5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
93 # if defined(BUGGY_MSC6)
94 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
95 # pragma optimize("a",off)
96 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
97 # pragma optimize("w",on )
98 # endif /* BUGGY_MSC6 */
102 #define STATIC static
105 typedef struct RExC_state_t {
106 U32 flags; /* are we folding, multilining? */
107 char *precomp; /* uncompiled string. */
108 REGEXP *rx_sv; /* The SV that is the regexp. */
109 regexp *rx; /* perl core regexp structure */
110 regexp_internal *rxi; /* internal data for regexp object pprivate field */
111 char *start; /* Start of input for compile */
112 char *end; /* End of input for compile */
113 char *parse; /* Input-scan pointer. */
114 I32 whilem_seen; /* number of WHILEM in this expr */
115 regnode *emit_start; /* Start of emitted-code area */
116 regnode *emit_bound; /* First regnode outside of the allocated space */
117 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
118 I32 naughty; /* How bad is this pattern? */
119 I32 sawback; /* Did we see \1, ...? */
121 I32 size; /* Code size. */
122 I32 npar; /* Capture buffer count, (OPEN). */
123 I32 cpar; /* Capture buffer count, (CLOSE). */
124 I32 nestroot; /* root parens we are in - used by accept */
128 regnode **open_parens; /* pointers to open parens */
129 regnode **close_parens; /* pointers to close parens */
130 regnode *opend; /* END node in program */
131 I32 utf8; /* whether the pattern is utf8 or not */
132 I32 orig_utf8; /* whether the pattern was originally in utf8 */
133 /* XXX use this for future optimisation of case
134 * where pattern must be upgraded to utf8. */
135 HV *charnames; /* cache of named sequences */
136 HV *paren_names; /* Paren names */
138 regnode **recurse; /* Recurse regops */
139 I32 recurse_count; /* Number of recurse regops */
141 char *starttry; /* -Dr: where regtry was called. */
142 #define RExC_starttry (pRExC_state->starttry)
145 const char *lastparse;
147 AV *paren_name_list; /* idx -> name */
148 #define RExC_lastparse (pRExC_state->lastparse)
149 #define RExC_lastnum (pRExC_state->lastnum)
150 #define RExC_paren_name_list (pRExC_state->paren_name_list)
154 #define RExC_flags (pRExC_state->flags)
155 #define RExC_precomp (pRExC_state->precomp)
156 #define RExC_rx_sv (pRExC_state->rx_sv)
157 #define RExC_rx (pRExC_state->rx)
158 #define RExC_rxi (pRExC_state->rxi)
159 #define RExC_start (pRExC_state->start)
160 #define RExC_end (pRExC_state->end)
161 #define RExC_parse (pRExC_state->parse)
162 #define RExC_whilem_seen (pRExC_state->whilem_seen)
163 #ifdef RE_TRACK_PATTERN_OFFSETS
164 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
166 #define RExC_emit (pRExC_state->emit)
167 #define RExC_emit_start (pRExC_state->emit_start)
168 #define RExC_emit_bound (pRExC_state->emit_bound)
169 #define RExC_naughty (pRExC_state->naughty)
170 #define RExC_sawback (pRExC_state->sawback)
171 #define RExC_seen (pRExC_state->seen)
172 #define RExC_size (pRExC_state->size)
173 #define RExC_npar (pRExC_state->npar)
174 #define RExC_nestroot (pRExC_state->nestroot)
175 #define RExC_extralen (pRExC_state->extralen)
176 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
177 #define RExC_seen_evals (pRExC_state->seen_evals)
178 #define RExC_utf8 (pRExC_state->utf8)
179 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
180 #define RExC_charnames (pRExC_state->charnames)
181 #define RExC_open_parens (pRExC_state->open_parens)
182 #define RExC_close_parens (pRExC_state->close_parens)
183 #define RExC_opend (pRExC_state->opend)
184 #define RExC_paren_names (pRExC_state->paren_names)
185 #define RExC_recurse (pRExC_state->recurse)
186 #define RExC_recurse_count (pRExC_state->recurse_count)
189 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
190 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
191 ((*s) == '{' && regcurly(s)))
194 #undef SPSTART /* dratted cpp namespace... */
197 * Flags to be passed up and down.
199 #define WORST 0 /* Worst case. */
200 #define HASWIDTH 0x01 /* Known to match non-null strings. */
201 #define SIMPLE 0x02 /* Simple enough to be STAR/PLUS operand. */
202 #define SPSTART 0x04 /* Starts with * or +. */
203 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
204 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
206 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
208 /* whether trie related optimizations are enabled */
209 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
210 #define TRIE_STUDY_OPT
211 #define FULL_TRIE_STUDY
217 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
218 #define PBITVAL(paren) (1 << ((paren) & 7))
219 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
220 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
221 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
224 /* About scan_data_t.
226 During optimisation we recurse through the regexp program performing
227 various inplace (keyhole style) optimisations. In addition study_chunk
228 and scan_commit populate this data structure with information about
229 what strings MUST appear in the pattern. We look for the longest
230 string that must appear for at a fixed location, and we look for the
231 longest string that may appear at a floating location. So for instance
236 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
237 strings (because they follow a .* construct). study_chunk will identify
238 both FOO and BAR as being the longest fixed and floating strings respectively.
240 The strings can be composites, for instance
244 will result in a composite fixed substring 'foo'.
246 For each string some basic information is maintained:
248 - offset or min_offset
249 This is the position the string must appear at, or not before.
250 It also implicitly (when combined with minlenp) tells us how many
251 character must match before the string we are searching.
252 Likewise when combined with minlenp and the length of the string
253 tells us how many characters must appear after the string we have
257 Only used for floating strings. This is the rightmost point that
258 the string can appear at. Ifset to I32 max it indicates that the
259 string can occur infinitely far to the right.
262 A pointer to the minimum length of the pattern that the string
263 was found inside. This is important as in the case of positive
264 lookahead or positive lookbehind we can have multiple patterns
269 The minimum length of the pattern overall is 3, the minimum length
270 of the lookahead part is 3, but the minimum length of the part that
271 will actually match is 1. So 'FOO's minimum length is 3, but the
272 minimum length for the F is 1. This is important as the minimum length
273 is used to determine offsets in front of and behind the string being
274 looked for. Since strings can be composites this is the length of the
275 pattern at the time it was commited with a scan_commit. Note that
276 the length is calculated by study_chunk, so that the minimum lengths
277 are not known until the full pattern has been compiled, thus the
278 pointer to the value.
282 In the case of lookbehind the string being searched for can be
283 offset past the start point of the final matching string.
284 If this value was just blithely removed from the min_offset it would
285 invalidate some of the calculations for how many chars must match
286 before or after (as they are derived from min_offset and minlen and
287 the length of the string being searched for).
288 When the final pattern is compiled and the data is moved from the
289 scan_data_t structure into the regexp structure the information
290 about lookbehind is factored in, with the information that would
291 have been lost precalculated in the end_shift field for the
294 The fields pos_min and pos_delta are used to store the minimum offset
295 and the delta to the maximum offset at the current point in the pattern.
299 typedef struct scan_data_t {
300 /*I32 len_min; unused */
301 /*I32 len_delta; unused */
305 I32 last_end; /* min value, <0 unless valid. */
308 SV **longest; /* Either &l_fixed, or &l_float. */
309 SV *longest_fixed; /* longest fixed string found in pattern */
310 I32 offset_fixed; /* offset where it starts */
311 I32 *minlen_fixed; /* pointer to the minlen relevent to the string */
312 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
313 SV *longest_float; /* longest floating string found in pattern */
314 I32 offset_float_min; /* earliest point in string it can appear */
315 I32 offset_float_max; /* latest point in string it can appear */
316 I32 *minlen_float; /* pointer to the minlen relevent to the string */
317 I32 lookbehind_float; /* is the position of the string modified by LB */
321 struct regnode_charclass_class *start_class;
325 * Forward declarations for pregcomp()'s friends.
328 static const scan_data_t zero_scan_data =
329 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
331 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
332 #define SF_BEFORE_SEOL 0x0001
333 #define SF_BEFORE_MEOL 0x0002
334 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
335 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
338 # define SF_FIX_SHIFT_EOL (0+2)
339 # define SF_FL_SHIFT_EOL (0+4)
341 # define SF_FIX_SHIFT_EOL (+2)
342 # define SF_FL_SHIFT_EOL (+4)
345 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
346 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
348 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
349 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
350 #define SF_IS_INF 0x0040
351 #define SF_HAS_PAR 0x0080
352 #define SF_IN_PAR 0x0100
353 #define SF_HAS_EVAL 0x0200
354 #define SCF_DO_SUBSTR 0x0400
355 #define SCF_DO_STCLASS_AND 0x0800
356 #define SCF_DO_STCLASS_OR 0x1000
357 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
358 #define SCF_WHILEM_VISITED_POS 0x2000
360 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
361 #define SCF_SEEN_ACCEPT 0x8000
363 #define UTF (RExC_utf8 != 0)
364 #define LOC ((RExC_flags & RXf_PMf_LOCALE) != 0)
365 #define FOLD ((RExC_flags & RXf_PMf_FOLD) != 0)
367 #define OOB_UNICODE 12345678
368 #define OOB_NAMEDCLASS -1
370 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
371 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
374 /* length of regex to show in messages that don't mark a position within */
375 #define RegexLengthToShowInErrorMessages 127
378 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
379 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
380 * op/pragma/warn/regcomp.
382 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
383 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
385 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
388 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
389 * arg. Show regex, up to a maximum length. If it's too long, chop and add
392 #define _FAIL(code) STMT_START { \
393 const char *ellipses = ""; \
394 IV len = RExC_end - RExC_precomp; \
397 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
398 if (len > RegexLengthToShowInErrorMessages) { \
399 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
400 len = RegexLengthToShowInErrorMessages - 10; \
406 #define FAIL(msg) _FAIL( \
407 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
408 msg, (int)len, RExC_precomp, ellipses))
410 #define FAIL2(msg,arg) _FAIL( \
411 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
412 arg, (int)len, RExC_precomp, ellipses))
415 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
417 #define Simple_vFAIL(m) STMT_START { \
418 const IV offset = RExC_parse - RExC_precomp; \
419 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
420 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
424 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
426 #define vFAIL(m) STMT_START { \
428 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
433 * Like Simple_vFAIL(), but accepts two arguments.
435 #define Simple_vFAIL2(m,a1) STMT_START { \
436 const IV offset = RExC_parse - RExC_precomp; \
437 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
438 (int)offset, RExC_precomp, RExC_precomp + offset); \
442 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
444 #define vFAIL2(m,a1) STMT_START { \
446 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
447 Simple_vFAIL2(m, a1); \
452 * Like Simple_vFAIL(), but accepts three arguments.
454 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
455 const IV offset = RExC_parse - RExC_precomp; \
456 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
457 (int)offset, RExC_precomp, RExC_precomp + offset); \
461 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
463 #define vFAIL3(m,a1,a2) STMT_START { \
465 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
466 Simple_vFAIL3(m, a1, a2); \
470 * Like Simple_vFAIL(), but accepts four arguments.
472 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
473 const IV offset = RExC_parse - RExC_precomp; \
474 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
475 (int)offset, RExC_precomp, RExC_precomp + offset); \
478 #define vWARN(loc,m) STMT_START { \
479 const IV offset = loc - RExC_precomp; \
480 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
481 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
484 #define vWARNdep(loc,m) STMT_START { \
485 const IV offset = loc - RExC_precomp; \
486 Perl_warner(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
487 "%s" REPORT_LOCATION, \
488 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
492 #define vWARN2(loc, m, a1) STMT_START { \
493 const IV offset = loc - RExC_precomp; \
494 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
495 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
498 #define vWARN3(loc, m, a1, a2) STMT_START { \
499 const IV offset = loc - RExC_precomp; \
500 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
501 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
504 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
505 const IV offset = loc - RExC_precomp; \
506 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
507 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
510 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
511 const IV offset = loc - RExC_precomp; \
512 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
513 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
517 /* Allow for side effects in s */
518 #define REGC(c,s) STMT_START { \
519 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
522 /* Macros for recording node offsets. 20001227 mjd@plover.com
523 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
524 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
525 * Element 0 holds the number n.
526 * Position is 1 indexed.
528 #ifndef RE_TRACK_PATTERN_OFFSETS
529 #define Set_Node_Offset_To_R(node,byte)
530 #define Set_Node_Offset(node,byte)
531 #define Set_Cur_Node_Offset
532 #define Set_Node_Length_To_R(node,len)
533 #define Set_Node_Length(node,len)
534 #define Set_Node_Cur_Length(node)
535 #define Node_Offset(n)
536 #define Node_Length(n)
537 #define Set_Node_Offset_Length(node,offset,len)
538 #define ProgLen(ri) ri->u.proglen
539 #define SetProgLen(ri,x) ri->u.proglen = x
541 #define ProgLen(ri) ri->u.offsets[0]
542 #define SetProgLen(ri,x) ri->u.offsets[0] = x
543 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
545 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
546 __LINE__, (int)(node), (int)(byte))); \
548 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
550 RExC_offsets[2*(node)-1] = (byte); \
555 #define Set_Node_Offset(node,byte) \
556 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
557 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
559 #define Set_Node_Length_To_R(node,len) STMT_START { \
561 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
562 __LINE__, (int)(node), (int)(len))); \
564 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
566 RExC_offsets[2*(node)] = (len); \
571 #define Set_Node_Length(node,len) \
572 Set_Node_Length_To_R((node)-RExC_emit_start, len)
573 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
574 #define Set_Node_Cur_Length(node) \
575 Set_Node_Length(node, RExC_parse - parse_start)
577 /* Get offsets and lengths */
578 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
579 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
581 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
582 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
583 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
587 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
588 #define EXPERIMENTAL_INPLACESCAN
589 #endif /*RE_TRACK_PATTERN_OFFSETS*/
591 #define DEBUG_STUDYDATA(str,data,depth) \
592 DEBUG_OPTIMISE_MORE_r(if(data){ \
593 PerlIO_printf(Perl_debug_log, \
594 "%*s" str "Pos:%"IVdf"/%"IVdf \
595 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
596 (int)(depth)*2, "", \
597 (IV)((data)->pos_min), \
598 (IV)((data)->pos_delta), \
599 (UV)((data)->flags), \
600 (IV)((data)->whilem_c), \
601 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
602 is_inf ? "INF " : "" \
604 if ((data)->last_found) \
605 PerlIO_printf(Perl_debug_log, \
606 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
607 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
608 SvPVX_const((data)->last_found), \
609 (IV)((data)->last_end), \
610 (IV)((data)->last_start_min), \
611 (IV)((data)->last_start_max), \
612 ((data)->longest && \
613 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
614 SvPVX_const((data)->longest_fixed), \
615 (IV)((data)->offset_fixed), \
616 ((data)->longest && \
617 (data)->longest==&((data)->longest_float)) ? "*" : "", \
618 SvPVX_const((data)->longest_float), \
619 (IV)((data)->offset_float_min), \
620 (IV)((data)->offset_float_max) \
622 PerlIO_printf(Perl_debug_log,"\n"); \
625 static void clear_re(pTHX_ void *r);
627 /* Mark that we cannot extend a found fixed substring at this point.
628 Update the longest found anchored substring and the longest found
629 floating substrings if needed. */
632 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
634 const STRLEN l = CHR_SVLEN(data->last_found);
635 const STRLEN old_l = CHR_SVLEN(*data->longest);
636 GET_RE_DEBUG_FLAGS_DECL;
638 PERL_ARGS_ASSERT_SCAN_COMMIT;
640 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
641 SvSetMagicSV(*data->longest, data->last_found);
642 if (*data->longest == data->longest_fixed) {
643 data->offset_fixed = l ? data->last_start_min : data->pos_min;
644 if (data->flags & SF_BEFORE_EOL)
646 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
648 data->flags &= ~SF_FIX_BEFORE_EOL;
649 data->minlen_fixed=minlenp;
650 data->lookbehind_fixed=0;
652 else { /* *data->longest == data->longest_float */
653 data->offset_float_min = l ? data->last_start_min : data->pos_min;
654 data->offset_float_max = (l
655 ? data->last_start_max
656 : data->pos_min + data->pos_delta);
657 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
658 data->offset_float_max = I32_MAX;
659 if (data->flags & SF_BEFORE_EOL)
661 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
663 data->flags &= ~SF_FL_BEFORE_EOL;
664 data->minlen_float=minlenp;
665 data->lookbehind_float=0;
668 SvCUR_set(data->last_found, 0);
670 SV * const sv = data->last_found;
671 if (SvUTF8(sv) && SvMAGICAL(sv)) {
672 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
678 data->flags &= ~SF_BEFORE_EOL;
679 DEBUG_STUDYDATA("commit: ",data,0);
682 /* Can match anything (initialization) */
684 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
686 PERL_ARGS_ASSERT_CL_ANYTHING;
688 ANYOF_CLASS_ZERO(cl);
689 ANYOF_BITMAP_SETALL(cl);
690 cl->flags = ANYOF_EOS|ANYOF_UNICODE_ALL;
692 cl->flags |= ANYOF_LOCALE;
695 /* Can match anything (initialization) */
697 S_cl_is_anything(const struct regnode_charclass_class *cl)
701 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
703 for (value = 0; value <= ANYOF_MAX; value += 2)
704 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
706 if (!(cl->flags & ANYOF_UNICODE_ALL))
708 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
713 /* Can match anything (initialization) */
715 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
717 PERL_ARGS_ASSERT_CL_INIT;
719 Zero(cl, 1, struct regnode_charclass_class);
721 cl_anything(pRExC_state, cl);
725 S_cl_init_zero(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
727 PERL_ARGS_ASSERT_CL_INIT_ZERO;
729 Zero(cl, 1, struct regnode_charclass_class);
731 cl_anything(pRExC_state, cl);
733 cl->flags |= ANYOF_LOCALE;
736 /* 'And' a given class with another one. Can create false positives */
737 /* We assume that cl is not inverted */
739 S_cl_and(struct regnode_charclass_class *cl,
740 const struct regnode_charclass_class *and_with)
742 PERL_ARGS_ASSERT_CL_AND;
744 assert(and_with->type == ANYOF);
745 if (!(and_with->flags & ANYOF_CLASS)
746 && !(cl->flags & ANYOF_CLASS)
747 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
748 && !(and_with->flags & ANYOF_FOLD)
749 && !(cl->flags & ANYOF_FOLD)) {
752 if (and_with->flags & ANYOF_INVERT)
753 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
754 cl->bitmap[i] &= ~and_with->bitmap[i];
756 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
757 cl->bitmap[i] &= and_with->bitmap[i];
758 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
759 if (!(and_with->flags & ANYOF_EOS))
760 cl->flags &= ~ANYOF_EOS;
762 if (cl->flags & ANYOF_UNICODE_ALL && and_with->flags & ANYOF_UNICODE &&
763 !(and_with->flags & ANYOF_INVERT)) {
764 cl->flags &= ~ANYOF_UNICODE_ALL;
765 cl->flags |= ANYOF_UNICODE;
766 ARG_SET(cl, ARG(and_with));
768 if (!(and_with->flags & ANYOF_UNICODE_ALL) &&
769 !(and_with->flags & ANYOF_INVERT))
770 cl->flags &= ~ANYOF_UNICODE_ALL;
771 if (!(and_with->flags & (ANYOF_UNICODE|ANYOF_UNICODE_ALL)) &&
772 !(and_with->flags & ANYOF_INVERT))
773 cl->flags &= ~ANYOF_UNICODE;
776 /* 'OR' a given class with another one. Can create false positives */
777 /* We assume that cl is not inverted */
779 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
781 PERL_ARGS_ASSERT_CL_OR;
783 if (or_with->flags & ANYOF_INVERT) {
785 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
786 * <= (B1 | !B2) | (CL1 | !CL2)
787 * which is wasteful if CL2 is small, but we ignore CL2:
788 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
789 * XXXX Can we handle case-fold? Unclear:
790 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
791 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
793 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
794 && !(or_with->flags & ANYOF_FOLD)
795 && !(cl->flags & ANYOF_FOLD) ) {
798 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
799 cl->bitmap[i] |= ~or_with->bitmap[i];
800 } /* XXXX: logic is complicated otherwise */
802 cl_anything(pRExC_state, cl);
805 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
806 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
807 && (!(or_with->flags & ANYOF_FOLD)
808 || (cl->flags & ANYOF_FOLD)) ) {
811 /* OR char bitmap and class bitmap separately */
812 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
813 cl->bitmap[i] |= or_with->bitmap[i];
814 if (or_with->flags & ANYOF_CLASS) {
815 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
816 cl->classflags[i] |= or_with->classflags[i];
817 cl->flags |= ANYOF_CLASS;
820 else { /* XXXX: logic is complicated, leave it along for a moment. */
821 cl_anything(pRExC_state, cl);
824 if (or_with->flags & ANYOF_EOS)
825 cl->flags |= ANYOF_EOS;
827 if (cl->flags & ANYOF_UNICODE && or_with->flags & ANYOF_UNICODE &&
828 ARG(cl) != ARG(or_with)) {
829 cl->flags |= ANYOF_UNICODE_ALL;
830 cl->flags &= ~ANYOF_UNICODE;
832 if (or_with->flags & ANYOF_UNICODE_ALL) {
833 cl->flags |= ANYOF_UNICODE_ALL;
834 cl->flags &= ~ANYOF_UNICODE;
838 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
839 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
840 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
841 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
846 dump_trie(trie,widecharmap,revcharmap)
847 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
848 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
850 These routines dump out a trie in a somewhat readable format.
851 The _interim_ variants are used for debugging the interim
852 tables that are used to generate the final compressed
853 representation which is what dump_trie expects.
855 Part of the reason for their existance is to provide a form
856 of documentation as to how the different representations function.
861 Dumps the final compressed table form of the trie to Perl_debug_log.
862 Used for debugging make_trie().
866 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
867 AV *revcharmap, U32 depth)
870 SV *sv=sv_newmortal();
871 int colwidth= widecharmap ? 6 : 4;
872 GET_RE_DEBUG_FLAGS_DECL;
874 PERL_ARGS_ASSERT_DUMP_TRIE;
876 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
877 (int)depth * 2 + 2,"",
878 "Match","Base","Ofs" );
880 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
881 SV ** const tmp = av_fetch( revcharmap, state, 0);
883 PerlIO_printf( Perl_debug_log, "%*s",
885 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
886 PL_colors[0], PL_colors[1],
887 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
888 PERL_PV_ESCAPE_FIRSTCHAR
893 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
894 (int)depth * 2 + 2,"");
896 for( state = 0 ; state < trie->uniquecharcount ; state++ )
897 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
898 PerlIO_printf( Perl_debug_log, "\n");
900 for( state = 1 ; state < trie->statecount ; state++ ) {
901 const U32 base = trie->states[ state ].trans.base;
903 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
905 if ( trie->states[ state ].wordnum ) {
906 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
908 PerlIO_printf( Perl_debug_log, "%6s", "" );
911 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
916 while( ( base + ofs < trie->uniquecharcount ) ||
917 ( base + ofs - trie->uniquecharcount < trie->lasttrans
918 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
921 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
923 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
924 if ( ( base + ofs >= trie->uniquecharcount ) &&
925 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
926 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
928 PerlIO_printf( Perl_debug_log, "%*"UVXf,
930 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
932 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
936 PerlIO_printf( Perl_debug_log, "]");
939 PerlIO_printf( Perl_debug_log, "\n" );
943 Dumps a fully constructed but uncompressed trie in list form.
944 List tries normally only are used for construction when the number of
945 possible chars (trie->uniquecharcount) is very high.
946 Used for debugging make_trie().
949 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
950 HV *widecharmap, AV *revcharmap, U32 next_alloc,
954 SV *sv=sv_newmortal();
955 int colwidth= widecharmap ? 6 : 4;
956 GET_RE_DEBUG_FLAGS_DECL;
958 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
960 /* print out the table precompression. */
961 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
962 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
963 "------:-----+-----------------\n" );
965 for( state=1 ; state < next_alloc ; state ++ ) {
968 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
969 (int)depth * 2 + 2,"", (UV)state );
970 if ( ! trie->states[ state ].wordnum ) {
971 PerlIO_printf( Perl_debug_log, "%5s| ","");
973 PerlIO_printf( Perl_debug_log, "W%4x| ",
974 trie->states[ state ].wordnum
977 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
978 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
980 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
982 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
983 PL_colors[0], PL_colors[1],
984 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
985 PERL_PV_ESCAPE_FIRSTCHAR
987 TRIE_LIST_ITEM(state,charid).forid,
988 (UV)TRIE_LIST_ITEM(state,charid).newstate
991 PerlIO_printf(Perl_debug_log, "\n%*s| ",
992 (int)((depth * 2) + 14), "");
995 PerlIO_printf( Perl_debug_log, "\n");
1000 Dumps a fully constructed but uncompressed trie in table form.
1001 This is the normal DFA style state transition table, with a few
1002 twists to facilitate compression later.
1003 Used for debugging make_trie().
1006 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1007 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1012 SV *sv=sv_newmortal();
1013 int colwidth= widecharmap ? 6 : 4;
1014 GET_RE_DEBUG_FLAGS_DECL;
1016 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1019 print out the table precompression so that we can do a visual check
1020 that they are identical.
1023 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1025 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1026 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1028 PerlIO_printf( Perl_debug_log, "%*s",
1030 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1031 PL_colors[0], PL_colors[1],
1032 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1033 PERL_PV_ESCAPE_FIRSTCHAR
1039 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1041 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1042 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1045 PerlIO_printf( Perl_debug_log, "\n" );
1047 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1049 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1050 (int)depth * 2 + 2,"",
1051 (UV)TRIE_NODENUM( state ) );
1053 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1054 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1056 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1058 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1060 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1061 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1063 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1064 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1071 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1072 startbranch: the first branch in the whole branch sequence
1073 first : start branch of sequence of branch-exact nodes.
1074 May be the same as startbranch
1075 last : Thing following the last branch.
1076 May be the same as tail.
1077 tail : item following the branch sequence
1078 count : words in the sequence
1079 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1080 depth : indent depth
1082 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1084 A trie is an N'ary tree where the branches are determined by digital
1085 decomposition of the key. IE, at the root node you look up the 1st character and
1086 follow that branch repeat until you find the end of the branches. Nodes can be
1087 marked as "accepting" meaning they represent a complete word. Eg:
1091 would convert into the following structure. Numbers represent states, letters
1092 following numbers represent valid transitions on the letter from that state, if
1093 the number is in square brackets it represents an accepting state, otherwise it
1094 will be in parenthesis.
1096 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1100 (1) +-i->(6)-+-s->[7]
1102 +-s->(3)-+-h->(4)-+-e->[5]
1104 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1106 This shows that when matching against the string 'hers' we will begin at state 1
1107 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1108 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1109 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1110 single traverse. We store a mapping from accepting to state to which word was
1111 matched, and then when we have multiple possibilities we try to complete the
1112 rest of the regex in the order in which they occured in the alternation.
1114 The only prior NFA like behaviour that would be changed by the TRIE support is
1115 the silent ignoring of duplicate alternations which are of the form:
1117 / (DUPE|DUPE) X? (?{ ... }) Y /x
1119 Thus EVAL blocks follwing a trie may be called a different number of times with
1120 and without the optimisation. With the optimisations dupes will be silently
1121 ignored. This inconsistant behaviour of EVAL type nodes is well established as
1122 the following demonstrates:
1124 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1126 which prints out 'word' three times, but
1128 'words'=~/(word|word|word)(?{ print $1 })S/
1130 which doesnt print it out at all. This is due to other optimisations kicking in.
1132 Example of what happens on a structural level:
1134 The regexp /(ac|ad|ab)+/ will produce the folowing debug output:
1136 1: CURLYM[1] {1,32767}(18)
1147 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1148 and should turn into:
1150 1: CURLYM[1] {1,32767}(18)
1152 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1160 Cases where tail != last would be like /(?foo|bar)baz/:
1170 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1171 and would end up looking like:
1174 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1181 d = uvuni_to_utf8_flags(d, uv, 0);
1183 is the recommended Unicode-aware way of saying
1188 #define TRIE_STORE_REVCHAR \
1191 SV *zlopp = newSV(2); \
1192 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1193 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, uvc & 0xFF); \
1194 SvCUR_set(zlopp, kapow - flrbbbbb); \
1197 av_push(revcharmap, zlopp); \
1199 char ooooff = (char)uvc; \
1200 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1204 #define TRIE_READ_CHAR STMT_START { \
1208 if ( foldlen > 0 ) { \
1209 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1214 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1215 uvc = to_uni_fold( uvc, foldbuf, &foldlen ); \
1216 foldlen -= UNISKIP( uvc ); \
1217 scan = foldbuf + UNISKIP( uvc ); \
1220 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1230 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1231 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1232 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1233 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1235 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1236 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1237 TRIE_LIST_CUR( state )++; \
1240 #define TRIE_LIST_NEW(state) STMT_START { \
1241 Newxz( trie->states[ state ].trans.list, \
1242 4, reg_trie_trans_le ); \
1243 TRIE_LIST_CUR( state ) = 1; \
1244 TRIE_LIST_LEN( state ) = 4; \
1247 #define TRIE_HANDLE_WORD(state) STMT_START { \
1248 U16 dupe= trie->states[ state ].wordnum; \
1249 regnode * const noper_next = regnext( noper ); \
1251 if (trie->wordlen) \
1252 trie->wordlen[ curword ] = wordlen; \
1254 /* store the word for dumping */ \
1256 if (OP(noper) != NOTHING) \
1257 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1259 tmp = newSVpvn_utf8( "", 0, UTF ); \
1260 av_push( trie_words, tmp ); \
1265 if ( noper_next < tail ) { \
1267 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1268 trie->jump[curword] = (U16)(noper_next - convert); \
1270 jumper = noper_next; \
1272 nextbranch= regnext(cur); \
1276 /* So it's a dupe. This means we need to maintain a */\
1277 /* linked-list from the first to the next. */\
1278 /* we only allocate the nextword buffer when there */\
1279 /* a dupe, so first time we have to do the allocation */\
1280 if (!trie->nextword) \
1281 trie->nextword = (U16 *) \
1282 PerlMemShared_calloc( word_count + 1, sizeof(U16)); \
1283 while ( trie->nextword[dupe] ) \
1284 dupe= trie->nextword[dupe]; \
1285 trie->nextword[dupe]= curword; \
1287 /* we haven't inserted this word yet. */ \
1288 trie->states[ state ].wordnum = curword; \
1293 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1294 ( ( base + charid >= ucharcount \
1295 && base + charid < ubound \
1296 && state == trie->trans[ base - ucharcount + charid ].check \
1297 && trie->trans[ base - ucharcount + charid ].next ) \
1298 ? trie->trans[ base - ucharcount + charid ].next \
1299 : ( state==1 ? special : 0 ) \
1303 #define MADE_JUMP_TRIE 2
1304 #define MADE_EXACT_TRIE 4
1307 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1310 /* first pass, loop through and scan words */
1311 reg_trie_data *trie;
1312 HV *widecharmap = NULL;
1313 AV *revcharmap = newAV();
1315 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1320 regnode *jumper = NULL;
1321 regnode *nextbranch = NULL;
1322 regnode *convert = NULL;
1323 /* we just use folder as a flag in utf8 */
1324 const U8 * const folder = ( flags == EXACTF
1326 : ( flags == EXACTFL
1333 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1334 AV *trie_words = NULL;
1335 /* along with revcharmap, this only used during construction but both are
1336 * useful during debugging so we store them in the struct when debugging.
1339 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1340 STRLEN trie_charcount=0;
1342 SV *re_trie_maxbuff;
1343 GET_RE_DEBUG_FLAGS_DECL;
1345 PERL_ARGS_ASSERT_MAKE_TRIE;
1347 PERL_UNUSED_ARG(depth);
1350 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1352 trie->startstate = 1;
1353 trie->wordcount = word_count;
1354 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1355 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1356 if (!(UTF && folder))
1357 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1359 trie_words = newAV();
1362 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1363 if (!SvIOK(re_trie_maxbuff)) {
1364 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1367 PerlIO_printf( Perl_debug_log,
1368 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1369 (int)depth * 2 + 2, "",
1370 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1371 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1375 /* Find the node we are going to overwrite */
1376 if ( first == startbranch && OP( last ) != BRANCH ) {
1377 /* whole branch chain */
1380 /* branch sub-chain */
1381 convert = NEXTOPER( first );
1384 /* -- First loop and Setup --
1386 We first traverse the branches and scan each word to determine if it
1387 contains widechars, and how many unique chars there are, this is
1388 important as we have to build a table with at least as many columns as we
1391 We use an array of integers to represent the character codes 0..255
1392 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1393 native representation of the character value as the key and IV's for the
1396 *TODO* If we keep track of how many times each character is used we can
1397 remap the columns so that the table compression later on is more
1398 efficient in terms of memory by ensuring most common value is in the
1399 middle and the least common are on the outside. IMO this would be better
1400 than a most to least common mapping as theres a decent chance the most
1401 common letter will share a node with the least common, meaning the node
1402 will not be compressable. With a middle is most common approach the worst
1403 case is when we have the least common nodes twice.
1407 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1408 regnode * const noper = NEXTOPER( cur );
1409 const U8 *uc = (U8*)STRING( noper );
1410 const U8 * const e = uc + STR_LEN( noper );
1412 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1413 const U8 *scan = (U8*)NULL;
1414 U32 wordlen = 0; /* required init */
1416 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1418 if (OP(noper) == NOTHING) {
1422 if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
1423 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1424 regardless of encoding */
1426 for ( ; uc < e ; uc += len ) {
1427 TRIE_CHARCOUNT(trie)++;
1431 if ( !trie->charmap[ uvc ] ) {
1432 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1434 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
1438 /* store the codepoint in the bitmap, and if its ascii
1439 also store its folded equivelent. */
1440 TRIE_BITMAP_SET(trie,uvc);
1442 /* store the folded codepoint */
1443 if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);
1446 /* store first byte of utf8 representation of
1447 codepoints in the 127 < uvc < 256 range */
1448 if (127 < uvc && uvc < 192) {
1449 TRIE_BITMAP_SET(trie,194);
1450 } else if (191 < uvc ) {
1451 TRIE_BITMAP_SET(trie,195);
1452 /* && uvc < 256 -- we know uvc is < 256 already */
1455 set_bit = 0; /* We've done our bit :-) */
1460 widecharmap = newHV();
1462 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1465 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1467 if ( !SvTRUE( *svpp ) ) {
1468 sv_setiv( *svpp, ++trie->uniquecharcount );
1473 if( cur == first ) {
1476 } else if (chars < trie->minlen) {
1478 } else if (chars > trie->maxlen) {
1482 } /* end first pass */
1483 DEBUG_TRIE_COMPILE_r(
1484 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1485 (int)depth * 2 + 2,"",
1486 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1487 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1488 (int)trie->minlen, (int)trie->maxlen )
1490 trie->wordlen = (U32 *) PerlMemShared_calloc( word_count, sizeof(U32) );
1493 We now know what we are dealing with in terms of unique chars and
1494 string sizes so we can calculate how much memory a naive
1495 representation using a flat table will take. If it's over a reasonable
1496 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1497 conservative but potentially much slower representation using an array
1500 At the end we convert both representations into the same compressed
1501 form that will be used in regexec.c for matching with. The latter
1502 is a form that cannot be used to construct with but has memory
1503 properties similar to the list form and access properties similar
1504 to the table form making it both suitable for fast searches and
1505 small enough that its feasable to store for the duration of a program.
1507 See the comment in the code where the compressed table is produced
1508 inplace from the flat tabe representation for an explanation of how
1509 the compression works.
1514 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1516 Second Pass -- Array Of Lists Representation
1518 Each state will be represented by a list of charid:state records
1519 (reg_trie_trans_le) the first such element holds the CUR and LEN
1520 points of the allocated array. (See defines above).
1522 We build the initial structure using the lists, and then convert
1523 it into the compressed table form which allows faster lookups
1524 (but cant be modified once converted).
1527 STRLEN transcount = 1;
1529 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1530 "%*sCompiling trie using list compiler\n",
1531 (int)depth * 2 + 2, ""));
1533 trie->states = (reg_trie_state *)
1534 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1535 sizeof(reg_trie_state) );
1539 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1541 regnode * const noper = NEXTOPER( cur );
1542 U8 *uc = (U8*)STRING( noper );
1543 const U8 * const e = uc + STR_LEN( noper );
1544 U32 state = 1; /* required init */
1545 U16 charid = 0; /* sanity init */
1546 U8 *scan = (U8*)NULL; /* sanity init */
1547 STRLEN foldlen = 0; /* required init */
1548 U32 wordlen = 0; /* required init */
1549 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1551 if (OP(noper) != NOTHING) {
1552 for ( ; uc < e ; uc += len ) {
1557 charid = trie->charmap[ uvc ];
1559 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1563 charid=(U16)SvIV( *svpp );
1566 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1573 if ( !trie->states[ state ].trans.list ) {
1574 TRIE_LIST_NEW( state );
1576 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1577 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1578 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1583 newstate = next_alloc++;
1584 TRIE_LIST_PUSH( state, charid, newstate );
1589 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1593 TRIE_HANDLE_WORD(state);
1595 } /* end second pass */
1597 /* next alloc is the NEXT state to be allocated */
1598 trie->statecount = next_alloc;
1599 trie->states = (reg_trie_state *)
1600 PerlMemShared_realloc( trie->states,
1602 * sizeof(reg_trie_state) );
1604 /* and now dump it out before we compress it */
1605 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1606 revcharmap, next_alloc,
1610 trie->trans = (reg_trie_trans *)
1611 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1618 for( state=1 ; state < next_alloc ; state ++ ) {
1622 DEBUG_TRIE_COMPILE_MORE_r(
1623 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1627 if (trie->states[state].trans.list) {
1628 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1632 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1633 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1634 if ( forid < minid ) {
1636 } else if ( forid > maxid ) {
1640 if ( transcount < tp + maxid - minid + 1) {
1642 trie->trans = (reg_trie_trans *)
1643 PerlMemShared_realloc( trie->trans,
1645 * sizeof(reg_trie_trans) );
1646 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1648 base = trie->uniquecharcount + tp - minid;
1649 if ( maxid == minid ) {
1651 for ( ; zp < tp ; zp++ ) {
1652 if ( ! trie->trans[ zp ].next ) {
1653 base = trie->uniquecharcount + zp - minid;
1654 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1655 trie->trans[ zp ].check = state;
1661 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1662 trie->trans[ tp ].check = state;
1667 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1668 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1669 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1670 trie->trans[ tid ].check = state;
1672 tp += ( maxid - minid + 1 );
1674 Safefree(trie->states[ state ].trans.list);
1677 DEBUG_TRIE_COMPILE_MORE_r(
1678 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1681 trie->states[ state ].trans.base=base;
1683 trie->lasttrans = tp + 1;
1687 Second Pass -- Flat Table Representation.
1689 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1690 We know that we will need Charcount+1 trans at most to store the data
1691 (one row per char at worst case) So we preallocate both structures
1692 assuming worst case.
1694 We then construct the trie using only the .next slots of the entry
1697 We use the .check field of the first entry of the node temporarily to
1698 make compression both faster and easier by keeping track of how many non
1699 zero fields are in the node.
1701 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1704 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1705 number representing the first entry of the node, and state as a
1706 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1707 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1708 are 2 entrys per node. eg:
1716 The table is internally in the right hand, idx form. However as we also
1717 have to deal with the states array which is indexed by nodenum we have to
1718 use TRIE_NODENUM() to convert.
1721 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1722 "%*sCompiling trie using table compiler\n",
1723 (int)depth * 2 + 2, ""));
1725 trie->trans = (reg_trie_trans *)
1726 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1727 * trie->uniquecharcount + 1,
1728 sizeof(reg_trie_trans) );
1729 trie->states = (reg_trie_state *)
1730 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1731 sizeof(reg_trie_state) );
1732 next_alloc = trie->uniquecharcount + 1;
1735 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1737 regnode * const noper = NEXTOPER( cur );
1738 const U8 *uc = (U8*)STRING( noper );
1739 const U8 * const e = uc + STR_LEN( noper );
1741 U32 state = 1; /* required init */
1743 U16 charid = 0; /* sanity init */
1744 U32 accept_state = 0; /* sanity init */
1745 U8 *scan = (U8*)NULL; /* sanity init */
1747 STRLEN foldlen = 0; /* required init */
1748 U32 wordlen = 0; /* required init */
1749 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1751 if ( OP(noper) != NOTHING ) {
1752 for ( ; uc < e ; uc += len ) {
1757 charid = trie->charmap[ uvc ];
1759 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1760 charid = svpp ? (U16)SvIV(*svpp) : 0;
1764 if ( !trie->trans[ state + charid ].next ) {
1765 trie->trans[ state + charid ].next = next_alloc;
1766 trie->trans[ state ].check++;
1767 next_alloc += trie->uniquecharcount;
1769 state = trie->trans[ state + charid ].next;
1771 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1773 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1776 accept_state = TRIE_NODENUM( state );
1777 TRIE_HANDLE_WORD(accept_state);
1779 } /* end second pass */
1781 /* and now dump it out before we compress it */
1782 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1784 next_alloc, depth+1));
1788 * Inplace compress the table.*
1790 For sparse data sets the table constructed by the trie algorithm will
1791 be mostly 0/FAIL transitions or to put it another way mostly empty.
1792 (Note that leaf nodes will not contain any transitions.)
1794 This algorithm compresses the tables by eliminating most such
1795 transitions, at the cost of a modest bit of extra work during lookup:
1797 - Each states[] entry contains a .base field which indicates the
1798 index in the state[] array wheres its transition data is stored.
1800 - If .base is 0 there are no valid transitions from that node.
1802 - If .base is nonzero then charid is added to it to find an entry in
1805 -If trans[states[state].base+charid].check!=state then the
1806 transition is taken to be a 0/Fail transition. Thus if there are fail
1807 transitions at the front of the node then the .base offset will point
1808 somewhere inside the previous nodes data (or maybe even into a node
1809 even earlier), but the .check field determines if the transition is
1813 The following process inplace converts the table to the compressed
1814 table: We first do not compress the root node 1,and mark its all its
1815 .check pointers as 1 and set its .base pointer as 1 as well. This
1816 allows to do a DFA construction from the compressed table later, and
1817 ensures that any .base pointers we calculate later are greater than
1820 - We set 'pos' to indicate the first entry of the second node.
1822 - We then iterate over the columns of the node, finding the first and
1823 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
1824 and set the .check pointers accordingly, and advance pos
1825 appropriately and repreat for the next node. Note that when we copy
1826 the next pointers we have to convert them from the original
1827 NODEIDX form to NODENUM form as the former is not valid post
1830 - If a node has no transitions used we mark its base as 0 and do not
1831 advance the pos pointer.
1833 - If a node only has one transition we use a second pointer into the
1834 structure to fill in allocated fail transitions from other states.
1835 This pointer is independent of the main pointer and scans forward
1836 looking for null transitions that are allocated to a state. When it
1837 finds one it writes the single transition into the "hole". If the
1838 pointer doesnt find one the single transition is appended as normal.
1840 - Once compressed we can Renew/realloc the structures to release the
1843 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
1844 specifically Fig 3.47 and the associated pseudocode.
1848 const U32 laststate = TRIE_NODENUM( next_alloc );
1851 trie->statecount = laststate;
1853 for ( state = 1 ; state < laststate ; state++ ) {
1855 const U32 stateidx = TRIE_NODEIDX( state );
1856 const U32 o_used = trie->trans[ stateidx ].check;
1857 U32 used = trie->trans[ stateidx ].check;
1858 trie->trans[ stateidx ].check = 0;
1860 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
1861 if ( flag || trie->trans[ stateidx + charid ].next ) {
1862 if ( trie->trans[ stateidx + charid ].next ) {
1864 for ( ; zp < pos ; zp++ ) {
1865 if ( ! trie->trans[ zp ].next ) {
1869 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
1870 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1871 trie->trans[ zp ].check = state;
1872 if ( ++zp > pos ) pos = zp;
1879 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
1881 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1882 trie->trans[ pos ].check = state;
1887 trie->lasttrans = pos + 1;
1888 trie->states = (reg_trie_state *)
1889 PerlMemShared_realloc( trie->states, laststate
1890 * sizeof(reg_trie_state) );
1891 DEBUG_TRIE_COMPILE_MORE_r(
1892 PerlIO_printf( Perl_debug_log,
1893 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
1894 (int)depth * 2 + 2,"",
1895 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
1898 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
1901 } /* end table compress */
1903 DEBUG_TRIE_COMPILE_MORE_r(
1904 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
1905 (int)depth * 2 + 2, "",
1906 (UV)trie->statecount,
1907 (UV)trie->lasttrans)
1909 /* resize the trans array to remove unused space */
1910 trie->trans = (reg_trie_trans *)
1911 PerlMemShared_realloc( trie->trans, trie->lasttrans
1912 * sizeof(reg_trie_trans) );
1914 /* and now dump out the compressed format */
1915 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
1917 { /* Modify the program and insert the new TRIE node*/
1918 U8 nodetype =(U8)(flags & 0xFF);
1922 regnode *optimize = NULL;
1923 #ifdef RE_TRACK_PATTERN_OFFSETS
1926 U32 mjd_nodelen = 0;
1927 #endif /* RE_TRACK_PATTERN_OFFSETS */
1928 #endif /* DEBUGGING */
1930 This means we convert either the first branch or the first Exact,
1931 depending on whether the thing following (in 'last') is a branch
1932 or not and whther first is the startbranch (ie is it a sub part of
1933 the alternation or is it the whole thing.)
1934 Assuming its a sub part we conver the EXACT otherwise we convert
1935 the whole branch sequence, including the first.
1937 /* Find the node we are going to overwrite */
1938 if ( first != startbranch || OP( last ) == BRANCH ) {
1939 /* branch sub-chain */
1940 NEXT_OFF( first ) = (U16)(last - first);
1941 #ifdef RE_TRACK_PATTERN_OFFSETS
1943 mjd_offset= Node_Offset((convert));
1944 mjd_nodelen= Node_Length((convert));
1947 /* whole branch chain */
1949 #ifdef RE_TRACK_PATTERN_OFFSETS
1952 const regnode *nop = NEXTOPER( convert );
1953 mjd_offset= Node_Offset((nop));
1954 mjd_nodelen= Node_Length((nop));
1958 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
1959 (int)depth * 2 + 2, "",
1960 (UV)mjd_offset, (UV)mjd_nodelen)
1963 /* But first we check to see if there is a common prefix we can
1964 split out as an EXACT and put in front of the TRIE node. */
1965 trie->startstate= 1;
1966 if ( trie->bitmap && !widecharmap && !trie->jump ) {
1968 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
1972 const U32 base = trie->states[ state ].trans.base;
1974 if ( trie->states[state].wordnum )
1977 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1978 if ( ( base + ofs >= trie->uniquecharcount ) &&
1979 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1980 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1982 if ( ++count > 1 ) {
1983 SV **tmp = av_fetch( revcharmap, ofs, 0);
1984 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
1985 if ( state == 1 ) break;
1987 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
1989 PerlIO_printf(Perl_debug_log,
1990 "%*sNew Start State=%"UVuf" Class: [",
1991 (int)depth * 2 + 2, "",
1994 SV ** const tmp = av_fetch( revcharmap, idx, 0);
1995 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
1997 TRIE_BITMAP_SET(trie,*ch);
1999 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2001 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2005 TRIE_BITMAP_SET(trie,*ch);
2007 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2008 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2014 SV **tmp = av_fetch( revcharmap, idx, 0);
2016 char *ch = SvPV( *tmp, len );
2018 SV *sv=sv_newmortal();
2019 PerlIO_printf( Perl_debug_log,
2020 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2021 (int)depth * 2 + 2, "",
2023 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2024 PL_colors[0], PL_colors[1],
2025 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2026 PERL_PV_ESCAPE_FIRSTCHAR
2031 OP( convert ) = nodetype;
2032 str=STRING(convert);
2035 STR_LEN(convert) += len;
2041 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2047 regnode *n = convert+NODE_SZ_STR(convert);
2048 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2049 trie->startstate = state;
2050 trie->minlen -= (state - 1);
2051 trie->maxlen -= (state - 1);
2053 /* At least the UNICOS C compiler choked on this
2054 * being argument to DEBUG_r(), so let's just have
2057 #ifdef PERL_EXT_RE_BUILD
2063 regnode *fix = convert;
2064 U32 word = trie->wordcount;
2066 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2067 while( ++fix < n ) {
2068 Set_Node_Offset_Length(fix, 0, 0);
2071 SV ** const tmp = av_fetch( trie_words, word, 0 );
2073 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2074 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2076 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2084 NEXT_OFF(convert) = (U16)(tail - convert);
2085 DEBUG_r(optimize= n);
2091 if ( trie->maxlen ) {
2092 NEXT_OFF( convert ) = (U16)(tail - convert);
2093 ARG_SET( convert, data_slot );
2094 /* Store the offset to the first unabsorbed branch in
2095 jump[0], which is otherwise unused by the jump logic.
2096 We use this when dumping a trie and during optimisation. */
2098 trie->jump[0] = (U16)(nextbranch - convert);
2101 if ( !trie->states[trie->startstate].wordnum && trie->bitmap &&
2102 ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2104 OP( convert ) = TRIEC;
2105 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2106 PerlMemShared_free(trie->bitmap);
2109 OP( convert ) = TRIE;
2111 /* store the type in the flags */
2112 convert->flags = nodetype;
2116 + regarglen[ OP( convert ) ];
2118 /* XXX We really should free up the resource in trie now,
2119 as we won't use them - (which resources?) dmq */
2121 /* needed for dumping*/
2122 DEBUG_r(if (optimize) {
2123 regnode *opt = convert;
2125 while ( ++opt < optimize) {
2126 Set_Node_Offset_Length(opt,0,0);
2129 Try to clean up some of the debris left after the
2132 while( optimize < jumper ) {
2133 mjd_nodelen += Node_Length((optimize));
2134 OP( optimize ) = OPTIMIZED;
2135 Set_Node_Offset_Length(optimize,0,0);
2138 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2140 } /* end node insert */
2141 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2143 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2144 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2146 SvREFCNT_dec(revcharmap);
2150 : trie->startstate>1
2156 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2158 /* The Trie is constructed and compressed now so we can build a fail array now if its needed
2160 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2161 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2164 We find the fail state for each state in the trie, this state is the longest proper
2165 suffix of the current states 'word' that is also a proper prefix of another word in our
2166 trie. State 1 represents the word '' and is the thus the default fail state. This allows
2167 the DFA not to have to restart after its tried and failed a word at a given point, it
2168 simply continues as though it had been matching the other word in the first place.
2170 'abcdgu'=~/abcdefg|cdgu/
2171 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2172 fail, which would bring use to the state representing 'd' in the second word where we would
2173 try 'g' and succeed, prodceding to match 'cdgu'.
2175 /* add a fail transition */
2176 const U32 trie_offset = ARG(source);
2177 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2179 const U32 ucharcount = trie->uniquecharcount;
2180 const U32 numstates = trie->statecount;
2181 const U32 ubound = trie->lasttrans + ucharcount;
2185 U32 base = trie->states[ 1 ].trans.base;
2188 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2189 GET_RE_DEBUG_FLAGS_DECL;
2191 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2193 PERL_UNUSED_ARG(depth);
2197 ARG_SET( stclass, data_slot );
2198 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2199 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2200 aho->trie=trie_offset;
2201 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2202 Copy( trie->states, aho->states, numstates, reg_trie_state );
2203 Newxz( q, numstates, U32);
2204 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2207 /* initialize fail[0..1] to be 1 so that we always have
2208 a valid final fail state */
2209 fail[ 0 ] = fail[ 1 ] = 1;
2211 for ( charid = 0; charid < ucharcount ; charid++ ) {
2212 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2214 q[ q_write ] = newstate;
2215 /* set to point at the root */
2216 fail[ q[ q_write++ ] ]=1;
2219 while ( q_read < q_write) {
2220 const U32 cur = q[ q_read++ % numstates ];
2221 base = trie->states[ cur ].trans.base;
2223 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2224 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2226 U32 fail_state = cur;
2229 fail_state = fail[ fail_state ];
2230 fail_base = aho->states[ fail_state ].trans.base;
2231 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2233 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2234 fail[ ch_state ] = fail_state;
2235 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2237 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2239 q[ q_write++ % numstates] = ch_state;
2243 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2244 when we fail in state 1, this allows us to use the
2245 charclass scan to find a valid start char. This is based on the principle
2246 that theres a good chance the string being searched contains lots of stuff
2247 that cant be a start char.
2249 fail[ 0 ] = fail[ 1 ] = 0;
2250 DEBUG_TRIE_COMPILE_r({
2251 PerlIO_printf(Perl_debug_log,
2252 "%*sStclass Failtable (%"UVuf" states): 0",
2253 (int)(depth * 2), "", (UV)numstates
2255 for( q_read=1; q_read<numstates; q_read++ ) {
2256 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2258 PerlIO_printf(Perl_debug_log, "\n");
2261 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2266 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2267 * These need to be revisited when a newer toolchain becomes available.
2269 #if defined(__sparc64__) && defined(__GNUC__)
2270 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2271 # undef SPARC64_GCC_WORKAROUND
2272 # define SPARC64_GCC_WORKAROUND 1
2276 #define DEBUG_PEEP(str,scan,depth) \
2277 DEBUG_OPTIMISE_r({if (scan){ \
2278 SV * const mysv=sv_newmortal(); \
2279 regnode *Next = regnext(scan); \
2280 regprop(RExC_rx, mysv, scan); \
2281 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2282 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2283 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2290 #define JOIN_EXACT(scan,min,flags) \
2291 if (PL_regkind[OP(scan)] == EXACT) \
2292 join_exact(pRExC_state,(scan),(min),(flags),NULL,depth+1)
2295 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, I32 *min, U32 flags,regnode *val, U32 depth) {
2296 /* Merge several consecutive EXACTish nodes into one. */
2297 regnode *n = regnext(scan);
2299 regnode *next = scan + NODE_SZ_STR(scan);
2303 regnode *stop = scan;
2304 GET_RE_DEBUG_FLAGS_DECL;
2306 PERL_UNUSED_ARG(depth);
2309 PERL_ARGS_ASSERT_JOIN_EXACT;
2310 #ifndef EXPERIMENTAL_INPLACESCAN
2311 PERL_UNUSED_ARG(flags);
2312 PERL_UNUSED_ARG(val);
2314 DEBUG_PEEP("join",scan,depth);
2316 /* Skip NOTHING, merge EXACT*. */
2318 ( PL_regkind[OP(n)] == NOTHING ||
2319 (stringok && (OP(n) == OP(scan))))
2321 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX) {
2323 if (OP(n) == TAIL || n > next)
2325 if (PL_regkind[OP(n)] == NOTHING) {
2326 DEBUG_PEEP("skip:",n,depth);
2327 NEXT_OFF(scan) += NEXT_OFF(n);
2328 next = n + NODE_STEP_REGNODE;
2335 else if (stringok) {
2336 const unsigned int oldl = STR_LEN(scan);
2337 regnode * const nnext = regnext(n);
2339 DEBUG_PEEP("merg",n,depth);
2342 if (oldl + STR_LEN(n) > U8_MAX)
2344 NEXT_OFF(scan) += NEXT_OFF(n);
2345 STR_LEN(scan) += STR_LEN(n);
2346 next = n + NODE_SZ_STR(n);
2347 /* Now we can overwrite *n : */
2348 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2356 #ifdef EXPERIMENTAL_INPLACESCAN
2357 if (flags && !NEXT_OFF(n)) {
2358 DEBUG_PEEP("atch", val, depth);
2359 if (reg_off_by_arg[OP(n)]) {
2360 ARG_SET(n, val - n);
2363 NEXT_OFF(n) = val - n;
2370 if (UTF && ( OP(scan) == EXACTF ) && ( STR_LEN(scan) >= 6 ) ) {
2372 Two problematic code points in Unicode casefolding of EXACT nodes:
2374 U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2375 U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2381 U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2382 U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2384 This means that in case-insensitive matching (or "loose matching",
2385 as Unicode calls it), an EXACTF of length six (the UTF-8 encoded byte
2386 length of the above casefolded versions) can match a target string
2387 of length two (the byte length of UTF-8 encoded U+0390 or U+03B0).
2388 This would rather mess up the minimum length computation.
2390 What we'll do is to look for the tail four bytes, and then peek
2391 at the preceding two bytes to see whether we need to decrease
2392 the minimum length by four (six minus two).
2394 Thanks to the design of UTF-8, there cannot be false matches:
2395 A sequence of valid UTF-8 bytes cannot be a subsequence of
2396 another valid sequence of UTF-8 bytes.
2399 char * const s0 = STRING(scan), *s, *t;
2400 char * const s1 = s0 + STR_LEN(scan) - 1;
2401 char * const s2 = s1 - 4;
2402 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2403 const char t0[] = "\xaf\x49\xaf\x42";
2405 const char t0[] = "\xcc\x88\xcc\x81";
2407 const char * const t1 = t0 + 3;
2410 s < s2 && (t = ninstr(s, s1, t0, t1));
2413 if (((U8)t[-1] == 0x68 && (U8)t[-2] == 0xB4) ||
2414 ((U8)t[-1] == 0x46 && (U8)t[-2] == 0xB5))
2416 if (((U8)t[-1] == 0xB9 && (U8)t[-2] == 0xCE) ||
2417 ((U8)t[-1] == 0x85 && (U8)t[-2] == 0xCF))
2425 n = scan + NODE_SZ_STR(scan);
2427 if (PL_regkind[OP(n)] != NOTHING || OP(n) == NOTHING) {
2434 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2438 /* REx optimizer. Converts nodes into quickier variants "in place".
2439 Finds fixed substrings. */
2441 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2442 to the position after last scanned or to NULL. */
2444 #define INIT_AND_WITHP \
2445 assert(!and_withp); \
2446 Newx(and_withp,1,struct regnode_charclass_class); \
2447 SAVEFREEPV(and_withp)
2449 /* this is a chain of data about sub patterns we are processing that
2450 need to be handled seperately/specially in study_chunk. Its so
2451 we can simulate recursion without losing state. */
2453 typedef struct scan_frame {
2454 regnode *last; /* last node to process in this frame */
2455 regnode *next; /* next node to process when last is reached */
2456 struct scan_frame *prev; /*previous frame*/
2457 I32 stop; /* what stopparen do we use */
2461 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2463 #define CASE_SYNST_FNC(nAmE) \
2465 if (flags & SCF_DO_STCLASS_AND) { \
2466 for (value = 0; value < 256; value++) \
2467 if (!is_ ## nAmE ## _cp(value)) \
2468 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2471 for (value = 0; value < 256; value++) \
2472 if (is_ ## nAmE ## _cp(value)) \
2473 ANYOF_BITMAP_SET(data->start_class, value); \
2477 if (flags & SCF_DO_STCLASS_AND) { \
2478 for (value = 0; value < 256; value++) \
2479 if (is_ ## nAmE ## _cp(value)) \
2480 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2483 for (value = 0; value < 256; value++) \
2484 if (!is_ ## nAmE ## _cp(value)) \
2485 ANYOF_BITMAP_SET(data->start_class, value); \
2492 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2493 I32 *minlenp, I32 *deltap,
2498 struct regnode_charclass_class *and_withp,
2499 U32 flags, U32 depth)
2500 /* scanp: Start here (read-write). */
2501 /* deltap: Write maxlen-minlen here. */
2502 /* last: Stop before this one. */
2503 /* data: string data about the pattern */
2504 /* stopparen: treat close N as END */
2505 /* recursed: which subroutines have we recursed into */
2506 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
2509 I32 min = 0, pars = 0, code;
2510 regnode *scan = *scanp, *next;
2512 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2513 int is_inf_internal = 0; /* The studied chunk is infinite */
2514 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2515 scan_data_t data_fake;
2516 SV *re_trie_maxbuff = NULL;
2517 regnode *first_non_open = scan;
2518 I32 stopmin = I32_MAX;
2519 scan_frame *frame = NULL;
2520 GET_RE_DEBUG_FLAGS_DECL;
2522 PERL_ARGS_ASSERT_STUDY_CHUNK;
2525 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
2529 while (first_non_open && OP(first_non_open) == OPEN)
2530 first_non_open=regnext(first_non_open);
2535 while ( scan && OP(scan) != END && scan < last ){
2536 /* Peephole optimizer: */
2537 DEBUG_STUDYDATA("Peep:", data,depth);
2538 DEBUG_PEEP("Peep",scan,depth);
2539 JOIN_EXACT(scan,&min,0);
2541 /* Follow the next-chain of the current node and optimize
2542 away all the NOTHINGs from it. */
2543 if (OP(scan) != CURLYX) {
2544 const int max = (reg_off_by_arg[OP(scan)]
2546 /* I32 may be smaller than U16 on CRAYs! */
2547 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
2548 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
2552 /* Skip NOTHING and LONGJMP. */
2553 while ((n = regnext(n))
2554 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
2555 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
2556 && off + noff < max)
2558 if (reg_off_by_arg[OP(scan)])
2561 NEXT_OFF(scan) = off;
2566 /* The principal pseudo-switch. Cannot be a switch, since we
2567 look into several different things. */
2568 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
2569 || OP(scan) == IFTHEN) {
2570 next = regnext(scan);
2572 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
2574 if (OP(next) == code || code == IFTHEN) {
2575 /* NOTE - There is similar code to this block below for handling
2576 TRIE nodes on a re-study. If you change stuff here check there
2578 I32 max1 = 0, min1 = I32_MAX, num = 0;
2579 struct regnode_charclass_class accum;
2580 regnode * const startbranch=scan;
2582 if (flags & SCF_DO_SUBSTR)
2583 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
2584 if (flags & SCF_DO_STCLASS)
2585 cl_init_zero(pRExC_state, &accum);
2587 while (OP(scan) == code) {
2588 I32 deltanext, minnext, f = 0, fake;
2589 struct regnode_charclass_class this_class;
2592 data_fake.flags = 0;
2594 data_fake.whilem_c = data->whilem_c;
2595 data_fake.last_closep = data->last_closep;
2598 data_fake.last_closep = &fake;
2600 data_fake.pos_delta = delta;
2601 next = regnext(scan);
2602 scan = NEXTOPER(scan);
2604 scan = NEXTOPER(scan);
2605 if (flags & SCF_DO_STCLASS) {
2606 cl_init(pRExC_state, &this_class);
2607 data_fake.start_class = &this_class;
2608 f = SCF_DO_STCLASS_AND;
2610 if (flags & SCF_WHILEM_VISITED_POS)
2611 f |= SCF_WHILEM_VISITED_POS;
2613 /* we suppose the run is continuous, last=next...*/
2614 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
2616 stopparen, recursed, NULL, f,depth+1);
2619 if (max1 < minnext + deltanext)
2620 max1 = minnext + deltanext;
2621 if (deltanext == I32_MAX)
2622 is_inf = is_inf_internal = 1;
2624 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
2626 if (data_fake.flags & SCF_SEEN_ACCEPT) {
2627 if ( stopmin > minnext)
2628 stopmin = min + min1;
2629 flags &= ~SCF_DO_SUBSTR;
2631 data->flags |= SCF_SEEN_ACCEPT;
2634 if (data_fake.flags & SF_HAS_EVAL)
2635 data->flags |= SF_HAS_EVAL;
2636 data->whilem_c = data_fake.whilem_c;
2638 if (flags & SCF_DO_STCLASS)
2639 cl_or(pRExC_state, &accum, &this_class);
2641 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
2643 if (flags & SCF_DO_SUBSTR) {
2644 data->pos_min += min1;
2645 data->pos_delta += max1 - min1;
2646 if (max1 != min1 || is_inf)
2647 data->longest = &(data->longest_float);
2650 delta += max1 - min1;
2651 if (flags & SCF_DO_STCLASS_OR) {
2652 cl_or(pRExC_state, data->start_class, &accum);
2654 cl_and(data->start_class, and_withp);
2655 flags &= ~SCF_DO_STCLASS;
2658 else if (flags & SCF_DO_STCLASS_AND) {
2660 cl_and(data->start_class, &accum);
2661 flags &= ~SCF_DO_STCLASS;
2664 /* Switch to OR mode: cache the old value of
2665 * data->start_class */
2667 StructCopy(data->start_class, and_withp,
2668 struct regnode_charclass_class);
2669 flags &= ~SCF_DO_STCLASS_AND;
2670 StructCopy(&accum, data->start_class,
2671 struct regnode_charclass_class);
2672 flags |= SCF_DO_STCLASS_OR;
2673 data->start_class->flags |= ANYOF_EOS;
2677 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
2680 Assuming this was/is a branch we are dealing with: 'scan' now
2681 points at the item that follows the branch sequence, whatever
2682 it is. We now start at the beginning of the sequence and look
2689 which would be constructed from a pattern like /A|LIST|OF|WORDS/
2691 If we can find such a subseqence we need to turn the first
2692 element into a trie and then add the subsequent branch exact
2693 strings to the trie.
2697 1. patterns where the whole set of branch can be converted.
2699 2. patterns where only a subset can be converted.
2701 In case 1 we can replace the whole set with a single regop
2702 for the trie. In case 2 we need to keep the start and end
2705 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
2706 becomes BRANCH TRIE; BRANCH X;
2708 There is an additional case, that being where there is a
2709 common prefix, which gets split out into an EXACT like node
2710 preceding the TRIE node.
2712 If x(1..n)==tail then we can do a simple trie, if not we make
2713 a "jump" trie, such that when we match the appropriate word
2714 we "jump" to the appopriate tail node. Essentailly we turn
2715 a nested if into a case structure of sorts.
2720 if (!re_trie_maxbuff) {
2721 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2722 if (!SvIOK(re_trie_maxbuff))
2723 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2725 if ( SvIV(re_trie_maxbuff)>=0 ) {
2727 regnode *first = (regnode *)NULL;
2728 regnode *last = (regnode *)NULL;
2729 regnode *tail = scan;
2734 SV * const mysv = sv_newmortal(); /* for dumping */
2736 /* var tail is used because there may be a TAIL
2737 regop in the way. Ie, the exacts will point to the
2738 thing following the TAIL, but the last branch will
2739 point at the TAIL. So we advance tail. If we
2740 have nested (?:) we may have to move through several
2744 while ( OP( tail ) == TAIL ) {
2745 /* this is the TAIL generated by (?:) */
2746 tail = regnext( tail );
2751 regprop(RExC_rx, mysv, tail );
2752 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
2753 (int)depth * 2 + 2, "",
2754 "Looking for TRIE'able sequences. Tail node is: ",
2755 SvPV_nolen_const( mysv )
2761 step through the branches, cur represents each
2762 branch, noper is the first thing to be matched
2763 as part of that branch and noper_next is the
2764 regnext() of that node. if noper is an EXACT
2765 and noper_next is the same as scan (our current
2766 position in the regex) then the EXACT branch is
2767 a possible optimization target. Once we have
2768 two or more consequetive such branches we can
2769 create a trie of the EXACT's contents and stich
2770 it in place. If the sequence represents all of
2771 the branches we eliminate the whole thing and
2772 replace it with a single TRIE. If it is a
2773 subsequence then we need to stitch it in. This
2774 means the first branch has to remain, and needs
2775 to be repointed at the item on the branch chain
2776 following the last branch optimized. This could
2777 be either a BRANCH, in which case the
2778 subsequence is internal, or it could be the
2779 item following the branch sequence in which
2780 case the subsequence is at the end.
2784 /* dont use tail as the end marker for this traverse */
2785 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
2786 regnode * const noper = NEXTOPER( cur );
2787 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
2788 regnode * const noper_next = regnext( noper );
2792 regprop(RExC_rx, mysv, cur);
2793 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
2794 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
2796 regprop(RExC_rx, mysv, noper);
2797 PerlIO_printf( Perl_debug_log, " -> %s",
2798 SvPV_nolen_const(mysv));
2801 regprop(RExC_rx, mysv, noper_next );
2802 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
2803 SvPV_nolen_const(mysv));
2805 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
2806 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
2808 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
2809 : PL_regkind[ OP( noper ) ] == EXACT )
2810 || OP(noper) == NOTHING )
2812 && noper_next == tail
2817 if ( !first || optype == NOTHING ) {
2818 if (!first) first = cur;
2819 optype = OP( noper );
2825 Currently we assume that the trie can handle unicode and ascii
2826 matches fold cased matches. If this proves true then the following
2827 define will prevent tries in this situation.
2829 #define TRIE_TYPE_IS_SAFE (UTF || optype==EXACT)
2831 #define TRIE_TYPE_IS_SAFE 1
2832 if ( last && TRIE_TYPE_IS_SAFE ) {
2833 make_trie( pRExC_state,
2834 startbranch, first, cur, tail, count,
2837 if ( PL_regkind[ OP( noper ) ] == EXACT
2839 && noper_next == tail
2844 optype = OP( noper );
2854 regprop(RExC_rx, mysv, cur);
2855 PerlIO_printf( Perl_debug_log,
2856 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
2857 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
2861 if ( last && TRIE_TYPE_IS_SAFE ) {
2862 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
2863 #ifdef TRIE_STUDY_OPT
2864 if ( ((made == MADE_EXACT_TRIE &&
2865 startbranch == first)
2866 || ( first_non_open == first )) &&
2868 flags |= SCF_TRIE_RESTUDY;
2869 if ( startbranch == first
2872 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
2882 else if ( code == BRANCHJ ) { /* single branch is optimized. */
2883 scan = NEXTOPER(NEXTOPER(scan));
2884 } else /* single branch is optimized. */
2885 scan = NEXTOPER(scan);
2887 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
2888 scan_frame *newframe = NULL;
2893 if (OP(scan) != SUSPEND) {
2894 /* set the pointer */
2895 if (OP(scan) == GOSUB) {
2897 RExC_recurse[ARG2L(scan)] = scan;
2898 start = RExC_open_parens[paren-1];
2899 end = RExC_close_parens[paren-1];
2902 start = RExC_rxi->program + 1;
2906 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
2907 SAVEFREEPV(recursed);
2909 if (!PAREN_TEST(recursed,paren+1)) {
2910 PAREN_SET(recursed,paren+1);
2911 Newx(newframe,1,scan_frame);
2913 if (flags & SCF_DO_SUBSTR) {
2914 SCAN_COMMIT(pRExC_state,data,minlenp);
2915 data->longest = &(data->longest_float);
2917 is_inf = is_inf_internal = 1;
2918 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
2919 cl_anything(pRExC_state, data->start_class);
2920 flags &= ~SCF_DO_STCLASS;
2923 Newx(newframe,1,scan_frame);
2926 end = regnext(scan);
2931 SAVEFREEPV(newframe);
2932 newframe->next = regnext(scan);
2933 newframe->last = last;
2934 newframe->stop = stopparen;
2935 newframe->prev = frame;
2945 else if (OP(scan) == EXACT) {
2946 I32 l = STR_LEN(scan);
2949 const U8 * const s = (U8*)STRING(scan);
2950 l = utf8_length(s, s + l);
2951 uc = utf8_to_uvchr(s, NULL);
2953 uc = *((U8*)STRING(scan));
2956 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
2957 /* The code below prefers earlier match for fixed
2958 offset, later match for variable offset. */
2959 if (data->last_end == -1) { /* Update the start info. */
2960 data->last_start_min = data->pos_min;
2961 data->last_start_max = is_inf
2962 ? I32_MAX : data->pos_min + data->pos_delta;
2964 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
2966 SvUTF8_on(data->last_found);
2968 SV * const sv = data->last_found;
2969 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
2970 mg_find(sv, PERL_MAGIC_utf8) : NULL;
2971 if (mg && mg->mg_len >= 0)
2972 mg->mg_len += utf8_length((U8*)STRING(scan),
2973 (U8*)STRING(scan)+STR_LEN(scan));
2975 data->last_end = data->pos_min + l;
2976 data->pos_min += l; /* As in the first entry. */
2977 data->flags &= ~SF_BEFORE_EOL;
2979 if (flags & SCF_DO_STCLASS_AND) {
2980 /* Check whether it is compatible with what we know already! */
2984 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
2985 && !ANYOF_BITMAP_TEST(data->start_class, uc)
2986 && (!(data->start_class->flags & ANYOF_FOLD)
2987 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold[uc])))
2990 ANYOF_CLASS_ZERO(data->start_class);
2991 ANYOF_BITMAP_ZERO(data->start_class);
2993 ANYOF_BITMAP_SET(data->start_class, uc);
2994 data->start_class->flags &= ~ANYOF_EOS;
2996 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
2998 else if (flags & SCF_DO_STCLASS_OR) {
2999 /* false positive possible if the class is case-folded */
3001 ANYOF_BITMAP_SET(data->start_class, uc);
3003 data->start_class->flags |= ANYOF_UNICODE_ALL;
3004 data->start_class->flags &= ~ANYOF_EOS;
3005 cl_and(data->start_class, and_withp);
3007 flags &= ~SCF_DO_STCLASS;
3009 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3010 I32 l = STR_LEN(scan);
3011 UV uc = *((U8*)STRING(scan));
3013 /* Search for fixed substrings supports EXACT only. */
3014 if (flags & SCF_DO_SUBSTR) {
3016 SCAN_COMMIT(pRExC_state, data, minlenp);
3019 const U8 * const s = (U8 *)STRING(scan);
3020 l = utf8_length(s, s + l);
3021 uc = utf8_to_uvchr(s, NULL);
3024 if (flags & SCF_DO_SUBSTR)
3026 if (flags & SCF_DO_STCLASS_AND) {
3027 /* Check whether it is compatible with what we know already! */
3031 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3032 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3033 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold[uc])))
3035 ANYOF_CLASS_ZERO(data->start_class);
3036 ANYOF_BITMAP_ZERO(data->start_class);
3038 ANYOF_BITMAP_SET(data->start_class, uc);
3039 data->start_class->flags &= ~ANYOF_EOS;
3040 data->start_class->flags |= ANYOF_FOLD;
3041 if (OP(scan) == EXACTFL)
3042 data->start_class->flags |= ANYOF_LOCALE;
3045 else if (flags & SCF_DO_STCLASS_OR) {
3046 if (data->start_class->flags & ANYOF_FOLD) {
3047 /* false positive possible if the class is case-folded.
3048 Assume that the locale settings are the same... */
3050 ANYOF_BITMAP_SET(data->start_class, uc);
3051 data->start_class->flags &= ~ANYOF_EOS;
3053 cl_and(data->start_class, and_withp);
3055 flags &= ~SCF_DO_STCLASS;
3057 else if (strchr((const char*)PL_varies,OP(scan))) {
3058 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3059 I32 f = flags, pos_before = 0;
3060 regnode * const oscan = scan;
3061 struct regnode_charclass_class this_class;
3062 struct regnode_charclass_class *oclass = NULL;
3063 I32 next_is_eval = 0;
3065 switch (PL_regkind[OP(scan)]) {
3066 case WHILEM: /* End of (?:...)* . */
3067 scan = NEXTOPER(scan);
3070 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3071 next = NEXTOPER(scan);
3072 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3074 maxcount = REG_INFTY;
3075 next = regnext(scan);
3076 scan = NEXTOPER(scan);
3080 if (flags & SCF_DO_SUBSTR)
3085 if (flags & SCF_DO_STCLASS) {
3087 maxcount = REG_INFTY;
3088 next = regnext(scan);
3089 scan = NEXTOPER(scan);
3092 is_inf = is_inf_internal = 1;
3093 scan = regnext(scan);
3094 if (flags & SCF_DO_SUBSTR) {
3095 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3096 data->longest = &(data->longest_float);
3098 goto optimize_curly_tail;
3100 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3101 && (scan->flags == stopparen))
3106 mincount = ARG1(scan);
3107 maxcount = ARG2(scan);
3109 next = regnext(scan);
3110 if (OP(scan) == CURLYX) {
3111 I32 lp = (data ? *(data->last_closep) : 0);
3112 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3114 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3115 next_is_eval = (OP(scan) == EVAL);
3117 if (flags & SCF_DO_SUBSTR) {
3118 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3119 pos_before = data->pos_min;
3123 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3125 data->flags |= SF_IS_INF;
3127 if (flags & SCF_DO_STCLASS) {
3128 cl_init(pRExC_state, &this_class);
3129 oclass = data->start_class;
3130 data->start_class = &this_class;
3131 f |= SCF_DO_STCLASS_AND;
3132 f &= ~SCF_DO_STCLASS_OR;
3134 /* These are the cases when once a subexpression
3135 fails at a particular position, it cannot succeed
3136 even after backtracking at the enclosing scope.
3138 XXXX what if minimal match and we are at the
3139 initial run of {n,m}? */
3140 if ((mincount != maxcount - 1) && (maxcount != REG_INFTY))
3141 f &= ~SCF_WHILEM_VISITED_POS;
3143 /* This will finish on WHILEM, setting scan, or on NULL: */
3144 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3145 last, data, stopparen, recursed, NULL,
3147 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3149 if (flags & SCF_DO_STCLASS)
3150 data->start_class = oclass;
3151 if (mincount == 0 || minnext == 0) {
3152 if (flags & SCF_DO_STCLASS_OR) {
3153 cl_or(pRExC_state, data->start_class, &this_class);
3155 else if (flags & SCF_DO_STCLASS_AND) {
3156 /* Switch to OR mode: cache the old value of
3157 * data->start_class */
3159 StructCopy(data->start_class, and_withp,
3160 struct regnode_charclass_class);
3161 flags &= ~SCF_DO_STCLASS_AND;
3162 StructCopy(&this_class, data->start_class,
3163 struct regnode_charclass_class);
3164 flags |= SCF_DO_STCLASS_OR;
3165 data->start_class->flags |= ANYOF_EOS;
3167 } else { /* Non-zero len */
3168 if (flags & SCF_DO_STCLASS_OR) {
3169 cl_or(pRExC_state, data->start_class, &this_class);
3170 cl_and(data->start_class, and_withp);
3172 else if (flags & SCF_DO_STCLASS_AND)
3173 cl_and(data->start_class, &this_class);
3174 flags &= ~SCF_DO_STCLASS;
3176 if (!scan) /* It was not CURLYX, but CURLY. */
3178 if ( /* ? quantifier ok, except for (?{ ... }) */
3179 (next_is_eval || !(mincount == 0 && maxcount == 1))
3180 && (minnext == 0) && (deltanext == 0)
3181 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3182 && maxcount <= REG_INFTY/3 /* Complement check for big count */
3183 && ckWARN(WARN_REGEXP))
3186 "Quantifier unexpected on zero-length expression");
3189 min += minnext * mincount;
3190 is_inf_internal |= ((maxcount == REG_INFTY
3191 && (minnext + deltanext) > 0)
3192 || deltanext == I32_MAX);
3193 is_inf |= is_inf_internal;
3194 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3196 /* Try powerful optimization CURLYX => CURLYN. */
3197 if ( OP(oscan) == CURLYX && data
3198 && data->flags & SF_IN_PAR
3199 && !(data->flags & SF_HAS_EVAL)
3200 && !deltanext && minnext == 1 ) {
3201 /* Try to optimize to CURLYN. */
3202 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3203 regnode * const nxt1 = nxt;
3210 if (!strchr((const char*)PL_simple,OP(nxt))
3211 && !(PL_regkind[OP(nxt)] == EXACT
3212 && STR_LEN(nxt) == 1))
3218 if (OP(nxt) != CLOSE)
3220 if (RExC_open_parens) {
3221 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3222 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3224 /* Now we know that nxt2 is the only contents: */
3225 oscan->flags = (U8)ARG(nxt);
3227 OP(nxt1) = NOTHING; /* was OPEN. */
3230 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3231 NEXT_OFF(nxt1+ 1) = 0; /* just for consistancy. */
3232 NEXT_OFF(nxt2) = 0; /* just for consistancy with CURLY. */
3233 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3234 OP(nxt + 1) = OPTIMIZED; /* was count. */
3235 NEXT_OFF(nxt+ 1) = 0; /* just for consistancy. */
3240 /* Try optimization CURLYX => CURLYM. */
3241 if ( OP(oscan) == CURLYX && data
3242 && !(data->flags & SF_HAS_PAR)
3243 && !(data->flags & SF_HAS_EVAL)
3244 && !deltanext /* atom is fixed width */
3245 && minnext != 0 /* CURLYM can't handle zero width */
3247 /* XXXX How to optimize if data == 0? */
3248 /* Optimize to a simpler form. */
3249 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3253 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3254 && (OP(nxt2) != WHILEM))
3256 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3257 /* Need to optimize away parenths. */
3258 if (data->flags & SF_IN_PAR) {
3259 /* Set the parenth number. */
3260 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3262 if (OP(nxt) != CLOSE)
3263 FAIL("Panic opt close");
3264 oscan->flags = (U8)ARG(nxt);
3265 if (RExC_open_parens) {
3266 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3267 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3269 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3270 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3273 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3274 OP(nxt + 1) = OPTIMIZED; /* was count. */
3275 NEXT_OFF(nxt1 + 1) = 0; /* just for consistancy. */
3276 NEXT_OFF(nxt + 1) = 0; /* just for consistancy. */
3279 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3280 regnode *nnxt = regnext(nxt1);
3283 if (reg_off_by_arg[OP(nxt1)])
3284 ARG_SET(nxt1, nxt2 - nxt1);
3285 else if (nxt2 - nxt1 < U16_MAX)
3286 NEXT_OFF(nxt1) = nxt2 - nxt1;
3288 OP(nxt) = NOTHING; /* Cannot beautify */
3293 /* Optimize again: */
3294 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
3295 NULL, stopparen, recursed, NULL, 0,depth+1);
3300 else if ((OP(oscan) == CURLYX)
3301 && (flags & SCF_WHILEM_VISITED_POS)
3302 /* See the comment on a similar expression above.
3303 However, this time it not a subexpression
3304 we care about, but the expression itself. */
3305 && (maxcount == REG_INFTY)
3306 && data && ++data->whilem_c < 16) {
3307 /* This stays as CURLYX, we can put the count/of pair. */
3308 /* Find WHILEM (as in regexec.c) */
3309 regnode *nxt = oscan + NEXT_OFF(oscan);
3311 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
3313 PREVOPER(nxt)->flags = (U8)(data->whilem_c
3314 | (RExC_whilem_seen << 4)); /* On WHILEM */
3316 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
3318 if (flags & SCF_DO_SUBSTR) {
3319 SV *last_str = NULL;
3320 int counted = mincount != 0;
3322 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
3323 #if defined(SPARC64_GCC_WORKAROUND)
3326 const char *s = NULL;
3329 if (pos_before >= data->last_start_min)
3332 b = data->last_start_min;
3335 s = SvPV_const(data->last_found, l);
3336 old = b - data->last_start_min;
3339 I32 b = pos_before >= data->last_start_min
3340 ? pos_before : data->last_start_min;
3342 const char * const s = SvPV_const(data->last_found, l);
3343 I32 old = b - data->last_start_min;
3347 old = utf8_hop((U8*)s, old) - (U8*)s;
3350 /* Get the added string: */
3351 last_str = newSVpvn_utf8(s + old, l, UTF);
3352 if (deltanext == 0 && pos_before == b) {
3353 /* What was added is a constant string */
3355 SvGROW(last_str, (mincount * l) + 1);
3356 repeatcpy(SvPVX(last_str) + l,
3357 SvPVX_const(last_str), l, mincount - 1);
3358 SvCUR_set(last_str, SvCUR(last_str) * mincount);
3359 /* Add additional parts. */
3360 SvCUR_set(data->last_found,
3361 SvCUR(data->last_found) - l);
3362 sv_catsv(data->last_found, last_str);
3364 SV * sv = data->last_found;
3366 SvUTF8(sv) && SvMAGICAL(sv) ?
3367 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3368 if (mg && mg->mg_len >= 0)
3369 mg->mg_len += CHR_SVLEN(last_str) - l;
3371 data->last_end += l * (mincount - 1);
3374 /* start offset must point into the last copy */
3375 data->last_start_min += minnext * (mincount - 1);
3376 data->last_start_max += is_inf ? I32_MAX
3377 : (maxcount - 1) * (minnext + data->pos_delta);
3380 /* It is counted once already... */
3381 data->pos_min += minnext * (mincount - counted);
3382 data->pos_delta += - counted * deltanext +
3383 (minnext + deltanext) * maxcount - minnext * mincount;
3384 if (mincount != maxcount) {
3385 /* Cannot extend fixed substrings found inside
3387 SCAN_COMMIT(pRExC_state,data,minlenp);
3388 if (mincount && last_str) {
3389 SV * const sv = data->last_found;
3390 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3391 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3395 sv_setsv(sv, last_str);
3396 data->last_end = data->pos_min;
3397 data->last_start_min =
3398 data->pos_min - CHR_SVLEN(last_str);
3399 data->last_start_max = is_inf
3401 : data->pos_min + data->pos_delta
3402 - CHR_SVLEN(last_str);
3404 data->longest = &(data->longest_float);
3406 SvREFCNT_dec(last_str);
3408 if (data && (fl & SF_HAS_EVAL))
3409 data->flags |= SF_HAS_EVAL;
3410 optimize_curly_tail:
3411 if (OP(oscan) != CURLYX) {
3412 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
3414 NEXT_OFF(oscan) += NEXT_OFF(next);
3417 default: /* REF and CLUMP only? */
3418 if (flags & SCF_DO_SUBSTR) {
3419 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3420 data->longest = &(data->longest_float);
3422 is_inf = is_inf_internal = 1;
3423 if (flags & SCF_DO_STCLASS_OR)
3424 cl_anything(pRExC_state, data->start_class);
3425 flags &= ~SCF_DO_STCLASS;
3429 else if (OP(scan) == LNBREAK) {
3430 if (flags & SCF_DO_STCLASS) {
3432 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3433 if (flags & SCF_DO_STCLASS_AND) {
3434 for (value = 0; value < 256; value++)
3435 if (!is_VERTWS_cp(value))
3436 ANYOF_BITMAP_CLEAR(data->start_class, value);
3439 for (value = 0; value < 256; value++)
3440 if (is_VERTWS_cp(value))
3441 ANYOF_BITMAP_SET(data->start_class, value);
3443 if (flags & SCF_DO_STCLASS_OR)
3444 cl_and(data->start_class, and_withp);
3445 flags &= ~SCF_DO_STCLASS;
3449 if (flags & SCF_DO_SUBSTR) {
3450 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3452 data->pos_delta += 1;
3453 data->longest = &(data->longest_float);
3457 else if (OP(scan) == FOLDCHAR) {
3458 int d = ARG(scan)==0xDF ? 1 : 2;
3459 flags &= ~SCF_DO_STCLASS;
3462 if (flags & SCF_DO_SUBSTR) {
3463 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3465 data->pos_delta += d;
3466 data->longest = &(data->longest_float);
3469 else if (strchr((const char*)PL_simple,OP(scan))) {
3472 if (flags & SCF_DO_SUBSTR) {
3473 SCAN_COMMIT(pRExC_state,data,minlenp);
3477 if (flags & SCF_DO_STCLASS) {
3478 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3480 /* Some of the logic below assumes that switching
3481 locale on will only add false positives. */
3482 switch (PL_regkind[OP(scan)]) {
3486 /* Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan)); */
3487 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3488 cl_anything(pRExC_state, data->start_class);
3491 if (OP(scan) == SANY)
3493 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
3494 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
3495 || (data->start_class->flags & ANYOF_CLASS));
3496 cl_anything(pRExC_state, data->start_class);
3498 if (flags & SCF_DO_STCLASS_AND || !value)
3499 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
3502 if (flags & SCF_DO_STCLASS_AND)
3503 cl_and(data->start_class,
3504 (struct regnode_charclass_class*)scan);
3506 cl_or(pRExC_state, data->start_class,
3507 (struct regnode_charclass_class*)scan);
3510 if (flags & SCF_DO_STCLASS_AND) {
3511 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3512 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3513 for (value = 0; value < 256; value++)
3514 if (!isALNUM(value))
3515 ANYOF_BITMAP_CLEAR(data->start_class, value);
3519 if (data->start_class->flags & ANYOF_LOCALE)
3520 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3522 for (value = 0; value < 256; value++)
3524 ANYOF_BITMAP_SET(data->start_class, value);
3529 if (flags & SCF_DO_STCLASS_AND) {
3530 if (data->start_class->flags & ANYOF_LOCALE)
3531 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3534 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3535 data->start_class->flags |= ANYOF_LOCALE;
3539 if (flags & SCF_DO_STCLASS_AND) {
3540 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3541 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3542 for (value = 0; value < 256; value++)
3544 ANYOF_BITMAP_CLEAR(data->start_class, value);
3548 if (data->start_class->flags & ANYOF_LOCALE)
3549 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3551 for (value = 0; value < 256; value++)
3552 if (!isALNUM(value))
3553 ANYOF_BITMAP_SET(data->start_class, value);
3558 if (flags & SCF_DO_STCLASS_AND) {
3559 if (data->start_class->flags & ANYOF_LOCALE)
3560 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3563 data->start_class->flags |= ANYOF_LOCALE;
3564 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3568 if (flags & SCF_DO_STCLASS_AND) {
3569 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3570 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3571 for (value = 0; value < 256; value++)
3572 if (!isSPACE(value))
3573 ANYOF_BITMAP_CLEAR(data->start_class, value);
3577 if (data->start_class->flags & ANYOF_LOCALE)
3578 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3580 for (value = 0; value < 256; value++)
3582 ANYOF_BITMAP_SET(data->start_class, value);
3587 if (flags & SCF_DO_STCLASS_AND) {
3588 if (data->start_class->flags & ANYOF_LOCALE)
3589 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3592 data->start_class->flags |= ANYOF_LOCALE;
3593 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3597 if (flags & SCF_DO_STCLASS_AND) {
3598 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3599 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3600 for (value = 0; value < 256; value++)
3602 ANYOF_BITMAP_CLEAR(data->start_class, value);
3606 if (data->start_class->flags & ANYOF_LOCALE)
3607 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3609 for (value = 0; value < 256; value++)
3610 if (!isSPACE(value))
3611 ANYOF_BITMAP_SET(data->start_class, value);
3616 if (flags & SCF_DO_STCLASS_AND) {
3617 if (data->start_class->flags & ANYOF_LOCALE) {
3618 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3619 for (value = 0; value < 256; value++)
3620 if (!isSPACE(value))
3621 ANYOF_BITMAP_CLEAR(data->start_class, value);
3625 data->start_class->flags |= ANYOF_LOCALE;
3626 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3630 if (flags & SCF_DO_STCLASS_AND) {
3631 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NDIGIT);
3632 for (value = 0; value < 256; value++)
3633 if (!isDIGIT(value))
3634 ANYOF_BITMAP_CLEAR(data->start_class, value);
3637 if (data->start_class->flags & ANYOF_LOCALE)
3638 ANYOF_CLASS_SET(data->start_class,ANYOF_DIGIT);
3640 for (value = 0; value < 256; value++)
3642 ANYOF_BITMAP_SET(data->start_class, value);
3647 if (flags & SCF_DO_STCLASS_AND) {
3648 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_DIGIT);
3649 for (value = 0; value < 256; value++)
3651 ANYOF_BITMAP_CLEAR(data->start_class, value);
3654 if (data->start_class->flags & ANYOF_LOCALE)
3655 ANYOF_CLASS_SET(data->start_class,ANYOF_NDIGIT);
3657 for (value = 0; value < 256; value++)
3658 if (!isDIGIT(value))
3659 ANYOF_BITMAP_SET(data->start_class, value);
3663 CASE_SYNST_FNC(VERTWS);
3664 CASE_SYNST_FNC(HORIZWS);
3667 if (flags & SCF_DO_STCLASS_OR)
3668 cl_and(data->start_class, and_withp);
3669 flags &= ~SCF_DO_STCLASS;
3672 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
3673 data->flags |= (OP(scan) == MEOL
3677 else if ( PL_regkind[OP(scan)] == BRANCHJ
3678 /* Lookbehind, or need to calculate parens/evals/stclass: */
3679 && (scan->flags || data || (flags & SCF_DO_STCLASS))
3680 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
3681 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
3682 || OP(scan) == UNLESSM )
3684 /* Negative Lookahead/lookbehind
3685 In this case we can't do fixed string optimisation.
3688 I32 deltanext, minnext, fake = 0;
3690 struct regnode_charclass_class intrnl;
3693 data_fake.flags = 0;
3695 data_fake.whilem_c = data->whilem_c;
3696 data_fake.last_closep = data->last_closep;
3699 data_fake.last_closep = &fake;
3700 data_fake.pos_delta = delta;
3701 if ( flags & SCF_DO_STCLASS && !scan->flags
3702 && OP(scan) == IFMATCH ) { /* Lookahead */
3703 cl_init(pRExC_state, &intrnl);
3704 data_fake.start_class = &intrnl;
3705 f |= SCF_DO_STCLASS_AND;
3707 if (flags & SCF_WHILEM_VISITED_POS)
3708 f |= SCF_WHILEM_VISITED_POS;
3709 next = regnext(scan);
3710 nscan = NEXTOPER(NEXTOPER(scan));
3711 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
3712 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
3715 FAIL("Variable length lookbehind not implemented");
3717 else if (minnext > (I32)U8_MAX) {
3718 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
3720 scan->flags = (U8)minnext;
3723 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3725 if (data_fake.flags & SF_HAS_EVAL)
3726 data->flags |= SF_HAS_EVAL;
3727 data->whilem_c = data_fake.whilem_c;
3729 if (f & SCF_DO_STCLASS_AND) {
3730 if (flags & SCF_DO_STCLASS_OR) {
3731 /* OR before, AND after: ideally we would recurse with
3732 * data_fake to get the AND applied by study of the
3733 * remainder of the pattern, and then derecurse;
3734 * *** HACK *** for now just treat as "no information".
3735 * See [perl #56690].
3737 cl_init(pRExC_state, data->start_class);
3739 /* AND before and after: combine and continue */
3740 const int was = (data->start_class->flags & ANYOF_EOS);
3742 cl_and(data->start_class, &intrnl);
3744 data->start_class->flags |= ANYOF_EOS;
3748 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
3750 /* Positive Lookahead/lookbehind
3751 In this case we can do fixed string optimisation,
3752 but we must be careful about it. Note in the case of
3753 lookbehind the positions will be offset by the minimum
3754 length of the pattern, something we won't know about
3755 until after the recurse.
3757 I32 deltanext, fake = 0;
3759 struct regnode_charclass_class intrnl;
3761 /* We use SAVEFREEPV so that when the full compile
3762 is finished perl will clean up the allocated
3763 minlens when its all done. This was we don't
3764 have to worry about freeing them when we know
3765 they wont be used, which would be a pain.
3768 Newx( minnextp, 1, I32 );
3769 SAVEFREEPV(minnextp);
3772 StructCopy(data, &data_fake, scan_data_t);
3773 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
3776 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
3777 data_fake.last_found=newSVsv(data->last_found);
3781 data_fake.last_closep = &fake;
3782 data_fake.flags = 0;
3783 data_fake.pos_delta = delta;
3785 data_fake.flags |= SF_IS_INF;
3786 if ( flags & SCF_DO_STCLASS && !scan->flags
3787 && OP(scan) == IFMATCH ) { /* Lookahead */
3788 cl_init(pRExC_state, &intrnl);
3789 data_fake.start_class = &intrnl;
3790 f |= SCF_DO_STCLASS_AND;
3792 if (flags & SCF_WHILEM_VISITED_POS)
3793 f |= SCF_WHILEM_VISITED_POS;
3794 next = regnext(scan);
3795 nscan = NEXTOPER(NEXTOPER(scan));
3797 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
3798 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
3801 FAIL("Variable length lookbehind not implemented");
3803 else if (*minnextp > (I32)U8_MAX) {
3804 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
3806 scan->flags = (U8)*minnextp;
3811 if (f & SCF_DO_STCLASS_AND) {
3812 const int was = (data->start_class->flags & ANYOF_EOS);
3814 cl_and(data->start_class, &intrnl);
3816 data->start_class->flags |= ANYOF_EOS;
3819 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3821 if (data_fake.flags & SF_HAS_EVAL)
3822 data->flags |= SF_HAS_EVAL;
3823 data->whilem_c = data_fake.whilem_c;
3824 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
3825 if (RExC_rx->minlen<*minnextp)
3826 RExC_rx->minlen=*minnextp;
3827 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
3828 SvREFCNT_dec(data_fake.last_found);
3830 if ( data_fake.minlen_fixed != minlenp )
3832 data->offset_fixed= data_fake.offset_fixed;
Perl 5 / perl.git / blob