3724 /* All folds except under /iaa that include s, S,
3725 * and sharp_s also may include the others */
3726 if (OP(scan) != EXACTFA) {
3727 if (uc == 's' || uc == 'S') {
3728 ANYOF_BITMAP_SET(data->start_class,
3729 LATIN_SMALL_LETTER_SHARP_S);
3731 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3732 ANYOF_BITMAP_SET(data->start_class, 's');
3733 ANYOF_BITMAP_SET(data->start_class, 'S');
3738 data->start_class->flags &= ~ANYOF_EOS;
3740 cl_and(data->start_class, and_withp);
3742 flags &= ~SCF_DO_STCLASS;
3744 else if (REGNODE_VARIES(OP(scan))) {
3745 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3746 I32 f = flags, pos_before = 0;
3747 regnode * const oscan = scan;
3748 struct regnode_charclass_class this_class;
3749 struct regnode_charclass_class *oclass = NULL;
3750 I32 next_is_eval = 0;
3752 switch (PL_regkind[OP(scan)]) {
3753 case WHILEM: /* End of (?:...)* . */
3754 scan = NEXTOPER(scan);
3757 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3758 next = NEXTOPER(scan);
3759 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3761 maxcount = REG_INFTY;
3762 next = regnext(scan);
3763 scan = NEXTOPER(scan);
3767 if (flags & SCF_DO_SUBSTR)
3772 if (flags & SCF_DO_STCLASS) {
3774 maxcount = REG_INFTY;
3775 next = regnext(scan);
3776 scan = NEXTOPER(scan);
3779 is_inf = is_inf_internal = 1;
3780 scan = regnext(scan);
3781 if (flags & SCF_DO_SUBSTR) {
3782 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3783 data->longest = &(data->longest_float);
3785 goto optimize_curly_tail;
3787 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3788 && (scan->flags == stopparen))
3793 mincount = ARG1(scan);
3794 maxcount = ARG2(scan);
3796 next = regnext(scan);
3797 if (OP(scan) == CURLYX) {
3798 I32 lp = (data ? *(data->last_closep) : 0);
3799 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3801 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3802 next_is_eval = (OP(scan) == EVAL);
3804 if (flags & SCF_DO_SUBSTR) {
3805 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3806 pos_before = data->pos_min;
3810 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3812 data->flags |= SF_IS_INF;
3814 if (flags & SCF_DO_STCLASS) {
3815 cl_init(pRExC_state, &this_class);
3816 oclass = data->start_class;
3817 data->start_class = &this_class;
3818 f |= SCF_DO_STCLASS_AND;
3819 f &= ~SCF_DO_STCLASS_OR;
3821 /* Exclude from super-linear cache processing any {n,m}
3822 regops for which the combination of input pos and regex
3823 pos is not enough information to determine if a match
3826 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3827 regex pos at the \s*, the prospects for a match depend not
3828 only on the input position but also on how many (bar\s*)
3829 repeats into the {4,8} we are. */
https://perl5.git.perl.org / perl5.git / blob