5 * "A fair jaw-cracker dwarf-language must be." --Samwise Gamgee
8 /* This file contains functions for compiling a regular expression. See
9 * also regexec.c which funnily enough, contains functions for executing
10 * a regular expression.
12 * This file is also copied at build time to ext/re/re_comp.c, where
13 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
14 * This causes the main functions to be compiled under new names and with
15 * debugging support added, which makes "use re 'debug'" work.
18 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
19 * confused with the original package (see point 3 below). Thanks, Henry!
22 /* Additional note: this code is very heavily munged from Henry's version
23 * in places. In some spots I've traded clarity for efficiency, so don't
24 * blame Henry for some of the lack of readability.
27 /* The names of the functions have been changed from regcomp and
28 * regexec to pregcomp and pregexec in order to avoid conflicts
29 * with the POSIX routines of the same names.
32 #ifdef PERL_EXT_RE_BUILD
37 * pregcomp and pregexec -- regsub and regerror are not used in perl
39 * Copyright (c) 1986 by University of Toronto.
40 * Written by Henry Spencer. Not derived from licensed software.
42 * Permission is granted to anyone to use this software for any
43 * purpose on any computer system, and to redistribute it freely,
44 * subject to the following restrictions:
46 * 1. The author is not responsible for the consequences of use of
47 * this software, no matter how awful, even if they arise
50 * 2. The origin of this software must not be misrepresented, either
51 * by explicit claim or by omission.
53 * 3. Altered versions must be plainly marked as such, and must not
54 * be misrepresented as being the original software.
57 **** Alterations to Henry's code are...
59 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
60 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, by Larry Wall and others
62 **** You may distribute under the terms of either the GNU General Public
63 **** License or the Artistic License, as specified in the README file.
66 * Beware that some of this code is subtly aware of the way operator
67 * precedence is structured in regular expressions. Serious changes in
68 * regular-expression syntax might require a total rethink.
71 #define PERL_IN_REGCOMP_C
74 #ifndef PERL_IN_XSUB_RE
79 #ifdef PERL_IN_XSUB_RE
90 # if defined(BUGGY_MSC6)
91 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
92 # pragma optimize("a",off)
93 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
94 # pragma optimize("w",on )
95 # endif /* BUGGY_MSC6 */
102 typedef struct RExC_state_t {
103 U32 flags; /* are we folding, multilining? */
104 char *precomp; /* uncompiled string. */
105 regexp *rx; /* perl core regexp structure */
106 regexp_internal *rxi; /* internal data for regexp object pprivate field */
107 char *start; /* Start of input for compile */
108 char *end; /* End of input for compile */
109 char *parse; /* Input-scan pointer. */
110 I32 whilem_seen; /* number of WHILEM in this expr */
111 regnode *emit_start; /* Start of emitted-code area */
112 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
113 I32 naughty; /* How bad is this pattern? */
114 I32 sawback; /* Did we see \1, ...? */
116 I32 size; /* Code size. */
117 I32 npar; /* Capture buffer count, (OPEN). */
118 I32 cpar; /* Capture buffer count, (CLOSE). */
119 I32 nestroot; /* root parens we are in - used by accept */
123 regnode **open_parens; /* pointers to open parens */
124 regnode **close_parens; /* pointers to close parens */
125 regnode *opend; /* END node in program */
127 HV *charnames; /* cache of named sequences */
128 HV *paren_names; /* Paren names */
129 regnode **recurse; /* Recurse regops */
130 I32 recurse_count; /* Number of recurse regops */
132 char *starttry; /* -Dr: where regtry was called. */
133 #define RExC_starttry (pRExC_state->starttry)
136 const char *lastparse;
138 #define RExC_lastparse (pRExC_state->lastparse)
139 #define RExC_lastnum (pRExC_state->lastnum)
143 #define RExC_flags (pRExC_state->flags)
144 #define RExC_precomp (pRExC_state->precomp)
145 #define RExC_rx (pRExC_state->rx)
146 #define RExC_rxi (pRExC_state->rxi)
147 #define RExC_start (pRExC_state->start)
148 #define RExC_end (pRExC_state->end)
149 #define RExC_parse (pRExC_state->parse)
150 #define RExC_whilem_seen (pRExC_state->whilem_seen)
151 #define RExC_offsets (pRExC_state->rxi->offsets) /* I am not like the others */
152 #define RExC_emit (pRExC_state->emit)
153 #define RExC_emit_start (pRExC_state->emit_start)
154 #define RExC_naughty (pRExC_state->naughty)
155 #define RExC_sawback (pRExC_state->sawback)
156 #define RExC_seen (pRExC_state->seen)
157 #define RExC_size (pRExC_state->size)
158 #define RExC_npar (pRExC_state->npar)
159 #define RExC_nestroot (pRExC_state->nestroot)
160 #define RExC_extralen (pRExC_state->extralen)
161 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
162 #define RExC_seen_evals (pRExC_state->seen_evals)
163 #define RExC_utf8 (pRExC_state->utf8)
164 #define RExC_charnames (pRExC_state->charnames)
165 #define RExC_open_parens (pRExC_state->open_parens)
166 #define RExC_close_parens (pRExC_state->close_parens)
167 #define RExC_opend (pRExC_state->opend)
168 #define RExC_paren_names (pRExC_state->paren_names)
169 #define RExC_recurse (pRExC_state->recurse)
170 #define RExC_recurse_count (pRExC_state->recurse_count)
172 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
173 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
174 ((*s) == '{' && regcurly(s)))
177 #undef SPSTART /* dratted cpp namespace... */
180 * Flags to be passed up and down.
182 #define WORST 0 /* Worst case. */
183 #define HASWIDTH 0x1 /* Known to match non-null strings. */
184 #define SIMPLE 0x2 /* Simple enough to be STAR/PLUS operand. */
185 #define SPSTART 0x4 /* Starts with * or +. */
186 #define TRYAGAIN 0x8 /* Weeded out a declaration. */
188 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
190 /* whether trie related optimizations are enabled */
191 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
192 #define TRIE_STUDY_OPT
193 #define FULL_TRIE_STUDY
199 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
200 #define PBITVAL(paren) (1 << ((paren) & 7))
201 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
202 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
203 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
206 /* About scan_data_t.
208 During optimisation we recurse through the regexp program performing
209 various inplace (keyhole style) optimisations. In addition study_chunk
210 and scan_commit populate this data structure with information about
211 what strings MUST appear in the pattern. We look for the longest
212 string that must appear for at a fixed location, and we look for the
213 longest string that may appear at a floating location. So for instance
218 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
219 strings (because they follow a .* construct). study_chunk will identify
220 both FOO and BAR as being the longest fixed and floating strings respectively.
222 The strings can be composites, for instance
226 will result in a composite fixed substring 'foo'.
228 For each string some basic information is maintained:
230 - offset or min_offset
231 This is the position the string must appear at, or not before.
232 It also implicitly (when combined with minlenp) tells us how many
233 character must match before the string we are searching.
234 Likewise when combined with minlenp and the length of the string
235 tells us how many characters must appear after the string we have
239 Only used for floating strings. This is the rightmost point that
240 the string can appear at. Ifset to I32 max it indicates that the
241 string can occur infinitely far to the right.
244 A pointer to the minimum length of the pattern that the string
245 was found inside. This is important as in the case of positive
246 lookahead or positive lookbehind we can have multiple patterns
251 The minimum length of the pattern overall is 3, the minimum length
252 of the lookahead part is 3, but the minimum length of the part that
253 will actually match is 1. So 'FOO's minimum length is 3, but the
254 minimum length for the F is 1. This is important as the minimum length
255 is used to determine offsets in front of and behind the string being
256 looked for. Since strings can be composites this is the length of the
257 pattern at the time it was commited with a scan_commit. Note that
258 the length is calculated by study_chunk, so that the minimum lengths
259 are not known until the full pattern has been compiled, thus the
260 pointer to the value.
264 In the case of lookbehind the string being searched for can be
265 offset past the start point of the final matching string.
266 If this value was just blithely removed from the min_offset it would
267 invalidate some of the calculations for how many chars must match
268 before or after (as they are derived from min_offset and minlen and
269 the length of the string being searched for).
270 When the final pattern is compiled and the data is moved from the
271 scan_data_t structure into the regexp structure the information
272 about lookbehind is factored in, with the information that would
273 have been lost precalculated in the end_shift field for the
276 The fields pos_min and pos_delta are used to store the minimum offset
277 and the delta to the maximum offset at the current point in the pattern.
281 typedef struct scan_data_t {
282 /*I32 len_min; unused */
283 /*I32 len_delta; unused */
287 I32 last_end; /* min value, <0 unless valid. */
290 SV **longest; /* Either &l_fixed, or &l_float. */
291 SV *longest_fixed; /* longest fixed string found in pattern */
292 I32 offset_fixed; /* offset where it starts */
293 I32 *minlen_fixed; /* pointer to the minlen relevent to the string */
294 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
295 SV *longest_float; /* longest floating string found in pattern */
296 I32 offset_float_min; /* earliest point in string it can appear */
297 I32 offset_float_max; /* latest point in string it can appear */
298 I32 *minlen_float; /* pointer to the minlen relevent to the string */
299 I32 lookbehind_float; /* is the position of the string modified by LB */
303 struct regnode_charclass_class *start_class;
307 * Forward declarations for pregcomp()'s friends.
310 static const scan_data_t zero_scan_data =
311 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
313 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
314 #define SF_BEFORE_SEOL 0x0001
315 #define SF_BEFORE_MEOL 0x0002
316 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
317 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
320 # define SF_FIX_SHIFT_EOL (0+2)
321 # define SF_FL_SHIFT_EOL (0+4)
323 # define SF_FIX_SHIFT_EOL (+2)
324 # define SF_FL_SHIFT_EOL (+4)
327 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
328 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
330 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
331 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
332 #define SF_IS_INF 0x0040
333 #define SF_HAS_PAR 0x0080
334 #define SF_IN_PAR 0x0100
335 #define SF_HAS_EVAL 0x0200
336 #define SCF_DO_SUBSTR 0x0400
337 #define SCF_DO_STCLASS_AND 0x0800
338 #define SCF_DO_STCLASS_OR 0x1000
339 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
340 #define SCF_WHILEM_VISITED_POS 0x2000
342 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
343 #define SCF_SEEN_ACCEPT 0x8000
345 #define UTF (RExC_utf8 != 0)
346 #define LOC ((RExC_flags & RXf_PMf_LOCALE) != 0)
347 #define FOLD ((RExC_flags & RXf_PMf_FOLD) != 0)
349 #define OOB_UNICODE 12345678
350 #define OOB_NAMEDCLASS -1
352 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
353 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
356 /* length of regex to show in messages that don't mark a position within */
357 #define RegexLengthToShowInErrorMessages 127
360 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
361 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
362 * op/pragma/warn/regcomp.
364 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
365 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
367 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
370 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
371 * arg. Show regex, up to a maximum length. If it's too long, chop and add
374 #define _FAIL(code) STMT_START { \
375 const char *ellipses = ""; \
376 IV len = RExC_end - RExC_precomp; \
379 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx); \
380 if (len > RegexLengthToShowInErrorMessages) { \
381 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
382 len = RegexLengthToShowInErrorMessages - 10; \
388 #define FAIL(msg) _FAIL( \
389 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
390 msg, (int)len, RExC_precomp, ellipses))
392 #define FAIL2(msg,arg) _FAIL( \
393 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
394 arg, (int)len, RExC_precomp, ellipses))
397 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
399 #define Simple_vFAIL(m) STMT_START { \
400 const IV offset = RExC_parse - RExC_precomp; \
401 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
402 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
406 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
408 #define vFAIL(m) STMT_START { \
410 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx); \
415 * Like Simple_vFAIL(), but accepts two arguments.
417 #define Simple_vFAIL2(m,a1) STMT_START { \
418 const IV offset = RExC_parse - RExC_precomp; \
419 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
420 (int)offset, RExC_precomp, RExC_precomp + offset); \
424 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
426 #define vFAIL2(m,a1) STMT_START { \
428 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx); \
429 Simple_vFAIL2(m, a1); \
434 * Like Simple_vFAIL(), but accepts three arguments.
436 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
437 const IV offset = RExC_parse - RExC_precomp; \
438 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
439 (int)offset, RExC_precomp, RExC_precomp + offset); \
443 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
445 #define vFAIL3(m,a1,a2) STMT_START { \
447 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx); \
448 Simple_vFAIL3(m, a1, a2); \
452 * Like Simple_vFAIL(), but accepts four arguments.
454 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
455 const IV offset = RExC_parse - RExC_precomp; \
456 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
457 (int)offset, RExC_precomp, RExC_precomp + offset); \
460 #define vWARN(loc,m) STMT_START { \
461 const IV offset = loc - RExC_precomp; \
462 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
463 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
466 #define vWARNdep(loc,m) STMT_START { \
467 const IV offset = loc - RExC_precomp; \
468 Perl_warner(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
469 "%s" REPORT_LOCATION, \
470 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
474 #define vWARN2(loc, m, a1) STMT_START { \
475 const IV offset = loc - RExC_precomp; \
476 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
477 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
480 #define vWARN3(loc, m, a1, a2) STMT_START { \
481 const IV offset = loc - RExC_precomp; \
482 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
483 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
486 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
487 const IV offset = loc - RExC_precomp; \
488 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
489 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
492 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
493 const IV offset = loc - RExC_precomp; \
494 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
495 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
499 /* Allow for side effects in s */
500 #define REGC(c,s) STMT_START { \
501 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
504 /* Macros for recording node offsets. 20001227 mjd@plover.com
505 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
506 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
507 * Element 0 holds the number n.
508 * Position is 1 indexed.
511 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
513 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
514 __LINE__, (int)(node), (int)(byte))); \
516 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
518 RExC_offsets[2*(node)-1] = (byte); \
523 #define Set_Node_Offset(node,byte) \
524 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
525 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
527 #define Set_Node_Length_To_R(node,len) STMT_START { \
529 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
530 __LINE__, (int)(node), (int)(len))); \
532 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
534 RExC_offsets[2*(node)] = (len); \
539 #define Set_Node_Length(node,len) \
540 Set_Node_Length_To_R((node)-RExC_emit_start, len)
541 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
542 #define Set_Node_Cur_Length(node) \
543 Set_Node_Length(node, RExC_parse - parse_start)
545 /* Get offsets and lengths */
546 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
547 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
549 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
550 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
551 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
555 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
556 #define EXPERIMENTAL_INPLACESCAN
559 #define DEBUG_STUDYDATA(str,data,depth) \
560 DEBUG_OPTIMISE_MORE_r(if(data){ \
561 PerlIO_printf(Perl_debug_log, \
562 "%*s" str "Pos:%"IVdf"/%"IVdf \
563 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
564 (int)(depth)*2, "", \
565 (IV)((data)->pos_min), \
566 (IV)((data)->pos_delta), \
567 (UV)((data)->flags), \
568 (IV)((data)->whilem_c), \
569 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
570 is_inf ? "INF " : "" \
572 if ((data)->last_found) \
573 PerlIO_printf(Perl_debug_log, \
574 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
575 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
576 SvPVX_const((data)->last_found), \
577 (IV)((data)->last_end), \
578 (IV)((data)->last_start_min), \
579 (IV)((data)->last_start_max), \
580 ((data)->longest && \
581 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
582 SvPVX_const((data)->longest_fixed), \
583 (IV)((data)->offset_fixed), \
584 ((data)->longest && \
585 (data)->longest==&((data)->longest_float)) ? "*" : "", \
586 SvPVX_const((data)->longest_float), \
587 (IV)((data)->offset_float_min), \
588 (IV)((data)->offset_float_max) \
590 PerlIO_printf(Perl_debug_log,"\n"); \
593 static void clear_re(pTHX_ void *r);
595 /* Mark that we cannot extend a found fixed substring at this point.
596 Update the longest found anchored substring and the longest found
597 floating substrings if needed. */
600 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
602 const STRLEN l = CHR_SVLEN(data->last_found);
603 const STRLEN old_l = CHR_SVLEN(*data->longest);
604 GET_RE_DEBUG_FLAGS_DECL;
606 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
607 SvSetMagicSV(*data->longest, data->last_found);
608 if (*data->longest == data->longest_fixed) {
609 data->offset_fixed = l ? data->last_start_min : data->pos_min;
610 if (data->flags & SF_BEFORE_EOL)
612 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
614 data->flags &= ~SF_FIX_BEFORE_EOL;
615 data->minlen_fixed=minlenp;
616 data->lookbehind_fixed=0;
618 else { /* *data->longest == data->longest_float */
619 data->offset_float_min = l ? data->last_start_min : data->pos_min;
620 data->offset_float_max = (l
621 ? data->last_start_max
622 : data->pos_min + data->pos_delta);
623 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
624 data->offset_float_max = I32_MAX;
625 if (data->flags & SF_BEFORE_EOL)
627 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
629 data->flags &= ~SF_FL_BEFORE_EOL;
630 data->minlen_float=minlenp;
631 data->lookbehind_float=0;
634 SvCUR_set(data->last_found, 0);
636 SV * const sv = data->last_found;
637 if (SvUTF8(sv) && SvMAGICAL(sv)) {
638 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
644 data->flags &= ~SF_BEFORE_EOL;
645 DEBUG_STUDYDATA("cl_anything: ",data,0);
648 /* Can match anything (initialization) */
650 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
652 ANYOF_CLASS_ZERO(cl);
653 ANYOF_BITMAP_SETALL(cl);
654 cl->flags = ANYOF_EOS|ANYOF_UNICODE_ALL;
656 cl->flags |= ANYOF_LOCALE;
659 /* Can match anything (initialization) */
661 S_cl_is_anything(const struct regnode_charclass_class *cl)
665 for (value = 0; value <= ANYOF_MAX; value += 2)
666 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
668 if (!(cl->flags & ANYOF_UNICODE_ALL))
670 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
675 /* Can match anything (initialization) */
677 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
679 Zero(cl, 1, struct regnode_charclass_class);
681 cl_anything(pRExC_state, cl);
685 S_cl_init_zero(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
687 Zero(cl, 1, struct regnode_charclass_class);
689 cl_anything(pRExC_state, cl);
691 cl->flags |= ANYOF_LOCALE;
694 /* 'And' a given class with another one. Can create false positives */
695 /* We assume that cl is not inverted */
697 S_cl_and(struct regnode_charclass_class *cl,
698 const struct regnode_charclass_class *and_with)
701 assert(and_with->type == ANYOF);
702 if (!(and_with->flags & ANYOF_CLASS)
703 && !(cl->flags & ANYOF_CLASS)
704 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
705 && !(and_with->flags & ANYOF_FOLD)
706 && !(cl->flags & ANYOF_FOLD)) {
709 if (and_with->flags & ANYOF_INVERT)
710 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
711 cl->bitmap[i] &= ~and_with->bitmap[i];
713 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
714 cl->bitmap[i] &= and_with->bitmap[i];
715 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
716 if (!(and_with->flags & ANYOF_EOS))
717 cl->flags &= ~ANYOF_EOS;
719 if (cl->flags & ANYOF_UNICODE_ALL && and_with->flags & ANYOF_UNICODE &&
720 !(and_with->flags & ANYOF_INVERT)) {
721 cl->flags &= ~ANYOF_UNICODE_ALL;
722 cl->flags |= ANYOF_UNICODE;
723 ARG_SET(cl, ARG(and_with));
725 if (!(and_with->flags & ANYOF_UNICODE_ALL) &&
726 !(and_with->flags & ANYOF_INVERT))
727 cl->flags &= ~ANYOF_UNICODE_ALL;
728 if (!(and_with->flags & (ANYOF_UNICODE|ANYOF_UNICODE_ALL)) &&
729 !(and_with->flags & ANYOF_INVERT))
730 cl->flags &= ~ANYOF_UNICODE;
733 /* 'OR' a given class with another one. Can create false positives */
734 /* We assume that cl is not inverted */
736 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
738 if (or_with->flags & ANYOF_INVERT) {
740 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
741 * <= (B1 | !B2) | (CL1 | !CL2)
742 * which is wasteful if CL2 is small, but we ignore CL2:
743 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
744 * XXXX Can we handle case-fold? Unclear:
745 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
746 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
748 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
749 && !(or_with->flags & ANYOF_FOLD)
750 && !(cl->flags & ANYOF_FOLD) ) {
753 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
754 cl->bitmap[i] |= ~or_with->bitmap[i];
755 } /* XXXX: logic is complicated otherwise */
757 cl_anything(pRExC_state, cl);
760 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
761 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
762 && (!(or_with->flags & ANYOF_FOLD)
763 || (cl->flags & ANYOF_FOLD)) ) {
766 /* OR char bitmap and class bitmap separately */
767 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
768 cl->bitmap[i] |= or_with->bitmap[i];
769 if (or_with->flags & ANYOF_CLASS) {
770 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
771 cl->classflags[i] |= or_with->classflags[i];
772 cl->flags |= ANYOF_CLASS;
775 else { /* XXXX: logic is complicated, leave it along for a moment. */
776 cl_anything(pRExC_state, cl);
779 if (or_with->flags & ANYOF_EOS)
780 cl->flags |= ANYOF_EOS;
782 if (cl->flags & ANYOF_UNICODE && or_with->flags & ANYOF_UNICODE &&
783 ARG(cl) != ARG(or_with)) {
784 cl->flags |= ANYOF_UNICODE_ALL;
785 cl->flags &= ~ANYOF_UNICODE;
787 if (or_with->flags & ANYOF_UNICODE_ALL) {
788 cl->flags |= ANYOF_UNICODE_ALL;
789 cl->flags &= ~ANYOF_UNICODE;
793 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
794 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
795 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
796 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
801 dump_trie(trie,widecharmap,revcharmap)
802 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
803 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
805 These routines dump out a trie in a somewhat readable format.
806 The _interim_ variants are used for debugging the interim
807 tables that are used to generate the final compressed
808 representation which is what dump_trie expects.
810 Part of the reason for their existance is to provide a form
811 of documentation as to how the different representations function.
816 Dumps the final compressed table form of the trie to Perl_debug_log.
817 Used for debugging make_trie().
821 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
822 AV *revcharmap, U32 depth)
825 SV *sv=sv_newmortal();
826 int colwidth= widecharmap ? 6 : 4;
827 GET_RE_DEBUG_FLAGS_DECL;
830 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
831 (int)depth * 2 + 2,"",
832 "Match","Base","Ofs" );
834 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
835 SV ** const tmp = av_fetch( revcharmap, state, 0);
837 PerlIO_printf( Perl_debug_log, "%*s",
839 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
840 PL_colors[0], PL_colors[1],
841 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
842 PERL_PV_ESCAPE_FIRSTCHAR
847 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
848 (int)depth * 2 + 2,"");
850 for( state = 0 ; state < trie->uniquecharcount ; state++ )
851 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
852 PerlIO_printf( Perl_debug_log, "\n");
854 for( state = 1 ; state < trie->statecount ; state++ ) {
855 const U32 base = trie->states[ state ].trans.base;
857 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
859 if ( trie->states[ state ].wordnum ) {
860 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
862 PerlIO_printf( Perl_debug_log, "%6s", "" );
865 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
870 while( ( base + ofs < trie->uniquecharcount ) ||
871 ( base + ofs - trie->uniquecharcount < trie->lasttrans
872 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
875 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
877 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
878 if ( ( base + ofs >= trie->uniquecharcount ) &&
879 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
880 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
882 PerlIO_printf( Perl_debug_log, "%*"UVXf,
884 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
886 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
890 PerlIO_printf( Perl_debug_log, "]");
893 PerlIO_printf( Perl_debug_log, "\n" );
897 Dumps a fully constructed but uncompressed trie in list form.
898 List tries normally only are used for construction when the number of
899 possible chars (trie->uniquecharcount) is very high.
900 Used for debugging make_trie().
903 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
904 HV *widecharmap, AV *revcharmap, U32 next_alloc,
908 SV *sv=sv_newmortal();
909 int colwidth= widecharmap ? 6 : 4;
910 GET_RE_DEBUG_FLAGS_DECL;
911 /* print out the table precompression. */
912 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
913 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
914 "------:-----+-----------------\n" );
916 for( state=1 ; state < next_alloc ; state ++ ) {
919 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
920 (int)depth * 2 + 2,"", (UV)state );
921 if ( ! trie->states[ state ].wordnum ) {
922 PerlIO_printf( Perl_debug_log, "%5s| ","");
924 PerlIO_printf( Perl_debug_log, "W%4x| ",
925 trie->states[ state ].wordnum
928 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
929 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
931 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
933 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
934 PL_colors[0], PL_colors[1],
935 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
936 PERL_PV_ESCAPE_FIRSTCHAR
938 TRIE_LIST_ITEM(state,charid).forid,
939 (UV)TRIE_LIST_ITEM(state,charid).newstate
942 PerlIO_printf(Perl_debug_log, "\n%*s| ",
943 (int)((depth * 2) + 14), "");
946 PerlIO_printf( Perl_debug_log, "\n");
951 Dumps a fully constructed but uncompressed trie in table form.
952 This is the normal DFA style state transition table, with a few
953 twists to facilitate compression later.
954 Used for debugging make_trie().
957 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
958 HV *widecharmap, AV *revcharmap, U32 next_alloc,
963 SV *sv=sv_newmortal();
964 int colwidth= widecharmap ? 6 : 4;
965 GET_RE_DEBUG_FLAGS_DECL;
968 print out the table precompression so that we can do a visual check
969 that they are identical.
972 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
974 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
975 SV ** const tmp = av_fetch( revcharmap, charid, 0);
977 PerlIO_printf( Perl_debug_log, "%*s",
979 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
980 PL_colors[0], PL_colors[1],
981 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
982 PERL_PV_ESCAPE_FIRSTCHAR
988 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
990 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
991 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
994 PerlIO_printf( Perl_debug_log, "\n" );
996 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
998 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
999 (int)depth * 2 + 2,"",
1000 (UV)TRIE_NODENUM( state ) );
1002 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1003 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1005 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1007 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1009 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1010 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1012 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1013 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1020 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1021 startbranch: the first branch in the whole branch sequence
1022 first : start branch of sequence of branch-exact nodes.
1023 May be the same as startbranch
1024 last : Thing following the last branch.
1025 May be the same as tail.
1026 tail : item following the branch sequence
1027 count : words in the sequence
1028 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1029 depth : indent depth
1031 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1033 A trie is an N'ary tree where the branches are determined by digital
1034 decomposition of the key. IE, at the root node you look up the 1st character and
1035 follow that branch repeat until you find the end of the branches. Nodes can be
1036 marked as "accepting" meaning they represent a complete word. Eg:
1040 would convert into the following structure. Numbers represent states, letters
1041 following numbers represent valid transitions on the letter from that state, if
1042 the number is in square brackets it represents an accepting state, otherwise it
1043 will be in parenthesis.
1045 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1049 (1) +-i->(6)-+-s->[7]
1051 +-s->(3)-+-h->(4)-+-e->[5]
1053 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1055 This shows that when matching against the string 'hers' we will begin at state 1
1056 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1057 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1058 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1059 single traverse. We store a mapping from accepting to state to which word was
1060 matched, and then when we have multiple possibilities we try to complete the
1061 rest of the regex in the order in which they occured in the alternation.
1063 The only prior NFA like behaviour that would be changed by the TRIE support is
1064 the silent ignoring of duplicate alternations which are of the form:
1066 / (DUPE|DUPE) X? (?{ ... }) Y /x
1068 Thus EVAL blocks follwing a trie may be called a different number of times with
1069 and without the optimisation. With the optimisations dupes will be silently
1070 ignored. This inconsistant behaviour of EVAL type nodes is well established as
1071 the following demonstrates:
1073 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1075 which prints out 'word' three times, but
1077 'words'=~/(word|word|word)(?{ print $1 })S/
1079 which doesnt print it out at all. This is due to other optimisations kicking in.
1081 Example of what happens on a structural level:
1083 The regexp /(ac|ad|ab)+/ will produce the folowing debug output:
1085 1: CURLYM[1] {1,32767}(18)
1096 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1097 and should turn into:
1099 1: CURLYM[1] {1,32767}(18)
1101 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1109 Cases where tail != last would be like /(?foo|bar)baz/:
1119 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1120 and would end up looking like:
1123 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1130 d = uvuni_to_utf8_flags(d, uv, 0);
1132 is the recommended Unicode-aware way of saying
1137 #define TRIE_STORE_REVCHAR \
1139 SV *tmp = newSVpvs(""); \
1140 if (UTF) SvUTF8_on(tmp); \
1141 Perl_sv_catpvf( aTHX_ tmp, "%c", (int)uvc ); \
1142 av_push( revcharmap, tmp ); \
1145 #define TRIE_READ_CHAR STMT_START { \
1149 if ( foldlen > 0 ) { \
1150 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1155 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1156 uvc = to_uni_fold( uvc, foldbuf, &foldlen ); \
1157 foldlen -= UNISKIP( uvc ); \
1158 scan = foldbuf + UNISKIP( uvc ); \
1161 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1171 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1172 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1173 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1174 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1176 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1177 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1178 TRIE_LIST_CUR( state )++; \
1181 #define TRIE_LIST_NEW(state) STMT_START { \
1182 Newxz( trie->states[ state ].trans.list, \
1183 4, reg_trie_trans_le ); \
1184 TRIE_LIST_CUR( state ) = 1; \
1185 TRIE_LIST_LEN( state ) = 4; \
1188 #define TRIE_HANDLE_WORD(state) STMT_START { \
1189 U16 dupe= trie->states[ state ].wordnum; \
1190 regnode * const noper_next = regnext( noper ); \
1192 if (trie->wordlen) \
1193 trie->wordlen[ curword ] = wordlen; \
1195 /* store the word for dumping */ \
1197 if (OP(noper) != NOTHING) \
1198 tmp = newSVpvn(STRING(noper), STR_LEN(noper)); \
1200 tmp = newSVpvn( "", 0 ); \
1201 if ( UTF ) SvUTF8_on( tmp ); \
1202 av_push( trie_words, tmp ); \
1207 if ( noper_next < tail ) { \
1209 trie->jump = PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1210 trie->jump[curword] = (U16)(noper_next - convert); \
1212 jumper = noper_next; \
1214 nextbranch= regnext(cur); \
1218 /* So it's a dupe. This means we need to maintain a */\
1219 /* linked-list from the first to the next. */\
1220 /* we only allocate the nextword buffer when there */\
1221 /* a dupe, so first time we have to do the allocation */\
1222 if (!trie->nextword) \
1224 PerlMemShared_calloc( word_count + 1, sizeof(U16)); \
1225 while ( trie->nextword[dupe] ) \
1226 dupe= trie->nextword[dupe]; \
1227 trie->nextword[dupe]= curword; \
1229 /* we haven't inserted this word yet. */ \
1230 trie->states[ state ].wordnum = curword; \
1235 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1236 ( ( base + charid >= ucharcount \
1237 && base + charid < ubound \
1238 && state == trie->trans[ base - ucharcount + charid ].check \
1239 && trie->trans[ base - ucharcount + charid ].next ) \
1240 ? trie->trans[ base - ucharcount + charid ].next \
1241 : ( state==1 ? special : 0 ) \
1245 #define MADE_JUMP_TRIE 2
1246 #define MADE_EXACT_TRIE 4
1249 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1252 /* first pass, loop through and scan words */
1253 reg_trie_data *trie;
1254 HV *widecharmap = NULL;
1255 AV *revcharmap = newAV();
1257 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1262 regnode *jumper = NULL;
1263 regnode *nextbranch = NULL;
1264 regnode *convert = NULL;
1265 /* we just use folder as a flag in utf8 */
1266 const U8 * const folder = ( flags == EXACTF
1268 : ( flags == EXACTFL
1275 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1276 AV *trie_words = NULL;
1277 /* along with revcharmap, this only used during construction but both are
1278 * useful during debugging so we store them in the struct when debugging.
1281 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1282 STRLEN trie_charcount=0;
1284 SV *re_trie_maxbuff;
1285 GET_RE_DEBUG_FLAGS_DECL;
1287 PERL_UNUSED_ARG(depth);
1290 trie = PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1292 trie->startstate = 1;
1293 trie->wordcount = word_count;
1294 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1295 trie->charmap = PerlMemShared_calloc( 256, sizeof(U16) );
1296 if (!(UTF && folder))
1297 trie->bitmap = PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1299 trie_words = newAV();
1302 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1303 if (!SvIOK(re_trie_maxbuff)) {
1304 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1307 PerlIO_printf( Perl_debug_log,
1308 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1309 (int)depth * 2 + 2, "",
1310 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1311 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1315 /* Find the node we are going to overwrite */
1316 if ( first == startbranch && OP( last ) != BRANCH ) {
1317 /* whole branch chain */
1320 /* branch sub-chain */
1321 convert = NEXTOPER( first );
1324 /* -- First loop and Setup --
1326 We first traverse the branches and scan each word to determine if it
1327 contains widechars, and how many unique chars there are, this is
1328 important as we have to build a table with at least as many columns as we
1331 We use an array of integers to represent the character codes 0..255
1332 (trie->charmap) and we use a an HV* to store unicode characters. We use the
1333 native representation of the character value as the key and IV's for the
1336 *TODO* If we keep track of how many times each character is used we can
1337 remap the columns so that the table compression later on is more
1338 efficient in terms of memory by ensuring most common value is in the
1339 middle and the least common are on the outside. IMO this would be better
1340 than a most to least common mapping as theres a decent chance the most
1341 common letter will share a node with the least common, meaning the node
1342 will not be compressable. With a middle is most common approach the worst
1343 case is when we have the least common nodes twice.
1347 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1348 regnode * const noper = NEXTOPER( cur );
1349 const U8 *uc = (U8*)STRING( noper );
1350 const U8 * const e = uc + STR_LEN( noper );
1352 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1353 const U8 *scan = (U8*)NULL;
1354 U32 wordlen = 0; /* required init */
1357 if (OP(noper) == NOTHING) {
1362 TRIE_BITMAP_SET(trie,*uc);
1363 if ( folder ) TRIE_BITMAP_SET(trie,folder[ *uc ]);
1365 for ( ; uc < e ; uc += len ) {
1366 TRIE_CHARCOUNT(trie)++;
1370 if ( !trie->charmap[ uvc ] ) {
1371 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1373 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
1379 widecharmap = newHV();
1381 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1384 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1386 if ( !SvTRUE( *svpp ) ) {
1387 sv_setiv( *svpp, ++trie->uniquecharcount );
1392 if( cur == first ) {
1395 } else if (chars < trie->minlen) {
1397 } else if (chars > trie->maxlen) {
1401 } /* end first pass */
1402 DEBUG_TRIE_COMPILE_r(
1403 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1404 (int)depth * 2 + 2,"",
1405 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1406 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1407 (int)trie->minlen, (int)trie->maxlen )
1409 trie->wordlen = PerlMemShared_calloc( word_count, sizeof(U32) );
1412 We now know what we are dealing with in terms of unique chars and
1413 string sizes so we can calculate how much memory a naive
1414 representation using a flat table will take. If it's over a reasonable
1415 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1416 conservative but potentially much slower representation using an array
1419 At the end we convert both representations into the same compressed
1420 form that will be used in regexec.c for matching with. The latter
1421 is a form that cannot be used to construct with but has memory
1422 properties similar to the list form and access properties similar
1423 to the table form making it both suitable for fast searches and
1424 small enough that its feasable to store for the duration of a program.
1426 See the comment in the code where the compressed table is produced
1427 inplace from the flat tabe representation for an explanation of how
1428 the compression works.
1433 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1435 Second Pass -- Array Of Lists Representation
1437 Each state will be represented by a list of charid:state records
1438 (reg_trie_trans_le) the first such element holds the CUR and LEN
1439 points of the allocated array. (See defines above).
1441 We build the initial structure using the lists, and then convert
1442 it into the compressed table form which allows faster lookups
1443 (but cant be modified once converted).
1446 STRLEN transcount = 1;
1448 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1449 "%*sCompiling trie using list compiler\n",
1450 (int)depth * 2 + 2, ""));
1452 trie->states = PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1453 sizeof(reg_trie_state) );
1457 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1459 regnode * const noper = NEXTOPER( cur );
1460 U8 *uc = (U8*)STRING( noper );
1461 const U8 * const e = uc + STR_LEN( noper );
1462 U32 state = 1; /* required init */
1463 U16 charid = 0; /* sanity init */
1464 U8 *scan = (U8*)NULL; /* sanity init */
1465 STRLEN foldlen = 0; /* required init */
1466 U32 wordlen = 0; /* required init */
1467 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1469 if (OP(noper) != NOTHING) {
1470 for ( ; uc < e ; uc += len ) {
1475 charid = trie->charmap[ uvc ];
1477 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1481 charid=(U16)SvIV( *svpp );
1484 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1491 if ( !trie->states[ state ].trans.list ) {
1492 TRIE_LIST_NEW( state );
1494 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1495 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1496 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1501 newstate = next_alloc++;
1502 TRIE_LIST_PUSH( state, charid, newstate );
1507 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1511 TRIE_HANDLE_WORD(state);
1513 } /* end second pass */
1515 /* next alloc is the NEXT state to be allocated */
1516 trie->statecount = next_alloc;
1517 trie->states = PerlMemShared_realloc( trie->states, next_alloc
1518 * sizeof(reg_trie_state) );
1520 /* and now dump it out before we compress it */
1521 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1522 revcharmap, next_alloc,
1527 = PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1534 for( state=1 ; state < next_alloc ; state ++ ) {
1538 DEBUG_TRIE_COMPILE_MORE_r(
1539 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1543 if (trie->states[state].trans.list) {
1544 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1548 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1549 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1550 if ( forid < minid ) {
1552 } else if ( forid > maxid ) {
1556 if ( transcount < tp + maxid - minid + 1) {
1559 = PerlMemShared_realloc( trie->trans,
1561 * sizeof(reg_trie_trans) );
1562 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1564 base = trie->uniquecharcount + tp - minid;
1565 if ( maxid == minid ) {
1567 for ( ; zp < tp ; zp++ ) {
1568 if ( ! trie->trans[ zp ].next ) {
1569 base = trie->uniquecharcount + zp - minid;
1570 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1571 trie->trans[ zp ].check = state;
1577 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1578 trie->trans[ tp ].check = state;
1583 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1584 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1585 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1586 trie->trans[ tid ].check = state;
1588 tp += ( maxid - minid + 1 );
1590 Safefree(trie->states[ state ].trans.list);
1593 DEBUG_TRIE_COMPILE_MORE_r(
1594 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1597 trie->states[ state ].trans.base=base;
1599 trie->lasttrans = tp + 1;
1603 Second Pass -- Flat Table Representation.
1605 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1606 We know that we will need Charcount+1 trans at most to store the data
1607 (one row per char at worst case) So we preallocate both structures
1608 assuming worst case.
1610 We then construct the trie using only the .next slots of the entry
1613 We use the .check field of the first entry of the node temporarily to
1614 make compression both faster and easier by keeping track of how many non
1615 zero fields are in the node.
1617 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1620 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1621 number representing the first entry of the node, and state as a
1622 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1623 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1624 are 2 entrys per node. eg:
1632 The table is internally in the right hand, idx form. However as we also
1633 have to deal with the states array which is indexed by nodenum we have to
1634 use TRIE_NODENUM() to convert.
1637 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1638 "%*sCompiling trie using table compiler\n",
1639 (int)depth * 2 + 2, ""));
1641 trie->trans = PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1642 * trie->uniquecharcount + 1,
1643 sizeof(reg_trie_trans) );
1644 trie->states = PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1645 sizeof(reg_trie_state) );
1646 next_alloc = trie->uniquecharcount + 1;
1649 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1651 regnode * const noper = NEXTOPER( cur );
1652 const U8 *uc = (U8*)STRING( noper );
1653 const U8 * const e = uc + STR_LEN( noper );
1655 U32 state = 1; /* required init */
1657 U16 charid = 0; /* sanity init */
1658 U32 accept_state = 0; /* sanity init */
1659 U8 *scan = (U8*)NULL; /* sanity init */
1661 STRLEN foldlen = 0; /* required init */
1662 U32 wordlen = 0; /* required init */
1663 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1665 if ( OP(noper) != NOTHING ) {
1666 for ( ; uc < e ; uc += len ) {
1671 charid = trie->charmap[ uvc ];
1673 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1674 charid = svpp ? (U16)SvIV(*svpp) : 0;
1678 if ( !trie->trans[ state + charid ].next ) {
1679 trie->trans[ state + charid ].next = next_alloc;
1680 trie->trans[ state ].check++;
1681 next_alloc += trie->uniquecharcount;
1683 state = trie->trans[ state + charid ].next;
1685 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1687 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1690 accept_state = TRIE_NODENUM( state );
1691 TRIE_HANDLE_WORD(accept_state);
1693 } /* end second pass */
1695 /* and now dump it out before we compress it */
1696 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1698 next_alloc, depth+1));
1702 * Inplace compress the table.*
1704 For sparse data sets the table constructed by the trie algorithm will
1705 be mostly 0/FAIL transitions or to put it another way mostly empty.
1706 (Note that leaf nodes will not contain any transitions.)
1708 This algorithm compresses the tables by eliminating most such
1709 transitions, at the cost of a modest bit of extra work during lookup:
1711 - Each states[] entry contains a .base field which indicates the
1712 index in the state[] array wheres its transition data is stored.
1714 - If .base is 0 there are no valid transitions from that node.
1716 - If .base is nonzero then charid is added to it to find an entry in
1719 -If trans[states[state].base+charid].check!=state then the
1720 transition is taken to be a 0/Fail transition. Thus if there are fail
1721 transitions at the front of the node then the .base offset will point
1722 somewhere inside the previous nodes data (or maybe even into a node
1723 even earlier), but the .check field determines if the transition is
1727 The following process inplace converts the table to the compressed
1728 table: We first do not compress the root node 1,and mark its all its
1729 .check pointers as 1 and set its .base pointer as 1 as well. This
1730 allows to do a DFA construction from the compressed table later, and
1731 ensures that any .base pointers we calculate later are greater than
1734 - We set 'pos' to indicate the first entry of the second node.
1736 - We then iterate over the columns of the node, finding the first and
1737 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
1738 and set the .check pointers accordingly, and advance pos
1739 appropriately and repreat for the next node. Note that when we copy
1740 the next pointers we have to convert them from the original
1741 NODEIDX form to NODENUM form as the former is not valid post
1744 - If a node has no transitions used we mark its base as 0 and do not
1745 advance the pos pointer.
1747 - If a node only has one transition we use a second pointer into the
1748 structure to fill in allocated fail transitions from other states.
1749 This pointer is independent of the main pointer and scans forward
1750 looking for null transitions that are allocated to a state. When it
1751 finds one it writes the single transition into the "hole". If the
1752 pointer doesnt find one the single transition is appended as normal.
1754 - Once compressed we can Renew/realloc the structures to release the
1757 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
1758 specifically Fig 3.47 and the associated pseudocode.
1762 const U32 laststate = TRIE_NODENUM( next_alloc );
1765 trie->statecount = laststate;
1767 for ( state = 1 ; state < laststate ; state++ ) {
1769 const U32 stateidx = TRIE_NODEIDX( state );
1770 const U32 o_used = trie->trans[ stateidx ].check;
1771 U32 used = trie->trans[ stateidx ].check;
1772 trie->trans[ stateidx ].check = 0;
1774 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
1775 if ( flag || trie->trans[ stateidx + charid ].next ) {
1776 if ( trie->trans[ stateidx + charid ].next ) {
1778 for ( ; zp < pos ; zp++ ) {
1779 if ( ! trie->trans[ zp ].next ) {
1783 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
1784 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1785 trie->trans[ zp ].check = state;
1786 if ( ++zp > pos ) pos = zp;
1793 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
1795 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1796 trie->trans[ pos ].check = state;
1801 trie->lasttrans = pos + 1;
1802 trie->states = PerlMemShared_realloc( trie->states, laststate
1803 * sizeof(reg_trie_state) );
1804 DEBUG_TRIE_COMPILE_MORE_r(
1805 PerlIO_printf( Perl_debug_log,
1806 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
1807 (int)depth * 2 + 2,"",
1808 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
1811 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
1814 } /* end table compress */
1816 DEBUG_TRIE_COMPILE_MORE_r(
1817 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
1818 (int)depth * 2 + 2, "",
1819 (UV)trie->statecount,
1820 (UV)trie->lasttrans)
1822 /* resize the trans array to remove unused space */
1823 trie->trans = PerlMemShared_realloc( trie->trans, trie->lasttrans
1824 * sizeof(reg_trie_trans) );
1826 /* and now dump out the compressed format */
1827 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
1829 { /* Modify the program and insert the new TRIE node*/
1830 U8 nodetype =(U8)(flags & 0xFF);
1834 regnode *optimize = NULL;
1836 U32 mjd_nodelen = 0;
1839 This means we convert either the first branch or the first Exact,
1840 depending on whether the thing following (in 'last') is a branch
1841 or not and whther first is the startbranch (ie is it a sub part of
1842 the alternation or is it the whole thing.)
1843 Assuming its a sub part we conver the EXACT otherwise we convert
1844 the whole branch sequence, including the first.
1846 /* Find the node we are going to overwrite */
1847 if ( first != startbranch || OP( last ) == BRANCH ) {
1848 /* branch sub-chain */
1849 NEXT_OFF( first ) = (U16)(last - first);
1851 mjd_offset= Node_Offset((convert));
1852 mjd_nodelen= Node_Length((convert));
1854 /* whole branch chain */
1857 const regnode *nop = NEXTOPER( convert );
1858 mjd_offset= Node_Offset((nop));
1859 mjd_nodelen= Node_Length((nop));
1864 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
1865 (int)depth * 2 + 2, "",
1866 (UV)mjd_offset, (UV)mjd_nodelen)
1869 /* But first we check to see if there is a common prefix we can
1870 split out as an EXACT and put in front of the TRIE node. */
1871 trie->startstate= 1;
1872 if ( trie->bitmap && !widecharmap && !trie->jump ) {
1874 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
1878 const U32 base = trie->states[ state ].trans.base;
1880 if ( trie->states[state].wordnum )
1883 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1884 if ( ( base + ofs >= trie->uniquecharcount ) &&
1885 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1886 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1888 if ( ++count > 1 ) {
1889 SV **tmp = av_fetch( revcharmap, ofs, 0);
1890 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
1891 if ( state == 1 ) break;
1893 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
1895 PerlIO_printf(Perl_debug_log,
1896 "%*sNew Start State=%"UVuf" Class: [",
1897 (int)depth * 2 + 2, "",
1900 SV ** const tmp = av_fetch( revcharmap, idx, 0);
1901 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
1903 TRIE_BITMAP_SET(trie,*ch);
1905 TRIE_BITMAP_SET(trie, folder[ *ch ]);
1907 PerlIO_printf(Perl_debug_log, (char*)ch)
1911 TRIE_BITMAP_SET(trie,*ch);
1913 TRIE_BITMAP_SET(trie,folder[ *ch ]);
1914 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
1920 SV **tmp = av_fetch( revcharmap, idx, 0);
1921 char *ch = SvPV_nolen( *tmp );
1923 SV *sv=sv_newmortal();
1924 PerlIO_printf( Perl_debug_log,
1925 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
1926 (int)depth * 2 + 2, "",
1928 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
1929 PL_colors[0], PL_colors[1],
1930 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1931 PERL_PV_ESCAPE_FIRSTCHAR
1936 OP( convert ) = nodetype;
1937 str=STRING(convert);
1948 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
1954 regnode *n = convert+NODE_SZ_STR(convert);
1955 NEXT_OFF(convert) = NODE_SZ_STR(convert);
1956 trie->startstate = state;
1957 trie->minlen -= (state - 1);
1958 trie->maxlen -= (state - 1);
1960 regnode *fix = convert;
1961 U32 word = trie->wordcount;
1963 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
1964 while( ++fix < n ) {
1965 Set_Node_Offset_Length(fix, 0, 0);
1968 SV ** const tmp = av_fetch( trie_words, word, 0 );
1970 if ( STR_LEN(convert) <= SvCUR(*tmp) )
1971 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
1973 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
1980 NEXT_OFF(convert) = (U16)(tail - convert);
1981 DEBUG_r(optimize= n);
1987 if ( trie->maxlen ) {
1988 NEXT_OFF( convert ) = (U16)(tail - convert);
1989 ARG_SET( convert, data_slot );
1990 /* Store the offset to the first unabsorbed branch in
1991 jump[0], which is otherwise unused by the jump logic.
1992 We use this when dumping a trie and during optimisation. */
1994 trie->jump[0] = (U16)(nextbranch - convert);
1997 if ( !trie->states[trie->startstate].wordnum && trie->bitmap &&
1998 ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2000 OP( convert ) = TRIEC;
2001 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2002 PerlMemShared_free(trie->bitmap);
2005 OP( convert ) = TRIE;
2007 /* store the type in the flags */
2008 convert->flags = nodetype;
2012 + regarglen[ OP( convert ) ];
2014 /* XXX We really should free up the resource in trie now,
2015 as we won't use them - (which resources?) dmq */
2017 /* needed for dumping*/
2018 DEBUG_r(if (optimize) {
2019 regnode *opt = convert;
2020 while ( ++opt < optimize) {
2021 Set_Node_Offset_Length(opt,0,0);
2024 Try to clean up some of the debris left after the
2027 while( optimize < jumper ) {
2028 mjd_nodelen += Node_Length((optimize));
2029 OP( optimize ) = OPTIMIZED;
2030 Set_Node_Offset_Length(optimize,0,0);
2033 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2035 } /* end node insert */
2036 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2038 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2039 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2041 SvREFCNT_dec(revcharmap);
2045 : trie->startstate>1
2051 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2053 /* The Trie is constructed and compressed now so we can build a fail array now if its needed
2055 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2056 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2059 We find the fail state for each state in the trie, this state is the longest proper
2060 suffix of the current states 'word' that is also a proper prefix of another word in our
2061 trie. State 1 represents the word '' and is the thus the default fail state. This allows
2062 the DFA not to have to restart after its tried and failed a word at a given point, it
2063 simply continues as though it had been matching the other word in the first place.
2065 'abcdgu'=~/abcdefg|cdgu/
2066 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2067 fail, which would bring use to the state representing 'd' in the second word where we would
2068 try 'g' and succeed, prodceding to match 'cdgu'.
2070 /* add a fail transition */
2071 const U32 trie_offset = ARG(source);
2072 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2074 const U32 ucharcount = trie->uniquecharcount;
2075 const U32 numstates = trie->statecount;
2076 const U32 ubound = trie->lasttrans + ucharcount;
2080 U32 base = trie->states[ 1 ].trans.base;
2083 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2084 GET_RE_DEBUG_FLAGS_DECL;
2086 PERL_UNUSED_ARG(depth);
2090 ARG_SET( stclass, data_slot );
2091 aho = PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2092 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2093 aho->trie=trie_offset;
2094 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2095 Copy( trie->states, aho->states, numstates, reg_trie_state );
2096 Newxz( q, numstates, U32);
2097 aho->fail = PerlMemShared_calloc( numstates, sizeof(U32) );
2100 /* initialize fail[0..1] to be 1 so that we always have
2101 a valid final fail state */
2102 fail[ 0 ] = fail[ 1 ] = 1;
2104 for ( charid = 0; charid < ucharcount ; charid++ ) {
2105 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2107 q[ q_write ] = newstate;
2108 /* set to point at the root */
2109 fail[ q[ q_write++ ] ]=1;
2112 while ( q_read < q_write) {
2113 const U32 cur = q[ q_read++ % numstates ];
2114 base = trie->states[ cur ].trans.base;
2116 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2117 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2119 U32 fail_state = cur;
2122 fail_state = fail[ fail_state ];
2123 fail_base = aho->states[ fail_state ].trans.base;
2124 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2126 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2127 fail[ ch_state ] = fail_state;
2128 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2130 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2132 q[ q_write++ % numstates] = ch_state;
2136 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2137 when we fail in state 1, this allows us to use the
2138 charclass scan to find a valid start char. This is based on the principle
2139 that theres a good chance the string being searched contains lots of stuff
2140 that cant be a start char.
2142 fail[ 0 ] = fail[ 1 ] = 0;
2143 DEBUG_TRIE_COMPILE_r({
2144 PerlIO_printf(Perl_debug_log,
2145 "%*sStclass Failtable (%"UVuf" states): 0",
2146 (int)(depth * 2), "", (UV)numstates
2148 for( q_read=1; q_read<numstates; q_read++ ) {
2149 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2151 PerlIO_printf(Perl_debug_log, "\n");
2154 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2159 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2160 * These need to be revisited when a newer toolchain becomes available.
2162 #if defined(__sparc64__) && defined(__GNUC__)
2163 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2164 # undef SPARC64_GCC_WORKAROUND
2165 # define SPARC64_GCC_WORKAROUND 1
2169 #define DEBUG_PEEP(str,scan,depth) \
2170 DEBUG_OPTIMISE_r({if (scan){ \
2171 SV * const mysv=sv_newmortal(); \
2172 regnode *Next = regnext(scan); \
2173 regprop(RExC_rx, mysv, scan); \
2174 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2175 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2176 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2183 #define JOIN_EXACT(scan,min,flags) \
2184 if (PL_regkind[OP(scan)] == EXACT) \
2185 join_exact(pRExC_state,(scan),(min),(flags),NULL,depth+1)
2188 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, I32 *min, U32 flags,regnode *val, U32 depth) {
2189 /* Merge several consecutive EXACTish nodes into one. */
2190 regnode *n = regnext(scan);
2192 regnode *next = scan + NODE_SZ_STR(scan);
2196 regnode *stop = scan;
2197 GET_RE_DEBUG_FLAGS_DECL;
2199 PERL_UNUSED_ARG(depth);
2201 #ifndef EXPERIMENTAL_INPLACESCAN
2202 PERL_UNUSED_ARG(flags);
2203 PERL_UNUSED_ARG(val);
2205 DEBUG_PEEP("join",scan,depth);
2207 /* Skip NOTHING, merge EXACT*. */
2209 ( PL_regkind[OP(n)] == NOTHING ||
2210 (stringok && (OP(n) == OP(scan))))
2212 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX) {
2214 if (OP(n) == TAIL || n > next)
2216 if (PL_regkind[OP(n)] == NOTHING) {
2217 DEBUG_PEEP("skip:",n,depth);
2218 NEXT_OFF(scan) += NEXT_OFF(n);
2219 next = n + NODE_STEP_REGNODE;
2226 else if (stringok) {
2227 const unsigned int oldl = STR_LEN(scan);
2228 regnode * const nnext = regnext(n);
2230 DEBUG_PEEP("merg",n,depth);
2233 if (oldl + STR_LEN(n) > U8_MAX)
2235 NEXT_OFF(scan) += NEXT_OFF(n);
2236 STR_LEN(scan) += STR_LEN(n);
2237 next = n + NODE_SZ_STR(n);
2238 /* Now we can overwrite *n : */
2239 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2247 #ifdef EXPERIMENTAL_INPLACESCAN
2248 if (flags && !NEXT_OFF(n)) {
2249 DEBUG_PEEP("atch", val, depth);
2250 if (reg_off_by_arg[OP(n)]) {
2251 ARG_SET(n, val - n);
2254 NEXT_OFF(n) = val - n;
2261 if (UTF && ( OP(scan) == EXACTF ) && ( STR_LEN(scan) >= 6 ) ) {
2263 Two problematic code points in Unicode casefolding of EXACT nodes:
2265 U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2266 U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2272 U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2273 U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2275 This means that in case-insensitive matching (or "loose matching",
2276 as Unicode calls it), an EXACTF of length six (the UTF-8 encoded byte
2277 length of the above casefolded versions) can match a target string
2278 of length two (the byte length of UTF-8 encoded U+0390 or U+03B0).
2279 This would rather mess up the minimum length computation.
2281 What we'll do is to look for the tail four bytes, and then peek
2282 at the preceding two bytes to see whether we need to decrease
2283 the minimum length by four (six minus two).
2285 Thanks to the design of UTF-8, there cannot be false matches:
2286 A sequence of valid UTF-8 bytes cannot be a subsequence of
2287 another valid sequence of UTF-8 bytes.
2290 char * const s0 = STRING(scan), *s, *t;
2291 char * const s1 = s0 + STR_LEN(scan) - 1;
2292 char * const s2 = s1 - 4;
2293 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2294 const char t0[] = "\xaf\x49\xaf\x42";
2296 const char t0[] = "\xcc\x88\xcc\x81";
2298 const char * const t1 = t0 + 3;
2301 s < s2 && (t = ninstr(s, s1, t0, t1));
2304 if (((U8)t[-1] == 0x68 && (U8)t[-2] == 0xB4) ||
2305 ((U8)t[-1] == 0x46 && (U8)t[-2] == 0xB5))
2307 if (((U8)t[-1] == 0xB9 && (U8)t[-2] == 0xCE) ||
2308 ((U8)t[-1] == 0x85 && (U8)t[-2] == 0xCF))
2316 n = scan + NODE_SZ_STR(scan);
2318 if (PL_regkind[OP(n)] != NOTHING || OP(n) == NOTHING) {
2325 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2329 /* REx optimizer. Converts nodes into quickier variants "in place".
2330 Finds fixed substrings. */
2332 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2333 to the position after last scanned or to NULL. */
2335 #define INIT_AND_WITHP \
2336 assert(!and_withp); \
2337 Newx(and_withp,1,struct regnode_charclass_class); \
2338 SAVEFREEPV(and_withp)
2340 /* this is a chain of data about sub patterns we are processing that
2341 need to be handled seperately/specially in study_chunk. Its so
2342 we can simulate recursion without losing state. */
2344 typedef struct scan_frame {
2345 regnode *last; /* last node to process in this frame */
2346 regnode *next; /* next node to process when last is reached */
2347 struct scan_frame *prev; /*previous frame*/
2348 I32 stop; /* what stopparen do we use */
2352 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2355 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2356 I32 *minlenp, I32 *deltap,
2361 struct regnode_charclass_class *and_withp,
2362 U32 flags, U32 depth)
2363 /* scanp: Start here (read-write). */
2364 /* deltap: Write maxlen-minlen here. */
2365 /* last: Stop before this one. */
2366 /* data: string data about the pattern */
2367 /* stopparen: treat close N as END */
2368 /* recursed: which subroutines have we recursed into */
2369 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
2372 I32 min = 0, pars = 0, code;
2373 regnode *scan = *scanp, *next;
2375 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2376 int is_inf_internal = 0; /* The studied chunk is infinite */
2377 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2378 scan_data_t data_fake;
2379 SV *re_trie_maxbuff = NULL;
2380 regnode *first_non_open = scan;
2381 I32 stopmin = I32_MAX;
2382 scan_frame *frame = NULL;
2384 GET_RE_DEBUG_FLAGS_DECL;
2387 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
2391 while (first_non_open && OP(first_non_open) == OPEN)
2392 first_non_open=regnext(first_non_open);
2397 while ( scan && OP(scan) != END && scan < last ){
2398 /* Peephole optimizer: */
2399 DEBUG_STUDYDATA("Peep:", data,depth);
2400 DEBUG_PEEP("Peep",scan,depth);
2401 JOIN_EXACT(scan,&min,0);
2403 /* Follow the next-chain of the current node and optimize
2404 away all the NOTHINGs from it. */
2405 if (OP(scan) != CURLYX) {
2406 const int max = (reg_off_by_arg[OP(scan)]
2408 /* I32 may be smaller than U16 on CRAYs! */
2409 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
2410 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
2414 /* Skip NOTHING and LONGJMP. */
2415 while ((n = regnext(n))
2416 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
2417 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
2418 && off + noff < max)
2420 if (reg_off_by_arg[OP(scan)])
2423 NEXT_OFF(scan) = off;
2428 /* The principal pseudo-switch. Cannot be a switch, since we
2429 look into several different things. */
2430 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
2431 || OP(scan) == IFTHEN) {
2432 next = regnext(scan);
2434 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
2436 if (OP(next) == code || code == IFTHEN) {
2437 /* NOTE - There is similar code to this block below for handling
2438 TRIE nodes on a re-study. If you change stuff here check there
2440 I32 max1 = 0, min1 = I32_MAX, num = 0;
2441 struct regnode_charclass_class accum;
2442 regnode * const startbranch=scan;
2444 if (flags & SCF_DO_SUBSTR)
2445 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
2446 if (flags & SCF_DO_STCLASS)
2447 cl_init_zero(pRExC_state, &accum);
2449 while (OP(scan) == code) {
2450 I32 deltanext, minnext, f = 0, fake;
2451 struct regnode_charclass_class this_class;
2454 data_fake.flags = 0;
2456 data_fake.whilem_c = data->whilem_c;
2457 data_fake.last_closep = data->last_closep;
2460 data_fake.last_closep = &fake;
2462 data_fake.pos_delta = delta;
2463 next = regnext(scan);
2464 scan = NEXTOPER(scan);
2466 scan = NEXTOPER(scan);
2467 if (flags & SCF_DO_STCLASS) {
2468 cl_init(pRExC_state, &this_class);
2469 data_fake.start_class = &this_class;
2470 f = SCF_DO_STCLASS_AND;
2472 if (flags & SCF_WHILEM_VISITED_POS)
2473 f |= SCF_WHILEM_VISITED_POS;
2475 /* we suppose the run is continuous, last=next...*/
2476 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
2478 stopparen, recursed, NULL, f,depth+1);
2481 if (max1 < minnext + deltanext)
2482 max1 = minnext + deltanext;
2483 if (deltanext == I32_MAX)
2484 is_inf = is_inf_internal = 1;
2486 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
2488 if (data_fake.flags & SCF_SEEN_ACCEPT) {
2489 if ( stopmin > minnext)
2490 stopmin = min + min1;
2491 flags &= ~SCF_DO_SUBSTR;
2493 data->flags |= SCF_SEEN_ACCEPT;
2496 if (data_fake.flags & SF_HAS_EVAL)
2497 data->flags |= SF_HAS_EVAL;
2498 data->whilem_c = data_fake.whilem_c;
2500 if (flags & SCF_DO_STCLASS)
2501 cl_or(pRExC_state, &accum, &this_class);
2503 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
2505 if (flags & SCF_DO_SUBSTR) {
2506 data->pos_min += min1;
2507 data->pos_delta += max1 - min1;
2508 if (max1 != min1 || is_inf)
2509 data->longest = &(data->longest_float);
2512 delta += max1 - min1;
2513 if (flags & SCF_DO_STCLASS_OR) {
2514 cl_or(pRExC_state, data->start_class, &accum);
2516 cl_and(data->start_class, and_withp);
2517 flags &= ~SCF_DO_STCLASS;
2520 else if (flags & SCF_DO_STCLASS_AND) {
2522 cl_and(data->start_class, &accum);
2523 flags &= ~SCF_DO_STCLASS;
2526 /* Switch to OR mode: cache the old value of
2527 * data->start_class */
2529 StructCopy(data->start_class, and_withp,
2530 struct regnode_charclass_class);
2531 flags &= ~SCF_DO_STCLASS_AND;
2532 StructCopy(&accum, data->start_class,
2533 struct regnode_charclass_class);
2534 flags |= SCF_DO_STCLASS_OR;
2535 data->start_class->flags |= ANYOF_EOS;
2539 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
2542 Assuming this was/is a branch we are dealing with: 'scan' now
2543 points at the item that follows the branch sequence, whatever
2544 it is. We now start at the beginning of the sequence and look
2551 which would be constructed from a pattern like /A|LIST|OF|WORDS/
2553 If we can find such a subseqence we need to turn the first
2554 element into a trie and then add the subsequent branch exact
2555 strings to the trie.
2559 1. patterns where the whole set of branch can be converted.
2561 2. patterns where only a subset can be converted.
2563 In case 1 we can replace the whole set with a single regop
2564 for the trie. In case 2 we need to keep the start and end
2567 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
2568 becomes BRANCH TRIE; BRANCH X;
2570 There is an additional case, that being where there is a
2571 common prefix, which gets split out into an EXACT like node
2572 preceding the TRIE node.
2574 If x(1..n)==tail then we can do a simple trie, if not we make
2575 a "jump" trie, such that when we match the appropriate word
2576 we "jump" to the appopriate tail node. Essentailly we turn
2577 a nested if into a case structure of sorts.
2582 if (!re_trie_maxbuff) {
2583 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2584 if (!SvIOK(re_trie_maxbuff))
2585 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2587 if ( SvIV(re_trie_maxbuff)>=0 ) {
2589 regnode *first = (regnode *)NULL;
2590 regnode *last = (regnode *)NULL;
2591 regnode *tail = scan;
2596 SV * const mysv = sv_newmortal(); /* for dumping */
2598 /* var tail is used because there may be a TAIL
2599 regop in the way. Ie, the exacts will point to the
2600 thing following the TAIL, but the last branch will
2601 point at the TAIL. So we advance tail. If we
2602 have nested (?:) we may have to move through several
2606 while ( OP( tail ) == TAIL ) {
2607 /* this is the TAIL generated by (?:) */
2608 tail = regnext( tail );
2613 regprop(RExC_rx, mysv, tail );
2614 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
2615 (int)depth * 2 + 2, "",
2616 "Looking for TRIE'able sequences. Tail node is: ",
2617 SvPV_nolen_const( mysv )
2623 step through the branches, cur represents each
2624 branch, noper is the first thing to be matched
2625 as part of that branch and noper_next is the
2626 regnext() of that node. if noper is an EXACT
2627 and noper_next is the same as scan (our current
2628 position in the regex) then the EXACT branch is
2629 a possible optimization target. Once we have
2630 two or more consequetive such branches we can
2631 create a trie of the EXACT's contents and stich
2632 it in place. If the sequence represents all of
2633 the branches we eliminate the whole thing and
2634 replace it with a single TRIE. If it is a
2635 subsequence then we need to stitch it in. This
2636 means the first branch has to remain, and needs
2637 to be repointed at the item on the branch chain
2638 following the last branch optimized. This could
2639 be either a BRANCH, in which case the
2640 subsequence is internal, or it could be the
2641 item following the branch sequence in which
2642 case the subsequence is at the end.
2646 /* dont use tail as the end marker for this traverse */
2647 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
2648 regnode * const noper = NEXTOPER( cur );
2649 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
2650 regnode * const noper_next = regnext( noper );
2654 regprop(RExC_rx, mysv, cur);
2655 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
2656 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
2658 regprop(RExC_rx, mysv, noper);
2659 PerlIO_printf( Perl_debug_log, " -> %s",
2660 SvPV_nolen_const(mysv));
2663 regprop(RExC_rx, mysv, noper_next );
2664 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
2665 SvPV_nolen_const(mysv));
2667 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
2668 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
2670 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
2671 : PL_regkind[ OP( noper ) ] == EXACT )
2672 || OP(noper) == NOTHING )
2674 && noper_next == tail
2679 if ( !first || optype == NOTHING ) {
2680 if (!first) first = cur;
2681 optype = OP( noper );
2687 make_trie( pRExC_state,
2688 startbranch, first, cur, tail, count,
2691 if ( PL_regkind[ OP( noper ) ] == EXACT
2693 && noper_next == tail
2698 optype = OP( noper );
2708 regprop(RExC_rx, mysv, cur);
2709 PerlIO_printf( Perl_debug_log,
2710 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
2711 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
2715 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
2716 #ifdef TRIE_STUDY_OPT
2717 if ( ((made == MADE_EXACT_TRIE &&
2718 startbranch == first)
2719 || ( first_non_open == first )) &&
2721 flags |= SCF_TRIE_RESTUDY;
2722 if ( startbranch == first
2725 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
2735 else if ( code == BRANCHJ ) { /* single branch is optimized. */
2736 scan = NEXTOPER(NEXTOPER(scan));
2737 } else /* single branch is optimized. */
2738 scan = NEXTOPER(scan);
2740 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
2741 scan_frame *newframe = NULL;
2746 if (OP(scan) != SUSPEND) {
2747 /* set the pointer */
2748 if (OP(scan) == GOSUB) {
2750 RExC_recurse[ARG2L(scan)] = scan;
2751 start = RExC_open_parens[paren-1];
2752 end = RExC_close_parens[paren-1];
2755 start = RExC_rxi->program + 1;
2759 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
2760 SAVEFREEPV(recursed);
2762 if (!PAREN_TEST(recursed,paren+1)) {
2763 PAREN_SET(recursed,paren+1);
2764 Newx(newframe,1,scan_frame);
2766 if (flags & SCF_DO_SUBSTR) {
2767 SCAN_COMMIT(pRExC_state,data,minlenp);
2768 data->longest = &(data->longest_float);
2770 is_inf = is_inf_internal = 1;
2771 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
2772 cl_anything(pRExC_state, data->start_class);
2773 flags &= ~SCF_DO_STCLASS;
2776 Newx(newframe,1,scan_frame);
2779 end = regnext(scan);
2784 SAVEFREEPV(newframe);
2785 newframe->next = regnext(scan);
2786 newframe->last = last;
2787 newframe->stop = stopparen;
2788 newframe->prev = frame;
2798 else if (OP(scan) == EXACT) {
2799 I32 l = STR_LEN(scan);
2802 const U8 * const s = (U8*)STRING(scan);
2803 l = utf8_length(s, s + l);
2804 uc = utf8_to_uvchr(s, NULL);
2806 uc = *((U8*)STRING(scan));
2809 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
2810 /* The code below prefers earlier match for fixed
2811 offset, later match for variable offset. */
2812 if (data->last_end == -1) { /* Update the start info. */
2813 data->last_start_min = data->pos_min;
2814 data->last_start_max = is_inf
2815 ? I32_MAX : data->pos_min + data->pos_delta;
2817 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
2819 SvUTF8_on(data->last_found);
2821 SV * const sv = data->last_found;
2822 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
2823 mg_find(sv, PERL_MAGIC_utf8) : NULL;
2824 if (mg && mg->mg_len >= 0)
2825 mg->mg_len += utf8_length((U8*)STRING(scan),
2826 (U8*)STRING(scan)+STR_LEN(scan));
2828 data->last_end = data->pos_min + l;
2829 data->pos_min += l; /* As in the first entry. */
2830 data->flags &= ~SF_BEFORE_EOL;
2832 if (flags & SCF_DO_STCLASS_AND) {
2833 /* Check whether it is compatible with what we know already! */
2837 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
2838 && !ANYOF_BITMAP_TEST(data->start_class, uc)
2839 && (!(data->start_class->flags & ANYOF_FOLD)
2840 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold[uc])))
2843 ANYOF_CLASS_ZERO(data->start_class);
2844 ANYOF_BITMAP_ZERO(data->start_class);
2846 ANYOF_BITMAP_SET(data->start_class, uc);
2847 data->start_class->flags &= ~ANYOF_EOS;
2849 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
2851 else if (flags & SCF_DO_STCLASS_OR) {
2852 /* false positive possible if the class is case-folded */
2854 ANYOF_BITMAP_SET(data->start_class, uc);
2856 data->start_class->flags |= ANYOF_UNICODE_ALL;
2857 data->start_class->flags &= ~ANYOF_EOS;
2858 cl_and(data->start_class, and_withp);
2860 flags &= ~SCF_DO_STCLASS;
2862 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
2863 I32 l = STR_LEN(scan);
2864 UV uc = *((U8*)STRING(scan));
2866 /* Search for fixed substrings supports EXACT only. */
2867 if (flags & SCF_DO_SUBSTR) {
2869 SCAN_COMMIT(pRExC_state, data, minlenp);
2872 const U8 * const s = (U8 *)STRING(scan);
2873 l = utf8_length(s, s + l);
2874 uc = utf8_to_uvchr(s, NULL);
2877 if (flags & SCF_DO_SUBSTR)
2879 if (flags & SCF_DO_STCLASS_AND) {
2880 /* Check whether it is compatible with what we know already! */
2884 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
2885 && !ANYOF_BITMAP_TEST(data->start_class, uc)
2886 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold[uc])))
2888 ANYOF_CLASS_ZERO(data->start_class);
2889 ANYOF_BITMAP_ZERO(data->start_class);
2891 ANYOF_BITMAP_SET(data->start_class, uc);
2892 data->start_class->flags &= ~ANYOF_EOS;
2893 data->start_class->flags |= ANYOF_FOLD;
2894 if (OP(scan) == EXACTFL)
2895 data->start_class->flags |= ANYOF_LOCALE;
2898 else if (flags & SCF_DO_STCLASS_OR) {
2899 if (data->start_class->flags & ANYOF_FOLD) {
2900 /* false positive possible if the class is case-folded.
2901 Assume that the locale settings are the same... */
2903 ANYOF_BITMAP_SET(data->start_class, uc);
2904 data->start_class->flags &= ~ANYOF_EOS;
2906 cl_and(data->start_class, and_withp);
2908 flags &= ~SCF_DO_STCLASS;
2910 else if (strchr((const char*)PL_varies,OP(scan))) {
2911 I32 mincount, maxcount, minnext, deltanext, fl = 0;
2912 I32 f = flags, pos_before = 0;
2913 regnode * const oscan = scan;
2914 struct regnode_charclass_class this_class;
2915 struct regnode_charclass_class *oclass = NULL;
2916 I32 next_is_eval = 0;
2918 switch (PL_regkind[OP(scan)]) {
2919 case WHILEM: /* End of (?:...)* . */
2920 scan = NEXTOPER(scan);
2923 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
2924 next = NEXTOPER(scan);
2925 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
2927 maxcount = REG_INFTY;
2928 next = regnext(scan);
2929 scan = NEXTOPER(scan);
2933 if (flags & SCF_DO_SUBSTR)
2938 if (flags & SCF_DO_STCLASS) {
2940 maxcount = REG_INFTY;
2941 next = regnext(scan);
2942 scan = NEXTOPER(scan);
2945 is_inf = is_inf_internal = 1;
2946 scan = regnext(scan);
2947 if (flags & SCF_DO_SUBSTR) {
2948 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
2949 data->longest = &(data->longest_float);
2951 goto optimize_curly_tail;
2953 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
2954 && (scan->flags == stopparen))
2959 mincount = ARG1(scan);
2960 maxcount = ARG2(scan);
2962 next = regnext(scan);
2963 if (OP(scan) == CURLYX) {
2964 I32 lp = (data ? *(data->last_closep) : 0);
2965 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
2967 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
2968 next_is_eval = (OP(scan) == EVAL);
2970 if (flags & SCF_DO_SUBSTR) {
2971 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
2972 pos_before = data->pos_min;
2976 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
2978 data->flags |= SF_IS_INF;
2980 if (flags & SCF_DO_STCLASS) {
2981 cl_init(pRExC_state, &this_class);
2982 oclass = data->start_class;
2983 data->start_class = &this_class;
2984 f |= SCF_DO_STCLASS_AND;
2985 f &= ~SCF_DO_STCLASS_OR;
2987 /* These are the cases when once a subexpression
2988 fails at a particular position, it cannot succeed
2989 even after backtracking at the enclosing scope.
2991 XXXX what if minimal match and we are at the
2992 initial run of {n,m}? */
2993 if ((mincount != maxcount - 1) && (maxcount != REG_INFTY))
2994 f &= ~SCF_WHILEM_VISITED_POS;
2996 /* This will finish on WHILEM, setting scan, or on NULL: */
2997 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
2998 last, data, stopparen, recursed, NULL,
3000 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3002 if (flags & SCF_DO_STCLASS)
3003 data->start_class = oclass;
3004 if (mincount == 0 || minnext == 0) {
3005 if (flags & SCF_DO_STCLASS_OR) {
3006 cl_or(pRExC_state, data->start_class, &this_class);
3008 else if (flags & SCF_DO_STCLASS_AND) {
3009 /* Switch to OR mode: cache the old value of
3010 * data->start_class */
3012 StructCopy(data->start_class, and_withp,
3013 struct regnode_charclass_class);
3014 flags &= ~SCF_DO_STCLASS_AND;
3015 StructCopy(&this_class, data->start_class,
3016 struct regnode_charclass_class);
3017 flags |= SCF_DO_STCLASS_OR;
3018 data->start_class->flags |= ANYOF_EOS;
3020 } else { /* Non-zero len */
3021 if (flags & SCF_DO_STCLASS_OR) {
3022 cl_or(pRExC_state, data->start_class, &this_class);
3023 cl_and(data->start_class, and_withp);
3025 else if (flags & SCF_DO_STCLASS_AND)
3026 cl_and(data->start_class, &this_class);
3027 flags &= ~SCF_DO_STCLASS;
3029 if (!scan) /* It was not CURLYX, but CURLY. */
3031 if ( /* ? quantifier ok, except for (?{ ... }) */
3032 (next_is_eval || !(mincount == 0 && maxcount == 1))
3033 && (minnext == 0) && (deltanext == 0)
3034 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3035 && maxcount <= REG_INFTY/3 /* Complement check for big count */
3036 && ckWARN(WARN_REGEXP))
3039 "Quantifier unexpected on zero-length expression");
3042 min += minnext * mincount;
3043 is_inf_internal |= ((maxcount == REG_INFTY
3044 && (minnext + deltanext) > 0)
3045 || deltanext == I32_MAX);
3046 is_inf |= is_inf_internal;
3047 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3049 /* Try powerful optimization CURLYX => CURLYN. */
3050 if ( OP(oscan) == CURLYX && data
3051 && data->flags & SF_IN_PAR
3052 && !(data->flags & SF_HAS_EVAL)
3053 && !deltanext && minnext == 1 ) {
3054 /* Try to optimize to CURLYN. */
3055 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3056 regnode * const nxt1 = nxt;
3063 if (!strchr((const char*)PL_simple,OP(nxt))
3064 && !(PL_regkind[OP(nxt)] == EXACT
3065 && STR_LEN(nxt) == 1))
3071 if (OP(nxt) != CLOSE)
3073 if (RExC_open_parens) {
3074 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3075 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3077 /* Now we know that nxt2 is the only contents: */
3078 oscan->flags = (U8)ARG(nxt);
3080 OP(nxt1) = NOTHING; /* was OPEN. */
3083 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3084 NEXT_OFF(nxt1+ 1) = 0; /* just for consistancy. */
3085 NEXT_OFF(nxt2) = 0; /* just for consistancy with CURLY. */
3086 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3087 OP(nxt + 1) = OPTIMIZED; /* was count. */
3088 NEXT_OFF(nxt+ 1) = 0; /* just for consistancy. */
3093 /* Try optimization CURLYX => CURLYM. */
3094 if ( OP(oscan) == CURLYX && data
3095 && !(data->flags & SF_HAS_PAR)
3096 && !(data->flags & SF_HAS_EVAL)
3097 && !deltanext /* atom is fixed width */
3098 && minnext != 0 /* CURLYM can't handle zero width */
3100 /* XXXX How to optimize if data == 0? */
3101 /* Optimize to a simpler form. */
3102 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3106 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3107 && (OP(nxt2) != WHILEM))
3109 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3110 /* Need to optimize away parenths. */
3111 if (data->flags & SF_IN_PAR) {
3112 /* Set the parenth number. */
3113 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3115 if (OP(nxt) != CLOSE)
3116 FAIL("Panic opt close");
3117 oscan->flags = (U8)ARG(nxt);
3118 if (RExC_open_parens) {
3119 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3120 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3122 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3123 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3126 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3127 OP(nxt + 1) = OPTIMIZED; /* was count. */
3128 NEXT_OFF(nxt1 + 1) = 0; /* just for consistancy. */
3129 NEXT_OFF(nxt + 1) = 0; /* just for consistancy. */
3132 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3133 regnode *nnxt = regnext(nxt1);
3136 if (reg_off_by_arg[OP(nxt1)])
3137 ARG_SET(nxt1, nxt2 - nxt1);
3138 else if (nxt2 - nxt1 < U16_MAX)
3139 NEXT_OFF(nxt1) = nxt2 - nxt1;
3141 OP(nxt) = NOTHING; /* Cannot beautify */
3146 /* Optimize again: */
3147 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
3148 NULL, stopparen, recursed, NULL, 0,depth+1);
3153 else if ((OP(oscan) == CURLYX)
3154 && (flags & SCF_WHILEM_VISITED_POS)
3155 /* See the comment on a similar expression above.
3156 However, this time it not a subexpression
3157 we care about, but the expression itself. */
3158 && (maxcount == REG_INFTY)
3159 && data && ++data->whilem_c < 16) {
3160 /* This stays as CURLYX, we can put the count/of pair. */
3161 /* Find WHILEM (as in regexec.c) */
3162 regnode *nxt = oscan + NEXT_OFF(oscan);
3164 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
3166 PREVOPER(nxt)->flags = (U8)(data->whilem_c
3167 | (RExC_whilem_seen << 4)); /* On WHILEM */
3169 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
3171 if (flags & SCF_DO_SUBSTR) {
3172 SV *last_str = NULL;
3173 int counted = mincount != 0;
3175 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
3176 #if defined(SPARC64_GCC_WORKAROUND)
3179 const char *s = NULL;
3182 if (pos_before >= data->last_start_min)
3185 b = data->last_start_min;
3188 s = SvPV_const(data->last_found, l);
3189 old = b - data->last_start_min;
3192 I32 b = pos_before >= data->last_start_min
3193 ? pos_before : data->last_start_min;
3195 const char * const s = SvPV_const(data->last_found, l);
3196 I32 old = b - data->last_start_min;
3200 old = utf8_hop((U8*)s, old) - (U8*)s;
3203 /* Get the added string: */
3204 last_str = newSVpvn(s + old, l);
3206 SvUTF8_on(last_str);
3207 if (deltanext == 0 && pos_before == b) {
3208 /* What was added is a constant string */
3210 SvGROW(last_str, (mincount * l) + 1);
3211 repeatcpy(SvPVX(last_str) + l,
3212 SvPVX_const(last_str), l, mincount - 1);
3213 SvCUR_set(last_str, SvCUR(last_str) * mincount);
3214 /* Add additional parts. */
3215 SvCUR_set(data->last_found,
3216 SvCUR(data->last_found) - l);
3217 sv_catsv(data->last_found, last_str);
3219 SV * sv = data->last_found;
3221 SvUTF8(sv) && SvMAGICAL(sv) ?
3222 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3223 if (mg && mg->mg_len >= 0)
3224 mg->mg_len += CHR_SVLEN(last_str);
3226 data->last_end += l * (mincount - 1);
3229 /* start offset must point into the last copy */
3230 data->last_start_min += minnext * (mincount - 1);
3231 data->last_start_max += is_inf ? I32_MAX
3232 : (maxcount - 1) * (minnext + data->pos_delta);
3235 /* It is counted once already... */
3236 data->pos_min += minnext * (mincount - counted);
3237 data->pos_delta += - counted * deltanext +
3238 (minnext + deltanext) * maxcount - minnext * mincount;
3239 if (mincount != maxcount) {
3240 /* Cannot extend fixed substrings found inside
3242 SCAN_COMMIT(pRExC_state,data,minlenp);
3243 if (mincount && last_str) {
3244 SV * const sv = data->last_found;
3245 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3246 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3250 sv_setsv(sv, last_str);
3251 data->last_end = data->pos_min;
3252 data->last_start_min =
3253 data->pos_min - CHR_SVLEN(last_str);
3254 data->last_start_max = is_inf
3256 : data->pos_min + data->pos_delta
3257 - CHR_SVLEN(last_str);
3259 data->longest = &(data->longest_float);
3261 SvREFCNT_dec(last_str);
3263 if (data && (fl & SF_HAS_EVAL))
3264 data->flags |= SF_HAS_EVAL;
3265 optimize_curly_tail:
3266 if (OP(oscan) != CURLYX) {
3267 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
3269 NEXT_OFF(oscan) += NEXT_OFF(next);
3272 default: /* REF and CLUMP only? */
3273 if (flags & SCF_DO_SUBSTR) {
3274 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3275 data->longest = &(data->longest_float);
3277 is_inf = is_inf_internal = 1;
3278 if (flags & SCF_DO_STCLASS_OR)
3279 cl_anything(pRExC_state, data->start_class);
3280 flags &= ~SCF_DO_STCLASS;
3284 else if (strchr((const char*)PL_simple,OP(scan))) {
3287 if (flags & SCF_DO_SUBSTR) {
3288 SCAN_COMMIT(pRExC_state,data,minlenp);
3292 if (flags & SCF_DO_STCLASS) {
3293 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3295 /* Some of the logic below assumes that switching
3296 locale on will only add false positives. */
3297 switch (PL_regkind[OP(scan)]) {
3301 /* Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan)); */
3302 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3303 cl_anything(pRExC_state, data->start_class);
3306 if (OP(scan) == SANY)
3308 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
3309 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
3310 || (data->start_class->flags & ANYOF_CLASS));
3311 cl_anything(pRExC_state, data->start_class);
3313 if (flags & SCF_DO_STCLASS_AND || !value)
3314 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
3317 if (flags & SCF_DO_STCLASS_AND)
3318 cl_and(data->start_class,
3319 (struct regnode_charclass_class*)scan);
3321 cl_or(pRExC_state, data->start_class,
3322 (struct regnode_charclass_class*)scan);
3325 if (flags & SCF_DO_STCLASS_AND) {
3326 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3327 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3328 for (value = 0; value < 256; value++)
3329 if (!isALNUM(value))
3330 ANYOF_BITMAP_CLEAR(data->start_class, value);
3334 if (data->start_class->flags & ANYOF_LOCALE)
3335 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3337 for (value = 0; value < 256; value++)
3339 ANYOF_BITMAP_SET(data->start_class, value);
3344 if (flags & SCF_DO_STCLASS_AND) {
3345 if (data->start_class->flags & ANYOF_LOCALE)
3346 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3349 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3350 data->start_class->flags |= ANYOF_LOCALE;
3354 if (flags & SCF_DO_STCLASS_AND) {
3355 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3356 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3357 for (value = 0; value < 256; value++)
3359 ANYOF_BITMAP_CLEAR(data->start_class, value);
3363 if (data->start_class->flags & ANYOF_LOCALE)
3364 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3366 for (value = 0; value < 256; value++)
3367 if (!isALNUM(value))
3368 ANYOF_BITMAP_SET(data->start_class, value);
3373 if (flags & SCF_DO_STCLASS_AND) {
3374 if (data->start_class->flags & ANYOF_LOCALE)
3375 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3378 data->start_class->flags |= ANYOF_LOCALE;
3379 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3383 if (flags & SCF_DO_STCLASS_AND) {
3384 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3385 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3386 for (value = 0; value < 256; value++)
3387 if (!isSPACE(value))
3388 ANYOF_BITMAP_CLEAR(data->start_class, value);
3392 if (data->start_class->flags & ANYOF_LOCALE)
3393 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3395 for (value = 0; value < 256; value++)
3397 ANYOF_BITMAP_SET(data->start_class, value);
3402 if (flags & SCF_DO_STCLASS_AND) {
3403 if (data->start_class->flags & ANYOF_LOCALE)
3404 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3407 data->start_class->flags |= ANYOF_LOCALE;
3408 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3412 if (flags & SCF_DO_STCLASS_AND) {
3413 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3414 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3415 for (value = 0; value < 256; value++)
3417 ANYOF_BITMAP_CLEAR(data->start_class, value);
3421 if (data->start_class->flags & ANYOF_LOCALE)
3422 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3424 for (value = 0; value < 256; value++)
3425 if (!isSPACE(value))
3426 ANYOF_BITMAP_SET(data->start_class, value);
3431 if (flags & SCF_DO_STCLASS_AND) {
3432 if (data->start_class->flags & ANYOF_LOCALE) {
3433 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3434 for (value = 0; value < 256; value++)
3435 if (!isSPACE(value))
3436 ANYOF_BITMAP_CLEAR(data->start_class, value);
3440 data->start_class->flags |= ANYOF_LOCALE;
3441 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3445 if (flags & SCF_DO_STCLASS_AND) {
3446 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NDIGIT);
3447 for (value = 0; value < 256; value++)
3448 if (!isDIGIT(value))
3449 ANYOF_BITMAP_CLEAR(data->start_class, value);
3452 if (data->start_class->flags & ANYOF_LOCALE)
3453 ANYOF_CLASS_SET(data->start_class,ANYOF_DIGIT);
3455 for (value = 0; value < 256; value++)
3457 ANYOF_BITMAP_SET(data->start_class, value);
3462 if (flags & SCF_DO_STCLASS_AND) {
3463 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_DIGIT);
3464 for (value = 0; value < 256; value++)
3466 ANYOF_BITMAP_CLEAR(data->start_class, value);
3469 if (data->start_class->flags & ANYOF_LOCALE)
3470 ANYOF_CLASS_SET(data->start_class,ANYOF_NDIGIT);
3472 for (value = 0; value < 256; value++)
3473 if (!isDIGIT(value))
3474 ANYOF_BITMAP_SET(data->start_class, value);
3479 if (flags & SCF_DO_STCLASS_OR)
3480 cl_and(data->start_class, and_withp);
3481 flags &= ~SCF_DO_STCLASS;
3484 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
3485 data->flags |= (OP(scan) == MEOL
3489 else if ( PL_regkind[OP(scan)] == BRANCHJ
3490 /* Lookbehind, or need to calculate parens/evals/stclass: */
3491 && (scan->flags || data || (flags & SCF_DO_STCLASS))
3492 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
3493 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
3494 || OP(scan) == UNLESSM )
3496 /* Negative Lookahead/lookbehind
3497 In this case we can't do fixed string optimisation.
3500 I32 deltanext, minnext, fake = 0;
3502 struct regnode_charclass_class intrnl;
3505 data_fake.flags = 0;
3507 data_fake.whilem_c = data->whilem_c;
3508 data_fake.last_closep = data->last_closep;
3511 data_fake.last_closep = &fake;
3512 data_fake.pos_delta = delta;
3513 if ( flags & SCF_DO_STCLASS && !scan->flags
3514 && OP(scan) == IFMATCH ) { /* Lookahead */
3515 cl_init(pRExC_state, &intrnl);
3516 data_fake.start_class = &intrnl;
3517 f |= SCF_DO_STCLASS_AND;
3519 if (flags & SCF_WHILEM_VISITED_POS)
3520 f |= SCF_WHILEM_VISITED_POS;
3521 next = regnext(scan);
3522 nscan = NEXTOPER(NEXTOPER(scan));
3523 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
3524 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
3527 FAIL("Variable length lookbehind not implemented");
3529 else if (minnext > (I32)U8_MAX) {
3530 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
3532 scan->flags = (U8)minnext;
3535 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3537 if (data_fake.flags & SF_HAS_EVAL)
3538 data->flags |= SF_HAS_EVAL;
3539 data->whilem_c = data_fake.whilem_c;
3541 if (f & SCF_DO_STCLASS_AND) {
3542 const int was = (data->start_class->flags & ANYOF_EOS);
3544 cl_and(data->start_class, &intrnl);
3546 data->start_class->flags |= ANYOF_EOS;
3549 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
3551 /* Positive Lookahead/lookbehind
3552 In this case we can do fixed string optimisation,
3553 but we must be careful about it. Note in the case of
3554 lookbehind the positions will be offset by the minimum
3555 length of the pattern, something we won't know about
3556 until after the recurse.
3558 I32 deltanext, fake = 0;
3560 struct regnode_charclass_class intrnl;
3562 /* We use SAVEFREEPV so that when the full compile
3563 is finished perl will clean up the allocated
3564 minlens when its all done. This was we don't
3565 have to worry about freeing them when we know
3566 they wont be used, which would be a pain.
3569 Newx( minnextp, 1, I32 );
3570 SAVEFREEPV(minnextp);
3573 StructCopy(data, &data_fake, scan_data_t);
3574 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
3577 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
3578 data_fake.last_found=newSVsv(data->last_found);
3582 data_fake.last_closep = &fake;
3583 data_fake.flags = 0;
3584 data_fake.pos_delta = delta;
3586 data_fake.flags |= SF_IS_INF;
3587 if ( flags & SCF_DO_STCLASS && !scan->flags
3588 && OP(scan) == IFMATCH ) { /* Lookahead */
3589 cl_init(pRExC_state, &intrnl);
3590 data_fake.start_class = &intrnl;
3591 f |= SCF_DO_STCLASS_AND;
3593 if (flags & SCF_WHILEM_VISITED_POS)
3594 f |= SCF_WHILEM_VISITED_POS;
3595 next = regnext(scan);
3596 nscan = NEXTOPER(NEXTOPER(scan));
3598 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
3599 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
3602 FAIL("Variable length lookbehind not implemented");
3604 else if (*minnextp > (I32)U8_MAX) {
3605 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
3607 scan->flags = (U8)*minnextp;
3612 if (f & SCF_DO_STCLASS_AND) {
3613 const int was = (data->start_class->flags & ANYOF_EOS);
3615 cl_and(data->start_class, &intrnl);
3617 data->start_class->flags |= ANYOF_EOS;
3620 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3622 if (data_fake.flags & SF_HAS_EVAL)
3623 data->flags |= SF_HAS_EVAL;
3624 data->whilem_c = data_fake.whilem_c;
3625 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
3626 if (RExC_rx->minlen<*minnextp)
3627 RExC_rx->minlen=*minnextp;
3628 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
3629 SvREFCNT_dec(data_fake.last_found);
3631 if ( data_fake.minlen_fixed != minlenp )
3633 data->offset_fixed= data_fake.offset_fixed;
3634 data->minlen_fixed= data_fake.minlen_fixed;
3635 data->lookbehind_fixed+= scan->flags;
3637 if ( data_fake.minlen_float != minlenp )
3639 data->minlen_float= data_fake.minlen_float;
3640 data->offset_float_min=data_fake.offset_float_min;
3641 data->offset_float_max=data_fake.offset_float_max;
3642 data->lookbehind_float+= scan->flags;
3651 else if (OP(scan) == OPEN) {
3652 if (stopparen != (I32)ARG(scan))
3655 else if (OP(scan) == CLOSE) {
3656 if (stopparen == (I32)ARG(scan)) {
3659 if ((I32)ARG(scan) == is_par) {
3660 next = regnext(scan);
3662 if ( next && (OP(next) != WHILEM) && next < last)
3663 is_par = 0; /* Disable optimization */
3666 *(data->last_closep) = ARG(scan);
3668 else if (OP(scan) == EVAL) {
3670 data->flags |= SF_HAS_EVAL;
3672 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
3673 if (flags & SCF_DO_SUBSTR) {
3674 SCAN_COMMIT(pRExC_state,data,minlenp);
3675 flags &= ~SCF_DO_SUBSTR;
3677 if (data && OP(scan)==ACCEPT) {
3678 data->flags |= SCF_SEEN_ACCEPT;
3683 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
3685 if (flags & SCF_DO_SUBSTR) {
3686 SCAN_COMMIT(pRExC_state,data,minlenp);
3687 data->longest = &(data->longest_float);
3689 is_inf = is_inf_internal = 1;
3690 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3691 cl_anything(pRExC_state, data->start_class);
3692 flags &= ~SCF_DO_STCLASS;
3694 else if (OP(scan) == GPOS) {
3695 if (!(RExC_rx->extflags & RXf_GPOS_FLOAT) &&
3696 !(delta || is_inf || (data && data->pos_delta)))
3698 if (!(RExC_rx->extflags & RXf_ANCH) && (flags & SCF_DO_SUBSTR))
3699 RExC_rx->extflags |= RXf_ANCH_GPOS;
3700 if (RExC_rx->gofs < (U32)min)
3701 RExC_rx->gofs = min;
3703 RExC_rx->extflags |= RXf_GPOS_FLOAT;
3707 #ifdef TRIE_STUDY_OPT
3708 #ifdef FULL_TRIE_STUDY
3709 else if (PL_regkind[OP(scan)] == TRIE) {
3710 /* NOTE - There is similar code to this block above for handling
3711 BRANCH nodes on the initial study. If you change stuff here
3713 regnode *trie_node= scan;
3714 regnode *tail= regnext(scan);
3715 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
3716 I32 max1 = 0, min1 = I32_MAX;
3717 struct regnode_charclass_class accum;
3719 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
3720 SCAN_COMMIT(pRExC_state, data,minlenp); /* Cannot merge strings after this. */
3721 if (flags & SCF_DO_STCLASS)
3722 cl_init_zero(pRExC_state, &accum);
3728 const regnode *nextbranch= NULL;
3731 for ( word=1 ; word <= trie->wordcount ; word++)
3733 I32 deltanext=0, minnext=0, f = 0, fake;
3734 struct regnode_charclass_class this_class;
3736 data_fake.flags = 0;
3738 data_fake.whilem_c = data->whilem_c;
3739 data_fake.last_closep = data->last_closep;
3742 data_fake.last_closep = &fake;
3743 data_fake.pos_delta = delta;
3744 if (flags & SCF_DO_STCLASS) {
3745 cl_init(pRExC_state, &this_class);
3746 data_fake.start_class = &this_class;
3747 f = SCF_DO_STCLASS_AND;
3749 if (flags & SCF_WHILEM_VISITED_POS)
3750 f |= SCF_WHILEM_VISITED_POS;
3752 if (trie->jump[word]) {
3754 nextbranch = trie_node + trie->jump[0];
3755 scan= trie_node + trie->jump[word];
3756 /* We go from the jump point to the branch that follows
3757 it. Note this means we need the vestigal unused branches
3758 even though they arent otherwise used.
3760 minnext = study_chunk(pRExC_state, &scan, minlenp,
3761 &deltanext, (regnode *)nextbranch, &data_fake,
3762 stopparen, recursed, NULL, f,depth+1);
3764 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
3765 nextbranch= regnext((regnode*)nextbranch);
3767 if (min1 > (I32)(minnext + trie->minlen))
3768 min1 = minnext + trie->minlen;
3769 if (max1 < (I32)(minnext + deltanext + trie->maxlen))
3770 max1 = minnext + deltanext + trie->maxlen;
3771 if (deltanext == I32_MAX)
3772 is_inf = is_inf_internal = 1;
3774 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3776 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3777 if ( stopmin > min + min1)
3778 stopmin = min + min1;
3779 flags &= ~SCF_DO_SUBSTR;
3781 data->flags |= SCF_SEEN_ACCEPT;
3784 if (data_fake.flags & SF_HAS_EVAL)
3785 data->flags |= SF_HAS_EVAL;
3786 data->whilem_c = data_fake.whilem_c;
3788 if (flags & SCF_DO_STCLASS)
3789 cl_or(pRExC_state, &accum, &this_class);
3792 if (flags & SCF_DO_SUBSTR) {
3793 data->pos_min += min1;
3794 data->pos_delta += max1 - min1;
3795 if (max1 != min1 || is_inf)
3796 data->longest = &(data->longest_float);
3799 delta += max1 - min1;
3800 if (flags & SCF_DO_STCLASS_OR) {
3801 cl_or(pRExC_state, data->start_class, &accum);
3803 cl_and(data->start_class, and_withp);
3804 flags &= ~SCF_DO_STCLASS;
3807 else if (flags & SCF_DO_STCLASS_AND) {
3809 cl_and(data->start_class, &accum);
3810 flags &= ~SCF_DO_STCLASS;
3813 /* Switch to OR mode: cache the old value of
3814 * data->start_class */
3816 StructCopy(data->start_class, and_withp,
3817 struct regnode_charclass_class);
3818 flags &= ~SCF_DO_STCLASS_AND;
https://perl5.git.perl.org / perl5.git / blob