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 I32 uni_semantics; /* If a d charset modifier should use unicode
138 rules, even if the pattern is not in
140 HV *paren_names; /* Paren names */
142 regnode **recurse; /* Recurse regops */
143 I32 recurse_count; /* Number of recurse regops */
146 I32 override_recoding;
148 char *starttry; /* -Dr: where regtry was called. */
149 #define RExC_starttry (pRExC_state->starttry)
152 const char *lastparse;
154 AV *paren_name_list; /* idx -> name */
155 #define RExC_lastparse (pRExC_state->lastparse)
156 #define RExC_lastnum (pRExC_state->lastnum)
157 #define RExC_paren_name_list (pRExC_state->paren_name_list)
161 #define RExC_flags (pRExC_state->flags)
162 #define RExC_precomp (pRExC_state->precomp)
163 #define RExC_rx_sv (pRExC_state->rx_sv)
164 #define RExC_rx (pRExC_state->rx)
165 #define RExC_rxi (pRExC_state->rxi)
166 #define RExC_start (pRExC_state->start)
167 #define RExC_end (pRExC_state->end)
168 #define RExC_parse (pRExC_state->parse)
169 #define RExC_whilem_seen (pRExC_state->whilem_seen)
170 #ifdef RE_TRACK_PATTERN_OFFSETS
171 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
173 #define RExC_emit (pRExC_state->emit)
174 #define RExC_emit_start (pRExC_state->emit_start)
175 #define RExC_emit_bound (pRExC_state->emit_bound)
176 #define RExC_naughty (pRExC_state->naughty)
177 #define RExC_sawback (pRExC_state->sawback)
178 #define RExC_seen (pRExC_state->seen)
179 #define RExC_size (pRExC_state->size)
180 #define RExC_npar (pRExC_state->npar)
181 #define RExC_nestroot (pRExC_state->nestroot)
182 #define RExC_extralen (pRExC_state->extralen)
183 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
184 #define RExC_seen_evals (pRExC_state->seen_evals)
185 #define RExC_utf8 (pRExC_state->utf8)
186 #define RExC_uni_semantics (pRExC_state->uni_semantics)
187 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
188 #define RExC_open_parens (pRExC_state->open_parens)
189 #define RExC_close_parens (pRExC_state->close_parens)
190 #define RExC_opend (pRExC_state->opend)
191 #define RExC_paren_names (pRExC_state->paren_names)
192 #define RExC_recurse (pRExC_state->recurse)
193 #define RExC_recurse_count (pRExC_state->recurse_count)
194 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
195 #define RExC_contains_locale (pRExC_state->contains_locale)
196 #define RExC_override_recoding (pRExC_state->override_recoding)
199 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
200 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
201 ((*s) == '{' && regcurly(s)))
204 #undef SPSTART /* dratted cpp namespace... */
207 * Flags to be passed up and down.
209 #define WORST 0 /* Worst case. */
210 #define HASWIDTH 0x01 /* Known to match non-null strings. */
212 /* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
213 * character, and if utf8, must be invariant. Note that this is not the same thing as REGNODE_SIMPLE */
215 #define SPSTART 0x04 /* Starts with * or +. */
216 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
217 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
219 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
221 /* whether trie related optimizations are enabled */
222 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
223 #define TRIE_STUDY_OPT
224 #define FULL_TRIE_STUDY
230 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
231 #define PBITVAL(paren) (1 << ((paren) & 7))
232 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
233 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
234 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
236 /* If not already in utf8, do a longjmp back to the beginning */
237 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
238 #define REQUIRE_UTF8 STMT_START { \
239 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
242 /* About scan_data_t.
244 During optimisation we recurse through the regexp program performing
245 various inplace (keyhole style) optimisations. In addition study_chunk
246 and scan_commit populate this data structure with information about
247 what strings MUST appear in the pattern. We look for the longest
248 string that must appear at a fixed location, and we look for the
249 longest string that may appear at a floating location. So for instance
254 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
255 strings (because they follow a .* construct). study_chunk will identify
256 both FOO and BAR as being the longest fixed and floating strings respectively.
258 The strings can be composites, for instance
262 will result in a composite fixed substring 'foo'.
264 For each string some basic information is maintained:
266 - offset or min_offset
267 This is the position the string must appear at, or not before.
268 It also implicitly (when combined with minlenp) tells us how many
269 characters must match before the string we are searching for.
270 Likewise when combined with minlenp and the length of the string it
271 tells us how many characters must appear after the string we have
275 Only used for floating strings. This is the rightmost point that
276 the string can appear at. If set to I32 max it indicates that the
277 string can occur infinitely far to the right.
280 A pointer to the minimum length of the pattern that the string
281 was found inside. This is important as in the case of positive
282 lookahead or positive lookbehind we can have multiple patterns
287 The minimum length of the pattern overall is 3, the minimum length
288 of the lookahead part is 3, but the minimum length of the part that
289 will actually match is 1. So 'FOO's minimum length is 3, but the
290 minimum length for the F is 1. This is important as the minimum length
291 is used to determine offsets in front of and behind the string being
292 looked for. Since strings can be composites this is the length of the
293 pattern at the time it was committed with a scan_commit. Note that
294 the length is calculated by study_chunk, so that the minimum lengths
295 are not known until the full pattern has been compiled, thus the
296 pointer to the value.
300 In the case of lookbehind the string being searched for can be
301 offset past the start point of the final matching string.
302 If this value was just blithely removed from the min_offset it would
303 invalidate some of the calculations for how many chars must match
304 before or after (as they are derived from min_offset and minlen and
305 the length of the string being searched for).
306 When the final pattern is compiled and the data is moved from the
307 scan_data_t structure into the regexp structure the information
308 about lookbehind is factored in, with the information that would
309 have been lost precalculated in the end_shift field for the
312 The fields pos_min and pos_delta are used to store the minimum offset
313 and the delta to the maximum offset at the current point in the pattern.
317 typedef struct scan_data_t {
318 /*I32 len_min; unused */
319 /*I32 len_delta; unused */
323 I32 last_end; /* min value, <0 unless valid. */
326 SV **longest; /* Either &l_fixed, or &l_float. */
327 SV *longest_fixed; /* longest fixed string found in pattern */
328 I32 offset_fixed; /* offset where it starts */
329 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
330 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
331 SV *longest_float; /* longest floating string found in pattern */
332 I32 offset_float_min; /* earliest point in string it can appear */
333 I32 offset_float_max; /* latest point in string it can appear */
334 I32 *minlen_float; /* pointer to the minlen relevant to the string */
335 I32 lookbehind_float; /* is the position of the string modified by LB */
339 struct regnode_charclass_class *start_class;
343 * Forward declarations for pregcomp()'s friends.
346 static const scan_data_t zero_scan_data =
347 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
349 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
350 #define SF_BEFORE_SEOL 0x0001
351 #define SF_BEFORE_MEOL 0x0002
352 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
353 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
356 # define SF_FIX_SHIFT_EOL (0+2)
357 # define SF_FL_SHIFT_EOL (0+4)
359 # define SF_FIX_SHIFT_EOL (+2)
360 # define SF_FL_SHIFT_EOL (+4)
363 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
364 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
366 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
367 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
368 #define SF_IS_INF 0x0040
369 #define SF_HAS_PAR 0x0080
370 #define SF_IN_PAR 0x0100
371 #define SF_HAS_EVAL 0x0200
372 #define SCF_DO_SUBSTR 0x0400
373 #define SCF_DO_STCLASS_AND 0x0800
374 #define SCF_DO_STCLASS_OR 0x1000
375 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
376 #define SCF_WHILEM_VISITED_POS 0x2000
378 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
379 #define SCF_SEEN_ACCEPT 0x8000
381 #define UTF cBOOL(RExC_utf8)
382 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
383 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
384 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
385 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
386 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
387 #define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
388 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
390 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
392 #define OOB_UNICODE 12345678
393 #define OOB_NAMEDCLASS -1
395 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
396 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
399 /* length of regex to show in messages that don't mark a position within */
400 #define RegexLengthToShowInErrorMessages 127
403 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
404 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
405 * op/pragma/warn/regcomp.
407 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
408 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
410 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
413 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
414 * arg. Show regex, up to a maximum length. If it's too long, chop and add
417 #define _FAIL(code) STMT_START { \
418 const char *ellipses = ""; \
419 IV len = RExC_end - RExC_precomp; \
422 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
423 if (len > RegexLengthToShowInErrorMessages) { \
424 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
425 len = RegexLengthToShowInErrorMessages - 10; \
431 #define FAIL(msg) _FAIL( \
432 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
433 msg, (int)len, RExC_precomp, ellipses))
435 #define FAIL2(msg,arg) _FAIL( \
436 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
437 arg, (int)len, RExC_precomp, ellipses))
440 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
442 #define Simple_vFAIL(m) STMT_START { \
443 const IV offset = RExC_parse - RExC_precomp; \
444 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
445 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
449 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
451 #define vFAIL(m) STMT_START { \
453 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
458 * Like Simple_vFAIL(), but accepts two arguments.
460 #define Simple_vFAIL2(m,a1) STMT_START { \
461 const IV offset = RExC_parse - RExC_precomp; \
462 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
463 (int)offset, RExC_precomp, RExC_precomp + offset); \
467 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
469 #define vFAIL2(m,a1) STMT_START { \
471 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
472 Simple_vFAIL2(m, a1); \
477 * Like Simple_vFAIL(), but accepts three arguments.
479 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
480 const IV offset = RExC_parse - RExC_precomp; \
481 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
482 (int)offset, RExC_precomp, RExC_precomp + offset); \
486 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
488 #define vFAIL3(m,a1,a2) STMT_START { \
490 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
491 Simple_vFAIL3(m, a1, a2); \
495 * Like Simple_vFAIL(), but accepts four arguments.
497 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
498 const IV offset = RExC_parse - RExC_precomp; \
499 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
500 (int)offset, RExC_precomp, RExC_precomp + offset); \
503 #define ckWARNreg(loc,m) STMT_START { \
504 const IV offset = loc - RExC_precomp; \
505 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
506 (int)offset, RExC_precomp, RExC_precomp + offset); \
509 #define ckWARNregdep(loc,m) STMT_START { \
510 const IV offset = loc - RExC_precomp; \
511 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
513 (int)offset, RExC_precomp, RExC_precomp + offset); \
516 #define ckWARN2regdep(loc,m, a1) STMT_START { \
517 const IV offset = loc - RExC_precomp; \
518 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
520 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
523 #define ckWARN2reg(loc, m, a1) STMT_START { \
524 const IV offset = loc - RExC_precomp; \
525 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
526 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
529 #define vWARN3(loc, m, a1, a2) STMT_START { \
530 const IV offset = loc - RExC_precomp; \
531 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
532 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
535 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
536 const IV offset = loc - RExC_precomp; \
537 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
538 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
541 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
542 const IV offset = loc - RExC_precomp; \
543 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
544 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
547 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
548 const IV offset = loc - RExC_precomp; \
549 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
550 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
553 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
554 const IV offset = loc - RExC_precomp; \
555 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
556 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
560 /* Allow for side effects in s */
561 #define REGC(c,s) STMT_START { \
562 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
565 /* Macros for recording node offsets. 20001227 mjd@plover.com
566 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
567 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
568 * Element 0 holds the number n.
569 * Position is 1 indexed.
571 #ifndef RE_TRACK_PATTERN_OFFSETS
572 #define Set_Node_Offset_To_R(node,byte)
573 #define Set_Node_Offset(node,byte)
574 #define Set_Cur_Node_Offset
575 #define Set_Node_Length_To_R(node,len)
576 #define Set_Node_Length(node,len)
577 #define Set_Node_Cur_Length(node)
578 #define Node_Offset(n)
579 #define Node_Length(n)
580 #define Set_Node_Offset_Length(node,offset,len)
581 #define ProgLen(ri) ri->u.proglen
582 #define SetProgLen(ri,x) ri->u.proglen = x
584 #define ProgLen(ri) ri->u.offsets[0]
585 #define SetProgLen(ri,x) ri->u.offsets[0] = x
586 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
588 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
589 __LINE__, (int)(node), (int)(byte))); \
591 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
593 RExC_offsets[2*(node)-1] = (byte); \
598 #define Set_Node_Offset(node,byte) \
599 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
600 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
602 #define Set_Node_Length_To_R(node,len) STMT_START { \
604 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
605 __LINE__, (int)(node), (int)(len))); \
607 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
609 RExC_offsets[2*(node)] = (len); \
614 #define Set_Node_Length(node,len) \
615 Set_Node_Length_To_R((node)-RExC_emit_start, len)
616 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
617 #define Set_Node_Cur_Length(node) \
618 Set_Node_Length(node, RExC_parse - parse_start)
620 /* Get offsets and lengths */
621 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
622 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
624 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
625 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
626 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
630 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
631 #define EXPERIMENTAL_INPLACESCAN
632 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
634 #define DEBUG_STUDYDATA(str,data,depth) \
635 DEBUG_OPTIMISE_MORE_r(if(data){ \
636 PerlIO_printf(Perl_debug_log, \
637 "%*s" str "Pos:%"IVdf"/%"IVdf \
638 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
639 (int)(depth)*2, "", \
640 (IV)((data)->pos_min), \
641 (IV)((data)->pos_delta), \
642 (UV)((data)->flags), \
643 (IV)((data)->whilem_c), \
644 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
645 is_inf ? "INF " : "" \
647 if ((data)->last_found) \
648 PerlIO_printf(Perl_debug_log, \
649 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
650 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
651 SvPVX_const((data)->last_found), \
652 (IV)((data)->last_end), \
653 (IV)((data)->last_start_min), \
654 (IV)((data)->last_start_max), \
655 ((data)->longest && \
656 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
657 SvPVX_const((data)->longest_fixed), \
658 (IV)((data)->offset_fixed), \
659 ((data)->longest && \
660 (data)->longest==&((data)->longest_float)) ? "*" : "", \
661 SvPVX_const((data)->longest_float), \
662 (IV)((data)->offset_float_min), \
663 (IV)((data)->offset_float_max) \
665 PerlIO_printf(Perl_debug_log,"\n"); \
668 static void clear_re(pTHX_ void *r);
670 /* Mark that we cannot extend a found fixed substring at this point.
671 Update the longest found anchored substring and the longest found
672 floating substrings if needed. */
675 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
677 const STRLEN l = CHR_SVLEN(data->last_found);
678 const STRLEN old_l = CHR_SVLEN(*data->longest);
679 GET_RE_DEBUG_FLAGS_DECL;
681 PERL_ARGS_ASSERT_SCAN_COMMIT;
683 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
684 SvSetMagicSV(*data->longest, data->last_found);
685 if (*data->longest == data->longest_fixed) {
686 data->offset_fixed = l ? data->last_start_min : data->pos_min;
687 if (data->flags & SF_BEFORE_EOL)
689 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
691 data->flags &= ~SF_FIX_BEFORE_EOL;
692 data->minlen_fixed=minlenp;
693 data->lookbehind_fixed=0;
695 else { /* *data->longest == data->longest_float */
696 data->offset_float_min = l ? data->last_start_min : data->pos_min;
697 data->offset_float_max = (l
698 ? data->last_start_max
699 : data->pos_min + data->pos_delta);
700 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
701 data->offset_float_max = I32_MAX;
702 if (data->flags & SF_BEFORE_EOL)
704 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
706 data->flags &= ~SF_FL_BEFORE_EOL;
707 data->minlen_float=minlenp;
708 data->lookbehind_float=0;
711 SvCUR_set(data->last_found, 0);
713 SV * const sv = data->last_found;
714 if (SvUTF8(sv) && SvMAGICAL(sv)) {
715 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
721 data->flags &= ~SF_BEFORE_EOL;
722 DEBUG_STUDYDATA("commit: ",data,0);
725 /* Can match anything (initialization) */
727 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
729 PERL_ARGS_ASSERT_CL_ANYTHING;
731 ANYOF_BITMAP_SETALL(cl);
732 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
733 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
735 /* If any portion of the regex is to operate under locale rules,
736 * initialization includes it. The reason this isn't done for all regexes
737 * is that the optimizer was written under the assumption that locale was
738 * all-or-nothing. Given the complexity and lack of documentation in the
739 * optimizer, and that there are inadequate test cases for locale, so many
740 * parts of it may not work properly, it is safest to avoid locale unless
742 if (RExC_contains_locale) {
743 ANYOF_CLASS_SETALL(cl); /* /l uses class */
744 cl->flags |= ANYOF_LOCALE;
747 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
751 /* Can match anything (initialization) */
753 S_cl_is_anything(const struct regnode_charclass_class *cl)
757 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
759 for (value = 0; value <= ANYOF_MAX; value += 2)
760 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
762 if (!(cl->flags & ANYOF_UNICODE_ALL))
764 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
769 /* Can match anything (initialization) */
771 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
773 PERL_ARGS_ASSERT_CL_INIT;
775 Zero(cl, 1, struct regnode_charclass_class);
777 cl_anything(pRExC_state, cl);
778 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
781 /* These two functions currently do the exact same thing */
782 #define cl_init_zero S_cl_init
784 /* 'AND' a given class with another one. Can create false positives. 'cl'
785 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
786 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
788 S_cl_and(struct regnode_charclass_class *cl,
789 const struct regnode_charclass_class *and_with)
791 PERL_ARGS_ASSERT_CL_AND;
793 assert(and_with->type == ANYOF);
795 /* I (khw) am not sure all these restrictions are necessary XXX */
796 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
797 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
798 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
799 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
800 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
803 if (and_with->flags & ANYOF_INVERT)
804 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
805 cl->bitmap[i] &= ~and_with->bitmap[i];
807 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
808 cl->bitmap[i] &= and_with->bitmap[i];
809 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
811 if (and_with->flags & ANYOF_INVERT) {
813 /* Here, the and'ed node is inverted. Get the AND of the flags that
814 * aren't affected by the inversion. Those that are affected are
815 * handled individually below */
816 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
817 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
818 cl->flags |= affected_flags;
820 /* We currently don't know how to deal with things that aren't in the
821 * bitmap, but we know that the intersection is no greater than what
822 * is already in cl, so let there be false positives that get sorted
823 * out after the synthetic start class succeeds, and the node is
824 * matched for real. */
826 /* The inversion of these two flags indicate that the resulting
827 * intersection doesn't have them */
828 if (and_with->flags & ANYOF_UNICODE_ALL) {
829 cl->flags &= ~ANYOF_UNICODE_ALL;
831 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
832 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
835 else { /* and'd node is not inverted */
836 U8 outside_bitmap_but_not_utf8; /* Temp variable */
838 if (! ANYOF_NONBITMAP(and_with)) {
840 /* Here 'and_with' doesn't match anything outside the bitmap
841 * (except possibly ANYOF_UNICODE_ALL), which means the
842 * intersection can't either, except for ANYOF_UNICODE_ALL, in
843 * which case we don't know what the intersection is, but it's no
844 * greater than what cl already has, so can just leave it alone,
845 * with possible false positives */
846 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
847 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
848 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
851 else if (! ANYOF_NONBITMAP(cl)) {
853 /* Here, 'and_with' does match something outside the bitmap, and cl
854 * doesn't have a list of things to match outside the bitmap. If
855 * cl can match all code points above 255, the intersection will
856 * be those above-255 code points that 'and_with' matches. If cl
857 * can't match all Unicode code points, it means that it can't
858 * match anything outside the bitmap (since the 'if' that got us
859 * into this block tested for that), so we leave the bitmap empty.
861 if (cl->flags & ANYOF_UNICODE_ALL) {
862 ARG_SET(cl, ARG(and_with));
864 /* and_with's ARG may match things that don't require UTF8.
865 * And now cl's will too, in spite of this being an 'and'. See
866 * the comments below about the kludge */
867 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
871 /* Here, both 'and_with' and cl match something outside the
872 * bitmap. Currently we do not do the intersection, so just match
873 * whatever cl had at the beginning. */
877 /* Take the intersection of the two sets of flags. However, the
878 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
879 * kludge around the fact that this flag is not treated like the others
880 * which are initialized in cl_anything(). The way the optimizer works
881 * is that the synthetic start class (SSC) is initialized to match
882 * anything, and then the first time a real node is encountered, its
883 * values are AND'd with the SSC's with the result being the values of
884 * the real node. However, there are paths through the optimizer where
885 * the AND never gets called, so those initialized bits are set
886 * inappropriately, which is not usually a big deal, as they just cause
887 * false positives in the SSC, which will just mean a probably
888 * imperceptible slow down in execution. However this bit has a
889 * higher false positive consequence in that it can cause utf8.pm,
890 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
891 * bigger slowdown and also causes significant extra memory to be used.
892 * In order to prevent this, the code now takes a different tack. The
893 * bit isn't set unless some part of the regular expression needs it,
894 * but once set it won't get cleared. This means that these extra
895 * modules won't get loaded unless there was some path through the
896 * pattern that would have required them anyway, and so any false
897 * positives that occur by not ANDing them out when they could be
898 * aren't as severe as they would be if we treated this bit like all
900 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
901 & ANYOF_NONBITMAP_NON_UTF8;
902 cl->flags &= and_with->flags;
903 cl->flags |= outside_bitmap_but_not_utf8;
907 /* 'OR' a given class with another one. Can create false positives. 'cl'
908 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
909 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
911 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
913 PERL_ARGS_ASSERT_CL_OR;
915 if (or_with->flags & ANYOF_INVERT) {
917 /* Here, the or'd node is to be inverted. This means we take the
918 * complement of everything not in the bitmap, but currently we don't
919 * know what that is, so give up and match anything */
920 if (ANYOF_NONBITMAP(or_with)) {
921 cl_anything(pRExC_state, cl);
924 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
925 * <= (B1 | !B2) | (CL1 | !CL2)
926 * which is wasteful if CL2 is small, but we ignore CL2:
927 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
928 * XXXX Can we handle case-fold? Unclear:
929 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
930 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
932 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
933 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
934 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
937 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
938 cl->bitmap[i] |= ~or_with->bitmap[i];
939 } /* XXXX: logic is complicated otherwise */
941 cl_anything(pRExC_state, cl);
944 /* And, we can just take the union of the flags that aren't affected
945 * by the inversion */
946 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
948 /* For the remaining flags:
949 ANYOF_UNICODE_ALL and inverted means to not match anything above
950 255, which means that the union with cl should just be
951 what cl has in it, so can ignore this flag
952 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
953 is 127-255 to match them, but then invert that, so the
954 union with cl should just be what cl has in it, so can
957 } else { /* 'or_with' is not inverted */
958 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
959 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
960 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
961 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
964 /* OR char bitmap and class bitmap separately */
965 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
966 cl->bitmap[i] |= or_with->bitmap[i];
967 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
968 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
969 cl->classflags[i] |= or_with->classflags[i];
970 cl->flags |= ANYOF_CLASS;
973 else { /* XXXX: logic is complicated, leave it along for a moment. */
974 cl_anything(pRExC_state, cl);
977 if (ANYOF_NONBITMAP(or_with)) {
979 /* Use the added node's outside-the-bit-map match if there isn't a
980 * conflict. If there is a conflict (both nodes match something
981 * outside the bitmap, but what they match outside is not the same
982 * pointer, and hence not easily compared until XXX we extend
983 * inversion lists this far), give up and allow the start class to
984 * match everything outside the bitmap. If that stuff is all above
985 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
986 if (! ANYOF_NONBITMAP(cl)) {
987 ARG_SET(cl, ARG(or_with));
989 else if (ARG(cl) != ARG(or_with)) {
991 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
992 cl_anything(pRExC_state, cl);
995 cl->flags |= ANYOF_UNICODE_ALL;
1000 /* Take the union */
1001 cl->flags |= or_with->flags;
1005 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1006 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1007 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1008 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1013 dump_trie(trie,widecharmap,revcharmap)
1014 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1015 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1017 These routines dump out a trie in a somewhat readable format.
1018 The _interim_ variants are used for debugging the interim
1019 tables that are used to generate the final compressed
1020 representation which is what dump_trie expects.
1022 Part of the reason for their existence is to provide a form
1023 of documentation as to how the different representations function.
1028 Dumps the final compressed table form of the trie to Perl_debug_log.
1029 Used for debugging make_trie().
1033 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1034 AV *revcharmap, U32 depth)
1037 SV *sv=sv_newmortal();
1038 int colwidth= widecharmap ? 6 : 4;
1040 GET_RE_DEBUG_FLAGS_DECL;
1042 PERL_ARGS_ASSERT_DUMP_TRIE;
1044 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1045 (int)depth * 2 + 2,"",
1046 "Match","Base","Ofs" );
1048 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1049 SV ** const tmp = av_fetch( revcharmap, state, 0);
1051 PerlIO_printf( Perl_debug_log, "%*s",
1053 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1054 PL_colors[0], PL_colors[1],
1055 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1056 PERL_PV_ESCAPE_FIRSTCHAR
1061 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1062 (int)depth * 2 + 2,"");
1064 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1065 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1066 PerlIO_printf( Perl_debug_log, "\n");
1068 for( state = 1 ; state < trie->statecount ; state++ ) {
1069 const U32 base = trie->states[ state ].trans.base;
1071 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1073 if ( trie->states[ state ].wordnum ) {
1074 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1076 PerlIO_printf( Perl_debug_log, "%6s", "" );
1079 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1084 while( ( base + ofs < trie->uniquecharcount ) ||
1085 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1086 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1089 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1091 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1092 if ( ( base + ofs >= trie->uniquecharcount ) &&
1093 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1094 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1096 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1098 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1100 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1104 PerlIO_printf( Perl_debug_log, "]");
1107 PerlIO_printf( Perl_debug_log, "\n" );
1109 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1110 for (word=1; word <= trie->wordcount; word++) {
1111 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1112 (int)word, (int)(trie->wordinfo[word].prev),
1113 (int)(trie->wordinfo[word].len));
1115 PerlIO_printf(Perl_debug_log, "\n" );
1118 Dumps a fully constructed but uncompressed trie in list form.
1119 List tries normally only are used for construction when the number of
1120 possible chars (trie->uniquecharcount) is very high.
1121 Used for debugging make_trie().
1124 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1125 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1129 SV *sv=sv_newmortal();
1130 int colwidth= widecharmap ? 6 : 4;
1131 GET_RE_DEBUG_FLAGS_DECL;
1133 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1135 /* print out the table precompression. */
1136 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1137 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1138 "------:-----+-----------------\n" );
1140 for( state=1 ; state < next_alloc ; state ++ ) {
1143 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1144 (int)depth * 2 + 2,"", (UV)state );
1145 if ( ! trie->states[ state ].wordnum ) {
1146 PerlIO_printf( Perl_debug_log, "%5s| ","");
1148 PerlIO_printf( Perl_debug_log, "W%4x| ",
1149 trie->states[ state ].wordnum
1152 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1153 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1155 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1157 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1158 PL_colors[0], PL_colors[1],
1159 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1160 PERL_PV_ESCAPE_FIRSTCHAR
1162 TRIE_LIST_ITEM(state,charid).forid,
1163 (UV)TRIE_LIST_ITEM(state,charid).newstate
1166 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1167 (int)((depth * 2) + 14), "");
1170 PerlIO_printf( Perl_debug_log, "\n");
1175 Dumps a fully constructed but uncompressed trie in table form.
1176 This is the normal DFA style state transition table, with a few
1177 twists to facilitate compression later.
1178 Used for debugging make_trie().
1181 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1182 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1187 SV *sv=sv_newmortal();
1188 int colwidth= widecharmap ? 6 : 4;
1189 GET_RE_DEBUG_FLAGS_DECL;
1191 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1194 print out the table precompression so that we can do a visual check
1195 that they are identical.
1198 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1200 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1201 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1203 PerlIO_printf( Perl_debug_log, "%*s",
1205 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1206 PL_colors[0], PL_colors[1],
1207 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1208 PERL_PV_ESCAPE_FIRSTCHAR
1214 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1216 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1217 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1220 PerlIO_printf( Perl_debug_log, "\n" );
1222 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1224 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1225 (int)depth * 2 + 2,"",
1226 (UV)TRIE_NODENUM( state ) );
1228 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1229 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1231 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1233 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1235 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1236 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1238 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1239 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1247 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1248 startbranch: the first branch in the whole branch sequence
1249 first : start branch of sequence of branch-exact nodes.
1250 May be the same as startbranch
1251 last : Thing following the last branch.
1252 May be the same as tail.
1253 tail : item following the branch sequence
1254 count : words in the sequence
1255 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1256 depth : indent depth
1258 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1260 A trie is an N'ary tree where the branches are determined by digital
1261 decomposition of the key. IE, at the root node you look up the 1st character and
1262 follow that branch repeat until you find the end of the branches. Nodes can be
1263 marked as "accepting" meaning they represent a complete word. Eg:
1267 would convert into the following structure. Numbers represent states, letters
1268 following numbers represent valid transitions on the letter from that state, if
1269 the number is in square brackets it represents an accepting state, otherwise it
1270 will be in parenthesis.
1272 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1276 (1) +-i->(6)-+-s->[7]
1278 +-s->(3)-+-h->(4)-+-e->[5]
1280 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1282 This shows that when matching against the string 'hers' we will begin at state 1
1283 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1284 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1285 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1286 single traverse. We store a mapping from accepting to state to which word was
1287 matched, and then when we have multiple possibilities we try to complete the
1288 rest of the regex in the order in which they occured in the alternation.
1290 The only prior NFA like behaviour that would be changed by the TRIE support is
1291 the silent ignoring of duplicate alternations which are of the form:
1293 / (DUPE|DUPE) X? (?{ ... }) Y /x
1295 Thus EVAL blocks following a trie may be called a different number of times with
1296 and without the optimisation. With the optimisations dupes will be silently
1297 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1298 the following demonstrates:
1300 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1302 which prints out 'word' three times, but
1304 'words'=~/(word|word|word)(?{ print $1 })S/
1306 which doesnt print it out at all. This is due to other optimisations kicking in.
1308 Example of what happens on a structural level:
1310 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1312 1: CURLYM[1] {1,32767}(18)
1323 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1324 and should turn into:
1326 1: CURLYM[1] {1,32767}(18)
1328 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1336 Cases where tail != last would be like /(?foo|bar)baz/:
1346 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1347 and would end up looking like:
1350 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1357 d = uvuni_to_utf8_flags(d, uv, 0);
1359 is the recommended Unicode-aware way of saying
1364 #define TRIE_STORE_REVCHAR \
1367 SV *zlopp = newSV(2); \
1368 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1369 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, uvc & 0xFF); \
1370 SvCUR_set(zlopp, kapow - flrbbbbb); \
1373 av_push(revcharmap, zlopp); \
1375 char ooooff = (char)uvc; \
1376 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1380 #define TRIE_READ_CHAR STMT_START { \
1384 if ( foldlen > 0 ) { \
1385 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1390 len = UTF8SKIP(uc);\
1391 uvc = to_utf8_fold( uc, foldbuf, &foldlen); \
1392 foldlen -= UNISKIP( uvc ); \
1393 scan = foldbuf + UNISKIP( uvc ); \
1396 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1406 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1407 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1408 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1409 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1411 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1412 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1413 TRIE_LIST_CUR( state )++; \
1416 #define TRIE_LIST_NEW(state) STMT_START { \
1417 Newxz( trie->states[ state ].trans.list, \
1418 4, reg_trie_trans_le ); \
1419 TRIE_LIST_CUR( state ) = 1; \
1420 TRIE_LIST_LEN( state ) = 4; \
1423 #define TRIE_HANDLE_WORD(state) STMT_START { \
1424 U16 dupe= trie->states[ state ].wordnum; \
1425 regnode * const noper_next = regnext( noper ); \
1428 /* store the word for dumping */ \
1430 if (OP(noper) != NOTHING) \
1431 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1433 tmp = newSVpvn_utf8( "", 0, UTF ); \
1434 av_push( trie_words, tmp ); \
1438 trie->wordinfo[curword].prev = 0; \
1439 trie->wordinfo[curword].len = wordlen; \
1440 trie->wordinfo[curword].accept = state; \
1442 if ( noper_next < tail ) { \
1444 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1445 trie->jump[curword] = (U16)(noper_next - convert); \
1447 jumper = noper_next; \
1449 nextbranch= regnext(cur); \
1453 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1454 /* chain, so that when the bits of chain are later */\
1455 /* linked together, the dups appear in the chain */\
1456 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1457 trie->wordinfo[dupe].prev = curword; \
1459 /* we haven't inserted this word yet. */ \
1460 trie->states[ state ].wordnum = curword; \
1465 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1466 ( ( base + charid >= ucharcount \
1467 && base + charid < ubound \
1468 && state == trie->trans[ base - ucharcount + charid ].check \
1469 && trie->trans[ base - ucharcount + charid ].next ) \
1470 ? trie->trans[ base - ucharcount + charid ].next \
1471 : ( state==1 ? special : 0 ) \
1475 #define MADE_JUMP_TRIE 2
1476 #define MADE_EXACT_TRIE 4
1479 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1482 /* first pass, loop through and scan words */
1483 reg_trie_data *trie;
1484 HV *widecharmap = NULL;
1485 AV *revcharmap = newAV();
1487 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1492 regnode *jumper = NULL;
1493 regnode *nextbranch = NULL;
1494 regnode *convert = NULL;
1495 U32 *prev_states; /* temp array mapping each state to previous one */
1496 /* we just use folder as a flag in utf8 */
1497 const U8 * folder = NULL;
1500 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1501 AV *trie_words = NULL;
1502 /* along with revcharmap, this only used during construction but both are
1503 * useful during debugging so we store them in the struct when debugging.
1506 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1507 STRLEN trie_charcount=0;
1509 SV *re_trie_maxbuff;
1510 GET_RE_DEBUG_FLAGS_DECL;
1512 PERL_ARGS_ASSERT_MAKE_TRIE;
1514 PERL_UNUSED_ARG(depth);
1520 case EXACTFU: folder = PL_fold_latin1; break;
1521 case EXACTF: folder = PL_fold; break;
1522 case EXACTFL: folder = PL_fold_locale; break;
1523 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u", (unsigned) flags );
1526 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1528 trie->startstate = 1;
1529 trie->wordcount = word_count;
1530 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1531 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1532 if (!(UTF && folder))
1533 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1534 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1535 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1538 trie_words = newAV();
1541 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1542 if (!SvIOK(re_trie_maxbuff)) {
1543 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1546 PerlIO_printf( Perl_debug_log,
1547 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1548 (int)depth * 2 + 2, "",
1549 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1550 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1554 /* Find the node we are going to overwrite */
1555 if ( first == startbranch && OP( last ) != BRANCH ) {
1556 /* whole branch chain */
1559 /* branch sub-chain */
1560 convert = NEXTOPER( first );
1563 /* -- First loop and Setup --
1565 We first traverse the branches and scan each word to determine if it
1566 contains widechars, and how many unique chars there are, this is
1567 important as we have to build a table with at least as many columns as we
1570 We use an array of integers to represent the character codes 0..255
1571 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1572 native representation of the character value as the key and IV's for the
1575 *TODO* If we keep track of how many times each character is used we can
1576 remap the columns so that the table compression later on is more
1577 efficient in terms of memory by ensuring the most common value is in the
1578 middle and the least common are on the outside. IMO this would be better
1579 than a most to least common mapping as theres a decent chance the most
1580 common letter will share a node with the least common, meaning the node
1581 will not be compressible. With a middle is most common approach the worst
1582 case is when we have the least common nodes twice.
1586 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1587 regnode * const noper = NEXTOPER( cur );
1588 const U8 *uc = (U8*)STRING( noper );
1589 const U8 * const e = uc + STR_LEN( noper );
1591 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1592 const U8 *scan = (U8*)NULL;
1593 U32 wordlen = 0; /* required init */
1595 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1597 if (OP(noper) == NOTHING) {
1601 if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
1602 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1603 regardless of encoding */
1605 for ( ; uc < e ; uc += len ) {
1606 TRIE_CHARCOUNT(trie)++;
1610 if ( !trie->charmap[ uvc ] ) {
1611 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1613 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
1617 /* store the codepoint in the bitmap, and its folded
1619 TRIE_BITMAP_SET(trie,uvc);
1621 /* store the folded codepoint */
1622 if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);
1625 /* store first byte of utf8 representation of
1626 variant codepoints */
1627 if (! UNI_IS_INVARIANT(uvc)) {
1628 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1631 set_bit = 0; /* We've done our bit :-) */
1636 widecharmap = newHV();
1638 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1641 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1643 if ( !SvTRUE( *svpp ) ) {
1644 sv_setiv( *svpp, ++trie->uniquecharcount );
1649 if( cur == first ) {
1652 } else if (chars < trie->minlen) {
1654 } else if (chars > trie->maxlen) {
1658 } /* end first pass */
1659 DEBUG_TRIE_COMPILE_r(
1660 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1661 (int)depth * 2 + 2,"",
1662 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1663 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1664 (int)trie->minlen, (int)trie->maxlen )
1668 We now know what we are dealing with in terms of unique chars and
1669 string sizes so we can calculate how much memory a naive
1670 representation using a flat table will take. If it's over a reasonable
1671 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1672 conservative but potentially much slower representation using an array
1675 At the end we convert both representations into the same compressed
1676 form that will be used in regexec.c for matching with. The latter
1677 is a form that cannot be used to construct with but has memory
1678 properties similar to the list form and access properties similar
1679 to the table form making it both suitable for fast searches and
1680 small enough that its feasable to store for the duration of a program.
1682 See the comment in the code where the compressed table is produced
1683 inplace from the flat tabe representation for an explanation of how
1684 the compression works.
1689 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1692 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1694 Second Pass -- Array Of Lists Representation
1696 Each state will be represented by a list of charid:state records
1697 (reg_trie_trans_le) the first such element holds the CUR and LEN
1698 points of the allocated array. (See defines above).
1700 We build the initial structure using the lists, and then convert
1701 it into the compressed table form which allows faster lookups
1702 (but cant be modified once converted).
1705 STRLEN transcount = 1;
1707 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1708 "%*sCompiling trie using list compiler\n",
1709 (int)depth * 2 + 2, ""));
1711 trie->states = (reg_trie_state *)
1712 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1713 sizeof(reg_trie_state) );
1717 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1719 regnode * const noper = NEXTOPER( cur );
1720 U8 *uc = (U8*)STRING( noper );
1721 const U8 * const e = uc + STR_LEN( noper );
1722 U32 state = 1; /* required init */
1723 U16 charid = 0; /* sanity init */
1724 U8 *scan = (U8*)NULL; /* sanity init */
1725 STRLEN foldlen = 0; /* required init */
1726 U32 wordlen = 0; /* required init */
1727 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1729 if (OP(noper) != NOTHING) {
1730 for ( ; uc < e ; uc += len ) {
1735 charid = trie->charmap[ uvc ];
1737 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1741 charid=(U16)SvIV( *svpp );
1744 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1751 if ( !trie->states[ state ].trans.list ) {
1752 TRIE_LIST_NEW( state );
1754 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1755 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1756 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1761 newstate = next_alloc++;
1762 prev_states[newstate] = state;
1763 TRIE_LIST_PUSH( state, charid, newstate );
1768 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1772 TRIE_HANDLE_WORD(state);
1774 } /* end second pass */
1776 /* next alloc is the NEXT state to be allocated */
1777 trie->statecount = next_alloc;
1778 trie->states = (reg_trie_state *)
1779 PerlMemShared_realloc( trie->states,
1781 * sizeof(reg_trie_state) );
1783 /* and now dump it out before we compress it */
1784 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1785 revcharmap, next_alloc,
1789 trie->trans = (reg_trie_trans *)
1790 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1797 for( state=1 ; state < next_alloc ; state ++ ) {
1801 DEBUG_TRIE_COMPILE_MORE_r(
1802 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1806 if (trie->states[state].trans.list) {
1807 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1811 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1812 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1813 if ( forid < minid ) {
1815 } else if ( forid > maxid ) {
1819 if ( transcount < tp + maxid - minid + 1) {
1821 trie->trans = (reg_trie_trans *)
1822 PerlMemShared_realloc( trie->trans,
1824 * sizeof(reg_trie_trans) );
1825 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1827 base = trie->uniquecharcount + tp - minid;
1828 if ( maxid == minid ) {
1830 for ( ; zp < tp ; zp++ ) {
1831 if ( ! trie->trans[ zp ].next ) {
1832 base = trie->uniquecharcount + zp - minid;
1833 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1834 trie->trans[ zp ].check = state;
1840 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1841 trie->trans[ tp ].check = state;
1846 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1847 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1848 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1849 trie->trans[ tid ].check = state;
1851 tp += ( maxid - minid + 1 );
1853 Safefree(trie->states[ state ].trans.list);
1856 DEBUG_TRIE_COMPILE_MORE_r(
1857 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1860 trie->states[ state ].trans.base=base;
1862 trie->lasttrans = tp + 1;
1866 Second Pass -- Flat Table Representation.
1868 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1869 We know that we will need Charcount+1 trans at most to store the data
1870 (one row per char at worst case) So we preallocate both structures
1871 assuming worst case.
1873 We then construct the trie using only the .next slots of the entry
1876 We use the .check field of the first entry of the node temporarily to
1877 make compression both faster and easier by keeping track of how many non
1878 zero fields are in the node.
1880 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1883 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1884 number representing the first entry of the node, and state as a
1885 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1886 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1887 are 2 entrys per node. eg:
1895 The table is internally in the right hand, idx form. However as we also
1896 have to deal with the states array which is indexed by nodenum we have to
1897 use TRIE_NODENUM() to convert.
1900 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1901 "%*sCompiling trie using table compiler\n",
1902 (int)depth * 2 + 2, ""));
1904 trie->trans = (reg_trie_trans *)
1905 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1906 * trie->uniquecharcount + 1,
1907 sizeof(reg_trie_trans) );
1908 trie->states = (reg_trie_state *)
1909 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1910 sizeof(reg_trie_state) );
1911 next_alloc = trie->uniquecharcount + 1;
1914 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1916 regnode * const noper = NEXTOPER( cur );
1917 const U8 *uc = (U8*)STRING( noper );
1918 const U8 * const e = uc + STR_LEN( noper );
1920 U32 state = 1; /* required init */
1922 U16 charid = 0; /* sanity init */
1923 U32 accept_state = 0; /* sanity init */
1924 U8 *scan = (U8*)NULL; /* sanity init */
1926 STRLEN foldlen = 0; /* required init */
1927 U32 wordlen = 0; /* required init */
1928 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1930 if ( OP(noper) != NOTHING ) {
1931 for ( ; uc < e ; uc += len ) {
1936 charid = trie->charmap[ uvc ];
1938 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1939 charid = svpp ? (U16)SvIV(*svpp) : 0;
1943 if ( !trie->trans[ state + charid ].next ) {
1944 trie->trans[ state + charid ].next = next_alloc;
1945 trie->trans[ state ].check++;
1946 prev_states[TRIE_NODENUM(next_alloc)]
1947 = TRIE_NODENUM(state);
1948 next_alloc += trie->uniquecharcount;
1950 state = trie->trans[ state + charid ].next;
1952 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1954 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1957 accept_state = TRIE_NODENUM( state );
1958 TRIE_HANDLE_WORD(accept_state);
1960 } /* end second pass */
1962 /* and now dump it out before we compress it */
1963 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1965 next_alloc, depth+1));
1969 * Inplace compress the table.*
1971 For sparse data sets the table constructed by the trie algorithm will
1972 be mostly 0/FAIL transitions or to put it another way mostly empty.
1973 (Note that leaf nodes will not contain any transitions.)
1975 This algorithm compresses the tables by eliminating most such
1976 transitions, at the cost of a modest bit of extra work during lookup:
1978 - Each states[] entry contains a .base field which indicates the
1979 index in the state[] array wheres its transition data is stored.
1981 - If .base is 0 there are no valid transitions from that node.
1983 - If .base is nonzero then charid is added to it to find an entry in
1986 -If trans[states[state].base+charid].check!=state then the
1987 transition is taken to be a 0/Fail transition. Thus if there are fail
1988 transitions at the front of the node then the .base offset will point
1989 somewhere inside the previous nodes data (or maybe even into a node
1990 even earlier), but the .check field determines if the transition is
1994 The following process inplace converts the table to the compressed
1995 table: We first do not compress the root node 1,and mark all its
1996 .check pointers as 1 and set its .base pointer as 1 as well. This
1997 allows us to do a DFA construction from the compressed table later,
1998 and ensures that any .base pointers we calculate later are greater
2001 - We set 'pos' to indicate the first entry of the second node.
2003 - We then iterate over the columns of the node, finding the first and
2004 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2005 and set the .check pointers accordingly, and advance pos
2006 appropriately and repreat for the next node. Note that when we copy
2007 the next pointers we have to convert them from the original
2008 NODEIDX form to NODENUM form as the former is not valid post
2011 - If a node has no transitions used we mark its base as 0 and do not
2012 advance the pos pointer.
2014 - If a node only has one transition we use a second pointer into the
2015 structure to fill in allocated fail transitions from other states.
2016 This pointer is independent of the main pointer and scans forward
2017 looking for null transitions that are allocated to a state. When it
2018 finds one it writes the single transition into the "hole". If the
2019 pointer doesnt find one the single transition is appended as normal.
2021 - Once compressed we can Renew/realloc the structures to release the
2024 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2025 specifically Fig 3.47 and the associated pseudocode.
2029 const U32 laststate = TRIE_NODENUM( next_alloc );
2032 trie->statecount = laststate;
2034 for ( state = 1 ; state < laststate ; state++ ) {
2036 const U32 stateidx = TRIE_NODEIDX( state );
2037 const U32 o_used = trie->trans[ stateidx ].check;
2038 U32 used = trie->trans[ stateidx ].check;
2039 trie->trans[ stateidx ].check = 0;
2041 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2042 if ( flag || trie->trans[ stateidx + charid ].next ) {
2043 if ( trie->trans[ stateidx + charid ].next ) {
2045 for ( ; zp < pos ; zp++ ) {
2046 if ( ! trie->trans[ zp ].next ) {
2050 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2051 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2052 trie->trans[ zp ].check = state;
2053 if ( ++zp > pos ) pos = zp;
2060 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2062 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2063 trie->trans[ pos ].check = state;
2068 trie->lasttrans = pos + 1;
2069 trie->states = (reg_trie_state *)
2070 PerlMemShared_realloc( trie->states, laststate
2071 * sizeof(reg_trie_state) );
2072 DEBUG_TRIE_COMPILE_MORE_r(
2073 PerlIO_printf( Perl_debug_log,
2074 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2075 (int)depth * 2 + 2,"",
2076 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2079 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2082 } /* end table compress */
2084 DEBUG_TRIE_COMPILE_MORE_r(
2085 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2086 (int)depth * 2 + 2, "",
2087 (UV)trie->statecount,
2088 (UV)trie->lasttrans)
2090 /* resize the trans array to remove unused space */
2091 trie->trans = (reg_trie_trans *)
2092 PerlMemShared_realloc( trie->trans, trie->lasttrans
2093 * sizeof(reg_trie_trans) );
2095 { /* Modify the program and insert the new TRIE node */
2096 U8 nodetype =(U8)(flags & 0xFF);
2100 regnode *optimize = NULL;
2101 #ifdef RE_TRACK_PATTERN_OFFSETS
2104 U32 mjd_nodelen = 0;
2105 #endif /* RE_TRACK_PATTERN_OFFSETS */
2106 #endif /* DEBUGGING */
2108 This means we convert either the first branch or the first Exact,
2109 depending on whether the thing following (in 'last') is a branch
2110 or not and whther first is the startbranch (ie is it a sub part of
2111 the alternation or is it the whole thing.)
2112 Assuming its a sub part we convert the EXACT otherwise we convert
2113 the whole branch sequence, including the first.
2115 /* Find the node we are going to overwrite */
2116 if ( first != startbranch || OP( last ) == BRANCH ) {
2117 /* branch sub-chain */
2118 NEXT_OFF( first ) = (U16)(last - first);
2119 #ifdef RE_TRACK_PATTERN_OFFSETS
2121 mjd_offset= Node_Offset((convert));
2122 mjd_nodelen= Node_Length((convert));
2125 /* whole branch chain */
2127 #ifdef RE_TRACK_PATTERN_OFFSETS
2130 const regnode *nop = NEXTOPER( convert );
2131 mjd_offset= Node_Offset((nop));
2132 mjd_nodelen= Node_Length((nop));
2136 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2137 (int)depth * 2 + 2, "",
2138 (UV)mjd_offset, (UV)mjd_nodelen)
2141 /* But first we check to see if there is a common prefix we can
2142 split out as an EXACT and put in front of the TRIE node. */
2143 trie->startstate= 1;
2144 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2146 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2150 const U32 base = trie->states[ state ].trans.base;
2152 if ( trie->states[state].wordnum )
2155 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2156 if ( ( base + ofs >= trie->uniquecharcount ) &&
2157 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2158 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2160 if ( ++count > 1 ) {
2161 SV **tmp = av_fetch( revcharmap, ofs, 0);
2162 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2163 if ( state == 1 ) break;
2165 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2167 PerlIO_printf(Perl_debug_log,
2168 "%*sNew Start State=%"UVuf" Class: [",
2169 (int)depth * 2 + 2, "",
2172 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2173 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2175 TRIE_BITMAP_SET(trie,*ch);
2177 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2179 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2183 TRIE_BITMAP_SET(trie,*ch);
2185 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2186 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2192 SV **tmp = av_fetch( revcharmap, idx, 0);
2194 char *ch = SvPV( *tmp, len );
2196 SV *sv=sv_newmortal();
2197 PerlIO_printf( Perl_debug_log,
2198 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2199 (int)depth * 2 + 2, "",
2201 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2202 PL_colors[0], PL_colors[1],
2203 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2204 PERL_PV_ESCAPE_FIRSTCHAR
2209 OP( convert ) = nodetype;
2210 str=STRING(convert);
2213 STR_LEN(convert) += len;
2219 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2224 trie->prefixlen = (state-1);
2226 regnode *n = convert+NODE_SZ_STR(convert);
2227 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2228 trie->startstate = state;
2229 trie->minlen -= (state - 1);
2230 trie->maxlen -= (state - 1);
2232 /* At least the UNICOS C compiler choked on this
2233 * being argument to DEBUG_r(), so let's just have
2236 #ifdef PERL_EXT_RE_BUILD
2242 regnode *fix = convert;
2243 U32 word = trie->wordcount;
2245 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2246 while( ++fix < n ) {
2247 Set_Node_Offset_Length(fix, 0, 0);
2250 SV ** const tmp = av_fetch( trie_words, word, 0 );
2252 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2253 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2255 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2263 NEXT_OFF(convert) = (U16)(tail - convert);
2264 DEBUG_r(optimize= n);
2270 if ( trie->maxlen ) {
2271 NEXT_OFF( convert ) = (U16)(tail - convert);
2272 ARG_SET( convert, data_slot );
2273 /* Store the offset to the first unabsorbed branch in
2274 jump[0], which is otherwise unused by the jump logic.
2275 We use this when dumping a trie and during optimisation. */
2277 trie->jump[0] = (U16)(nextbranch - convert);
2279 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2280 * and there is a bitmap
2281 * and the first "jump target" node we found leaves enough room
2282 * then convert the TRIE node into a TRIEC node, with the bitmap
2283 * embedded inline in the opcode - this is hypothetically faster.
2285 if ( !trie->states[trie->startstate].wordnum
2287 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2289 OP( convert ) = TRIEC;
2290 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2291 PerlMemShared_free(trie->bitmap);
2294 OP( convert ) = TRIE;
2296 /* store the type in the flags */
2297 convert->flags = nodetype;
2301 + regarglen[ OP( convert ) ];
2303 /* XXX We really should free up the resource in trie now,
2304 as we won't use them - (which resources?) dmq */
2306 /* needed for dumping*/
2307 DEBUG_r(if (optimize) {
2308 regnode *opt = convert;
2310 while ( ++opt < optimize) {
2311 Set_Node_Offset_Length(opt,0,0);
2314 Try to clean up some of the debris left after the
2317 while( optimize < jumper ) {
2318 mjd_nodelen += Node_Length((optimize));
2319 OP( optimize ) = OPTIMIZED;
2320 Set_Node_Offset_Length(optimize,0,0);
2323 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2325 } /* end node insert */
2327 /* Finish populating the prev field of the wordinfo array. Walk back
2328 * from each accept state until we find another accept state, and if
2329 * so, point the first word's .prev field at the second word. If the
2330 * second already has a .prev field set, stop now. This will be the
2331 * case either if we've already processed that word's accept state,
2332 * or that state had multiple words, and the overspill words were
2333 * already linked up earlier.
2340 for (word=1; word <= trie->wordcount; word++) {
2342 if (trie->wordinfo[word].prev)
2344 state = trie->wordinfo[word].accept;
2346 state = prev_states[state];
2349 prev = trie->states[state].wordnum;
2353 trie->wordinfo[word].prev = prev;
2355 Safefree(prev_states);
2359 /* and now dump out the compressed format */
2360 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2362 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2364 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2365 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2367 SvREFCNT_dec(revcharmap);
2371 : trie->startstate>1
2377 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2379 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2381 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2382 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2385 We find the fail state for each state in the trie, this state is the longest proper
2386 suffix of the current state's 'word' that is also a proper prefix of another word in our
2387 trie. State 1 represents the word '' and is thus the default fail state. This allows
2388 the DFA not to have to restart after its tried and failed a word at a given point, it
2389 simply continues as though it had been matching the other word in the first place.
2391 'abcdgu'=~/abcdefg|cdgu/
2392 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2393 fail, which would bring us to the state representing 'd' in the second word where we would
2394 try 'g' and succeed, proceeding to match 'cdgu'.
2396 /* add a fail transition */
2397 const U32 trie_offset = ARG(source);
2398 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2400 const U32 ucharcount = trie->uniquecharcount;
2401 const U32 numstates = trie->statecount;
2402 const U32 ubound = trie->lasttrans + ucharcount;
2406 U32 base = trie->states[ 1 ].trans.base;
2409 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2410 GET_RE_DEBUG_FLAGS_DECL;
2412 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2414 PERL_UNUSED_ARG(depth);
2418 ARG_SET( stclass, data_slot );
2419 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2420 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2421 aho->trie=trie_offset;
2422 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2423 Copy( trie->states, aho->states, numstates, reg_trie_state );
2424 Newxz( q, numstates, U32);
2425 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2428 /* initialize fail[0..1] to be 1 so that we always have
2429 a valid final fail state */
2430 fail[ 0 ] = fail[ 1 ] = 1;
2432 for ( charid = 0; charid < ucharcount ; charid++ ) {
2433 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2435 q[ q_write ] = newstate;
2436 /* set to point at the root */
2437 fail[ q[ q_write++ ] ]=1;
2440 while ( q_read < q_write) {
2441 const U32 cur = q[ q_read++ % numstates ];
2442 base = trie->states[ cur ].trans.base;
2444 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2445 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2447 U32 fail_state = cur;
2450 fail_state = fail[ fail_state ];
2451 fail_base = aho->states[ fail_state ].trans.base;
2452 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2454 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2455 fail[ ch_state ] = fail_state;
2456 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2458 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2460 q[ q_write++ % numstates] = ch_state;
2464 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2465 when we fail in state 1, this allows us to use the
2466 charclass scan to find a valid start char. This is based on the principle
2467 that theres a good chance the string being searched contains lots of stuff
2468 that cant be a start char.
2470 fail[ 0 ] = fail[ 1 ] = 0;
2471 DEBUG_TRIE_COMPILE_r({
2472 PerlIO_printf(Perl_debug_log,
2473 "%*sStclass Failtable (%"UVuf" states): 0",
2474 (int)(depth * 2), "", (UV)numstates
2476 for( q_read=1; q_read<numstates; q_read++ ) {
2477 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2479 PerlIO_printf(Perl_debug_log, "\n");
2482 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2487 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2488 * These need to be revisited when a newer toolchain becomes available.
2490 #if defined(__sparc64__) && defined(__GNUC__)
2491 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2492 # undef SPARC64_GCC_WORKAROUND
2493 # define SPARC64_GCC_WORKAROUND 1
2497 #define DEBUG_PEEP(str,scan,depth) \
2498 DEBUG_OPTIMISE_r({if (scan){ \
2499 SV * const mysv=sv_newmortal(); \
2500 regnode *Next = regnext(scan); \
2501 regprop(RExC_rx, mysv, scan); \
2502 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2503 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2504 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2508 /* The below joins as many adjacent EXACTish nodes as possible into a single
2509 * one, and looks for problematic sequences of characters whose folds vs.
2510 * non-folds have sufficiently different lengths, that the optimizer would be
2511 * fooled into rejecting legitimate matches of them, and the trie construction
2512 * code can't cope with them. The joining is only done if:
2513 * 1) there is room in the current conglomerated node to entirely contain the
2515 * 2) they are the exact same node type
2517 * The adjacent nodes actually may be separated by NOTHING kind nodes, and
2518 * these get optimized out
2520 * If there are problematic code sequences, *min_subtract is set to the delta
2521 * that the minimum size of the node can be less than its actual size. And,
2522 * the node type of the result is changed to reflect that it contains these
2525 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2526 * and contains LATIN SMALL LETTER SHARP S
2528 * This is as good a place as any to discuss the design of handling these
2529 * problematic sequences. It's been wrong in Perl for a very long time. There
2530 * are three code points in Unicode whose folded lengths differ so much from
2531 * the un-folded lengths that it causes problems for the optimizer and trie
2532 * construction. Why only these are problematic, and not others where lengths
2533 * also differ is something I (khw) do not understand. New versions of Unicode
2534 * might add more such code points. Hopefully the logic in fold_grind.t that
2535 * figures out what to test (in part by verifying that each size-combination
2536 * gets tested) will catch any that do come along, so they can be added to the
2537 * special handling below. The chances of new ones are actually rather small,
2538 * as most, if not all, of the world's scripts that have casefolding have
2539 * already been encoded by Unicode. Also, a number of Unicode's decisions were
2540 * made to allow compatibility with pre-existing standards, and almost all of
2541 * those have already been dealt with. These would otherwise be the most
2542 * likely candidates for generating further tricky sequences. In other words,
2543 * Unicode by itself is unlikely to add new ones unless it is for compatibility
2544 * with pre-existing standards, and there aren't many of those left.
2546 * The previous designs for dealing with these involved assigning a special
2547 * node for them. This approach doesn't work, as evidenced by this example:
2548 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2549 * Both these fold to "sss", but if the pattern is parsed to create a node of
2550 * that would match just the \xDF, it won't be able to handle the case where a
2551 * successful match would have to cross the node's boundary. The new approach
2552 * that hopefully generally solves the problem generates an EXACTFU_SS node
2555 * There are a number of components to the approach (a lot of work for just
2556 * three code points!):
2557 * 1) This routine examines each EXACTFish node that could contain the
2558 * problematic sequences. It returns in *min_subtract how much to
2559 * subtract from the the actual length of the string to get a real minimum
2560 * for one that could match it. This number is usually 0 except for the
2561 * problematic sequences. This delta is used by the caller to adjust the
2562 * min length of the match, and the delta between min and max, so that the
2563 * optimizer doesn't reject these possibilities based on size constraints.
2564 * 2) These sequences are not currently correctly handled by the trie code
2565 * either, so it changes the joined node type to ops that are not handled
2566 * by trie's, those new ops being EXACTFU_SS and EXACTFU_NO_TRIE.
2567 * 3) This is sufficient for the two Greek sequences (described below), but
2568 * the one involving the Sharp s (\xDF) needs more. The node type
2569 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2570 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2571 * case where there is a possible fold length change. That means that a
2572 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2573 * itself with length changes, and so can be processed faster. regexec.c
2574 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2575 * is pre-folded by regcomp.c. This saves effort in regex matching.
2576 * However, probably mostly for historical reasons, the pre-folding isn't
2577 * done for non-UTF8 patterns (and it can't be for EXACTF and EXACTFL
2578 * nodes, as what they fold to isn't known until runtime.) The fold
2579 * possibilities for the non-UTF8 patterns are quite simple, except for
2580 * the sharp s. All the ones that don't involve a UTF-8 target string
2581 * are members of a fold-pair, and arrays are set up for all of them
2582 * that quickly find the other member of the pair. It might actually
2583 * be faster to pre-fold these, but it isn't currently done, except for
2584 * the sharp s. Code elsewhere in this file makes sure that it gets
2585 * folded to 'ss', even if the pattern isn't UTF-8. This avoids the
2586 * issues described in the next item.
2587 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2588 * 'ss' or not is not knowable at compile time. It will match iff the
2589 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2590 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2591 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2592 * described in item 3). An assumption that the optimizer part of
2593 * regexec.c (probably unwittingly) makes is that a character in the
2594 * pattern corresponds to at most a single character in the target string.
2595 * (And I do mean character, and not byte here, unlike other parts of the
2596 * documentation that have never been updated to account for multibyte
2597 * Unicode.) This assumption is wrong only in this case, as all other
2598 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2599 * virtue of having this file pre-fold UTF-8 patterns. I'm
2600 * reluctant to try to change this assumption, so instead the code punts.
2601 * This routine examines EXACTF nodes for the sharp s, and returns a
2602 * boolean indicating whether or not the node is an EXACTF node that
2603 * contains a sharp s. When it is true, the caller sets a flag that later
2604 * causes the optimizer in this file to not set values for the floating
2605 * and fixed string lengths, and thus avoids the optimizer code in
2606 * regexec.c that makes the invalid assumption. Thus, there is no
2607 * optimization based on string lengths for EXACTF nodes that contain the
2608 * sharp s. This only happens for /id rules (which means the pattern
2612 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2613 if (PL_regkind[OP(scan)] == EXACT) \
2614 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2617 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, UV *min_subtract, bool *has_exactf_sharp_s, U32 flags,regnode *val, U32 depth) {
2618 /* Merge several consecutive EXACTish nodes into one. */
2619 regnode *n = regnext(scan);
2621 regnode *next = scan + NODE_SZ_STR(scan);
2625 regnode *stop = scan;
2626 GET_RE_DEBUG_FLAGS_DECL;
2628 PERL_UNUSED_ARG(depth);
2631 PERL_ARGS_ASSERT_JOIN_EXACT;
2632 #ifndef EXPERIMENTAL_INPLACESCAN
2633 PERL_UNUSED_ARG(flags);
2634 PERL_UNUSED_ARG(val);
2636 DEBUG_PEEP("join",scan,depth);
2638 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2639 * EXACT ones that are mergeable to the current one. */
2641 && (PL_regkind[OP(n)] == NOTHING
2642 || (stringok && OP(n) == OP(scan)))
2644 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2647 if (OP(n) == TAIL || n > next)
2649 if (PL_regkind[OP(n)] == NOTHING) {
2650 DEBUG_PEEP("skip:",n,depth);
2651 NEXT_OFF(scan) += NEXT_OFF(n);
2652 next = n + NODE_STEP_REGNODE;
2659 else if (stringok) {
2660 const unsigned int oldl = STR_LEN(scan);
2661 regnode * const nnext = regnext(n);
2663 if (oldl + STR_LEN(n) > U8_MAX)
2666 DEBUG_PEEP("merg",n,depth);
2669 NEXT_OFF(scan) += NEXT_OFF(n);
2670 STR_LEN(scan) += STR_LEN(n);
2671 next = n + NODE_SZ_STR(n);
2672 /* Now we can overwrite *n : */
2673 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2681 #ifdef EXPERIMENTAL_INPLACESCAN
2682 if (flags && !NEXT_OFF(n)) {
2683 DEBUG_PEEP("atch", val, depth);
2684 if (reg_off_by_arg[OP(n)]) {
2685 ARG_SET(n, val - n);
2688 NEXT_OFF(n) = val - n;
2696 *has_exactf_sharp_s = FALSE;
2698 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2699 * can now analyze for sequences of problematic code points. (Prior to
2700 * this final joining, sequences could have been split over boundaries, and
2701 * hence missed). The sequences only happen in folding, hence for any
2702 * non-EXACT EXACTish node */
2703 if (OP(scan) != EXACT) {
2705 U8 * s0 = (U8*) STRING(scan);
2706 U8 * const s_end = s0 + STR_LEN(scan);
2708 /* The below is perhaps overboard, but this allows us to save a test
2709 * each time through the loop at the expense of a mask. This is
2710 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2711 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2712 * This uses an exclusive 'or' to find that bit and then inverts it to
2713 * form a mask, with just a single 0, in the bit position where 'S' and
2715 const U8 S_or_s_mask = ~ ('S' ^ 's');
2716 const U8 s_masked = 's' & S_or_s_mask;
2718 /* One pass is made over the node's string looking for all the
2719 * possibilities. to avoid some tests in the loop, there are two main
2720 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2724 /* There are two problematic Greek code points in Unicode
2727 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2728 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2734 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2735 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2737 * This means that in case-insensitive matching (or "loose
2738 * matching", as Unicode calls it), an EXACTF of length six (the
2739 * UTF-8 encoded byte length of the above casefolded versions) can
2740 * match a target string of length two (the byte length of UTF-8
2741 * encoded U+0390 or U+03B0). This would rather mess up the
2742 * minimum length computation. (there are other code points that
2743 * also fold to these two sequences, but the delta is smaller)
2745 * If these sequences are found, the minimum length is decreased by
2746 * four (six minus two).
2748 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2749 * LETTER SHARP S. We decrease the min length by 1 for each
2750 * occurrence of 'ss' found */
2752 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2753 # define U390_first_byte 0xb4
2754 const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
2755 # define U3B0_first_byte 0xb5
2756 const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
2758 # define U390_first_byte 0xce
2759 const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
2760 # define U3B0_first_byte 0xcf
2761 const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
2763 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2764 yields a net of 0 */
2765 /* Examine the string for one of the problematic sequences */
2767 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2768 * sequence we are looking for is 2 */
2772 /* Look for the first byte in each problematic sequence */
2774 /* We don't have to worry about other things that fold to
2775 * 's' (such as the long s, U+017F), as all above-latin1
2776 * code points have been pre-folded */
2780 /* Current character is an 's' or 'S'. If next one is
2781 * as well, we have the dreaded sequence */
2782 if (((*(s+1) & S_or_s_mask) == s_masked)
2783 /* These two node types don't have special handling
2785 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2788 OP(scan) = EXACTFU_SS;
2789 s++; /* No need to look at this character again */
2793 case U390_first_byte:
2794 if (s_end - s >= len
2796 /* The 1's are because are skipping comparing the
2798 && memEQ(s + 1, U390_tail, len - 1))
2800 goto greek_sequence;
2804 case U3B0_first_byte:
2805 if (! (s_end - s >= len
2806 && memEQ(s + 1, U3B0_tail, len - 1)))
2813 /* This can't currently be handled by trie's, so change
2814 * the node type to indicate this. If EXACTFA and
2815 * EXACTFL were ever to be handled by trie's, this
2816 * would have to be changed. If this node has already
2817 * been changed to EXACTFU_SS in this loop, leave it as
2818 * is. (I (khw) think it doesn't matter in regexec.c
2819 * for UTF patterns, but no need to change it */
2820 if (OP(scan) == EXACTFU) {
2821 OP(scan) = EXACTFU_NO_TRIE;
2823 s += 6; /* We already know what this sequence is. Skip
2829 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2831 /* Here, the pattern is not UTF-8. We need to look only for the
2832 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2833 * in the final position. Otherwise we can stop looking 1 byte
2834 * earlier because have to find both the first and second 's' */
2835 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2837 for (s = s0; s < upper; s++) {
2842 && ((*(s+1) & S_or_s_mask) == s_masked))
2846 /* EXACTF nodes need to know that the minimum
2847 * length changed so that a sharp s in the string
2848 * can match this ss in the pattern, but they
2849 * remain EXACTF nodes, as they are not trie'able,
2850 * so don't have to invent a new node type to
2851 * exclude them from the trie code */
2852 if (OP(scan) != EXACTF) {
2853 OP(scan) = EXACTFU_SS;
2858 case LATIN_SMALL_LETTER_SHARP_S:
2859 if (OP(scan) == EXACTF) {
2860 *has_exactf_sharp_s = TRUE;
2869 /* Allow dumping but overwriting the collection of skipped
2870 * ops and/or strings with fake optimized ops */
2871 n = scan + NODE_SZ_STR(scan);
2879 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2883 /* REx optimizer. Converts nodes into quicker variants "in place".
2884 Finds fixed substrings. */
2886 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2887 to the position after last scanned or to NULL. */
2889 #define INIT_AND_WITHP \
2890 assert(!and_withp); \
2891 Newx(and_withp,1,struct regnode_charclass_class); \
2892 SAVEFREEPV(and_withp)
2894 /* this is a chain of data about sub patterns we are processing that
2895 need to be handled separately/specially in study_chunk. Its so
2896 we can simulate recursion without losing state. */
2898 typedef struct scan_frame {
2899 regnode *last; /* last node to process in this frame */
2900 regnode *next; /* next node to process when last is reached */
2901 struct scan_frame *prev; /*previous frame*/
2902 I32 stop; /* what stopparen do we use */
2906 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2908 #define CASE_SYNST_FNC(nAmE) \
2910 if (flags & SCF_DO_STCLASS_AND) { \
2911 for (value = 0; value < 256; value++) \
2912 if (!is_ ## nAmE ## _cp(value)) \
2913 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2916 for (value = 0; value < 256; value++) \
2917 if (is_ ## nAmE ## _cp(value)) \
2918 ANYOF_BITMAP_SET(data->start_class, value); \
2922 if (flags & SCF_DO_STCLASS_AND) { \
2923 for (value = 0; value < 256; value++) \
2924 if (is_ ## nAmE ## _cp(value)) \
2925 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2928 for (value = 0; value < 256; value++) \
2929 if (!is_ ## nAmE ## _cp(value)) \
2930 ANYOF_BITMAP_SET(data->start_class, value); \
2937 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2938 I32 *minlenp, I32 *deltap,
2943 struct regnode_charclass_class *and_withp,
2944 U32 flags, U32 depth)
2945 /* scanp: Start here (read-write). */
2946 /* deltap: Write maxlen-minlen here. */
2947 /* last: Stop before this one. */
2948 /* data: string data about the pattern */
2949 /* stopparen: treat close N as END */
2950 /* recursed: which subroutines have we recursed into */
2951 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
2954 I32 min = 0, pars = 0, code;
2955 regnode *scan = *scanp, *next;
2957 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2958 int is_inf_internal = 0; /* The studied chunk is infinite */
2959 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2960 scan_data_t data_fake;
2961 SV *re_trie_maxbuff = NULL;
2962 regnode *first_non_open = scan;
2963 I32 stopmin = I32_MAX;
2964 scan_frame *frame = NULL;
2965 GET_RE_DEBUG_FLAGS_DECL;
2967 PERL_ARGS_ASSERT_STUDY_CHUNK;
2970 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
2974 while (first_non_open && OP(first_non_open) == OPEN)
2975 first_non_open=regnext(first_non_open);
2980 while ( scan && OP(scan) != END && scan < last ){
2981 UV min_subtract = 0; /* How much to subtract from the minimum node
2982 length to get a real minimum (because the
2983 folded version may be shorter) */
2984 bool has_exactf_sharp_s = FALSE;
2985 /* Peephole optimizer: */
2986 DEBUG_STUDYDATA("Peep:", data,depth);
2987 DEBUG_PEEP("Peep",scan,depth);
2989 /* Its not clear to khw or hv why this is done here, and not in the
2990 * clauses that deal with EXACT nodes. khw's guess is that it's
2991 * because of a previous design */
2992 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
2994 /* Follow the next-chain of the current node and optimize
2995 away all the NOTHINGs from it. */
2996 if (OP(scan) != CURLYX) {
2997 const int max = (reg_off_by_arg[OP(scan)]
2999 /* I32 may be smaller than U16 on CRAYs! */
3000 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3001 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3005 /* Skip NOTHING and LONGJMP. */
3006 while ((n = regnext(n))
3007 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3008 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3009 && off + noff < max)
3011 if (reg_off_by_arg[OP(scan)])
3014 NEXT_OFF(scan) = off;
3019 /* The principal pseudo-switch. Cannot be a switch, since we
3020 look into several different things. */
3021 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3022 || OP(scan) == IFTHEN) {
3023 next = regnext(scan);
3025 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3027 if (OP(next) == code || code == IFTHEN) {
3028 /* NOTE - There is similar code to this block below for handling
3029 TRIE nodes on a re-study. If you change stuff here check there
3031 I32 max1 = 0, min1 = I32_MAX, num = 0;
3032 struct regnode_charclass_class accum;
3033 regnode * const startbranch=scan;
3035 if (flags & SCF_DO_SUBSTR)
3036 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3037 if (flags & SCF_DO_STCLASS)
3038 cl_init_zero(pRExC_state, &accum);
3040 while (OP(scan) == code) {
3041 I32 deltanext, minnext, f = 0, fake;
3042 struct regnode_charclass_class this_class;
3045 data_fake.flags = 0;
3047 data_fake.whilem_c = data->whilem_c;
3048 data_fake.last_closep = data->last_closep;
3051 data_fake.last_closep = &fake;
3053 data_fake.pos_delta = delta;
3054 next = regnext(scan);
3055 scan = NEXTOPER(scan);
3057 scan = NEXTOPER(scan);
3058 if (flags & SCF_DO_STCLASS) {
3059 cl_init(pRExC_state, &this_class);
3060 data_fake.start_class = &this_class;
3061 f = SCF_DO_STCLASS_AND;
3063 if (flags & SCF_WHILEM_VISITED_POS)
3064 f |= SCF_WHILEM_VISITED_POS;
3066 /* we suppose the run is continuous, last=next...*/
3067 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3069 stopparen, recursed, NULL, f,depth+1);
3072 if (max1 < minnext + deltanext)
3073 max1 = minnext + deltanext;
3074 if (deltanext == I32_MAX)
3075 is_inf = is_inf_internal = 1;
3077 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3079 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3080 if ( stopmin > minnext)
3081 stopmin = min + min1;
3082 flags &= ~SCF_DO_SUBSTR;
3084 data->flags |= SCF_SEEN_ACCEPT;
3087 if (data_fake.flags & SF_HAS_EVAL)
3088 data->flags |= SF_HAS_EVAL;
3089 data->whilem_c = data_fake.whilem_c;
3091 if (flags & SCF_DO_STCLASS)
3092 cl_or(pRExC_state, &accum, &this_class);
3094 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3096 if (flags & SCF_DO_SUBSTR) {
3097 data->pos_min += min1;
3098 data->pos_delta += max1 - min1;
3099 if (max1 != min1 || is_inf)
3100 data->longest = &(data->longest_float);
3103 delta += max1 - min1;
3104 if (flags & SCF_DO_STCLASS_OR) {
3105 cl_or(pRExC_state, data->start_class, &accum);
3107 cl_and(data->start_class, and_withp);
3108 flags &= ~SCF_DO_STCLASS;
3111 else if (flags & SCF_DO_STCLASS_AND) {
3113 cl_and(data->start_class, &accum);
3114 flags &= ~SCF_DO_STCLASS;
3117 /* Switch to OR mode: cache the old value of
3118 * data->start_class */
3120 StructCopy(data->start_class, and_withp,
3121 struct regnode_charclass_class);
3122 flags &= ~SCF_DO_STCLASS_AND;
3123 StructCopy(&accum, data->start_class,
3124 struct regnode_charclass_class);
3125 flags |= SCF_DO_STCLASS_OR;
3126 data->start_class->flags |= ANYOF_EOS;
3130 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3133 Assuming this was/is a branch we are dealing with: 'scan' now
3134 points at the item that follows the branch sequence, whatever
3135 it is. We now start at the beginning of the sequence and look
3142 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3144 If we can find such a subsequence we need to turn the first
3145 element into a trie and then add the subsequent branch exact
3146 strings to the trie.
3150 1. patterns where the whole set of branches can be converted.
3152 2. patterns where only a subset can be converted.
3154 In case 1 we can replace the whole set with a single regop
3155 for the trie. In case 2 we need to keep the start and end
3158 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3159 becomes BRANCH TRIE; BRANCH X;
3161 There is an additional case, that being where there is a
3162 common prefix, which gets split out into an EXACT like node
3163 preceding the TRIE node.
3165 If x(1..n)==tail then we can do a simple trie, if not we make
3166 a "jump" trie, such that when we match the appropriate word
3167 we "jump" to the appropriate tail node. Essentially we turn
3168 a nested if into a case structure of sorts.
3173 if (!re_trie_maxbuff) {
3174 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3175 if (!SvIOK(re_trie_maxbuff))
3176 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3178 if ( SvIV(re_trie_maxbuff)>=0 ) {
3180 regnode *first = (regnode *)NULL;
3181 regnode *last = (regnode *)NULL;
3182 regnode *tail = scan;
3187 SV * const mysv = sv_newmortal(); /* for dumping */
3189 /* var tail is used because there may be a TAIL
3190 regop in the way. Ie, the exacts will point to the
3191 thing following the TAIL, but the last branch will
3192 point at the TAIL. So we advance tail. If we
3193 have nested (?:) we may have to move through several
3197 while ( OP( tail ) == TAIL ) {
3198 /* this is the TAIL generated by (?:) */
3199 tail = regnext( tail );
3204 regprop(RExC_rx, mysv, tail );
3205 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3206 (int)depth * 2 + 2, "",
3207 "Looking for TRIE'able sequences. Tail node is: ",
3208 SvPV_nolen_const( mysv )
3214 step through the branches, cur represents each
3215 branch, noper is the first thing to be matched
3216 as part of that branch and noper_next is the
3217 regnext() of that node. if noper is an EXACT
3218 and noper_next is the same as scan (our current
3219 position in the regex) then the EXACT branch is
3220 a possible optimization target. Once we have
3221 two or more consecutive such branches we can
3222 create a trie of the EXACT's contents and stich
3223 it in place. If the sequence represents all of
3224 the branches we eliminate the whole thing and
3225 replace it with a single TRIE. If it is a
3226 subsequence then we need to stitch it in. This
3227 means the first branch has to remain, and needs
3228 to be repointed at the item on the branch chain
3229 following the last branch optimized. This could
3230 be either a BRANCH, in which case the
3231 subsequence is internal, or it could be the
3232 item following the branch sequence in which
3233 case the subsequence is at the end.
3237 /* dont use tail as the end marker for this traverse */
3238 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3239 regnode * const noper = NEXTOPER( cur );
3240 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3241 regnode * const noper_next = regnext( noper );
3245 regprop(RExC_rx, mysv, cur);
3246 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3247 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3249 regprop(RExC_rx, mysv, noper);
3250 PerlIO_printf( Perl_debug_log, " -> %s",
3251 SvPV_nolen_const(mysv));
3254 regprop(RExC_rx, mysv, noper_next );
3255 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3256 SvPV_nolen_const(mysv));
3258 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
3259 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
3261 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
3262 : PL_regkind[ OP( noper ) ] == EXACT )
3263 || OP(noper) == NOTHING )
3265 && noper_next == tail
3270 if ( !first || optype == NOTHING ) {
3271 if (!first) first = cur;
3272 optype = OP( noper );
3278 Currently the trie logic handles case insensitive matching properly only
3279 when the pattern is UTF-8 and the node is EXACTFU (thus forcing unicode
3282 If/when this is fixed the following define can be swapped
3283 in below to fully enable trie logic.
3285 #define TRIE_TYPE_IS_SAFE 1
3287 Note that join_exact() assumes that the other types of EXACTFish nodes are not
3288 used in tries, so that would have to be updated if this changed
3291 #define TRIE_TYPE_IS_SAFE ((UTF && optype == EXACTFU) || optype==EXACT)
3293 if ( last && TRIE_TYPE_IS_SAFE ) {
3294 make_trie( pRExC_state,
3295 startbranch, first, cur, tail, count,
3298 if ( PL_regkind[ OP( noper ) ] == EXACT
3300 && noper_next == tail
3305 optype = OP( noper );
3315 regprop(RExC_rx, mysv, cur);
3316 PerlIO_printf( Perl_debug_log,
3317 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3318 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3322 if ( last && TRIE_TYPE_IS_SAFE ) {
3323 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
3324 #ifdef TRIE_STUDY_OPT
3325 if ( ((made == MADE_EXACT_TRIE &&
3326 startbranch == first)
3327 || ( first_non_open == first )) &&
3329 flags |= SCF_TRIE_RESTUDY;
3330 if ( startbranch == first
3333 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3343 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3344 scan = NEXTOPER(NEXTOPER(scan));
3345 } else /* single branch is optimized. */
3346 scan = NEXTOPER(scan);
3348 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3349 scan_frame *newframe = NULL;
3354 if (OP(scan) != SUSPEND) {
3355 /* set the pointer */
3356 if (OP(scan) == GOSUB) {
3358 RExC_recurse[ARG2L(scan)] = scan;
3359 start = RExC_open_parens[paren-1];
3360 end = RExC_close_parens[paren-1];
3363 start = RExC_rxi->program + 1;
3367 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3368 SAVEFREEPV(recursed);
3370 if (!PAREN_TEST(recursed,paren+1)) {
3371 PAREN_SET(recursed,paren+1);
3372 Newx(newframe,1,scan_frame);
3374 if (flags & SCF_DO_SUBSTR) {
3375 SCAN_COMMIT(pRExC_state,data,minlenp);
3376 data->longest = &(data->longest_float);
3378 is_inf = is_inf_internal = 1;
3379 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3380 cl_anything(pRExC_state, data->start_class);
3381 flags &= ~SCF_DO_STCLASS;
3384 Newx(newframe,1,scan_frame);
3387 end = regnext(scan);
3392 SAVEFREEPV(newframe);
3393 newframe->next = regnext(scan);
3394 newframe->last = last;
3395 newframe->stop = stopparen;
3396 newframe->prev = frame;
3406 else if (OP(scan) == EXACT) {
3407 I32 l = STR_LEN(scan);
3410 const U8 * const s = (U8*)STRING(scan);
3411 l = utf8_length(s, s + l);
3412 uc = utf8_to_uvchr(s, NULL);
3414 uc = *((U8*)STRING(scan));
3417 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3418 /* The code below prefers earlier match for fixed
3419 offset, later match for variable offset. */
3420 if (data->last_end == -1) { /* Update the start info. */
3421 data->last_start_min = data->pos_min;
3422 data->last_start_max = is_inf
3423 ? I32_MAX : data->pos_min + data->pos_delta;
3425 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3427 SvUTF8_on(data->last_found);
3429 SV * const sv = data->last_found;
3430 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3431 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3432 if (mg && mg->mg_len >= 0)
3433 mg->mg_len += utf8_length((U8*)STRING(scan),
3434 (U8*)STRING(scan)+STR_LEN(scan));
3436 data->last_end = data->pos_min + l;
3437 data->pos_min += l; /* As in the first entry. */
3438 data->flags &= ~SF_BEFORE_EOL;
3440 if (flags & SCF_DO_STCLASS_AND) {
3441 /* Check whether it is compatible with what we know already! */
3445 /* If compatible, we or it in below. It is compatible if is
3446 * in the bitmp and either 1) its bit or its fold is set, or 2)
3447 * it's for a locale. Even if there isn't unicode semantics
3448 * here, at runtime there may be because of matching against a
3449 * utf8 string, so accept a possible false positive for
3450 * latin1-range folds */
3452 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3453 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3454 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
3455 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3460 ANYOF_CLASS_ZERO(data->start_class);
3461 ANYOF_BITMAP_ZERO(data->start_class);
3463 ANYOF_BITMAP_SET(data->start_class, uc);
3464 else if (uc >= 0x100) {
3467 /* Some Unicode code points fold to the Latin1 range; as
3468 * XXX temporary code, instead of figuring out if this is
3469 * one, just assume it is and set all the start class bits
3470 * that could be some such above 255 code point's fold
3471 * which will generate fals positives. As the code
3472 * elsewhere that does compute the fold settles down, it
3473 * can be extracted out and re-used here */
3474 for (i = 0; i < 256; i++){
3475 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3476 ANYOF_BITMAP_SET(data->start_class, i);
3480 data->start_class->flags &= ~ANYOF_EOS;
3482 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3484 else if (flags & SCF_DO_STCLASS_OR) {
3485 /* false positive possible if the class is case-folded */
3487 ANYOF_BITMAP_SET(data->start_class, uc);
3489 data->start_class->flags |= ANYOF_UNICODE_ALL;
3490 data->start_class->flags &= ~ANYOF_EOS;
3491 cl_and(data->start_class, and_withp);
3493 flags &= ~SCF_DO_STCLASS;
3495 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3496 I32 l = STR_LEN(scan);
3497 UV uc = *((U8*)STRING(scan));
3499 /* Search for fixed substrings supports EXACT only. */
3500 if (flags & SCF_DO_SUBSTR) {
3502 SCAN_COMMIT(pRExC_state, data, minlenp);
3505 const U8 * const s = (U8 *)STRING(scan);
3506 l = utf8_length(s, s + l);
3507 uc = utf8_to_uvchr(s, NULL);
3509 else if (has_exactf_sharp_s) {
3510 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3512 min += l - min_subtract;
3516 delta += min_subtract;
3517 if (flags & SCF_DO_SUBSTR) {
3518 data->pos_min += l - min_subtract;
3519 if (data->pos_min < 0) {
3522 data->pos_delta += min_subtract;
3524 data->longest = &(data->longest_float);
3527 if (flags & SCF_DO_STCLASS_AND) {
3528 /* Check whether it is compatible with what we know already! */
3531 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3532 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3533 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3537 ANYOF_CLASS_ZERO(data->start_class);
3538 ANYOF_BITMAP_ZERO(data->start_class);
3540 ANYOF_BITMAP_SET(data->start_class, uc);
3541 data->start_class->flags &= ~ANYOF_EOS;
3542 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
3543 if (OP(scan) == EXACTFL) {
3544 /* XXX This set is probably no longer necessary, and
3545 * probably wrong as LOCALE now is on in the initial
3547 data->start_class->flags |= ANYOF_LOCALE;
3551 /* Also set the other member of the fold pair. In case
3552 * that unicode semantics is called for at runtime, use
3553 * the full latin1 fold. (Can't do this for locale,
3554 * because not known until runtime) */
3555 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3557 /* All other (EXACTFL handled above) folds except under
3558 * /iaa that include s, S, and sharp_s also may include
3560 if (OP(scan) != EXACTFA) {
3561 if (uc == 's' || uc == 'S') {
3562 ANYOF_BITMAP_SET(data->start_class,
3563 LATIN_SMALL_LETTER_SHARP_S);
3565 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3566 ANYOF_BITMAP_SET(data->start_class, 's');
3567 ANYOF_BITMAP_SET(data->start_class, 'S');
3572 else if (uc >= 0x100) {
3574 for (i = 0; i < 256; i++){
3575 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3576 ANYOF_BITMAP_SET(data->start_class, i);
3581 else if (flags & SCF_DO_STCLASS_OR) {
3582 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
3583 /* false positive possible if the class is case-folded.
3584 Assume that the locale settings are the same... */
3586 ANYOF_BITMAP_SET(data->start_class, uc);
3587 if (OP(scan) != EXACTFL) {
3589 /* And set the other member of the fold pair, but
3590 * can't do that in locale because not known until
3592 ANYOF_BITMAP_SET(data->start_class,
3593 PL_fold_latin1[uc]);
3595 /* All folds except under /iaa that include s, S,
3596 * and sharp_s also may include the others */
3597 if (OP(scan) != EXACTFA) {
3598 if (uc == 's' || uc == 'S') {
3599 ANYOF_BITMAP_SET(data->start_class,
3600 LATIN_SMALL_LETTER_SHARP_S);
3602 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3603 ANYOF_BITMAP_SET(data->start_class, 's');
3604 ANYOF_BITMAP_SET(data->start_class, 'S');
3609 data->start_class->flags &= ~ANYOF_EOS;
3611 cl_and(data->start_class, and_withp);
3613 flags &= ~SCF_DO_STCLASS;
3615 else if (REGNODE_VARIES(OP(scan))) {
3616 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3617 I32 f = flags, pos_before = 0;
3618 regnode * const oscan = scan;
3619 struct regnode_charclass_class this_class;
3620 struct regnode_charclass_class *oclass = NULL;
3621 I32 next_is_eval = 0;
3623 switch (PL_regkind[OP(scan)]) {
3624 case WHILEM: /* End of (?:...)* . */
3625 scan = NEXTOPER(scan);
3628 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3629 next = NEXTOPER(scan);
3630 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3632 maxcount = REG_INFTY;
3633 next = regnext(scan);
3634 scan = NEXTOPER(scan);
3638 if (flags & SCF_DO_SUBSTR)
3643 if (flags & SCF_DO_STCLASS) {
3645 maxcount = REG_INFTY;
3646 next = regnext(scan);
3647 scan = NEXTOPER(scan);
3650 is_inf = is_inf_internal = 1;
3651 scan = regnext(scan);
3652 if (flags & SCF_DO_SUBSTR) {
3653 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3654 data->longest = &(data->longest_float);
3656 goto optimize_curly_tail;
3658 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3659 && (scan->flags == stopparen))
3664 mincount = ARG1(scan);
3665 maxcount = ARG2(scan);
3667 next = regnext(scan);
3668 if (OP(scan) == CURLYX) {
3669 I32 lp = (data ? *(data->last_closep) : 0);
3670 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3672 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3673 next_is_eval = (OP(scan) == EVAL);
3675 if (flags & SCF_DO_SUBSTR) {
3676 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3677 pos_before = data->pos_min;
3681 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3683 data->flags |= SF_IS_INF;
3685 if (flags & SCF_DO_STCLASS) {
3686 cl_init(pRExC_state, &this_class);
3687 oclass = data->start_class;
3688 data->start_class = &this_class;
3689 f |= SCF_DO_STCLASS_AND;
3690 f &= ~SCF_DO_STCLASS_OR;
3692 /* Exclude from super-linear cache processing any {n,m}
3693 regops for which the combination of input pos and regex
3694 pos is not enough information to determine if a match
3697 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3698 regex pos at the \s*, the prospects for a match depend not
3699 only on the input position but also on how many (bar\s*)
3700 repeats into the {4,8} we are. */
3701 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3702 f &= ~SCF_WHILEM_VISITED_POS;
3704 /* This will finish on WHILEM, setting scan, or on NULL: */
3705 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3706 last, data, stopparen, recursed, NULL,
3708 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3710 if (flags & SCF_DO_STCLASS)
3711 data->start_class = oclass;
3712 if (mincount == 0 || minnext == 0) {
3713 if (flags & SCF_DO_STCLASS_OR) {
3714 cl_or(pRExC_state, data->start_class, &this_class);
3716 else if (flags & SCF_DO_STCLASS_AND) {
3717 /* Switch to OR mode: cache the old value of
3718 * data->start_class */
3720 StructCopy(data->start_class, and_withp,
3721 struct regnode_charclass_class);
3722 flags &= ~SCF_DO_STCLASS_AND;
3723 StructCopy(&this_class, data->start_class,
3724 struct regnode_charclass_class);
3725 flags |= SCF_DO_STCLASS_OR;
3726 data->start_class->flags |= ANYOF_EOS;
3728 } else { /* Non-zero len */
3729 if (flags & SCF_DO_STCLASS_OR) {
3730 cl_or(pRExC_state, data->start_class, &this_class);
3731 cl_and(data->start_class, and_withp);
3733 else if (flags & SCF_DO_STCLASS_AND)
3734 cl_and(data->start_class, &this_class);
3735 flags &= ~SCF_DO_STCLASS;
3737 if (!scan) /* It was not CURLYX, but CURLY. */
3739 if ( /* ? quantifier ok, except for (?{ ... }) */
3740 (next_is_eval || !(mincount == 0 && maxcount == 1))
3741 && (minnext == 0) && (deltanext == 0)
3742 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3743 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3745 ckWARNreg(RExC_parse,
3746 "Quantifier unexpected on zero-length expression");
3749 min += minnext * mincount;
3750 is_inf_internal |= ((maxcount == REG_INFTY
3751 && (minnext + deltanext) > 0)
3752 || deltanext == I32_MAX);
3753 is_inf |= is_inf_internal;
3754 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3756 /* Try powerful optimization CURLYX => CURLYN. */
3757 if ( OP(oscan) == CURLYX && data
3758 && data->flags & SF_IN_PAR
3759 && !(data->flags & SF_HAS_EVAL)
3760 && !deltanext && minnext == 1 ) {
3761 /* Try to optimize to CURLYN. */
3762 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3763 regnode * const nxt1 = nxt;
3770 if (!REGNODE_SIMPLE(OP(nxt))
3771 && !(PL_regkind[OP(nxt)] == EXACT
3772 && STR_LEN(nxt) == 1))
3778 if (OP(nxt) != CLOSE)
3780 if (RExC_open_parens) {
3781 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3782 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3784 /* Now we know that nxt2 is the only contents: */
3785 oscan->flags = (U8)ARG(nxt);
3787 OP(nxt1) = NOTHING; /* was OPEN. */
3790 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3791 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3792 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3793 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3794 OP(nxt + 1) = OPTIMIZED; /* was count. */
3795 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3800 /* Try optimization CURLYX => CURLYM. */
3801 if ( OP(oscan) == CURLYX && data
3802 && !(data->flags & SF_HAS_PAR)
3803 && !(data->flags & SF_HAS_EVAL)
3804 && !deltanext /* atom is fixed width */
3805 && minnext != 0 /* CURLYM can't handle zero width */
3807 /* XXXX How to optimize if data == 0? */
3808 /* Optimize to a simpler form. */
3809 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3813 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3814 && (OP(nxt2) != WHILEM))
3816 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3817 /* Need to optimize away parenths. */
3818 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
3819 /* Set the parenth number. */
3820 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3822 oscan->flags = (U8)ARG(nxt);
3823 if (RExC_open_parens) {
3824 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3825 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3827 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3828 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3831 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3832 OP(nxt + 1) = OPTIMIZED; /* was count. */
3833 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
3834 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
3837 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3838 regnode *nnxt = regnext(nxt1);
3840 if (reg_off_by_arg[OP(nxt1)])
3841 ARG_SET(nxt1, nxt2 - nxt1);
3842 else if (nxt2 - nxt1 < U16_MAX)
3843 NEXT_OFF(nxt1) = nxt2 - nxt1;
3845 OP(nxt) = NOTHING; /* Cannot beautify */
3850 /* Optimize again: */
3851 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
3852 NULL, stopparen, recursed, NULL, 0,depth+1);
3857 else if ((OP(oscan) == CURLYX)
https://perl5.git.perl.org / perl5.git / blob