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
88 #include "dquote_static.c"
95 # if defined(BUGGY_MSC6)
96 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
97 # pragma optimize("a",off)
98 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
99 # pragma optimize("w",on )
100 # endif /* BUGGY_MSC6 */
104 #define STATIC static
107 typedef struct RExC_state_t {
108 U32 flags; /* are we folding, multilining? */
109 char *precomp; /* uncompiled string. */
110 REGEXP *rx_sv; /* The SV that is the regexp. */
111 regexp *rx; /* perl core regexp structure */
112 regexp_internal *rxi; /* internal data for regexp object pprivate field */
113 char *start; /* Start of input for compile */
114 char *end; /* End of input for compile */
115 char *parse; /* Input-scan pointer. */
116 I32 whilem_seen; /* number of WHILEM in this expr */
117 regnode *emit_start; /* Start of emitted-code area */
118 regnode *emit_bound; /* First regnode outside of the allocated space */
119 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
120 I32 naughty; /* How bad is this pattern? */
121 I32 sawback; /* Did we see \1, ...? */
123 I32 size; /* Code size. */
124 I32 npar; /* Capture buffer count, (OPEN). */
125 I32 cpar; /* Capture buffer count, (CLOSE). */
126 I32 nestroot; /* root parens we are in - used by accept */
130 regnode **open_parens; /* pointers to open parens */
131 regnode **close_parens; /* pointers to close parens */
132 regnode *opend; /* END node in program */
133 I32 utf8; /* whether the pattern is utf8 or not */
134 I32 orig_utf8; /* whether the pattern was originally in utf8 */
135 /* XXX use this for future optimisation of case
136 * where pattern must be upgraded to utf8. */
137 HV *paren_names; /* Paren names */
139 regnode **recurse; /* Recurse regops */
140 I32 recurse_count; /* Number of recurse regops */
142 char *starttry; /* -Dr: where regtry was called. */
143 #define RExC_starttry (pRExC_state->starttry)
146 const char *lastparse;
148 AV *paren_name_list; /* idx -> name */
149 #define RExC_lastparse (pRExC_state->lastparse)
150 #define RExC_lastnum (pRExC_state->lastnum)
151 #define RExC_paren_name_list (pRExC_state->paren_name_list)
155 #define RExC_flags (pRExC_state->flags)
156 #define RExC_precomp (pRExC_state->precomp)
157 #define RExC_rx_sv (pRExC_state->rx_sv)
158 #define RExC_rx (pRExC_state->rx)
159 #define RExC_rxi (pRExC_state->rxi)
160 #define RExC_start (pRExC_state->start)
161 #define RExC_end (pRExC_state->end)
162 #define RExC_parse (pRExC_state->parse)
163 #define RExC_whilem_seen (pRExC_state->whilem_seen)
164 #ifdef RE_TRACK_PATTERN_OFFSETS
165 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
167 #define RExC_emit (pRExC_state->emit)
168 #define RExC_emit_start (pRExC_state->emit_start)
169 #define RExC_emit_bound (pRExC_state->emit_bound)
170 #define RExC_naughty (pRExC_state->naughty)
171 #define RExC_sawback (pRExC_state->sawback)
172 #define RExC_seen (pRExC_state->seen)
173 #define RExC_size (pRExC_state->size)
174 #define RExC_npar (pRExC_state->npar)
175 #define RExC_nestroot (pRExC_state->nestroot)
176 #define RExC_extralen (pRExC_state->extralen)
177 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
178 #define RExC_seen_evals (pRExC_state->seen_evals)
179 #define RExC_utf8 (pRExC_state->utf8)
180 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
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. */
202 /* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
203 * character, and if utf8, must be invariant. */
205 #define SPSTART 0x04 /* Starts with * or +. */
206 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
207 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
209 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
211 /* whether trie related optimizations are enabled */
212 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
213 #define TRIE_STUDY_OPT
214 #define FULL_TRIE_STUDY
220 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
221 #define PBITVAL(paren) (1 << ((paren) & 7))
222 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
223 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
224 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
226 /* If not already in utf8, do a longjmp back to the beginning */
227 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
228 #define REQUIRE_UTF8 STMT_START { \
229 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
232 /* About scan_data_t.
234 During optimisation we recurse through the regexp program performing
235 various inplace (keyhole style) optimisations. In addition study_chunk
236 and scan_commit populate this data structure with information about
237 what strings MUST appear in the pattern. We look for the longest
238 string that must appear at a fixed location, and we look for the
239 longest string that may appear at a floating location. So for instance
244 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
245 strings (because they follow a .* construct). study_chunk will identify
246 both FOO and BAR as being the longest fixed and floating strings respectively.
248 The strings can be composites, for instance
252 will result in a composite fixed substring 'foo'.
254 For each string some basic information is maintained:
256 - offset or min_offset
257 This is the position the string must appear at, or not before.
258 It also implicitly (when combined with minlenp) tells us how many
259 characters must match before the string we are searching for.
260 Likewise when combined with minlenp and the length of the string it
261 tells us how many characters must appear after the string we have
265 Only used for floating strings. This is the rightmost point that
266 the string can appear at. If set to I32 max it indicates that the
267 string can occur infinitely far to the right.
270 A pointer to the minimum length of the pattern that the string
271 was found inside. This is important as in the case of positive
272 lookahead or positive lookbehind we can have multiple patterns
277 The minimum length of the pattern overall is 3, the minimum length
278 of the lookahead part is 3, but the minimum length of the part that
279 will actually match is 1. So 'FOO's minimum length is 3, but the
280 minimum length for the F is 1. This is important as the minimum length
281 is used to determine offsets in front of and behind the string being
282 looked for. Since strings can be composites this is the length of the
283 pattern at the time it was commited with a scan_commit. Note that
284 the length is calculated by study_chunk, so that the minimum lengths
285 are not known until the full pattern has been compiled, thus the
286 pointer to the value.
290 In the case of lookbehind the string being searched for can be
291 offset past the start point of the final matching string.
292 If this value was just blithely removed from the min_offset it would
293 invalidate some of the calculations for how many chars must match
294 before or after (as they are derived from min_offset and minlen and
295 the length of the string being searched for).
296 When the final pattern is compiled and the data is moved from the
297 scan_data_t structure into the regexp structure the information
298 about lookbehind is factored in, with the information that would
299 have been lost precalculated in the end_shift field for the
302 The fields pos_min and pos_delta are used to store the minimum offset
303 and the delta to the maximum offset at the current point in the pattern.
307 typedef struct scan_data_t {
308 /*I32 len_min; unused */
309 /*I32 len_delta; unused */
313 I32 last_end; /* min value, <0 unless valid. */
316 SV **longest; /* Either &l_fixed, or &l_float. */
317 SV *longest_fixed; /* longest fixed string found in pattern */
318 I32 offset_fixed; /* offset where it starts */
319 I32 *minlen_fixed; /* pointer to the minlen relevent to the string */
320 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
321 SV *longest_float; /* longest floating string found in pattern */
322 I32 offset_float_min; /* earliest point in string it can appear */
323 I32 offset_float_max; /* latest point in string it can appear */
324 I32 *minlen_float; /* pointer to the minlen relevent to the string */
325 I32 lookbehind_float; /* is the position of the string modified by LB */
329 struct regnode_charclass_class *start_class;
333 * Forward declarations for pregcomp()'s friends.
336 static const scan_data_t zero_scan_data =
337 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
339 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
340 #define SF_BEFORE_SEOL 0x0001
341 #define SF_BEFORE_MEOL 0x0002
342 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
343 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
346 # define SF_FIX_SHIFT_EOL (0+2)
347 # define SF_FL_SHIFT_EOL (0+4)
349 # define SF_FIX_SHIFT_EOL (+2)
350 # define SF_FL_SHIFT_EOL (+4)
353 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
354 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
356 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
357 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
358 #define SF_IS_INF 0x0040
359 #define SF_HAS_PAR 0x0080
360 #define SF_IN_PAR 0x0100
361 #define SF_HAS_EVAL 0x0200
362 #define SCF_DO_SUBSTR 0x0400
363 #define SCF_DO_STCLASS_AND 0x0800
364 #define SCF_DO_STCLASS_OR 0x1000
365 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
366 #define SCF_WHILEM_VISITED_POS 0x2000
368 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
369 #define SCF_SEEN_ACCEPT 0x8000
371 #define UTF cBOOL(RExC_utf8)
372 #define LOC cBOOL(RExC_flags & RXf_PMf_LOCALE)
373 #define UNI_SEMANTICS cBOOL(RExC_flags & RXf_PMf_UNICODE)
374 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
376 #define OOB_UNICODE 12345678
377 #define OOB_NAMEDCLASS -1
379 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
380 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
383 /* length of regex to show in messages that don't mark a position within */
384 #define RegexLengthToShowInErrorMessages 127
387 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
388 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
389 * op/pragma/warn/regcomp.
391 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
392 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
394 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
397 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
398 * arg. Show regex, up to a maximum length. If it's too long, chop and add
401 #define _FAIL(code) STMT_START { \
402 const char *ellipses = ""; \
403 IV len = RExC_end - RExC_precomp; \
406 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
407 if (len > RegexLengthToShowInErrorMessages) { \
408 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
409 len = RegexLengthToShowInErrorMessages - 10; \
415 #define FAIL(msg) _FAIL( \
416 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
417 msg, (int)len, RExC_precomp, ellipses))
419 #define FAIL2(msg,arg) _FAIL( \
420 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
421 arg, (int)len, RExC_precomp, ellipses))
424 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
426 #define Simple_vFAIL(m) STMT_START { \
427 const IV offset = RExC_parse - RExC_precomp; \
428 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
429 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
433 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
435 #define vFAIL(m) STMT_START { \
437 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
442 * Like Simple_vFAIL(), but accepts two arguments.
444 #define Simple_vFAIL2(m,a1) STMT_START { \
445 const IV offset = RExC_parse - RExC_precomp; \
446 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
447 (int)offset, RExC_precomp, RExC_precomp + offset); \
451 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
453 #define vFAIL2(m,a1) STMT_START { \
455 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
456 Simple_vFAIL2(m, a1); \
461 * Like Simple_vFAIL(), but accepts three arguments.
463 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
464 const IV offset = RExC_parse - RExC_precomp; \
465 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
466 (int)offset, RExC_precomp, RExC_precomp + offset); \
470 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
472 #define vFAIL3(m,a1,a2) STMT_START { \
474 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
475 Simple_vFAIL3(m, a1, a2); \
479 * Like Simple_vFAIL(), but accepts four arguments.
481 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
482 const IV offset = RExC_parse - RExC_precomp; \
483 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
484 (int)offset, RExC_precomp, RExC_precomp + offset); \
487 #define ckWARNreg(loc,m) STMT_START { \
488 const IV offset = loc - RExC_precomp; \
489 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
490 (int)offset, RExC_precomp, RExC_precomp + offset); \
493 #define ckWARNregdep(loc,m) STMT_START { \
494 const IV offset = loc - RExC_precomp; \
495 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
497 (int)offset, RExC_precomp, RExC_precomp + offset); \
500 #define ckWARN2reg(loc, m, a1) STMT_START { \
501 const IV offset = loc - RExC_precomp; \
502 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
503 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
506 #define vWARN3(loc, m, a1, a2) STMT_START { \
507 const IV offset = loc - RExC_precomp; \
508 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
509 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
512 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
513 const IV offset = loc - RExC_precomp; \
514 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
515 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
518 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
519 const IV offset = loc - RExC_precomp; \
520 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
521 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
524 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
525 const IV offset = loc - RExC_precomp; \
526 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
527 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
530 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
531 const IV offset = loc - RExC_precomp; \
532 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
533 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
537 /* Allow for side effects in s */
538 #define REGC(c,s) STMT_START { \
539 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
542 /* Macros for recording node offsets. 20001227 mjd@plover.com
543 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
544 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
545 * Element 0 holds the number n.
546 * Position is 1 indexed.
548 #ifndef RE_TRACK_PATTERN_OFFSETS
549 #define Set_Node_Offset_To_R(node,byte)
550 #define Set_Node_Offset(node,byte)
551 #define Set_Cur_Node_Offset
552 #define Set_Node_Length_To_R(node,len)
553 #define Set_Node_Length(node,len)
554 #define Set_Node_Cur_Length(node)
555 #define Node_Offset(n)
556 #define Node_Length(n)
557 #define Set_Node_Offset_Length(node,offset,len)
558 #define ProgLen(ri) ri->u.proglen
559 #define SetProgLen(ri,x) ri->u.proglen = x
561 #define ProgLen(ri) ri->u.offsets[0]
562 #define SetProgLen(ri,x) ri->u.offsets[0] = x
563 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
565 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
566 __LINE__, (int)(node), (int)(byte))); \
568 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
570 RExC_offsets[2*(node)-1] = (byte); \
575 #define Set_Node_Offset(node,byte) \
576 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
577 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
579 #define Set_Node_Length_To_R(node,len) STMT_START { \
581 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
582 __LINE__, (int)(node), (int)(len))); \
584 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
586 RExC_offsets[2*(node)] = (len); \
591 #define Set_Node_Length(node,len) \
592 Set_Node_Length_To_R((node)-RExC_emit_start, len)
593 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
594 #define Set_Node_Cur_Length(node) \
595 Set_Node_Length(node, RExC_parse - parse_start)
597 /* Get offsets and lengths */
598 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
599 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
601 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
602 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
603 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
607 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
608 #define EXPERIMENTAL_INPLACESCAN
609 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
611 #define DEBUG_STUDYDATA(str,data,depth) \
612 DEBUG_OPTIMISE_MORE_r(if(data){ \
613 PerlIO_printf(Perl_debug_log, \
614 "%*s" str "Pos:%"IVdf"/%"IVdf \
615 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
616 (int)(depth)*2, "", \
617 (IV)((data)->pos_min), \
618 (IV)((data)->pos_delta), \
619 (UV)((data)->flags), \
620 (IV)((data)->whilem_c), \
621 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
622 is_inf ? "INF " : "" \
624 if ((data)->last_found) \
625 PerlIO_printf(Perl_debug_log, \
626 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
627 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
628 SvPVX_const((data)->last_found), \
629 (IV)((data)->last_end), \
630 (IV)((data)->last_start_min), \
631 (IV)((data)->last_start_max), \
632 ((data)->longest && \
633 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
634 SvPVX_const((data)->longest_fixed), \
635 (IV)((data)->offset_fixed), \
636 ((data)->longest && \
637 (data)->longest==&((data)->longest_float)) ? "*" : "", \
638 SvPVX_const((data)->longest_float), \
639 (IV)((data)->offset_float_min), \
640 (IV)((data)->offset_float_max) \
642 PerlIO_printf(Perl_debug_log,"\n"); \
645 static void clear_re(pTHX_ void *r);
647 /* Mark that we cannot extend a found fixed substring at this point.
648 Update the longest found anchored substring and the longest found
649 floating substrings if needed. */
652 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
654 const STRLEN l = CHR_SVLEN(data->last_found);
655 const STRLEN old_l = CHR_SVLEN(*data->longest);
656 GET_RE_DEBUG_FLAGS_DECL;
658 PERL_ARGS_ASSERT_SCAN_COMMIT;
660 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
661 SvSetMagicSV(*data->longest, data->last_found);
662 if (*data->longest == data->longest_fixed) {
663 data->offset_fixed = l ? data->last_start_min : data->pos_min;
664 if (data->flags & SF_BEFORE_EOL)
666 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
668 data->flags &= ~SF_FIX_BEFORE_EOL;
669 data->minlen_fixed=minlenp;
670 data->lookbehind_fixed=0;
672 else { /* *data->longest == data->longest_float */
673 data->offset_float_min = l ? data->last_start_min : data->pos_min;
674 data->offset_float_max = (l
675 ? data->last_start_max
676 : data->pos_min + data->pos_delta);
677 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
678 data->offset_float_max = I32_MAX;
679 if (data->flags & SF_BEFORE_EOL)
681 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
683 data->flags &= ~SF_FL_BEFORE_EOL;
684 data->minlen_float=minlenp;
685 data->lookbehind_float=0;
688 SvCUR_set(data->last_found, 0);
690 SV * const sv = data->last_found;
691 if (SvUTF8(sv) && SvMAGICAL(sv)) {
692 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
698 data->flags &= ~SF_BEFORE_EOL;
699 DEBUG_STUDYDATA("commit: ",data,0);
702 /* Can match anything (initialization) */
704 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
706 PERL_ARGS_ASSERT_CL_ANYTHING;
708 ANYOF_CLASS_ZERO(cl);
709 ANYOF_BITMAP_SETALL(cl);
710 cl->flags = ANYOF_EOS|ANYOF_UNICODE_ALL;
712 cl->flags |= ANYOF_LOCALE;
713 cl->flags |= ANYOF_FOLD;
716 /* Can match anything (initialization) */
718 S_cl_is_anything(const struct regnode_charclass_class *cl)
722 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
724 for (value = 0; value <= ANYOF_MAX; value += 2)
725 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
727 if (!(cl->flags & ANYOF_UNICODE_ALL))
729 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
734 /* Can match anything (initialization) */
736 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
738 PERL_ARGS_ASSERT_CL_INIT;
740 Zero(cl, 1, struct regnode_charclass_class);
742 cl_anything(pRExC_state, cl);
746 S_cl_init_zero(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
748 PERL_ARGS_ASSERT_CL_INIT_ZERO;
750 Zero(cl, 1, struct regnode_charclass_class);
752 cl_anything(pRExC_state, cl);
754 cl->flags |= ANYOF_LOCALE;
757 /* 'And' a given class with another one. Can create false positives */
758 /* We assume that cl is not inverted */
760 S_cl_and(struct regnode_charclass_class *cl,
761 const struct regnode_charclass_class *and_with)
763 PERL_ARGS_ASSERT_CL_AND;
765 assert(and_with->type == ANYOF);
766 if (!(and_with->flags & ANYOF_CLASS)
767 && !(cl->flags & ANYOF_CLASS)
768 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
769 && !(and_with->flags & ANYOF_FOLD)
770 && !(cl->flags & ANYOF_FOLD)) {
773 if (and_with->flags & ANYOF_INVERT)
774 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
775 cl->bitmap[i] &= ~and_with->bitmap[i];
777 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
778 cl->bitmap[i] &= and_with->bitmap[i];
779 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
780 if (!(and_with->flags & ANYOF_EOS))
781 cl->flags &= ~ANYOF_EOS;
783 if (!(and_with->flags & ANYOF_FOLD))
784 cl->flags &= ~ANYOF_FOLD;
786 if (cl->flags & ANYOF_UNICODE_ALL && and_with->flags & ANYOF_NONBITMAP &&
787 !(and_with->flags & ANYOF_INVERT)) {
788 cl->flags &= ~ANYOF_UNICODE_ALL;
789 cl->flags |= and_with->flags & ANYOF_NONBITMAP; /* field is 2 bits; use
792 ARG_SET(cl, ARG(and_with));
794 if (!(and_with->flags & ANYOF_UNICODE_ALL) &&
795 !(and_with->flags & ANYOF_INVERT))
796 cl->flags &= ~ANYOF_UNICODE_ALL;
797 if (!(and_with->flags & (ANYOF_NONBITMAP|ANYOF_UNICODE_ALL)) &&
798 !(and_with->flags & ANYOF_INVERT))
799 cl->flags &= ~ANYOF_NONBITMAP;
802 /* 'OR' a given class with another one. Can create false positives */
803 /* We assume that cl is not inverted */
805 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
807 PERL_ARGS_ASSERT_CL_OR;
809 if (or_with->flags & ANYOF_INVERT) {
811 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
812 * <= (B1 | !B2) | (CL1 | !CL2)
813 * which is wasteful if CL2 is small, but we ignore CL2:
814 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
815 * XXXX Can we handle case-fold? Unclear:
816 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
817 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
819 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
820 && !(or_with->flags & ANYOF_FOLD)
821 && !(cl->flags & ANYOF_FOLD) ) {
824 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
825 cl->bitmap[i] |= ~or_with->bitmap[i];
826 } /* XXXX: logic is complicated otherwise */
828 cl_anything(pRExC_state, cl);
831 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
832 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
833 && (!(or_with->flags & ANYOF_FOLD)
834 || (cl->flags & ANYOF_FOLD)) ) {
837 /* OR char bitmap and class bitmap separately */
838 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
839 cl->bitmap[i] |= or_with->bitmap[i];
840 if (or_with->flags & ANYOF_CLASS) {
841 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
842 cl->classflags[i] |= or_with->classflags[i];
843 cl->flags |= ANYOF_CLASS;
846 else { /* XXXX: logic is complicated, leave it along for a moment. */
847 cl_anything(pRExC_state, cl);
850 if (or_with->flags & ANYOF_EOS)
851 cl->flags |= ANYOF_EOS;
853 if (or_with->flags & ANYOF_FOLD)
854 cl->flags |= ANYOF_FOLD;
856 /* If both nodes match something outside the bitmap, but what they match
857 * outside is not the same pointer, and hence not easily compared, give up
858 * and allow the start class to match everything outside the bitmap */
859 if (cl->flags & ANYOF_NONBITMAP && or_with->flags & ANYOF_NONBITMAP &&
860 ARG(cl) != ARG(or_with)) {
861 cl->flags |= ANYOF_UNICODE_ALL;
864 if (or_with->flags & ANYOF_UNICODE_ALL) {
865 cl->flags |= ANYOF_UNICODE_ALL;
869 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
870 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
871 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
872 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
877 dump_trie(trie,widecharmap,revcharmap)
878 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
879 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
881 These routines dump out a trie in a somewhat readable format.
882 The _interim_ variants are used for debugging the interim
883 tables that are used to generate the final compressed
884 representation which is what dump_trie expects.
886 Part of the reason for their existance is to provide a form
887 of documentation as to how the different representations function.
892 Dumps the final compressed table form of the trie to Perl_debug_log.
893 Used for debugging make_trie().
897 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
898 AV *revcharmap, U32 depth)
901 SV *sv=sv_newmortal();
902 int colwidth= widecharmap ? 6 : 4;
904 GET_RE_DEBUG_FLAGS_DECL;
906 PERL_ARGS_ASSERT_DUMP_TRIE;
908 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
909 (int)depth * 2 + 2,"",
910 "Match","Base","Ofs" );
912 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
913 SV ** const tmp = av_fetch( revcharmap, state, 0);
915 PerlIO_printf( Perl_debug_log, "%*s",
917 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
918 PL_colors[0], PL_colors[1],
919 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
920 PERL_PV_ESCAPE_FIRSTCHAR
925 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
926 (int)depth * 2 + 2,"");
928 for( state = 0 ; state < trie->uniquecharcount ; state++ )
929 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
930 PerlIO_printf( Perl_debug_log, "\n");
932 for( state = 1 ; state < trie->statecount ; state++ ) {
933 const U32 base = trie->states[ state ].trans.base;
935 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
937 if ( trie->states[ state ].wordnum ) {
938 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
940 PerlIO_printf( Perl_debug_log, "%6s", "" );
943 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
948 while( ( base + ofs < trie->uniquecharcount ) ||
949 ( base + ofs - trie->uniquecharcount < trie->lasttrans
950 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
953 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
955 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
956 if ( ( base + ofs >= trie->uniquecharcount ) &&
957 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
958 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
960 PerlIO_printf( Perl_debug_log, "%*"UVXf,
962 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
964 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
968 PerlIO_printf( Perl_debug_log, "]");
971 PerlIO_printf( Perl_debug_log, "\n" );
973 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
974 for (word=1; word <= trie->wordcount; word++) {
975 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
976 (int)word, (int)(trie->wordinfo[word].prev),
977 (int)(trie->wordinfo[word].len));
979 PerlIO_printf(Perl_debug_log, "\n" );
982 Dumps a fully constructed but uncompressed trie in list form.
983 List tries normally only are used for construction when the number of
984 possible chars (trie->uniquecharcount) is very high.
985 Used for debugging make_trie().
988 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
989 HV *widecharmap, AV *revcharmap, U32 next_alloc,
993 SV *sv=sv_newmortal();
994 int colwidth= widecharmap ? 6 : 4;
995 GET_RE_DEBUG_FLAGS_DECL;
997 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
999 /* print out the table precompression. */
1000 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1001 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1002 "------:-----+-----------------\n" );
1004 for( state=1 ; state < next_alloc ; state ++ ) {
1007 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1008 (int)depth * 2 + 2,"", (UV)state );
1009 if ( ! trie->states[ state ].wordnum ) {
1010 PerlIO_printf( Perl_debug_log, "%5s| ","");
1012 PerlIO_printf( Perl_debug_log, "W%4x| ",
1013 trie->states[ state ].wordnum
1016 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1017 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1019 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1021 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1022 PL_colors[0], PL_colors[1],
1023 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1024 PERL_PV_ESCAPE_FIRSTCHAR
1026 TRIE_LIST_ITEM(state,charid).forid,
1027 (UV)TRIE_LIST_ITEM(state,charid).newstate
1030 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1031 (int)((depth * 2) + 14), "");
1034 PerlIO_printf( Perl_debug_log, "\n");
1039 Dumps a fully constructed but uncompressed trie in table form.
1040 This is the normal DFA style state transition table, with a few
1041 twists to facilitate compression later.
1042 Used for debugging make_trie().
1045 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1046 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1051 SV *sv=sv_newmortal();
1052 int colwidth= widecharmap ? 6 : 4;
1053 GET_RE_DEBUG_FLAGS_DECL;
1055 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1058 print out the table precompression so that we can do a visual check
1059 that they are identical.
1062 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1064 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1065 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1067 PerlIO_printf( Perl_debug_log, "%*s",
1069 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1070 PL_colors[0], PL_colors[1],
1071 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1072 PERL_PV_ESCAPE_FIRSTCHAR
1078 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1080 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1081 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1084 PerlIO_printf( Perl_debug_log, "\n" );
1086 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1088 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1089 (int)depth * 2 + 2,"",
1090 (UV)TRIE_NODENUM( state ) );
1092 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1093 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1095 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1097 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1099 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1100 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1102 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1103 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1111 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1112 startbranch: the first branch in the whole branch sequence
1113 first : start branch of sequence of branch-exact nodes.
1114 May be the same as startbranch
1115 last : Thing following the last branch.
1116 May be the same as tail.
1117 tail : item following the branch sequence
1118 count : words in the sequence
1119 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1120 depth : indent depth
1122 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1124 A trie is an N'ary tree where the branches are determined by digital
1125 decomposition of the key. IE, at the root node you look up the 1st character and
1126 follow that branch repeat until you find the end of the branches. Nodes can be
1127 marked as "accepting" meaning they represent a complete word. Eg:
1131 would convert into the following structure. Numbers represent states, letters
1132 following numbers represent valid transitions on the letter from that state, if
1133 the number is in square brackets it represents an accepting state, otherwise it
1134 will be in parenthesis.
1136 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1140 (1) +-i->(6)-+-s->[7]
1142 +-s->(3)-+-h->(4)-+-e->[5]
1144 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1146 This shows that when matching against the string 'hers' we will begin at state 1
1147 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1148 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1149 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1150 single traverse. We store a mapping from accepting to state to which word was
1151 matched, and then when we have multiple possibilities we try to complete the
1152 rest of the regex in the order in which they occured in the alternation.
1154 The only prior NFA like behaviour that would be changed by the TRIE support is
1155 the silent ignoring of duplicate alternations which are of the form:
1157 / (DUPE|DUPE) X? (?{ ... }) Y /x
1159 Thus EVAL blocks following a trie may be called a different number of times with
1160 and without the optimisation. With the optimisations dupes will be silently
1161 ignored. This inconsistant behaviour of EVAL type nodes is well established as
1162 the following demonstrates:
1164 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1166 which prints out 'word' three times, but
1168 'words'=~/(word|word|word)(?{ print $1 })S/
1170 which doesnt print it out at all. This is due to other optimisations kicking in.
1172 Example of what happens on a structural level:
1174 The regexp /(ac|ad|ab)+/ will produce the folowing debug output:
1176 1: CURLYM[1] {1,32767}(18)
1187 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1188 and should turn into:
1190 1: CURLYM[1] {1,32767}(18)
1192 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1200 Cases where tail != last would be like /(?foo|bar)baz/:
1210 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1211 and would end up looking like:
1214 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1221 d = uvuni_to_utf8_flags(d, uv, 0);
1223 is the recommended Unicode-aware way of saying
1228 #define TRIE_STORE_REVCHAR \
1231 SV *zlopp = newSV(2); \
1232 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1233 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, uvc & 0xFF); \
1234 SvCUR_set(zlopp, kapow - flrbbbbb); \
1237 av_push(revcharmap, zlopp); \
1239 char ooooff = (char)uvc; \
1240 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1244 #define TRIE_READ_CHAR STMT_START { \
1248 if ( foldlen > 0 ) { \
1249 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1254 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1255 uvc = to_uni_fold( uvc, foldbuf, &foldlen ); \
1256 foldlen -= UNISKIP( uvc ); \
1257 scan = foldbuf + UNISKIP( uvc ); \
1260 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1270 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1271 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1272 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1273 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1275 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1276 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1277 TRIE_LIST_CUR( state )++; \
1280 #define TRIE_LIST_NEW(state) STMT_START { \
1281 Newxz( trie->states[ state ].trans.list, \
1282 4, reg_trie_trans_le ); \
1283 TRIE_LIST_CUR( state ) = 1; \
1284 TRIE_LIST_LEN( state ) = 4; \
1287 #define TRIE_HANDLE_WORD(state) STMT_START { \
1288 U16 dupe= trie->states[ state ].wordnum; \
1289 regnode * const noper_next = regnext( noper ); \
1292 /* store the word for dumping */ \
1294 if (OP(noper) != NOTHING) \
1295 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1297 tmp = newSVpvn_utf8( "", 0, UTF ); \
1298 av_push( trie_words, tmp ); \
1302 trie->wordinfo[curword].prev = 0; \
1303 trie->wordinfo[curword].len = wordlen; \
1304 trie->wordinfo[curword].accept = state; \
1306 if ( noper_next < tail ) { \
1308 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1309 trie->jump[curword] = (U16)(noper_next - convert); \
1311 jumper = noper_next; \
1313 nextbranch= regnext(cur); \
1317 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1318 /* chain, so that when the bits of chain are later */\
1319 /* linked together, the dups appear in the chain */\
1320 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1321 trie->wordinfo[dupe].prev = curword; \
1323 /* we haven't inserted this word yet. */ \
1324 trie->states[ state ].wordnum = curword; \
1329 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1330 ( ( base + charid >= ucharcount \
1331 && base + charid < ubound \
1332 && state == trie->trans[ base - ucharcount + charid ].check \
1333 && trie->trans[ base - ucharcount + charid ].next ) \
1334 ? trie->trans[ base - ucharcount + charid ].next \
1335 : ( state==1 ? special : 0 ) \
1339 #define MADE_JUMP_TRIE 2
1340 #define MADE_EXACT_TRIE 4
1343 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1346 /* first pass, loop through and scan words */
1347 reg_trie_data *trie;
1348 HV *widecharmap = NULL;
1349 AV *revcharmap = newAV();
1351 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1356 regnode *jumper = NULL;
1357 regnode *nextbranch = NULL;
1358 regnode *convert = NULL;
1359 U32 *prev_states; /* temp array mapping each state to previous one */
1360 /* we just use folder as a flag in utf8 */
1361 const U8 * folder = NULL;
1364 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1365 AV *trie_words = NULL;
1366 /* along with revcharmap, this only used during construction but both are
1367 * useful during debugging so we store them in the struct when debugging.
1370 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1371 STRLEN trie_charcount=0;
1373 SV *re_trie_maxbuff;
1374 GET_RE_DEBUG_FLAGS_DECL;
1376 PERL_ARGS_ASSERT_MAKE_TRIE;
1378 PERL_UNUSED_ARG(depth);
1382 case EXACTFU: folder = PL_fold_latin1; break;
1383 case EXACTF: folder = PL_fold; break;
1384 case EXACTFL: folder = PL_fold_locale; break;
1387 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1389 trie->startstate = 1;
1390 trie->wordcount = word_count;
1391 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1392 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1393 if (!(UTF && folder))
1394 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1395 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1396 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1399 trie_words = newAV();
1402 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1403 if (!SvIOK(re_trie_maxbuff)) {
1404 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1407 PerlIO_printf( Perl_debug_log,
1408 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1409 (int)depth * 2 + 2, "",
1410 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1411 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1415 /* Find the node we are going to overwrite */
1416 if ( first == startbranch && OP( last ) != BRANCH ) {
1417 /* whole branch chain */
1420 /* branch sub-chain */
1421 convert = NEXTOPER( first );
1424 /* -- First loop and Setup --
1426 We first traverse the branches and scan each word to determine if it
1427 contains widechars, and how many unique chars there are, this is
1428 important as we have to build a table with at least as many columns as we
1431 We use an array of integers to represent the character codes 0..255
1432 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1433 native representation of the character value as the key and IV's for the
1436 *TODO* If we keep track of how many times each character is used we can
1437 remap the columns so that the table compression later on is more
1438 efficient in terms of memory by ensuring the most common value is in the
1439 middle and the least common are on the outside. IMO this would be better
1440 than a most to least common mapping as theres a decent chance the most
1441 common letter will share a node with the least common, meaning the node
1442 will not be compressable. With a middle is most common approach the worst
1443 case is when we have the least common nodes twice.
1447 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1448 regnode * const noper = NEXTOPER( cur );
1449 const U8 *uc = (U8*)STRING( noper );
1450 const U8 * const e = uc + STR_LEN( noper );
1452 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1453 const U8 *scan = (U8*)NULL;
1454 U32 wordlen = 0; /* required init */
1456 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1458 if (OP(noper) == NOTHING) {
1462 if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
1463 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1464 regardless of encoding */
1466 for ( ; uc < e ; uc += len ) {
1467 TRIE_CHARCOUNT(trie)++;
1471 if ( !trie->charmap[ uvc ] ) {
1472 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1474 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
1478 /* store the codepoint in the bitmap, and its folded
1480 TRIE_BITMAP_SET(trie,uvc);
1482 /* store the folded codepoint */
1483 if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);
1486 /* store first byte of utf8 representation of
1487 codepoints in the 127 < uvc < 256 range */
1488 if (127 < uvc && uvc < 192) {
1489 TRIE_BITMAP_SET(trie,194);
1490 } else if (191 < uvc ) {
1491 TRIE_BITMAP_SET(trie,195);
1492 /* && uvc < 256 -- we know uvc is < 256 already */
1495 set_bit = 0; /* We've done our bit :-) */
1500 widecharmap = newHV();
1502 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1505 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1507 if ( !SvTRUE( *svpp ) ) {
1508 sv_setiv( *svpp, ++trie->uniquecharcount );
1513 if( cur == first ) {
1516 } else if (chars < trie->minlen) {
1518 } else if (chars > trie->maxlen) {
1522 } /* end first pass */
1523 DEBUG_TRIE_COMPILE_r(
1524 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1525 (int)depth * 2 + 2,"",
1526 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1527 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1528 (int)trie->minlen, (int)trie->maxlen )
1532 We now know what we are dealing with in terms of unique chars and
1533 string sizes so we can calculate how much memory a naive
1534 representation using a flat table will take. If it's over a reasonable
1535 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1536 conservative but potentially much slower representation using an array
1539 At the end we convert both representations into the same compressed
1540 form that will be used in regexec.c for matching with. The latter
1541 is a form that cannot be used to construct with but has memory
1542 properties similar to the list form and access properties similar
1543 to the table form making it both suitable for fast searches and
1544 small enough that its feasable to store for the duration of a program.
1546 See the comment in the code where the compressed table is produced
1547 inplace from the flat tabe representation for an explanation of how
1548 the compression works.
1553 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1556 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1558 Second Pass -- Array Of Lists Representation
1560 Each state will be represented by a list of charid:state records
1561 (reg_trie_trans_le) the first such element holds the CUR and LEN
1562 points of the allocated array. (See defines above).
1564 We build the initial structure using the lists, and then convert
1565 it into the compressed table form which allows faster lookups
1566 (but cant be modified once converted).
1569 STRLEN transcount = 1;
1571 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1572 "%*sCompiling trie using list compiler\n",
1573 (int)depth * 2 + 2, ""));
1575 trie->states = (reg_trie_state *)
1576 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1577 sizeof(reg_trie_state) );
1581 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1583 regnode * const noper = NEXTOPER( cur );
1584 U8 *uc = (U8*)STRING( noper );
1585 const U8 * const e = uc + STR_LEN( noper );
1586 U32 state = 1; /* required init */
1587 U16 charid = 0; /* sanity init */
1588 U8 *scan = (U8*)NULL; /* sanity init */
1589 STRLEN foldlen = 0; /* required init */
1590 U32 wordlen = 0; /* required init */
1591 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1593 if (OP(noper) != NOTHING) {
1594 for ( ; uc < e ; uc += len ) {
1599 charid = trie->charmap[ uvc ];
1601 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1605 charid=(U16)SvIV( *svpp );
1608 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1615 if ( !trie->states[ state ].trans.list ) {
1616 TRIE_LIST_NEW( state );
1618 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1619 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1620 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1625 newstate = next_alloc++;
1626 prev_states[newstate] = state;
1627 TRIE_LIST_PUSH( state, charid, newstate );
1632 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1636 TRIE_HANDLE_WORD(state);
1638 } /* end second pass */
1640 /* next alloc is the NEXT state to be allocated */
1641 trie->statecount = next_alloc;
1642 trie->states = (reg_trie_state *)
1643 PerlMemShared_realloc( trie->states,
1645 * sizeof(reg_trie_state) );
1647 /* and now dump it out before we compress it */
1648 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1649 revcharmap, next_alloc,
1653 trie->trans = (reg_trie_trans *)
1654 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1661 for( state=1 ; state < next_alloc ; state ++ ) {
1665 DEBUG_TRIE_COMPILE_MORE_r(
1666 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1670 if (trie->states[state].trans.list) {
1671 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1675 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1676 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1677 if ( forid < minid ) {
1679 } else if ( forid > maxid ) {
1683 if ( transcount < tp + maxid - minid + 1) {
1685 trie->trans = (reg_trie_trans *)
1686 PerlMemShared_realloc( trie->trans,
1688 * sizeof(reg_trie_trans) );
1689 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1691 base = trie->uniquecharcount + tp - minid;
1692 if ( maxid == minid ) {
1694 for ( ; zp < tp ; zp++ ) {
1695 if ( ! trie->trans[ zp ].next ) {
1696 base = trie->uniquecharcount + zp - minid;
1697 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1698 trie->trans[ zp ].check = state;
1704 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1705 trie->trans[ tp ].check = state;
1710 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1711 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1712 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1713 trie->trans[ tid ].check = state;
1715 tp += ( maxid - minid + 1 );
1717 Safefree(trie->states[ state ].trans.list);
1720 DEBUG_TRIE_COMPILE_MORE_r(
1721 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1724 trie->states[ state ].trans.base=base;
1726 trie->lasttrans = tp + 1;
1730 Second Pass -- Flat Table Representation.
1732 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1733 We know that we will need Charcount+1 trans at most to store the data
1734 (one row per char at worst case) So we preallocate both structures
1735 assuming worst case.
1737 We then construct the trie using only the .next slots of the entry
1740 We use the .check field of the first entry of the node temporarily to
1741 make compression both faster and easier by keeping track of how many non
1742 zero fields are in the node.
1744 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1747 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1748 number representing the first entry of the node, and state as a
1749 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1750 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1751 are 2 entrys per node. eg:
1759 The table is internally in the right hand, idx form. However as we also
1760 have to deal with the states array which is indexed by nodenum we have to
1761 use TRIE_NODENUM() to convert.
1764 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1765 "%*sCompiling trie using table compiler\n",
1766 (int)depth * 2 + 2, ""));
1768 trie->trans = (reg_trie_trans *)
1769 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1770 * trie->uniquecharcount + 1,
1771 sizeof(reg_trie_trans) );
1772 trie->states = (reg_trie_state *)
1773 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1774 sizeof(reg_trie_state) );
1775 next_alloc = trie->uniquecharcount + 1;
1778 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1780 regnode * const noper = NEXTOPER( cur );
1781 const U8 *uc = (U8*)STRING( noper );
1782 const U8 * const e = uc + STR_LEN( noper );
1784 U32 state = 1; /* required init */
1786 U16 charid = 0; /* sanity init */
1787 U32 accept_state = 0; /* sanity init */
1788 U8 *scan = (U8*)NULL; /* sanity init */
1790 STRLEN foldlen = 0; /* required init */
1791 U32 wordlen = 0; /* required init */
1792 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1794 if ( OP(noper) != NOTHING ) {
1795 for ( ; uc < e ; uc += len ) {
1800 charid = trie->charmap[ uvc ];
1802 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1803 charid = svpp ? (U16)SvIV(*svpp) : 0;
1807 if ( !trie->trans[ state + charid ].next ) {
1808 trie->trans[ state + charid ].next = next_alloc;
1809 trie->trans[ state ].check++;
1810 prev_states[TRIE_NODENUM(next_alloc)]
1811 = TRIE_NODENUM(state);
1812 next_alloc += trie->uniquecharcount;
1814 state = trie->trans[ state + charid ].next;
1816 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1818 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1821 accept_state = TRIE_NODENUM( state );
1822 TRIE_HANDLE_WORD(accept_state);
1824 } /* end second pass */
1826 /* and now dump it out before we compress it */
1827 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1829 next_alloc, depth+1));
1833 * Inplace compress the table.*
1835 For sparse data sets the table constructed by the trie algorithm will
1836 be mostly 0/FAIL transitions or to put it another way mostly empty.
1837 (Note that leaf nodes will not contain any transitions.)
1839 This algorithm compresses the tables by eliminating most such
1840 transitions, at the cost of a modest bit of extra work during lookup:
1842 - Each states[] entry contains a .base field which indicates the
1843 index in the state[] array wheres its transition data is stored.
1845 - If .base is 0 there are no valid transitions from that node.
1847 - If .base is nonzero then charid is added to it to find an entry in
1850 -If trans[states[state].base+charid].check!=state then the
1851 transition is taken to be a 0/Fail transition. Thus if there are fail
1852 transitions at the front of the node then the .base offset will point
1853 somewhere inside the previous nodes data (or maybe even into a node
1854 even earlier), but the .check field determines if the transition is
1858 The following process inplace converts the table to the compressed
1859 table: We first do not compress the root node 1,and mark all its
1860 .check pointers as 1 and set its .base pointer as 1 as well. This
1861 allows us to do a DFA construction from the compressed table later,
1862 and ensures that any .base pointers we calculate later are greater
1865 - We set 'pos' to indicate the first entry of the second node.
1867 - We then iterate over the columns of the node, finding the first and
1868 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
1869 and set the .check pointers accordingly, and advance pos
1870 appropriately and repreat for the next node. Note that when we copy
1871 the next pointers we have to convert them from the original
1872 NODEIDX form to NODENUM form as the former is not valid post
1875 - If a node has no transitions used we mark its base as 0 and do not
1876 advance the pos pointer.
1878 - If a node only has one transition we use a second pointer into the
1879 structure to fill in allocated fail transitions from other states.
1880 This pointer is independent of the main pointer and scans forward
1881 looking for null transitions that are allocated to a state. When it
1882 finds one it writes the single transition into the "hole". If the
1883 pointer doesnt find one the single transition is appended as normal.
1885 - Once compressed we can Renew/realloc the structures to release the
1888 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
1889 specifically Fig 3.47 and the associated pseudocode.
1893 const U32 laststate = TRIE_NODENUM( next_alloc );
1896 trie->statecount = laststate;
1898 for ( state = 1 ; state < laststate ; state++ ) {
1900 const U32 stateidx = TRIE_NODEIDX( state );
1901 const U32 o_used = trie->trans[ stateidx ].check;
1902 U32 used = trie->trans[ stateidx ].check;
1903 trie->trans[ stateidx ].check = 0;
1905 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
1906 if ( flag || trie->trans[ stateidx + charid ].next ) {
1907 if ( trie->trans[ stateidx + charid ].next ) {
1909 for ( ; zp < pos ; zp++ ) {
1910 if ( ! trie->trans[ zp ].next ) {
1914 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
1915 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1916 trie->trans[ zp ].check = state;
1917 if ( ++zp > pos ) pos = zp;
1924 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
1926 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1927 trie->trans[ pos ].check = state;
1932 trie->lasttrans = pos + 1;
1933 trie->states = (reg_trie_state *)
1934 PerlMemShared_realloc( trie->states, laststate
1935 * sizeof(reg_trie_state) );
1936 DEBUG_TRIE_COMPILE_MORE_r(
1937 PerlIO_printf( Perl_debug_log,
1938 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
1939 (int)depth * 2 + 2,"",
1940 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
1943 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
1946 } /* end table compress */
1948 DEBUG_TRIE_COMPILE_MORE_r(
1949 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
1950 (int)depth * 2 + 2, "",
1951 (UV)trie->statecount,
1952 (UV)trie->lasttrans)
1954 /* resize the trans array to remove unused space */
1955 trie->trans = (reg_trie_trans *)
1956 PerlMemShared_realloc( trie->trans, trie->lasttrans
1957 * sizeof(reg_trie_trans) );
1959 { /* Modify the program and insert the new TRIE node */
1960 U8 nodetype =(U8)(flags & 0xFF);
1964 regnode *optimize = NULL;
1965 #ifdef RE_TRACK_PATTERN_OFFSETS
1968 U32 mjd_nodelen = 0;
1969 #endif /* RE_TRACK_PATTERN_OFFSETS */
1970 #endif /* DEBUGGING */
1972 This means we convert either the first branch or the first Exact,
1973 depending on whether the thing following (in 'last') is a branch
1974 or not and whther first is the startbranch (ie is it a sub part of
1975 the alternation or is it the whole thing.)
1976 Assuming its a sub part we convert the EXACT otherwise we convert
1977 the whole branch sequence, including the first.
1979 /* Find the node we are going to overwrite */
1980 if ( first != startbranch || OP( last ) == BRANCH ) {
1981 /* branch sub-chain */
1982 NEXT_OFF( first ) = (U16)(last - first);
1983 #ifdef RE_TRACK_PATTERN_OFFSETS
1985 mjd_offset= Node_Offset((convert));
1986 mjd_nodelen= Node_Length((convert));
1989 /* whole branch chain */
1991 #ifdef RE_TRACK_PATTERN_OFFSETS
1994 const regnode *nop = NEXTOPER( convert );
1995 mjd_offset= Node_Offset((nop));
1996 mjd_nodelen= Node_Length((nop));
2000 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2001 (int)depth * 2 + 2, "",
2002 (UV)mjd_offset, (UV)mjd_nodelen)
2005 /* But first we check to see if there is a common prefix we can
2006 split out as an EXACT and put in front of the TRIE node. */
2007 trie->startstate= 1;
2008 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2010 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2014 const U32 base = trie->states[ state ].trans.base;
2016 if ( trie->states[state].wordnum )
2019 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2020 if ( ( base + ofs >= trie->uniquecharcount ) &&
2021 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2022 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2024 if ( ++count > 1 ) {
2025 SV **tmp = av_fetch( revcharmap, ofs, 0);
2026 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2027 if ( state == 1 ) break;
2029 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2031 PerlIO_printf(Perl_debug_log,
2032 "%*sNew Start State=%"UVuf" Class: [",
2033 (int)depth * 2 + 2, "",
2036 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2037 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2039 TRIE_BITMAP_SET(trie,*ch);
2041 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2043 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2047 TRIE_BITMAP_SET(trie,*ch);
2049 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2050 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2056 SV **tmp = av_fetch( revcharmap, idx, 0);
2058 char *ch = SvPV( *tmp, len );
2060 SV *sv=sv_newmortal();
2061 PerlIO_printf( Perl_debug_log,
2062 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2063 (int)depth * 2 + 2, "",
2065 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2066 PL_colors[0], PL_colors[1],
2067 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2068 PERL_PV_ESCAPE_FIRSTCHAR
2073 OP( convert ) = nodetype;
2074 str=STRING(convert);
2077 STR_LEN(convert) += len;
2083 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2088 trie->prefixlen = (state-1);
2090 regnode *n = convert+NODE_SZ_STR(convert);
2091 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2092 trie->startstate = state;
2093 trie->minlen -= (state - 1);
2094 trie->maxlen -= (state - 1);
2096 /* At least the UNICOS C compiler choked on this
2097 * being argument to DEBUG_r(), so let's just have
2100 #ifdef PERL_EXT_RE_BUILD
2106 regnode *fix = convert;
2107 U32 word = trie->wordcount;
2109 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2110 while( ++fix < n ) {
2111 Set_Node_Offset_Length(fix, 0, 0);
2114 SV ** const tmp = av_fetch( trie_words, word, 0 );
2116 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2117 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2119 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2127 NEXT_OFF(convert) = (U16)(tail - convert);
2128 DEBUG_r(optimize= n);
2134 if ( trie->maxlen ) {
2135 NEXT_OFF( convert ) = (U16)(tail - convert);
2136 ARG_SET( convert, data_slot );
2137 /* Store the offset to the first unabsorbed branch in
2138 jump[0], which is otherwise unused by the jump logic.
2139 We use this when dumping a trie and during optimisation. */
2141 trie->jump[0] = (U16)(nextbranch - convert);
2143 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2144 * and there is a bitmap
2145 * and the first "jump target" node we found leaves enough room
2146 * then convert the TRIE node into a TRIEC node, with the bitmap
2147 * embedded inline in the opcode - this is hypothetically faster.
2149 if ( !trie->states[trie->startstate].wordnum
2151 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2153 OP( convert ) = TRIEC;
2154 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2155 PerlMemShared_free(trie->bitmap);
2158 OP( convert ) = TRIE;
2160 /* store the type in the flags */
2161 convert->flags = nodetype;
2165 + regarglen[ OP( convert ) ];
2167 /* XXX We really should free up the resource in trie now,
2168 as we won't use them - (which resources?) dmq */
2170 /* needed for dumping*/
2171 DEBUG_r(if (optimize) {
2172 regnode *opt = convert;
2174 while ( ++opt < optimize) {
2175 Set_Node_Offset_Length(opt,0,0);
2178 Try to clean up some of the debris left after the
2181 while( optimize < jumper ) {
2182 mjd_nodelen += Node_Length((optimize));
2183 OP( optimize ) = OPTIMIZED;
2184 Set_Node_Offset_Length(optimize,0,0);
2187 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2189 } /* end node insert */
2191 /* Finish populating the prev field of the wordinfo array. Walk back
2192 * from each accept state until we find another accept state, and if
2193 * so, point the first word's .prev field at the second word. If the
2194 * second already has a .prev field set, stop now. This will be the
2195 * case either if we've already processed that word's accept state,
2196 * or that state had multiple words, and the overspill words were
2197 * already linked up earlier.
2204 for (word=1; word <= trie->wordcount; word++) {
2206 if (trie->wordinfo[word].prev)
2208 state = trie->wordinfo[word].accept;
2210 state = prev_states[state];
2213 prev = trie->states[state].wordnum;
2217 trie->wordinfo[word].prev = prev;
2219 Safefree(prev_states);
2223 /* and now dump out the compressed format */
2224 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2226 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2228 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2229 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2231 SvREFCNT_dec(revcharmap);
2235 : trie->startstate>1
2241 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2243 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2245 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2246 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2249 We find the fail state for each state in the trie, this state is the longest proper
2250 suffix of the current state's 'word' that is also a proper prefix of another word in our
2251 trie. State 1 represents the word '' and is thus the default fail state. This allows
2252 the DFA not to have to restart after its tried and failed a word at a given point, it
2253 simply continues as though it had been matching the other word in the first place.
2255 'abcdgu'=~/abcdefg|cdgu/
2256 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2257 fail, which would bring us to the state representing 'd' in the second word where we would
2258 try 'g' and succeed, proceeding to match 'cdgu'.
2260 /* add a fail transition */
2261 const U32 trie_offset = ARG(source);
2262 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2264 const U32 ucharcount = trie->uniquecharcount;
2265 const U32 numstates = trie->statecount;
2266 const U32 ubound = trie->lasttrans + ucharcount;
2270 U32 base = trie->states[ 1 ].trans.base;
2273 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2274 GET_RE_DEBUG_FLAGS_DECL;
2276 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2278 PERL_UNUSED_ARG(depth);
2282 ARG_SET( stclass, data_slot );
2283 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2284 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2285 aho->trie=trie_offset;
2286 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2287 Copy( trie->states, aho->states, numstates, reg_trie_state );
2288 Newxz( q, numstates, U32);
2289 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2292 /* initialize fail[0..1] to be 1 so that we always have
2293 a valid final fail state */
2294 fail[ 0 ] = fail[ 1 ] = 1;
2296 for ( charid = 0; charid < ucharcount ; charid++ ) {
2297 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2299 q[ q_write ] = newstate;
2300 /* set to point at the root */
2301 fail[ q[ q_write++ ] ]=1;
2304 while ( q_read < q_write) {
2305 const U32 cur = q[ q_read++ % numstates ];
2306 base = trie->states[ cur ].trans.base;
2308 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2309 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2311 U32 fail_state = cur;
2314 fail_state = fail[ fail_state ];
2315 fail_base = aho->states[ fail_state ].trans.base;
2316 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2318 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2319 fail[ ch_state ] = fail_state;
2320 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2322 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2324 q[ q_write++ % numstates] = ch_state;
2328 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2329 when we fail in state 1, this allows us to use the
2330 charclass scan to find a valid start char. This is based on the principle
2331 that theres a good chance the string being searched contains lots of stuff
2332 that cant be a start char.
2334 fail[ 0 ] = fail[ 1 ] = 0;
2335 DEBUG_TRIE_COMPILE_r({
2336 PerlIO_printf(Perl_debug_log,
2337 "%*sStclass Failtable (%"UVuf" states): 0",
2338 (int)(depth * 2), "", (UV)numstates
2340 for( q_read=1; q_read<numstates; q_read++ ) {
2341 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2343 PerlIO_printf(Perl_debug_log, "\n");
2346 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2351 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2352 * These need to be revisited when a newer toolchain becomes available.
2354 #if defined(__sparc64__) && defined(__GNUC__)
2355 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2356 # undef SPARC64_GCC_WORKAROUND
2357 # define SPARC64_GCC_WORKAROUND 1
2361 #define DEBUG_PEEP(str,scan,depth) \
2362 DEBUG_OPTIMISE_r({if (scan){ \
2363 SV * const mysv=sv_newmortal(); \
2364 regnode *Next = regnext(scan); \
2365 regprop(RExC_rx, mysv, scan); \
2366 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2367 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2368 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2375 #define JOIN_EXACT(scan,min,flags) \
2376 if (PL_regkind[OP(scan)] == EXACT) \
2377 join_exact(pRExC_state,(scan),(min),(flags),NULL,depth+1)
2380 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, I32 *min, U32 flags,regnode *val, U32 depth) {
2381 /* Merge several consecutive EXACTish nodes into one. */
2382 regnode *n = regnext(scan);
2384 regnode *next = scan + NODE_SZ_STR(scan);
2388 regnode *stop = scan;
2389 GET_RE_DEBUG_FLAGS_DECL;
2391 PERL_UNUSED_ARG(depth);
2394 PERL_ARGS_ASSERT_JOIN_EXACT;
2395 #ifndef EXPERIMENTAL_INPLACESCAN
2396 PERL_UNUSED_ARG(flags);
2397 PERL_UNUSED_ARG(val);
2399 DEBUG_PEEP("join",scan,depth);
2401 /* Skip NOTHING, merge EXACT*. */
2403 ( PL_regkind[OP(n)] == NOTHING ||
2404 (stringok && (OP(n) == OP(scan))))
2406 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX) {
2408 if (OP(n) == TAIL || n > next)
2410 if (PL_regkind[OP(n)] == NOTHING) {
2411 DEBUG_PEEP("skip:",n,depth);
2412 NEXT_OFF(scan) += NEXT_OFF(n);
2413 next = n + NODE_STEP_REGNODE;
2420 else if (stringok) {
2421 const unsigned int oldl = STR_LEN(scan);
2422 regnode * const nnext = regnext(n);
2424 DEBUG_PEEP("merg",n,depth);
2427 if (oldl + STR_LEN(n) > U8_MAX)
2429 NEXT_OFF(scan) += NEXT_OFF(n);
2430 STR_LEN(scan) += STR_LEN(n);
2431 next = n + NODE_SZ_STR(n);
2432 /* Now we can overwrite *n : */
2433 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2441 #ifdef EXPERIMENTAL_INPLACESCAN
2442 if (flags && !NEXT_OFF(n)) {
2443 DEBUG_PEEP("atch", val, depth);
2444 if (reg_off_by_arg[OP(n)]) {
2445 ARG_SET(n, val - n);
2448 NEXT_OFF(n) = val - n;
2456 && ( OP(scan) == EXACTF || OP(scan) == EXACTFU)
2457 && ( STR_LEN(scan) >= 6 ) )
2460 Two problematic code points in Unicode casefolding of EXACT nodes:
2462 U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2463 U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2469 U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2470 U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2472 This means that in case-insensitive matching (or "loose matching",
2473 as Unicode calls it), an EXACTF of length six (the UTF-8 encoded byte
2474 length of the above casefolded versions) can match a target string
2475 of length two (the byte length of UTF-8 encoded U+0390 or U+03B0).
2476 This would rather mess up the minimum length computation.
2478 What we'll do is to look for the tail four bytes, and then peek
2479 at the preceding two bytes to see whether we need to decrease
2480 the minimum length by four (six minus two).
2482 Thanks to the design of UTF-8, there cannot be false matches:
2483 A sequence of valid UTF-8 bytes cannot be a subsequence of
2484 another valid sequence of UTF-8 bytes.
2487 char * const s0 = STRING(scan), *s, *t;
2488 char * const s1 = s0 + STR_LEN(scan) - 1;
2489 char * const s2 = s1 - 4;
2490 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2491 const char t0[] = "\xaf\x49\xaf\x42";
2493 const char t0[] = "\xcc\x88\xcc\x81";
2495 const char * const t1 = t0 + 3;
2498 s < s2 && (t = ninstr(s, s1, t0, t1));
2501 if (((U8)t[-1] == 0x68 && (U8)t[-2] == 0xB4) ||
2502 ((U8)t[-1] == 0x46 && (U8)t[-2] == 0xB5))
2504 if (((U8)t[-1] == 0xB9 && (U8)t[-2] == 0xCE) ||
2505 ((U8)t[-1] == 0x85 && (U8)t[-2] == 0xCF))
2513 n = scan + NODE_SZ_STR(scan);
2515 if (PL_regkind[OP(n)] != NOTHING || OP(n) == NOTHING) {
2522 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2526 /* REx optimizer. Converts nodes into quickier variants "in place".
2527 Finds fixed substrings. */
2529 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2530 to the position after last scanned or to NULL. */
2532 #define INIT_AND_WITHP \
2533 assert(!and_withp); \
2534 Newx(and_withp,1,struct regnode_charclass_class); \
2535 SAVEFREEPV(and_withp)
2537 /* this is a chain of data about sub patterns we are processing that
2538 need to be handled seperately/specially in study_chunk. Its so
2539 we can simulate recursion without losing state. */
2541 typedef struct scan_frame {
2542 regnode *last; /* last node to process in this frame */
2543 regnode *next; /* next node to process when last is reached */
2544 struct scan_frame *prev; /*previous frame*/
2545 I32 stop; /* what stopparen do we use */
2549 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2551 #define CASE_SYNST_FNC(nAmE) \
2553 if (flags & SCF_DO_STCLASS_AND) { \
2554 for (value = 0; value < 256; value++) \
2555 if (!is_ ## nAmE ## _cp(value)) \
2556 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2559 for (value = 0; value < 256; value++) \
2560 if (is_ ## nAmE ## _cp(value)) \
2561 ANYOF_BITMAP_SET(data->start_class, value); \
2565 if (flags & SCF_DO_STCLASS_AND) { \
2566 for (value = 0; value < 256; value++) \
2567 if (is_ ## nAmE ## _cp(value)) \
2568 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2571 for (value = 0; value < 256; value++) \
2572 if (!is_ ## nAmE ## _cp(value)) \
2573 ANYOF_BITMAP_SET(data->start_class, value); \
2580 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2581 I32 *minlenp, I32 *deltap,
2586 struct regnode_charclass_class *and_withp,
2587 U32 flags, U32 depth)
2588 /* scanp: Start here (read-write). */
2589 /* deltap: Write maxlen-minlen here. */
2590 /* last: Stop before this one. */
2591 /* data: string data about the pattern */
2592 /* stopparen: treat close N as END */
2593 /* recursed: which subroutines have we recursed into */
2594 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
2597 I32 min = 0, pars = 0, code;
2598 regnode *scan = *scanp, *next;
2600 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2601 int is_inf_internal = 0; /* The studied chunk is infinite */
2602 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2603 scan_data_t data_fake;
2604 SV *re_trie_maxbuff = NULL;
2605 regnode *first_non_open = scan;
2606 I32 stopmin = I32_MAX;
2607 scan_frame *frame = NULL;
2608 GET_RE_DEBUG_FLAGS_DECL;
2610 PERL_ARGS_ASSERT_STUDY_CHUNK;
2613 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
2617 while (first_non_open && OP(first_non_open) == OPEN)
2618 first_non_open=regnext(first_non_open);
2623 while ( scan && OP(scan) != END && scan < last ){
2624 /* Peephole optimizer: */
2625 DEBUG_STUDYDATA("Peep:", data,depth);
2626 DEBUG_PEEP("Peep",scan,depth);
2627 JOIN_EXACT(scan,&min,0);
2629 /* Follow the next-chain of the current node and optimize
2630 away all the NOTHINGs from it. */
2631 if (OP(scan) != CURLYX) {
2632 const int max = (reg_off_by_arg[OP(scan)]
2634 /* I32 may be smaller than U16 on CRAYs! */
2635 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
2636 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
2640 /* Skip NOTHING and LONGJMP. */
2641 while ((n = regnext(n))
2642 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
2643 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
2644 && off + noff < max)
2646 if (reg_off_by_arg[OP(scan)])
2649 NEXT_OFF(scan) = off;
2654 /* The principal pseudo-switch. Cannot be a switch, since we
2655 look into several different things. */
2656 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
2657 || OP(scan) == IFTHEN) {
2658 next = regnext(scan);
2660 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
2662 if (OP(next) == code || code == IFTHEN) {
2663 /* NOTE - There is similar code to this block below for handling
2664 TRIE nodes on a re-study. If you change stuff here check there
2666 I32 max1 = 0, min1 = I32_MAX, num = 0;
2667 struct regnode_charclass_class accum;
2668 regnode * const startbranch=scan;
2670 if (flags & SCF_DO_SUBSTR)
2671 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
2672 if (flags & SCF_DO_STCLASS)
2673 cl_init_zero(pRExC_state, &accum);
2675 while (OP(scan) == code) {
2676 I32 deltanext, minnext, f = 0, fake;
2677 struct regnode_charclass_class this_class;
2680 data_fake.flags = 0;
2682 data_fake.whilem_c = data->whilem_c;
2683 data_fake.last_closep = data->last_closep;
2686 data_fake.last_closep = &fake;
2688 data_fake.pos_delta = delta;
2689 next = regnext(scan);
2690 scan = NEXTOPER(scan);
2692 scan = NEXTOPER(scan);
2693 if (flags & SCF_DO_STCLASS) {
2694 cl_init(pRExC_state, &this_class);
2695 data_fake.start_class = &this_class;
2696 f = SCF_DO_STCLASS_AND;
2698 if (flags & SCF_WHILEM_VISITED_POS)
2699 f |= SCF_WHILEM_VISITED_POS;
2701 /* we suppose the run is continuous, last=next...*/
2702 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
2704 stopparen, recursed, NULL, f,depth+1);
2707 if (max1 < minnext + deltanext)
2708 max1 = minnext + deltanext;
2709 if (deltanext == I32_MAX)
2710 is_inf = is_inf_internal = 1;
2712 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
2714 if (data_fake.flags & SCF_SEEN_ACCEPT) {
2715 if ( stopmin > minnext)
2716 stopmin = min + min1;
2717 flags &= ~SCF_DO_SUBSTR;
2719 data->flags |= SCF_SEEN_ACCEPT;
2722 if (data_fake.flags & SF_HAS_EVAL)
2723 data->flags |= SF_HAS_EVAL;
2724 data->whilem_c = data_fake.whilem_c;
2726 if (flags & SCF_DO_STCLASS)
2727 cl_or(pRExC_state, &accum, &this_class);
2729 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
2731 if (flags & SCF_DO_SUBSTR) {
2732 data->pos_min += min1;
2733 data->pos_delta += max1 - min1;
2734 if (max1 != min1 || is_inf)
2735 data->longest = &(data->longest_float);
2738 delta += max1 - min1;
2739 if (flags & SCF_DO_STCLASS_OR) {
2740 cl_or(pRExC_state, data->start_class, &accum);
2742 cl_and(data->start_class, and_withp);
2743 flags &= ~SCF_DO_STCLASS;
2746 else if (flags & SCF_DO_STCLASS_AND) {
2748 cl_and(data->start_class, &accum);
2749 flags &= ~SCF_DO_STCLASS;
2752 /* Switch to OR mode: cache the old value of
2753 * data->start_class */
2755 StructCopy(data->start_class, and_withp,
2756 struct regnode_charclass_class);
2757 flags &= ~SCF_DO_STCLASS_AND;
2758 StructCopy(&accum, data->start_class,
2759 struct regnode_charclass_class);
2760 flags |= SCF_DO_STCLASS_OR;
2761 data->start_class->flags |= ANYOF_EOS;
2765 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
2768 Assuming this was/is a branch we are dealing with: 'scan' now
2769 points at the item that follows the branch sequence, whatever
2770 it is. We now start at the beginning of the sequence and look
2777 which would be constructed from a pattern like /A|LIST|OF|WORDS/
2779 If we can find such a subseqence we need to turn the first
2780 element into a trie and then add the subsequent branch exact
2781 strings to the trie.
2785 1. patterns where the whole set of branches can be converted.
2787 2. patterns where only a subset can be converted.
2789 In case 1 we can replace the whole set with a single regop
2790 for the trie. In case 2 we need to keep the start and end
2793 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
2794 becomes BRANCH TRIE; BRANCH X;
2796 There is an additional case, that being where there is a
2797 common prefix, which gets split out into an EXACT like node
2798 preceding the TRIE node.
2800 If x(1..n)==tail then we can do a simple trie, if not we make
2801 a "jump" trie, such that when we match the appropriate word
2802 we "jump" to the appopriate tail node. Essentailly we turn
2803 a nested if into a case structure of sorts.
2808 if (!re_trie_maxbuff) {
2809 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2810 if (!SvIOK(re_trie_maxbuff))
2811 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2813 if ( SvIV(re_trie_maxbuff)>=0 ) {
2815 regnode *first = (regnode *)NULL;
2816 regnode *last = (regnode *)NULL;
2817 regnode *tail = scan;
2822 SV * const mysv = sv_newmortal(); /* for dumping */
2824 /* var tail is used because there may be a TAIL
2825 regop in the way. Ie, the exacts will point to the
2826 thing following the TAIL, but the last branch will
2827 point at the TAIL. So we advance tail. If we
2828 have nested (?:) we may have to move through several
2832 while ( OP( tail ) == TAIL ) {
2833 /* this is the TAIL generated by (?:) */
2834 tail = regnext( tail );
2839 regprop(RExC_rx, mysv, tail );
2840 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
2841 (int)depth * 2 + 2, "",
2842 "Looking for TRIE'able sequences. Tail node is: ",
2843 SvPV_nolen_const( mysv )
2849 step through the branches, cur represents each
2850 branch, noper is the first thing to be matched
2851 as part of that branch and noper_next is the
2852 regnext() of that node. if noper is an EXACT
2853 and noper_next is the same as scan (our current
2854 position in the regex) then the EXACT branch is
2855 a possible optimization target. Once we have
2856 two or more consequetive such branches we can
2857 create a trie of the EXACT's contents and stich
2858 it in place. If the sequence represents all of
2859 the branches we eliminate the whole thing and
2860 replace it with a single TRIE. If it is a
2861 subsequence then we need to stitch it in. This
2862 means the first branch has to remain, and needs
2863 to be repointed at the item on the branch chain
2864 following the last branch optimized. This could
2865 be either a BRANCH, in which case the
2866 subsequence is internal, or it could be the
2867 item following the branch sequence in which
2868 case the subsequence is at the end.
2872 /* dont use tail as the end marker for this traverse */
2873 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
2874 regnode * const noper = NEXTOPER( cur );
2875 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
2876 regnode * const noper_next = regnext( noper );
2880 regprop(RExC_rx, mysv, cur);
2881 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
2882 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
2884 regprop(RExC_rx, mysv, noper);
2885 PerlIO_printf( Perl_debug_log, " -> %s",
2886 SvPV_nolen_const(mysv));
2889 regprop(RExC_rx, mysv, noper_next );
2890 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
2891 SvPV_nolen_const(mysv));
2893 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
2894 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
2896 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
2897 : PL_regkind[ OP( noper ) ] == EXACT )
2898 || OP(noper) == NOTHING )
2900 && noper_next == tail
2905 if ( !first || optype == NOTHING ) {
2906 if (!first) first = cur;
2907 optype = OP( noper );
2913 Currently we do not believe that the trie logic can
2914 handle case insensitive matching properly when the
2915 pattern is not unicode (thus forcing unicode semantics).
2917 If/when this is fixed the following define can be swapped
2918 in below to fully enable trie logic.
2920 #define TRIE_TYPE_IS_SAFE 1
2923 #define TRIE_TYPE_IS_SAFE (UTF || optype==EXACT)
2925 if ( last && TRIE_TYPE_IS_SAFE ) {
2926 make_trie( pRExC_state,
2927 startbranch, first, cur, tail, count,
2930 if ( PL_regkind[ OP( noper ) ] == EXACT
2932 && noper_next == tail
2937 optype = OP( noper );
2947 regprop(RExC_rx, mysv, cur);
2948 PerlIO_printf( Perl_debug_log,
2949 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
2950 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
2954 if ( last && TRIE_TYPE_IS_SAFE ) {
2955 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
2956 #ifdef TRIE_STUDY_OPT
2957 if ( ((made == MADE_EXACT_TRIE &&
2958 startbranch == first)
2959 || ( first_non_open == first )) &&
2961 flags |= SCF_TRIE_RESTUDY;
2962 if ( startbranch == first
2965 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
2975 else if ( code == BRANCHJ ) { /* single branch is optimized. */
2976 scan = NEXTOPER(NEXTOPER(scan));
2977 } else /* single branch is optimized. */
2978 scan = NEXTOPER(scan);
2980 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
2981 scan_frame *newframe = NULL;
2986 if (OP(scan) != SUSPEND) {
2987 /* set the pointer */
2988 if (OP(scan) == GOSUB) {
2990 RExC_recurse[ARG2L(scan)] = scan;
2991 start = RExC_open_parens[paren-1];
2992 end = RExC_close_parens[paren-1];
2995 start = RExC_rxi->program + 1;
2999 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3000 SAVEFREEPV(recursed);
3002 if (!PAREN_TEST(recursed,paren+1)) {
3003 PAREN_SET(recursed,paren+1);
3004 Newx(newframe,1,scan_frame);
3006 if (flags & SCF_DO_SUBSTR) {
3007 SCAN_COMMIT(pRExC_state,data,minlenp);
3008 data->longest = &(data->longest_float);
3010 is_inf = is_inf_internal = 1;
3011 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3012 cl_anything(pRExC_state, data->start_class);
3013 flags &= ~SCF_DO_STCLASS;
3016 Newx(newframe,1,scan_frame);
3019 end = regnext(scan);
3024 SAVEFREEPV(newframe);
3025 newframe->next = regnext(scan);
3026 newframe->last = last;
3027 newframe->stop = stopparen;
3028 newframe->prev = frame;
3038 else if (OP(scan) == EXACT) {
3039 I32 l = STR_LEN(scan);
3042 const U8 * const s = (U8*)STRING(scan);
3043 l = utf8_length(s, s + l);
3044 uc = utf8_to_uvchr(s, NULL);
3046 uc = *((U8*)STRING(scan));
3049 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3050 /* The code below prefers earlier match for fixed
3051 offset, later match for variable offset. */
3052 if (data->last_end == -1) { /* Update the start info. */
3053 data->last_start_min = data->pos_min;
3054 data->last_start_max = is_inf
3055 ? I32_MAX : data->pos_min + data->pos_delta;
3057 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3059 SvUTF8_on(data->last_found);
3061 SV * const sv = data->last_found;
3062 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3063 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3064 if (mg && mg->mg_len >= 0)
3065 mg->mg_len += utf8_length((U8*)STRING(scan),
3066 (U8*)STRING(scan)+STR_LEN(scan));
3068 data->last_end = data->pos_min + l;
3069 data->pos_min += l; /* As in the first entry. */
3070 data->flags &= ~SF_BEFORE_EOL;
3072 if (flags & SCF_DO_STCLASS_AND) {
3073 /* Check whether it is compatible with what we know already! */
3077 /* If compatibile, we or it in below. It is compatible if is
3078 * in the bitmp and either 1) its bit or its fold is set, or 2)
3079 * it's for a locale. Even if there isn't unicode semantics
3080 * here, at runtime there may be because of matching against a
3081 * utf8 string, so accept a possible false positive for
3082 * latin1-range folds */
3084 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3085 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3086 && (!(data->start_class->flags & ANYOF_FOLD)
3087 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3090 ANYOF_CLASS_ZERO(data->start_class);
3091 ANYOF_BITMAP_ZERO(data->start_class);
3093 ANYOF_BITMAP_SET(data->start_class, uc);
3094 data->start_class->flags &= ~ANYOF_EOS;
3096 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3098 else if (flags & SCF_DO_STCLASS_OR) {
3099 /* false positive possible if the class is case-folded */
3101 ANYOF_BITMAP_SET(data->start_class, uc);
3103 data->start_class->flags |= ANYOF_UNICODE_ALL;
3104 data->start_class->flags &= ~ANYOF_EOS;
3105 cl_and(data->start_class, and_withp);
3107 flags &= ~SCF_DO_STCLASS;
3109 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3110 I32 l = STR_LEN(scan);
3111 UV uc = *((U8*)STRING(scan));
3113 /* Search for fixed substrings supports EXACT only. */
3114 if (flags & SCF_DO_SUBSTR) {
3116 SCAN_COMMIT(pRExC_state, data, minlenp);
3119 const U8 * const s = (U8 *)STRING(scan);
3120 l = utf8_length(s, s + l);
3121 uc = utf8_to_uvchr(s, NULL);
3124 if (flags & SCF_DO_SUBSTR)
3126 if (flags & SCF_DO_STCLASS_AND) {
3127 /* Check whether it is compatible with what we know already! */
3130 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3131 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3132 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3136 ANYOF_CLASS_ZERO(data->start_class);
3137 ANYOF_BITMAP_ZERO(data->start_class);
3139 ANYOF_BITMAP_SET(data->start_class, uc);
3140 data->start_class->flags &= ~ANYOF_EOS;
3141 data->start_class->flags |= ANYOF_FOLD;
3142 if (OP(scan) == EXACTFL) {
3143 data->start_class->flags |= ANYOF_LOCALE;
3147 /* Also set the other member of the fold pair. In case
3148 * that unicode semantics is called for at runtime, use
3149 * the full latin1 fold. (Can't do this for locale,
3150 * because not known until runtime */
3151 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3155 else if (flags & SCF_DO_STCLASS_OR) {
3156 if (data->start_class->flags & ANYOF_FOLD) {
3157 /* false positive possible if the class is case-folded.
3158 Assume that the locale settings are the same... */
3160 ANYOF_BITMAP_SET(data->start_class, uc);
3161 if (OP(scan) != EXACTFL) {
3163 /* And set the other member of the fold pair, but
3164 * can't do that in locale because not known until
3166 ANYOF_BITMAP_SET(data->start_class,
3167 PL_fold_latin1[uc]);
3170 data->start_class->flags &= ~ANYOF_EOS;
3172 cl_and(data->start_class, and_withp);
3174 flags &= ~SCF_DO_STCLASS;
3176 else if (REGNODE_VARIES(OP(scan))) {
3177 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3178 I32 f = flags, pos_before = 0;
3179 regnode * const oscan = scan;
3180 struct regnode_charclass_class this_class;
3181 struct regnode_charclass_class *oclass = NULL;
3182 I32 next_is_eval = 0;
3184 switch (PL_regkind[OP(scan)]) {
3185 case WHILEM: /* End of (?:...)* . */
3186 scan = NEXTOPER(scan);
3189 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3190 next = NEXTOPER(scan);
3191 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3193 maxcount = REG_INFTY;
3194 next = regnext(scan);
3195 scan = NEXTOPER(scan);
3199 if (flags & SCF_DO_SUBSTR)
3204 if (flags & SCF_DO_STCLASS) {
3206 maxcount = REG_INFTY;
3207 next = regnext(scan);
3208 scan = NEXTOPER(scan);
3211 is_inf = is_inf_internal = 1;
3212 scan = regnext(scan);
3213 if (flags & SCF_DO_SUBSTR) {
3214 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3215 data->longest = &(data->longest_float);
3217 goto optimize_curly_tail;
3219 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3220 && (scan->flags == stopparen))
3225 mincount = ARG1(scan);
3226 maxcount = ARG2(scan);
3228 next = regnext(scan);
3229 if (OP(scan) == CURLYX) {
3230 I32 lp = (data ? *(data->last_closep) : 0);
3231 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3233 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3234 next_is_eval = (OP(scan) == EVAL);
3236 if (flags & SCF_DO_SUBSTR) {
3237 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3238 pos_before = data->pos_min;
3242 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3244 data->flags |= SF_IS_INF;
3246 if (flags & SCF_DO_STCLASS) {
3247 cl_init(pRExC_state, &this_class);
3248 oclass = data->start_class;
3249 data->start_class = &this_class;
3250 f |= SCF_DO_STCLASS_AND;
3251 f &= ~SCF_DO_STCLASS_OR;
3253 /* Exclude from super-linear cache processing any {n,m}
3254 regops for which the combination of input pos and regex
3255 pos is not enough information to determine if a match
3258 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3259 regex pos at the \s*, the prospects for a match depend not
3260 only on the input position but also on how many (bar\s*)
3261 repeats into the {4,8} we are. */
3262 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3263 f &= ~SCF_WHILEM_VISITED_POS;
3265 /* This will finish on WHILEM, setting scan, or on NULL: */
3266 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3267 last, data, stopparen, recursed, NULL,
3269 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3271 if (flags & SCF_DO_STCLASS)
3272 data->start_class = oclass;
3273 if (mincount == 0 || minnext == 0) {
3274 if (flags & SCF_DO_STCLASS_OR) {
3275 cl_or(pRExC_state, data->start_class, &this_class);
3277 else if (flags & SCF_DO_STCLASS_AND) {
3278 /* Switch to OR mode: cache the old value of
3279 * data->start_class */
3281 StructCopy(data->start_class, and_withp,
3282 struct regnode_charclass_class);
3283 flags &= ~SCF_DO_STCLASS_AND;
3284 StructCopy(&this_class, data->start_class,
3285 struct regnode_charclass_class);
3286 flags |= SCF_DO_STCLASS_OR;
3287 data->start_class->flags |= ANYOF_EOS;
3289 } else { /* Non-zero len */
3290 if (flags & SCF_DO_STCLASS_OR) {
3291 cl_or(pRExC_state, data->start_class, &this_class);
3292 cl_and(data->start_class, and_withp);
3294 else if (flags & SCF_DO_STCLASS_AND)
3295 cl_and(data->start_class, &this_class);
3296 flags &= ~SCF_DO_STCLASS;
3298 if (!scan) /* It was not CURLYX, but CURLY. */
3300 if ( /* ? quantifier ok, except for (?{ ... }) */
3301 (next_is_eval || !(mincount == 0 && maxcount == 1))
3302 && (minnext == 0) && (deltanext == 0)
3303 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3304 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3306 ckWARNreg(RExC_parse,
3307 "Quantifier unexpected on zero-length expression");
3310 min += minnext * mincount;
3311 is_inf_internal |= ((maxcount == REG_INFTY
3312 && (minnext + deltanext) > 0)
3313 || deltanext == I32_MAX);
3314 is_inf |= is_inf_internal;
3315 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3317 /* Try powerful optimization CURLYX => CURLYN. */
3318 if ( OP(oscan) == CURLYX && data
3319 && data->flags & SF_IN_PAR
3320 && !(data->flags & SF_HAS_EVAL)
3321 && !deltanext && minnext == 1 ) {
3322 /* Try to optimize to CURLYN. */
3323 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3324 regnode * const nxt1 = nxt;
3331 if (!REGNODE_SIMPLE(OP(nxt))
3332 && !(PL_regkind[OP(nxt)] == EXACT
3333 && STR_LEN(nxt) == 1))
3339 if (OP(nxt) != CLOSE)
3341 if (RExC_open_parens) {
3342 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3343 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3345 /* Now we know that nxt2 is the only contents: */
3346 oscan->flags = (U8)ARG(nxt);
3348 OP(nxt1) = NOTHING; /* was OPEN. */
3351 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3352 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3353 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3354 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3355 OP(nxt + 1) = OPTIMIZED; /* was count. */
3356 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3361 /* Try optimization CURLYX => CURLYM. */
3362 if ( OP(oscan) == CURLYX && data
3363 && !(data->flags & SF_HAS_PAR)
3364 && !(data->flags & SF_HAS_EVAL)
3365 && !deltanext /* atom is fixed width */
3366 && minnext != 0 /* CURLYM can't handle zero width */
3368 /* XXXX How to optimize if data == 0? */
3369 /* Optimize to a simpler form. */
3370 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3374 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3375 && (OP(nxt2) != WHILEM))
3377 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3378 /* Need to optimize away parenths. */
3379 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
3380 /* Set the parenth number. */
3381 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3383 oscan->flags = (U8)ARG(nxt);
3384 if (RExC_open_parens) {
3385 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3386 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3388 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3389 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3392 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3393 OP(nxt + 1) = OPTIMIZED; /* was count. */
3394 NEXT_OFF(nxt1 + 1) = 0; /* just for consistancy. */
3395 NEXT_OFF(nxt + 1) = 0; /* just for consistancy. */
3398 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3399 regnode *nnxt = regnext(nxt1);
3401 if (reg_off_by_arg[OP(nxt1)])
3402 ARG_SET(nxt1, nxt2 - nxt1);
3403 else if (nxt2 - nxt1 < U16_MAX)
3404 NEXT_OFF(nxt1) = nxt2 - nxt1;
3406 OP(nxt) = NOTHING; /* Cannot beautify */
3411 /* Optimize again: */
3412 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
3413 NULL, stopparen, recursed, NULL, 0,depth+1);
3418 else if ((OP(oscan) == CURLYX)
3419 && (flags & SCF_WHILEM_VISITED_POS)
3420 /* See the comment on a similar expression above.
3421 However, this time it's not a subexpression
3422 we care about, but the expression itself. */
3423 && (maxcount == REG_INFTY)
3424 && data && ++data->whilem_c < 16) {
3425 /* This stays as CURLYX, we can put the count/of pair. */
3426 /* Find WHILEM (as in regexec.c) */
3427 regnode *nxt = oscan + NEXT_OFF(oscan);
3429 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
3431 PREVOPER(nxt)->flags = (U8)(data->whilem_c
3432 | (RExC_whilem_seen << 4)); /* On WHILEM */
3434 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
3436 if (flags & SCF_DO_SUBSTR) {
3437 SV *last_str = NULL;
3438 int counted = mincount != 0;
3440 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
3441 #if defined(SPARC64_GCC_WORKAROUND)
3444 const char *s = NULL;
3447 if (pos_before >= data->last_start_min)
3450 b = data->last_start_min;
3453 s = SvPV_const(data->last_found, l);
3454 old = b - data->last_start_min;
3457 I32 b = pos_before >= data->last_start_min
3458 ? pos_before : data->last_start_min;
3460 const char * const s = SvPV_const(data->last_found, l);
3461 I32 old = b - data->last_start_min;
3465 old = utf8_hop((U8*)s, old) - (U8*)s;
3467 /* Get the added string: */
3468 last_str = newSVpvn_utf8(s + old, l, UTF);
3469 if (deltanext == 0 && pos_before == b) {
3470 /* What was added is a constant string */
3472 SvGROW(last_str, (mincount * l) + 1);
3473 repeatcpy(SvPVX(last_str) + l,
3474 SvPVX_const(last_str), l, mincount - 1);
3475 SvCUR_set(last_str, SvCUR(last_str) * mincount);
3476 /* Add additional parts. */
3477 SvCUR_set(data->last_found,
3478 SvCUR(data->last_found) - l);
3479 sv_catsv(data->last_found, last_str);
3481 SV * sv = data->last_found;
3483 SvUTF8(sv) && SvMAGICAL(sv) ?
3484 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3485 if (mg && mg->mg_len >= 0)
3486 mg->mg_len += CHR_SVLEN(last_str) - l;
3488 data->last_end += l * (mincount - 1);
3491 /* start offset must point into the last copy */
3492 data->last_start_min += minnext * (mincount - 1);
3493 data->last_start_max += is_inf ? I32_MAX
3494 : (maxcount - 1) * (minnext + data->pos_delta);
3497 /* It is counted once already... */
3498 data->pos_min += minnext * (mincount - counted);
3499 data->pos_delta += - counted * deltanext +
3500 (minnext + deltanext) * maxcount - minnext * mincount;
3501 if (mincount != maxcount) {
3502 /* Cannot extend fixed substrings found inside
3504 SCAN_COMMIT(pRExC_state,data,minlenp);
3505 if (mincount && last_str) {
3506 SV * const sv = data->last_found;
3507 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3508 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3512 sv_setsv(sv, last_str);
3513 data->last_end = data->pos_min;
3514 data->last_start_min =
3515 data->pos_min - CHR_SVLEN(last_str);
3516 data->last_start_max = is_inf
3518 : data->pos_min + data->pos_delta
3519 - CHR_SVLEN(last_str);
3521 data->longest = &(data->longest_float);
3523 SvREFCNT_dec(last_str);
3525 if (data && (fl & SF_HAS_EVAL))
3526 data->flags |= SF_HAS_EVAL;
3527 optimize_curly_tail:
3528 if (OP(oscan) != CURLYX) {
3529 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
3531 NEXT_OFF(oscan) += NEXT_OFF(next);
3534 default: /* REF and CLUMP only? */
3535 if (flags & SCF_DO_SUBSTR) {
3536 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3537 data->longest = &(data->longest_float);
3539 is_inf = is_inf_internal = 1;
3540 if (flags & SCF_DO_STCLASS_OR)
3541 cl_anything(pRExC_state, data->start_class);
3542 flags &= ~SCF_DO_STCLASS;
3546 else if (OP(scan) == LNBREAK) {
3547 if (flags & SCF_DO_STCLASS) {
3549 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3550 if (flags & SCF_DO_STCLASS_AND) {
3551 for (value = 0; value < 256; value++)
3552 if (!is_VERTWS_cp(value))
3553 ANYOF_BITMAP_CLEAR(data->start_class, value);
3556 for (value = 0; value < 256; value++)
3557 if (is_VERTWS_cp(value))
3558 ANYOF_BITMAP_SET(data->start_class, value);
3560 if (flags & SCF_DO_STCLASS_OR)
3561 cl_and(data->start_class, and_withp);
3562 flags &= ~SCF_DO_STCLASS;
3566 if (flags & SCF_DO_SUBSTR) {
3567 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3569 data->pos_delta += 1;
3570 data->longest = &(data->longest_float);
3573 else if (OP(scan) == FOLDCHAR) {
3574 int d = ARG(scan)==0xDF ? 1 : 2;
3575 flags &= ~SCF_DO_STCLASS;
3578 if (flags & SCF_DO_SUBSTR) {
3579 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3581 data->pos_delta += d;
3582 data->longest = &(data->longest_float);
3585 else if (REGNODE_SIMPLE(OP(scan))) {
3588 if (flags & SCF_DO_SUBSTR) {
3589 SCAN_COMMIT(pRExC_state,data,minlenp);
3593 if (flags & SCF_DO_STCLASS) {
3594 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3596 /* Some of the logic below assumes that switching
3597 locale on will only add false positives. */
3598 switch (PL_regkind[OP(scan)]) {
3602 /* Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan)); */
3603 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3604 cl_anything(pRExC_state, data->start_class);
3607 if (OP(scan) == SANY)
3609 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
3610 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
3611 || ((data->start_class->flags & ANYOF_CLASS)
3612 && ANYOF_CLASS_TEST_ANY_SET(data->start_class)));
3613 cl_anything(pRExC_state, data->start_class);
3615 if (flags & SCF_DO_STCLASS_AND || !value)
3616 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
3619 if (flags & SCF_DO_STCLASS_AND)
3620 cl_and(data->start_class,
3621 (struct regnode_charclass_class*)scan);
3623 cl_or(pRExC_state, data->start_class,
3624 (struct regnode_charclass_class*)scan);
3627 if (flags & SCF_DO_STCLASS_AND) {
3628 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3629 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3630 if (FLAGS(scan) & USE_UNI) {
3631 for (value = 0; value < 256; value++) {
3632 if (!isWORDCHAR_L1(value)) {
3633 ANYOF_BITMAP_CLEAR(data->start_class, value);
3637 for (value = 0; value < 256; value++) {
3638 if (!isALNUM(value)) {
3639 ANYOF_BITMAP_CLEAR(data->start_class, value);
3646 if (data->start_class->flags & ANYOF_LOCALE)
3647 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3648 else if (FLAGS(scan) & USE_UNI) {
3649 for (value = 0; value < 256; value++) {
3650 if (isWORDCHAR_L1(value)) {
3651 ANYOF_BITMAP_SET(data->start_class, value);
3655 for (value = 0; value < 256; value++) {
3656 if (isALNUM(value)) {
3657 ANYOF_BITMAP_SET(data->start_class, value);
3664 if (flags & SCF_DO_STCLASS_AND) {
3665 if (data->start_class->flags & ANYOF_LOCALE)
3666 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3669 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3670 data->start_class->flags |= ANYOF_LOCALE;
3674 if (flags & SCF_DO_STCLASS_AND) {
3675 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3676 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3677 if (FLAGS(scan) & USE_UNI) {
3678 for (value = 0; value < 256; value++) {
3679 if (isWORDCHAR_L1(value)) {
3680 ANYOF_BITMAP_CLEAR(data->start_class, value);
3684 for (value = 0; value < 256; value++) {
3685 if (isALNUM(value)) {
3686 ANYOF_BITMAP_CLEAR(data->start_class, value);
3693 if (data->start_class->flags & ANYOF_LOCALE)
3694 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3696 for (value = 0; value < 256; value++)
3697 if (!isALNUM(value))
3698 ANYOF_BITMAP_SET(data->start_class, value);
3703 if (flags & SCF_DO_STCLASS_AND) {
3704 if (data->start_class->flags & ANYOF_LOCALE)
3705 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3708 data->start_class->flags |= ANYOF_LOCALE;
3709 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3713 if (flags & SCF_DO_STCLASS_AND) {
3714 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3715 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3716 if (FLAGS(scan) & USE_UNI) {
3717 for (value = 0; value < 256; value++) {
3718 if (!isSPACE_L1(value)) {
3719 ANYOF_BITMAP_CLEAR(data->start_class, value);
3723 for (value = 0; value < 256; value++) {
3724 if (!isSPACE(value)) {
3725 ANYOF_BITMAP_CLEAR(data->start_class, value);
3732 if (data->start_class->flags & ANYOF_LOCALE) {
3733 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3735 else if (FLAGS(scan) & USE_UNI) {
3736 for (value = 0; value < 256; value++) {
3737 if (isSPACE_L1(value)) {
3738 ANYOF_BITMAP_SET(data->start_class, value);
3742 for (value = 0; value < 256; value++) {
3743 if (isSPACE(value)) {
3744 ANYOF_BITMAP_SET(data->start_class, value);
3751 if (flags & SCF_DO_STCLASS_AND) {
3752 if (data->start_class->flags & ANYOF_LOCALE)
3753 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3756 data->start_class->flags |= ANYOF_LOCALE;
3757 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3761 if (flags & SCF_DO_STCLASS_AND) {
3762 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3763 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3764 if (FLAGS(scan) & USE_UNI) {
3765 for (value = 0; value < 256; value++) {
3766 if (isSPACE_L1(value)) {
3767 ANYOF_BITMAP_CLEAR(data->start_class, value);
3771 for (value = 0; value < 256; value++) {
3772 if (isSPACE(value)) {
3773 ANYOF_BITMAP_CLEAR(data->start_class, value);
3780 if (data->start_class->flags & ANYOF_LOCALE)
3781 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3782 else if (FLAGS(scan) & USE_UNI) {
3783 for (value = 0; value < 256; value++) {
3784 if (!isSPACE_L1(value)) {
3785 ANYOF_BITMAP_SET(data->start_class, value);
3790 for (value = 0; value < 256; value++) {
3791 if (!isSPACE(value)) {
3792 ANYOF_BITMAP_SET(data->start_class, value);
3799 if (flags & SCF_DO_STCLASS_AND) {
3800 if (data->start_class->flags & ANYOF_LOCALE) {
3801 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3802 for (value = 0; value < 256; value++)
3803 if (!isSPACE(value))
3804 ANYOF_BITMAP_CLEAR(data->start_class, value);
3808 data->start_class->flags |= ANYOF_LOCALE;
3809 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3813 if (flags & SCF_DO_STCLASS_AND) {
3814 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NDIGIT);
3815 for (value = 0; value < 256; value++)
3816 if (!isDIGIT(value))
3817 ANYOF_BITMAP_CLEAR(data->start_class, value);
3820 if (data->start_class->flags & ANYOF_LOCALE)
3821 ANYOF_CLASS_SET(data->start_class,ANYOF_DIGIT);
3823 for (value = 0; value < 256; value++)
3825 ANYOF_BITMAP_SET(data->start_class, value);
https://perl5.git.perl.org / perl5.git / blob