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"
89 #ifndef PERL_IN_XSUB_RE
90 # include "charclass_invlists.h"
98 # if defined(BUGGY_MSC6)
99 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
100 # pragma optimize("a",off)
101 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
102 # pragma optimize("w",on )
103 # endif /* BUGGY_MSC6 */
107 #define STATIC static
110 typedef struct RExC_state_t {
111 U32 flags; /* are we folding, multilining? */
112 char *precomp; /* uncompiled string. */
113 REGEXP *rx_sv; /* The SV that is the regexp. */
114 regexp *rx; /* perl core regexp structure */
115 regexp_internal *rxi; /* internal data for regexp object pprivate field */
116 char *start; /* Start of input for compile */
117 char *end; /* End of input for compile */
118 char *parse; /* Input-scan pointer. */
119 I32 whilem_seen; /* number of WHILEM in this expr */
120 regnode *emit_start; /* Start of emitted-code area */
121 regnode *emit_bound; /* First regnode outside of the allocated space */
122 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
123 I32 naughty; /* How bad is this pattern? */
124 I32 sawback; /* Did we see \1, ...? */
126 I32 size; /* Code size. */
127 I32 npar; /* Capture buffer count, (OPEN). */
128 I32 cpar; /* Capture buffer count, (CLOSE). */
129 I32 nestroot; /* root parens we are in - used by accept */
133 regnode **open_parens; /* pointers to open parens */
134 regnode **close_parens; /* pointers to close parens */
135 regnode *opend; /* END node in program */
136 I32 utf8; /* whether the pattern is utf8 or not */
137 I32 orig_utf8; /* whether the pattern was originally in utf8 */
138 /* XXX use this for future optimisation of case
139 * where pattern must be upgraded to utf8. */
140 I32 uni_semantics; /* If a d charset modifier should use unicode
141 rules, even if the pattern is not in
143 HV *paren_names; /* Paren names */
145 regnode **recurse; /* Recurse regops */
146 I32 recurse_count; /* Number of recurse regops */
149 I32 override_recoding;
151 char *starttry; /* -Dr: where regtry was called. */
152 #define RExC_starttry (pRExC_state->starttry)
155 const char *lastparse;
157 AV *paren_name_list; /* idx -> name */
158 #define RExC_lastparse (pRExC_state->lastparse)
159 #define RExC_lastnum (pRExC_state->lastnum)
160 #define RExC_paren_name_list (pRExC_state->paren_name_list)
164 #define RExC_flags (pRExC_state->flags)
165 #define RExC_precomp (pRExC_state->precomp)
166 #define RExC_rx_sv (pRExC_state->rx_sv)
167 #define RExC_rx (pRExC_state->rx)
168 #define RExC_rxi (pRExC_state->rxi)
169 #define RExC_start (pRExC_state->start)
170 #define RExC_end (pRExC_state->end)
171 #define RExC_parse (pRExC_state->parse)
172 #define RExC_whilem_seen (pRExC_state->whilem_seen)
173 #ifdef RE_TRACK_PATTERN_OFFSETS
174 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
176 #define RExC_emit (pRExC_state->emit)
177 #define RExC_emit_start (pRExC_state->emit_start)
178 #define RExC_emit_bound (pRExC_state->emit_bound)
179 #define RExC_naughty (pRExC_state->naughty)
180 #define RExC_sawback (pRExC_state->sawback)
181 #define RExC_seen (pRExC_state->seen)
182 #define RExC_size (pRExC_state->size)
183 #define RExC_npar (pRExC_state->npar)
184 #define RExC_nestroot (pRExC_state->nestroot)
185 #define RExC_extralen (pRExC_state->extralen)
186 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
187 #define RExC_seen_evals (pRExC_state->seen_evals)
188 #define RExC_utf8 (pRExC_state->utf8)
189 #define RExC_uni_semantics (pRExC_state->uni_semantics)
190 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
191 #define RExC_open_parens (pRExC_state->open_parens)
192 #define RExC_close_parens (pRExC_state->close_parens)
193 #define RExC_opend (pRExC_state->opend)
194 #define RExC_paren_names (pRExC_state->paren_names)
195 #define RExC_recurse (pRExC_state->recurse)
196 #define RExC_recurse_count (pRExC_state->recurse_count)
197 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
198 #define RExC_contains_locale (pRExC_state->contains_locale)
199 #define RExC_override_recoding (pRExC_state->override_recoding)
202 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
203 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
204 ((*s) == '{' && regcurly(s)))
207 #undef SPSTART /* dratted cpp namespace... */
210 * Flags to be passed up and down.
212 #define WORST 0 /* Worst case. */
213 #define HASWIDTH 0x01 /* Known to match non-null strings. */
215 /* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
216 * character, and if utf8, must be invariant. Note that this is not the same thing as REGNODE_SIMPLE */
218 #define SPSTART 0x04 /* Starts with * or +. */
219 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
220 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
222 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
224 /* whether trie related optimizations are enabled */
225 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
226 #define TRIE_STUDY_OPT
227 #define FULL_TRIE_STUDY
233 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
234 #define PBITVAL(paren) (1 << ((paren) & 7))
235 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
236 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
237 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
239 /* If not already in utf8, do a longjmp back to the beginning */
240 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
241 #define REQUIRE_UTF8 STMT_START { \
242 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
245 /* About scan_data_t.
247 During optimisation we recurse through the regexp program performing
248 various inplace (keyhole style) optimisations. In addition study_chunk
249 and scan_commit populate this data structure with information about
250 what strings MUST appear in the pattern. We look for the longest
251 string that must appear at a fixed location, and we look for the
252 longest string that may appear at a floating location. So for instance
257 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
258 strings (because they follow a .* construct). study_chunk will identify
259 both FOO and BAR as being the longest fixed and floating strings respectively.
261 The strings can be composites, for instance
265 will result in a composite fixed substring 'foo'.
267 For each string some basic information is maintained:
269 - offset or min_offset
270 This is the position the string must appear at, or not before.
271 It also implicitly (when combined with minlenp) tells us how many
272 characters must match before the string we are searching for.
273 Likewise when combined with minlenp and the length of the string it
274 tells us how many characters must appear after the string we have
278 Only used for floating strings. This is the rightmost point that
279 the string can appear at. If set to I32 max it indicates that the
280 string can occur infinitely far to the right.
283 A pointer to the minimum length of the pattern that the string
284 was found inside. This is important as in the case of positive
285 lookahead or positive lookbehind we can have multiple patterns
290 The minimum length of the pattern overall is 3, the minimum length
291 of the lookahead part is 3, but the minimum length of the part that
292 will actually match is 1. So 'FOO's minimum length is 3, but the
293 minimum length for the F is 1. This is important as the minimum length
294 is used to determine offsets in front of and behind the string being
295 looked for. Since strings can be composites this is the length of the
296 pattern at the time it was committed with a scan_commit. Note that
297 the length is calculated by study_chunk, so that the minimum lengths
298 are not known until the full pattern has been compiled, thus the
299 pointer to the value.
303 In the case of lookbehind the string being searched for can be
304 offset past the start point of the final matching string.
305 If this value was just blithely removed from the min_offset it would
306 invalidate some of the calculations for how many chars must match
307 before or after (as they are derived from min_offset and minlen and
308 the length of the string being searched for).
309 When the final pattern is compiled and the data is moved from the
310 scan_data_t structure into the regexp structure the information
311 about lookbehind is factored in, with the information that would
312 have been lost precalculated in the end_shift field for the
315 The fields pos_min and pos_delta are used to store the minimum offset
316 and the delta to the maximum offset at the current point in the pattern.
320 typedef struct scan_data_t {
321 /*I32 len_min; unused */
322 /*I32 len_delta; unused */
326 I32 last_end; /* min value, <0 unless valid. */
329 SV **longest; /* Either &l_fixed, or &l_float. */
330 SV *longest_fixed; /* longest fixed string found in pattern */
331 I32 offset_fixed; /* offset where it starts */
332 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
333 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
334 SV *longest_float; /* longest floating string found in pattern */
335 I32 offset_float_min; /* earliest point in string it can appear */
336 I32 offset_float_max; /* latest point in string it can appear */
337 I32 *minlen_float; /* pointer to the minlen relevant to the string */
338 I32 lookbehind_float; /* is the position of the string modified by LB */
342 struct regnode_charclass_class *start_class;
346 * Forward declarations for pregcomp()'s friends.
349 static const scan_data_t zero_scan_data =
350 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
352 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
353 #define SF_BEFORE_SEOL 0x0001
354 #define SF_BEFORE_MEOL 0x0002
355 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
356 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
359 # define SF_FIX_SHIFT_EOL (0+2)
360 # define SF_FL_SHIFT_EOL (0+4)
362 # define SF_FIX_SHIFT_EOL (+2)
363 # define SF_FL_SHIFT_EOL (+4)
366 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
367 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
369 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
370 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
371 #define SF_IS_INF 0x0040
372 #define SF_HAS_PAR 0x0080
373 #define SF_IN_PAR 0x0100
374 #define SF_HAS_EVAL 0x0200
375 #define SCF_DO_SUBSTR 0x0400
376 #define SCF_DO_STCLASS_AND 0x0800
377 #define SCF_DO_STCLASS_OR 0x1000
378 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
379 #define SCF_WHILEM_VISITED_POS 0x2000
381 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
382 #define SCF_SEEN_ACCEPT 0x8000
384 #define UTF cBOOL(RExC_utf8)
385 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
386 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
387 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
388 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
389 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
390 #define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
391 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
393 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
395 #define OOB_UNICODE 12345678
396 #define OOB_NAMEDCLASS -1
398 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
399 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
402 /* length of regex to show in messages that don't mark a position within */
403 #define RegexLengthToShowInErrorMessages 127
406 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
407 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
408 * op/pragma/warn/regcomp.
410 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
411 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
413 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
416 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
417 * arg. Show regex, up to a maximum length. If it's too long, chop and add
420 #define _FAIL(code) STMT_START { \
421 const char *ellipses = ""; \
422 IV len = RExC_end - RExC_precomp; \
425 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
426 if (len > RegexLengthToShowInErrorMessages) { \
427 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
428 len = RegexLengthToShowInErrorMessages - 10; \
434 #define FAIL(msg) _FAIL( \
435 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
436 msg, (int)len, RExC_precomp, ellipses))
438 #define FAIL2(msg,arg) _FAIL( \
439 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
440 arg, (int)len, RExC_precomp, ellipses))
443 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
445 #define Simple_vFAIL(m) STMT_START { \
446 const IV offset = RExC_parse - RExC_precomp; \
447 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
448 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
452 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
454 #define vFAIL(m) STMT_START { \
456 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
461 * Like Simple_vFAIL(), but accepts two arguments.
463 #define Simple_vFAIL2(m,a1) STMT_START { \
464 const IV offset = RExC_parse - RExC_precomp; \
465 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
466 (int)offset, RExC_precomp, RExC_precomp + offset); \
470 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
472 #define vFAIL2(m,a1) STMT_START { \
474 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
475 Simple_vFAIL2(m, a1); \
480 * Like Simple_vFAIL(), but accepts three arguments.
482 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
483 const IV offset = RExC_parse - RExC_precomp; \
484 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
485 (int)offset, RExC_precomp, RExC_precomp + offset); \
489 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
491 #define vFAIL3(m,a1,a2) STMT_START { \
493 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
494 Simple_vFAIL3(m, a1, a2); \
498 * Like Simple_vFAIL(), but accepts four arguments.
500 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
501 const IV offset = RExC_parse - RExC_precomp; \
502 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
503 (int)offset, RExC_precomp, RExC_precomp + offset); \
506 #define ckWARNreg(loc,m) STMT_START { \
507 const IV offset = loc - RExC_precomp; \
508 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
509 (int)offset, RExC_precomp, RExC_precomp + offset); \
512 #define ckWARNregdep(loc,m) STMT_START { \
513 const IV offset = loc - RExC_precomp; \
514 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
516 (int)offset, RExC_precomp, RExC_precomp + offset); \
519 #define ckWARN2regdep(loc,m, a1) STMT_START { \
520 const IV offset = loc - RExC_precomp; \
521 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
523 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
526 #define ckWARN2reg(loc, m, a1) STMT_START { \
527 const IV offset = loc - RExC_precomp; \
528 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
529 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
532 #define vWARN3(loc, m, a1, a2) STMT_START { \
533 const IV offset = loc - RExC_precomp; \
534 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
535 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
538 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
539 const IV offset = loc - RExC_precomp; \
540 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
541 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
544 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
545 const IV offset = loc - RExC_precomp; \
546 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
547 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
550 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
551 const IV offset = loc - RExC_precomp; \
552 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
553 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
556 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
557 const IV offset = loc - RExC_precomp; \
558 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
559 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
563 /* Allow for side effects in s */
564 #define REGC(c,s) STMT_START { \
565 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
568 /* Macros for recording node offsets. 20001227 mjd@plover.com
569 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
570 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
571 * Element 0 holds the number n.
572 * Position is 1 indexed.
574 #ifndef RE_TRACK_PATTERN_OFFSETS
575 #define Set_Node_Offset_To_R(node,byte)
576 #define Set_Node_Offset(node,byte)
577 #define Set_Cur_Node_Offset
578 #define Set_Node_Length_To_R(node,len)
579 #define Set_Node_Length(node,len)
580 #define Set_Node_Cur_Length(node)
581 #define Node_Offset(n)
582 #define Node_Length(n)
583 #define Set_Node_Offset_Length(node,offset,len)
584 #define ProgLen(ri) ri->u.proglen
585 #define SetProgLen(ri,x) ri->u.proglen = x
587 #define ProgLen(ri) ri->u.offsets[0]
588 #define SetProgLen(ri,x) ri->u.offsets[0] = x
589 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
591 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
592 __LINE__, (int)(node), (int)(byte))); \
594 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
596 RExC_offsets[2*(node)-1] = (byte); \
601 #define Set_Node_Offset(node,byte) \
602 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
603 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
605 #define Set_Node_Length_To_R(node,len) STMT_START { \
607 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
608 __LINE__, (int)(node), (int)(len))); \
610 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
612 RExC_offsets[2*(node)] = (len); \
617 #define Set_Node_Length(node,len) \
618 Set_Node_Length_To_R((node)-RExC_emit_start, len)
619 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
620 #define Set_Node_Cur_Length(node) \
621 Set_Node_Length(node, RExC_parse - parse_start)
623 /* Get offsets and lengths */
624 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
625 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
627 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
628 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
629 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
633 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
634 #define EXPERIMENTAL_INPLACESCAN
635 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
637 #define DEBUG_STUDYDATA(str,data,depth) \
638 DEBUG_OPTIMISE_MORE_r(if(data){ \
639 PerlIO_printf(Perl_debug_log, \
640 "%*s" str "Pos:%"IVdf"/%"IVdf \
641 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
642 (int)(depth)*2, "", \
643 (IV)((data)->pos_min), \
644 (IV)((data)->pos_delta), \
645 (UV)((data)->flags), \
646 (IV)((data)->whilem_c), \
647 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
648 is_inf ? "INF " : "" \
650 if ((data)->last_found) \
651 PerlIO_printf(Perl_debug_log, \
652 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
653 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
654 SvPVX_const((data)->last_found), \
655 (IV)((data)->last_end), \
656 (IV)((data)->last_start_min), \
657 (IV)((data)->last_start_max), \
658 ((data)->longest && \
659 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
660 SvPVX_const((data)->longest_fixed), \
661 (IV)((data)->offset_fixed), \
662 ((data)->longest && \
663 (data)->longest==&((data)->longest_float)) ? "*" : "", \
664 SvPVX_const((data)->longest_float), \
665 (IV)((data)->offset_float_min), \
666 (IV)((data)->offset_float_max) \
668 PerlIO_printf(Perl_debug_log,"\n"); \
671 static void clear_re(pTHX_ void *r);
673 /* Mark that we cannot extend a found fixed substring at this point.
674 Update the longest found anchored substring and the longest found
675 floating substrings if needed. */
678 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
680 const STRLEN l = CHR_SVLEN(data->last_found);
681 const STRLEN old_l = CHR_SVLEN(*data->longest);
682 GET_RE_DEBUG_FLAGS_DECL;
684 PERL_ARGS_ASSERT_SCAN_COMMIT;
686 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
687 SvSetMagicSV(*data->longest, data->last_found);
688 if (*data->longest == data->longest_fixed) {
689 data->offset_fixed = l ? data->last_start_min : data->pos_min;
690 if (data->flags & SF_BEFORE_EOL)
692 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
694 data->flags &= ~SF_FIX_BEFORE_EOL;
695 data->minlen_fixed=minlenp;
696 data->lookbehind_fixed=0;
698 else { /* *data->longest == data->longest_float */
699 data->offset_float_min = l ? data->last_start_min : data->pos_min;
700 data->offset_float_max = (l
701 ? data->last_start_max
702 : data->pos_min + data->pos_delta);
703 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
704 data->offset_float_max = I32_MAX;
705 if (data->flags & SF_BEFORE_EOL)
707 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
709 data->flags &= ~SF_FL_BEFORE_EOL;
710 data->minlen_float=minlenp;
711 data->lookbehind_float=0;
714 SvCUR_set(data->last_found, 0);
716 SV * const sv = data->last_found;
717 if (SvUTF8(sv) && SvMAGICAL(sv)) {
718 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
724 data->flags &= ~SF_BEFORE_EOL;
725 DEBUG_STUDYDATA("commit: ",data,0);
728 /* Can match anything (initialization) */
730 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
732 PERL_ARGS_ASSERT_CL_ANYTHING;
734 ANYOF_BITMAP_SETALL(cl);
735 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
736 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
738 /* If any portion of the regex is to operate under locale rules,
739 * initialization includes it. The reason this isn't done for all regexes
740 * is that the optimizer was written under the assumption that locale was
741 * all-or-nothing. Given the complexity and lack of documentation in the
742 * optimizer, and that there are inadequate test cases for locale, so many
743 * parts of it may not work properly, it is safest to avoid locale unless
745 if (RExC_contains_locale) {
746 ANYOF_CLASS_SETALL(cl); /* /l uses class */
747 cl->flags |= ANYOF_LOCALE;
750 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
754 /* Can match anything (initialization) */
756 S_cl_is_anything(const struct regnode_charclass_class *cl)
760 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
762 for (value = 0; value <= ANYOF_MAX; value += 2)
763 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
765 if (!(cl->flags & ANYOF_UNICODE_ALL))
767 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
772 /* Can match anything (initialization) */
774 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
776 PERL_ARGS_ASSERT_CL_INIT;
778 Zero(cl, 1, struct regnode_charclass_class);
780 cl_anything(pRExC_state, cl);
781 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
784 /* These two functions currently do the exact same thing */
785 #define cl_init_zero S_cl_init
787 /* 'AND' a given class with another one. Can create false positives. 'cl'
788 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
789 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
791 S_cl_and(struct regnode_charclass_class *cl,
792 const struct regnode_charclass_class *and_with)
794 PERL_ARGS_ASSERT_CL_AND;
796 assert(and_with->type == ANYOF);
798 /* I (khw) am not sure all these restrictions are necessary XXX */
799 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
800 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
801 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
802 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
803 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
806 if (and_with->flags & ANYOF_INVERT)
807 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
808 cl->bitmap[i] &= ~and_with->bitmap[i];
810 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
811 cl->bitmap[i] &= and_with->bitmap[i];
812 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
814 if (and_with->flags & ANYOF_INVERT) {
816 /* Here, the and'ed node is inverted. Get the AND of the flags that
817 * aren't affected by the inversion. Those that are affected are
818 * handled individually below */
819 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
820 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
821 cl->flags |= affected_flags;
823 /* We currently don't know how to deal with things that aren't in the
824 * bitmap, but we know that the intersection is no greater than what
825 * is already in cl, so let there be false positives that get sorted
826 * out after the synthetic start class succeeds, and the node is
827 * matched for real. */
829 /* The inversion of these two flags indicate that the resulting
830 * intersection doesn't have them */
831 if (and_with->flags & ANYOF_UNICODE_ALL) {
832 cl->flags &= ~ANYOF_UNICODE_ALL;
834 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
835 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
838 else { /* and'd node is not inverted */
839 U8 outside_bitmap_but_not_utf8; /* Temp variable */
841 if (! ANYOF_NONBITMAP(and_with)) {
843 /* Here 'and_with' doesn't match anything outside the bitmap
844 * (except possibly ANYOF_UNICODE_ALL), which means the
845 * intersection can't either, except for ANYOF_UNICODE_ALL, in
846 * which case we don't know what the intersection is, but it's no
847 * greater than what cl already has, so can just leave it alone,
848 * with possible false positives */
849 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
850 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
851 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
854 else if (! ANYOF_NONBITMAP(cl)) {
856 /* Here, 'and_with' does match something outside the bitmap, and cl
857 * doesn't have a list of things to match outside the bitmap. If
858 * cl can match all code points above 255, the intersection will
859 * be those above-255 code points that 'and_with' matches. If cl
860 * can't match all Unicode code points, it means that it can't
861 * match anything outside the bitmap (since the 'if' that got us
862 * into this block tested for that), so we leave the bitmap empty.
864 if (cl->flags & ANYOF_UNICODE_ALL) {
865 ARG_SET(cl, ARG(and_with));
867 /* and_with's ARG may match things that don't require UTF8.
868 * And now cl's will too, in spite of this being an 'and'. See
869 * the comments below about the kludge */
870 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
874 /* Here, both 'and_with' and cl match something outside the
875 * bitmap. Currently we do not do the intersection, so just match
876 * whatever cl had at the beginning. */
880 /* Take the intersection of the two sets of flags. However, the
881 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
882 * kludge around the fact that this flag is not treated like the others
883 * which are initialized in cl_anything(). The way the optimizer works
884 * is that the synthetic start class (SSC) is initialized to match
885 * anything, and then the first time a real node is encountered, its
886 * values are AND'd with the SSC's with the result being the values of
887 * the real node. However, there are paths through the optimizer where
888 * the AND never gets called, so those initialized bits are set
889 * inappropriately, which is not usually a big deal, as they just cause
890 * false positives in the SSC, which will just mean a probably
891 * imperceptible slow down in execution. However this bit has a
892 * higher false positive consequence in that it can cause utf8.pm,
893 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
894 * bigger slowdown and also causes significant extra memory to be used.
895 * In order to prevent this, the code now takes a different tack. The
896 * bit isn't set unless some part of the regular expression needs it,
897 * but once set it won't get cleared. This means that these extra
898 * modules won't get loaded unless there was some path through the
899 * pattern that would have required them anyway, and so any false
900 * positives that occur by not ANDing them out when they could be
901 * aren't as severe as they would be if we treated this bit like all
903 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
904 & ANYOF_NONBITMAP_NON_UTF8;
905 cl->flags &= and_with->flags;
906 cl->flags |= outside_bitmap_but_not_utf8;
910 /* 'OR' a given class with another one. Can create false positives. 'cl'
911 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
912 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
914 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
916 PERL_ARGS_ASSERT_CL_OR;
918 if (or_with->flags & ANYOF_INVERT) {
920 /* Here, the or'd node is to be inverted. This means we take the
921 * complement of everything not in the bitmap, but currently we don't
922 * know what that is, so give up and match anything */
923 if (ANYOF_NONBITMAP(or_with)) {
924 cl_anything(pRExC_state, cl);
927 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
928 * <= (B1 | !B2) | (CL1 | !CL2)
929 * which is wasteful if CL2 is small, but we ignore CL2:
930 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
931 * XXXX Can we handle case-fold? Unclear:
932 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
933 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
935 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
936 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
937 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
940 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
941 cl->bitmap[i] |= ~or_with->bitmap[i];
942 } /* XXXX: logic is complicated otherwise */
944 cl_anything(pRExC_state, cl);
947 /* And, we can just take the union of the flags that aren't affected
948 * by the inversion */
949 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
951 /* For the remaining flags:
952 ANYOF_UNICODE_ALL and inverted means to not match anything above
953 255, which means that the union with cl should just be
954 what cl has in it, so can ignore this flag
955 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
956 is 127-255 to match them, but then invert that, so the
957 union with cl should just be what cl has in it, so can
960 } else { /* 'or_with' is not inverted */
961 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
962 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
963 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
964 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
967 /* OR char bitmap and class bitmap separately */
968 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
969 cl->bitmap[i] |= or_with->bitmap[i];
970 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
971 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
972 cl->classflags[i] |= or_with->classflags[i];
973 cl->flags |= ANYOF_CLASS;
976 else { /* XXXX: logic is complicated, leave it along for a moment. */
977 cl_anything(pRExC_state, cl);
980 if (ANYOF_NONBITMAP(or_with)) {
982 /* Use the added node's outside-the-bit-map match if there isn't a
983 * conflict. If there is a conflict (both nodes match something
984 * outside the bitmap, but what they match outside is not the same
985 * pointer, and hence not easily compared until XXX we extend
986 * inversion lists this far), give up and allow the start class to
987 * match everything outside the bitmap. If that stuff is all above
988 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
989 if (! ANYOF_NONBITMAP(cl)) {
990 ARG_SET(cl, ARG(or_with));
992 else if (ARG(cl) != ARG(or_with)) {
994 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
995 cl_anything(pRExC_state, cl);
998 cl->flags |= ANYOF_UNICODE_ALL;
1003 /* Take the union */
1004 cl->flags |= or_with->flags;
1008 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1009 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1010 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1011 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1016 dump_trie(trie,widecharmap,revcharmap)
1017 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1018 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1020 These routines dump out a trie in a somewhat readable format.
1021 The _interim_ variants are used for debugging the interim
1022 tables that are used to generate the final compressed
1023 representation which is what dump_trie expects.
1025 Part of the reason for their existence is to provide a form
1026 of documentation as to how the different representations function.
1031 Dumps the final compressed table form of the trie to Perl_debug_log.
1032 Used for debugging make_trie().
1036 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1037 AV *revcharmap, U32 depth)
1040 SV *sv=sv_newmortal();
1041 int colwidth= widecharmap ? 6 : 4;
1043 GET_RE_DEBUG_FLAGS_DECL;
1045 PERL_ARGS_ASSERT_DUMP_TRIE;
1047 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1048 (int)depth * 2 + 2,"",
1049 "Match","Base","Ofs" );
1051 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1052 SV ** const tmp = av_fetch( revcharmap, state, 0);
1054 PerlIO_printf( Perl_debug_log, "%*s",
1056 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1057 PL_colors[0], PL_colors[1],
1058 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1059 PERL_PV_ESCAPE_FIRSTCHAR
1064 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1065 (int)depth * 2 + 2,"");
1067 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1068 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1069 PerlIO_printf( Perl_debug_log, "\n");
1071 for( state = 1 ; state < trie->statecount ; state++ ) {
1072 const U32 base = trie->states[ state ].trans.base;
1074 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1076 if ( trie->states[ state ].wordnum ) {
1077 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1079 PerlIO_printf( Perl_debug_log, "%6s", "" );
1082 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1087 while( ( base + ofs < trie->uniquecharcount ) ||
1088 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1089 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1092 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1094 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1095 if ( ( base + ofs >= trie->uniquecharcount ) &&
1096 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1097 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1099 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1101 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1103 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1107 PerlIO_printf( Perl_debug_log, "]");
1110 PerlIO_printf( Perl_debug_log, "\n" );
1112 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1113 for (word=1; word <= trie->wordcount; word++) {
1114 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1115 (int)word, (int)(trie->wordinfo[word].prev),
1116 (int)(trie->wordinfo[word].len));
1118 PerlIO_printf(Perl_debug_log, "\n" );
1121 Dumps a fully constructed but uncompressed trie in list form.
1122 List tries normally only are used for construction when the number of
1123 possible chars (trie->uniquecharcount) is very high.
1124 Used for debugging make_trie().
1127 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1128 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1132 SV *sv=sv_newmortal();
1133 int colwidth= widecharmap ? 6 : 4;
1134 GET_RE_DEBUG_FLAGS_DECL;
1136 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1138 /* print out the table precompression. */
1139 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1140 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1141 "------:-----+-----------------\n" );
1143 for( state=1 ; state < next_alloc ; state ++ ) {
1146 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1147 (int)depth * 2 + 2,"", (UV)state );
1148 if ( ! trie->states[ state ].wordnum ) {
1149 PerlIO_printf( Perl_debug_log, "%5s| ","");
1151 PerlIO_printf( Perl_debug_log, "W%4x| ",
1152 trie->states[ state ].wordnum
1155 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1156 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1158 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1160 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1161 PL_colors[0], PL_colors[1],
1162 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1163 PERL_PV_ESCAPE_FIRSTCHAR
1165 TRIE_LIST_ITEM(state,charid).forid,
1166 (UV)TRIE_LIST_ITEM(state,charid).newstate
1169 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1170 (int)((depth * 2) + 14), "");
1173 PerlIO_printf( Perl_debug_log, "\n");
1178 Dumps a fully constructed but uncompressed trie in table form.
1179 This is the normal DFA style state transition table, with a few
1180 twists to facilitate compression later.
1181 Used for debugging make_trie().
1184 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1185 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1190 SV *sv=sv_newmortal();
1191 int colwidth= widecharmap ? 6 : 4;
1192 GET_RE_DEBUG_FLAGS_DECL;
1194 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1197 print out the table precompression so that we can do a visual check
1198 that they are identical.
1201 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1203 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1204 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1206 PerlIO_printf( Perl_debug_log, "%*s",
1208 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1209 PL_colors[0], PL_colors[1],
1210 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1211 PERL_PV_ESCAPE_FIRSTCHAR
1217 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1219 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1220 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1223 PerlIO_printf( Perl_debug_log, "\n" );
1225 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1227 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1228 (int)depth * 2 + 2,"",
1229 (UV)TRIE_NODENUM( state ) );
1231 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1232 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1234 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1236 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1238 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1239 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1241 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1242 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1250 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1251 startbranch: the first branch in the whole branch sequence
1252 first : start branch of sequence of branch-exact nodes.
1253 May be the same as startbranch
1254 last : Thing following the last branch.
1255 May be the same as tail.
1256 tail : item following the branch sequence
1257 count : words in the sequence
1258 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1259 depth : indent depth
1261 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1263 A trie is an N'ary tree where the branches are determined by digital
1264 decomposition of the key. IE, at the root node you look up the 1st character and
1265 follow that branch repeat until you find the end of the branches. Nodes can be
1266 marked as "accepting" meaning they represent a complete word. Eg:
1270 would convert into the following structure. Numbers represent states, letters
1271 following numbers represent valid transitions on the letter from that state, if
1272 the number is in square brackets it represents an accepting state, otherwise it
1273 will be in parenthesis.
1275 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1279 (1) +-i->(6)-+-s->[7]
1281 +-s->(3)-+-h->(4)-+-e->[5]
1283 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1285 This shows that when matching against the string 'hers' we will begin at state 1
1286 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1287 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1288 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1289 single traverse. We store a mapping from accepting to state to which word was
1290 matched, and then when we have multiple possibilities we try to complete the
1291 rest of the regex in the order in which they occured in the alternation.
1293 The only prior NFA like behaviour that would be changed by the TRIE support is
1294 the silent ignoring of duplicate alternations which are of the form:
1296 / (DUPE|DUPE) X? (?{ ... }) Y /x
1298 Thus EVAL blocks following a trie may be called a different number of times with
1299 and without the optimisation. With the optimisations dupes will be silently
1300 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1301 the following demonstrates:
1303 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1305 which prints out 'word' three times, but
1307 'words'=~/(word|word|word)(?{ print $1 })S/
1309 which doesnt print it out at all. This is due to other optimisations kicking in.
1311 Example of what happens on a structural level:
1313 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1315 1: CURLYM[1] {1,32767}(18)
1326 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1327 and should turn into:
1329 1: CURLYM[1] {1,32767}(18)
1331 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1339 Cases where tail != last would be like /(?foo|bar)baz/:
1349 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1350 and would end up looking like:
1353 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1360 d = uvuni_to_utf8_flags(d, uv, 0);
1362 is the recommended Unicode-aware way of saying
1367 #define TRIE_STORE_REVCHAR \
1370 SV *zlopp = newSV(2); \
1371 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1372 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, uvc & 0xFF); \
1373 SvCUR_set(zlopp, kapow - flrbbbbb); \
1376 av_push(revcharmap, zlopp); \
1378 char ooooff = (char)uvc; \
1379 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1383 #define TRIE_READ_CHAR STMT_START { \
1387 if ( foldlen > 0 ) { \
1388 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1393 len = UTF8SKIP(uc);\
1394 uvc = to_utf8_fold( uc, foldbuf, &foldlen); \
1395 foldlen -= UNISKIP( uvc ); \
1396 scan = foldbuf + UNISKIP( uvc ); \
1399 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1409 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1410 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1411 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1412 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1414 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1415 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1416 TRIE_LIST_CUR( state )++; \
1419 #define TRIE_LIST_NEW(state) STMT_START { \
1420 Newxz( trie->states[ state ].trans.list, \
1421 4, reg_trie_trans_le ); \
1422 TRIE_LIST_CUR( state ) = 1; \
1423 TRIE_LIST_LEN( state ) = 4; \
1426 #define TRIE_HANDLE_WORD(state) STMT_START { \
1427 U16 dupe= trie->states[ state ].wordnum; \
1428 regnode * const noper_next = regnext( noper ); \
1431 /* store the word for dumping */ \
1433 if (OP(noper) != NOTHING) \
1434 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1436 tmp = newSVpvn_utf8( "", 0, UTF ); \
1437 av_push( trie_words, tmp ); \
1441 trie->wordinfo[curword].prev = 0; \
1442 trie->wordinfo[curword].len = wordlen; \
1443 trie->wordinfo[curword].accept = state; \
1445 if ( noper_next < tail ) { \
1447 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1448 trie->jump[curword] = (U16)(noper_next - convert); \
1450 jumper = noper_next; \
1452 nextbranch= regnext(cur); \
1456 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1457 /* chain, so that when the bits of chain are later */\
1458 /* linked together, the dups appear in the chain */\
1459 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1460 trie->wordinfo[dupe].prev = curword; \
1462 /* we haven't inserted this word yet. */ \
1463 trie->states[ state ].wordnum = curword; \
1468 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1469 ( ( base + charid >= ucharcount \
1470 && base + charid < ubound \
1471 && state == trie->trans[ base - ucharcount + charid ].check \
1472 && trie->trans[ base - ucharcount + charid ].next ) \
1473 ? trie->trans[ base - ucharcount + charid ].next \
1474 : ( state==1 ? special : 0 ) \
1478 #define MADE_JUMP_TRIE 2
1479 #define MADE_EXACT_TRIE 4
1482 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1485 /* first pass, loop through and scan words */
1486 reg_trie_data *trie;
1487 HV *widecharmap = NULL;
1488 AV *revcharmap = newAV();
1490 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1495 regnode *jumper = NULL;
1496 regnode *nextbranch = NULL;
1497 regnode *convert = NULL;
1498 U32 *prev_states; /* temp array mapping each state to previous one */
1499 /* we just use folder as a flag in utf8 */
1500 const U8 * folder = NULL;
1503 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1504 AV *trie_words = NULL;
1505 /* along with revcharmap, this only used during construction but both are
1506 * useful during debugging so we store them in the struct when debugging.
1509 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1510 STRLEN trie_charcount=0;
1512 SV *re_trie_maxbuff;
1513 GET_RE_DEBUG_FLAGS_DECL;
1515 PERL_ARGS_ASSERT_MAKE_TRIE;
1517 PERL_UNUSED_ARG(depth);
1523 case EXACTFU: folder = PL_fold_latin1; break;
1524 case EXACTF: folder = PL_fold; break;
1525 case EXACTFL: folder = PL_fold_locale; break;
1526 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u", (unsigned) flags );
1529 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1531 trie->startstate = 1;
1532 trie->wordcount = word_count;
1533 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1534 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1535 if (!(UTF && folder))
1536 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1537 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1538 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1541 trie_words = newAV();
1544 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1545 if (!SvIOK(re_trie_maxbuff)) {
1546 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1549 PerlIO_printf( Perl_debug_log,
1550 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1551 (int)depth * 2 + 2, "",
1552 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1553 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1557 /* Find the node we are going to overwrite */
1558 if ( first == startbranch && OP( last ) != BRANCH ) {
1559 /* whole branch chain */
1562 /* branch sub-chain */
1563 convert = NEXTOPER( first );
1566 /* -- First loop and Setup --
1568 We first traverse the branches and scan each word to determine if it
1569 contains widechars, and how many unique chars there are, this is
1570 important as we have to build a table with at least as many columns as we
1573 We use an array of integers to represent the character codes 0..255
1574 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1575 native representation of the character value as the key and IV's for the
1578 *TODO* If we keep track of how many times each character is used we can
1579 remap the columns so that the table compression later on is more
1580 efficient in terms of memory by ensuring the most common value is in the
1581 middle and the least common are on the outside. IMO this would be better
1582 than a most to least common mapping as theres a decent chance the most
1583 common letter will share a node with the least common, meaning the node
1584 will not be compressible. With a middle is most common approach the worst
1585 case is when we have the least common nodes twice.
1589 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1590 regnode * const noper = NEXTOPER( cur );
1591 const U8 *uc = (U8*)STRING( noper );
1592 const U8 * const e = uc + STR_LEN( noper );
1594 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1595 const U8 *scan = (U8*)NULL;
1596 U32 wordlen = 0; /* required init */
1598 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1600 if (OP(noper) == NOTHING) {
1604 if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
1605 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1606 regardless of encoding */
1608 for ( ; uc < e ; uc += len ) {
1609 TRIE_CHARCOUNT(trie)++;
1613 if ( !trie->charmap[ uvc ] ) {
1614 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1616 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
1620 /* store the codepoint in the bitmap, and its folded
1622 TRIE_BITMAP_SET(trie,uvc);
1624 /* store the folded codepoint */
1625 if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);
1628 /* store first byte of utf8 representation of
1629 variant codepoints */
1630 if (! UNI_IS_INVARIANT(uvc)) {
1631 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1634 set_bit = 0; /* We've done our bit :-) */
1639 widecharmap = newHV();
1641 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1644 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1646 if ( !SvTRUE( *svpp ) ) {
1647 sv_setiv( *svpp, ++trie->uniquecharcount );
1652 if( cur == first ) {
1655 } else if (chars < trie->minlen) {
1657 } else if (chars > trie->maxlen) {
1661 } /* end first pass */
1662 DEBUG_TRIE_COMPILE_r(
1663 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1664 (int)depth * 2 + 2,"",
1665 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1666 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1667 (int)trie->minlen, (int)trie->maxlen )
1671 We now know what we are dealing with in terms of unique chars and
1672 string sizes so we can calculate how much memory a naive
1673 representation using a flat table will take. If it's over a reasonable
1674 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1675 conservative but potentially much slower representation using an array
1678 At the end we convert both representations into the same compressed
1679 form that will be used in regexec.c for matching with. The latter
1680 is a form that cannot be used to construct with but has memory
1681 properties similar to the list form and access properties similar
1682 to the table form making it both suitable for fast searches and
1683 small enough that its feasable to store for the duration of a program.
1685 See the comment in the code where the compressed table is produced
1686 inplace from the flat tabe representation for an explanation of how
1687 the compression works.
1692 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1695 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1697 Second Pass -- Array Of Lists Representation
1699 Each state will be represented by a list of charid:state records
1700 (reg_trie_trans_le) the first such element holds the CUR and LEN
1701 points of the allocated array. (See defines above).
1703 We build the initial structure using the lists, and then convert
1704 it into the compressed table form which allows faster lookups
1705 (but cant be modified once converted).
1708 STRLEN transcount = 1;
1710 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1711 "%*sCompiling trie using list compiler\n",
1712 (int)depth * 2 + 2, ""));
1714 trie->states = (reg_trie_state *)
1715 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1716 sizeof(reg_trie_state) );
1720 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1722 regnode * const noper = NEXTOPER( cur );
1723 U8 *uc = (U8*)STRING( noper );
1724 const U8 * const e = uc + STR_LEN( noper );
1725 U32 state = 1; /* required init */
1726 U16 charid = 0; /* sanity init */
1727 U8 *scan = (U8*)NULL; /* sanity init */
1728 STRLEN foldlen = 0; /* required init */
1729 U32 wordlen = 0; /* required init */
1730 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1732 if (OP(noper) != NOTHING) {
1733 for ( ; uc < e ; uc += len ) {
1738 charid = trie->charmap[ uvc ];
1740 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1744 charid=(U16)SvIV( *svpp );
1747 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1754 if ( !trie->states[ state ].trans.list ) {
1755 TRIE_LIST_NEW( state );
1757 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1758 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1759 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1764 newstate = next_alloc++;
1765 prev_states[newstate] = state;
1766 TRIE_LIST_PUSH( state, charid, newstate );
1771 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1775 TRIE_HANDLE_WORD(state);
1777 } /* end second pass */
1779 /* next alloc is the NEXT state to be allocated */
1780 trie->statecount = next_alloc;
1781 trie->states = (reg_trie_state *)
1782 PerlMemShared_realloc( trie->states,
1784 * sizeof(reg_trie_state) );
1786 /* and now dump it out before we compress it */
1787 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1788 revcharmap, next_alloc,
1792 trie->trans = (reg_trie_trans *)
1793 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1800 for( state=1 ; state < next_alloc ; state ++ ) {
1804 DEBUG_TRIE_COMPILE_MORE_r(
1805 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1809 if (trie->states[state].trans.list) {
1810 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1814 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1815 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1816 if ( forid < minid ) {
1818 } else if ( forid > maxid ) {
1822 if ( transcount < tp + maxid - minid + 1) {
1824 trie->trans = (reg_trie_trans *)
1825 PerlMemShared_realloc( trie->trans,
1827 * sizeof(reg_trie_trans) );
1828 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1830 base = trie->uniquecharcount + tp - minid;
1831 if ( maxid == minid ) {
1833 for ( ; zp < tp ; zp++ ) {
1834 if ( ! trie->trans[ zp ].next ) {
1835 base = trie->uniquecharcount + zp - minid;
1836 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1837 trie->trans[ zp ].check = state;
1843 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1844 trie->trans[ tp ].check = state;
1849 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1850 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1851 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1852 trie->trans[ tid ].check = state;
1854 tp += ( maxid - minid + 1 );
1856 Safefree(trie->states[ state ].trans.list);
1859 DEBUG_TRIE_COMPILE_MORE_r(
1860 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1863 trie->states[ state ].trans.base=base;
1865 trie->lasttrans = tp + 1;
1869 Second Pass -- Flat Table Representation.
1871 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1872 We know that we will need Charcount+1 trans at most to store the data
1873 (one row per char at worst case) So we preallocate both structures
1874 assuming worst case.
1876 We then construct the trie using only the .next slots of the entry
1879 We use the .check field of the first entry of the node temporarily to
1880 make compression both faster and easier by keeping track of how many non
1881 zero fields are in the node.
1883 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1886 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1887 number representing the first entry of the node, and state as a
1888 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1889 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1890 are 2 entrys per node. eg:
1898 The table is internally in the right hand, idx form. However as we also
1899 have to deal with the states array which is indexed by nodenum we have to
1900 use TRIE_NODENUM() to convert.
1903 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1904 "%*sCompiling trie using table compiler\n",
1905 (int)depth * 2 + 2, ""));
1907 trie->trans = (reg_trie_trans *)
1908 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1909 * trie->uniquecharcount + 1,
1910 sizeof(reg_trie_trans) );
1911 trie->states = (reg_trie_state *)
1912 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1913 sizeof(reg_trie_state) );
1914 next_alloc = trie->uniquecharcount + 1;
1917 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1919 regnode * const noper = NEXTOPER( cur );
1920 const U8 *uc = (U8*)STRING( noper );
1921 const U8 * const e = uc + STR_LEN( noper );
1923 U32 state = 1; /* required init */
1925 U16 charid = 0; /* sanity init */
1926 U32 accept_state = 0; /* sanity init */
1927 U8 *scan = (U8*)NULL; /* sanity init */
1929 STRLEN foldlen = 0; /* required init */
1930 U32 wordlen = 0; /* required init */
1931 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1933 if ( OP(noper) != NOTHING ) {
1934 for ( ; uc < e ; uc += len ) {
1939 charid = trie->charmap[ uvc ];
1941 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1942 charid = svpp ? (U16)SvIV(*svpp) : 0;
1946 if ( !trie->trans[ state + charid ].next ) {
1947 trie->trans[ state + charid ].next = next_alloc;
1948 trie->trans[ state ].check++;
1949 prev_states[TRIE_NODENUM(next_alloc)]
1950 = TRIE_NODENUM(state);
1951 next_alloc += trie->uniquecharcount;
1953 state = trie->trans[ state + charid ].next;
1955 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1957 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1960 accept_state = TRIE_NODENUM( state );
1961 TRIE_HANDLE_WORD(accept_state);
1963 } /* end second pass */
1965 /* and now dump it out before we compress it */
1966 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1968 next_alloc, depth+1));
1972 * Inplace compress the table.*
1974 For sparse data sets the table constructed by the trie algorithm will
1975 be mostly 0/FAIL transitions or to put it another way mostly empty.
1976 (Note that leaf nodes will not contain any transitions.)
1978 This algorithm compresses the tables by eliminating most such
1979 transitions, at the cost of a modest bit of extra work during lookup:
1981 - Each states[] entry contains a .base field which indicates the
1982 index in the state[] array wheres its transition data is stored.
1984 - If .base is 0 there are no valid transitions from that node.
1986 - If .base is nonzero then charid is added to it to find an entry in
1989 -If trans[states[state].base+charid].check!=state then the
1990 transition is taken to be a 0/Fail transition. Thus if there are fail
1991 transitions at the front of the node then the .base offset will point
1992 somewhere inside the previous nodes data (or maybe even into a node
1993 even earlier), but the .check field determines if the transition is
1997 The following process inplace converts the table to the compressed
1998 table: We first do not compress the root node 1,and mark all its
1999 .check pointers as 1 and set its .base pointer as 1 as well. This
2000 allows us to do a DFA construction from the compressed table later,
2001 and ensures that any .base pointers we calculate later are greater
2004 - We set 'pos' to indicate the first entry of the second node.
2006 - We then iterate over the columns of the node, finding the first and
2007 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2008 and set the .check pointers accordingly, and advance pos
2009 appropriately and repreat for the next node. Note that when we copy
2010 the next pointers we have to convert them from the original
2011 NODEIDX form to NODENUM form as the former is not valid post
2014 - If a node has no transitions used we mark its base as 0 and do not
2015 advance the pos pointer.
2017 - If a node only has one transition we use a second pointer into the
2018 structure to fill in allocated fail transitions from other states.
2019 This pointer is independent of the main pointer and scans forward
2020 looking for null transitions that are allocated to a state. When it
2021 finds one it writes the single transition into the "hole". If the
2022 pointer doesnt find one the single transition is appended as normal.
2024 - Once compressed we can Renew/realloc the structures to release the
2027 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2028 specifically Fig 3.47 and the associated pseudocode.
2032 const U32 laststate = TRIE_NODENUM( next_alloc );
2035 trie->statecount = laststate;
2037 for ( state = 1 ; state < laststate ; state++ ) {
2039 const U32 stateidx = TRIE_NODEIDX( state );
2040 const U32 o_used = trie->trans[ stateidx ].check;
2041 U32 used = trie->trans[ stateidx ].check;
2042 trie->trans[ stateidx ].check = 0;
2044 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2045 if ( flag || trie->trans[ stateidx + charid ].next ) {
2046 if ( trie->trans[ stateidx + charid ].next ) {
2048 for ( ; zp < pos ; zp++ ) {
2049 if ( ! trie->trans[ zp ].next ) {
2053 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2054 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2055 trie->trans[ zp ].check = state;
2056 if ( ++zp > pos ) pos = zp;
2063 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2065 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2066 trie->trans[ pos ].check = state;
2071 trie->lasttrans = pos + 1;
2072 trie->states = (reg_trie_state *)
2073 PerlMemShared_realloc( trie->states, laststate
2074 * sizeof(reg_trie_state) );
2075 DEBUG_TRIE_COMPILE_MORE_r(
2076 PerlIO_printf( Perl_debug_log,
2077 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2078 (int)depth * 2 + 2,"",
2079 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2082 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2085 } /* end table compress */
2087 DEBUG_TRIE_COMPILE_MORE_r(
2088 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2089 (int)depth * 2 + 2, "",
2090 (UV)trie->statecount,
2091 (UV)trie->lasttrans)
2093 /* resize the trans array to remove unused space */
2094 trie->trans = (reg_trie_trans *)
2095 PerlMemShared_realloc( trie->trans, trie->lasttrans
2096 * sizeof(reg_trie_trans) );
2098 { /* Modify the program and insert the new TRIE node */
2099 U8 nodetype =(U8)(flags & 0xFF);
2103 regnode *optimize = NULL;
2104 #ifdef RE_TRACK_PATTERN_OFFSETS
2107 U32 mjd_nodelen = 0;
2108 #endif /* RE_TRACK_PATTERN_OFFSETS */
2109 #endif /* DEBUGGING */
2111 This means we convert either the first branch or the first Exact,
2112 depending on whether the thing following (in 'last') is a branch
2113 or not and whther first is the startbranch (ie is it a sub part of
2114 the alternation or is it the whole thing.)
2115 Assuming its a sub part we convert the EXACT otherwise we convert
2116 the whole branch sequence, including the first.
2118 /* Find the node we are going to overwrite */
2119 if ( first != startbranch || OP( last ) == BRANCH ) {
2120 /* branch sub-chain */
2121 NEXT_OFF( first ) = (U16)(last - first);
2122 #ifdef RE_TRACK_PATTERN_OFFSETS
2124 mjd_offset= Node_Offset((convert));
2125 mjd_nodelen= Node_Length((convert));
2128 /* whole branch chain */
2130 #ifdef RE_TRACK_PATTERN_OFFSETS
2133 const regnode *nop = NEXTOPER( convert );
2134 mjd_offset= Node_Offset((nop));
2135 mjd_nodelen= Node_Length((nop));
2139 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2140 (int)depth * 2 + 2, "",
2141 (UV)mjd_offset, (UV)mjd_nodelen)
2144 /* But first we check to see if there is a common prefix we can
2145 split out as an EXACT and put in front of the TRIE node. */
2146 trie->startstate= 1;
2147 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2149 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2153 const U32 base = trie->states[ state ].trans.base;
2155 if ( trie->states[state].wordnum )
2158 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2159 if ( ( base + ofs >= trie->uniquecharcount ) &&
2160 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2161 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2163 if ( ++count > 1 ) {
2164 SV **tmp = av_fetch( revcharmap, ofs, 0);
2165 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2166 if ( state == 1 ) break;
2168 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2170 PerlIO_printf(Perl_debug_log,
2171 "%*sNew Start State=%"UVuf" Class: [",
2172 (int)depth * 2 + 2, "",
2175 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2176 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2178 TRIE_BITMAP_SET(trie,*ch);
2180 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2182 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2186 TRIE_BITMAP_SET(trie,*ch);
2188 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2189 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2195 SV **tmp = av_fetch( revcharmap, idx, 0);
2197 char *ch = SvPV( *tmp, len );
2199 SV *sv=sv_newmortal();
2200 PerlIO_printf( Perl_debug_log,
2201 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2202 (int)depth * 2 + 2, "",
2204 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2205 PL_colors[0], PL_colors[1],
2206 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2207 PERL_PV_ESCAPE_FIRSTCHAR
2212 OP( convert ) = nodetype;
2213 str=STRING(convert);
2216 STR_LEN(convert) += len;
2222 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2227 trie->prefixlen = (state-1);
2229 regnode *n = convert+NODE_SZ_STR(convert);
2230 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2231 trie->startstate = state;
2232 trie->minlen -= (state - 1);
2233 trie->maxlen -= (state - 1);
2235 /* At least the UNICOS C compiler choked on this
2236 * being argument to DEBUG_r(), so let's just have
2239 #ifdef PERL_EXT_RE_BUILD
2245 regnode *fix = convert;
2246 U32 word = trie->wordcount;
2248 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2249 while( ++fix < n ) {
2250 Set_Node_Offset_Length(fix, 0, 0);
2253 SV ** const tmp = av_fetch( trie_words, word, 0 );
2255 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2256 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2258 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2266 NEXT_OFF(convert) = (U16)(tail - convert);
2267 DEBUG_r(optimize= n);
2273 if ( trie->maxlen ) {
2274 NEXT_OFF( convert ) = (U16)(tail - convert);
2275 ARG_SET( convert, data_slot );
2276 /* Store the offset to the first unabsorbed branch in
2277 jump[0], which is otherwise unused by the jump logic.
2278 We use this when dumping a trie and during optimisation. */
2280 trie->jump[0] = (U16)(nextbranch - convert);
2282 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2283 * and there is a bitmap
2284 * and the first "jump target" node we found leaves enough room
2285 * then convert the TRIE node into a TRIEC node, with the bitmap
2286 * embedded inline in the opcode - this is hypothetically faster.
2288 if ( !trie->states[trie->startstate].wordnum
2290 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2292 OP( convert ) = TRIEC;
2293 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2294 PerlMemShared_free(trie->bitmap);
2297 OP( convert ) = TRIE;
2299 /* store the type in the flags */
2300 convert->flags = nodetype;
2304 + regarglen[ OP( convert ) ];
2306 /* XXX We really should free up the resource in trie now,
2307 as we won't use them - (which resources?) dmq */
2309 /* needed for dumping*/
2310 DEBUG_r(if (optimize) {
2311 regnode *opt = convert;
2313 while ( ++opt < optimize) {
2314 Set_Node_Offset_Length(opt,0,0);
2317 Try to clean up some of the debris left after the
2320 while( optimize < jumper ) {
2321 mjd_nodelen += Node_Length((optimize));
2322 OP( optimize ) = OPTIMIZED;
2323 Set_Node_Offset_Length(optimize,0,0);
2326 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2328 } /* end node insert */
2330 /* Finish populating the prev field of the wordinfo array. Walk back
2331 * from each accept state until we find another accept state, and if
2332 * so, point the first word's .prev field at the second word. If the
2333 * second already has a .prev field set, stop now. This will be the
2334 * case either if we've already processed that word's accept state,
2335 * or that state had multiple words, and the overspill words were
2336 * already linked up earlier.
2343 for (word=1; word <= trie->wordcount; word++) {
2345 if (trie->wordinfo[word].prev)
2347 state = trie->wordinfo[word].accept;
2349 state = prev_states[state];
2352 prev = trie->states[state].wordnum;
2356 trie->wordinfo[word].prev = prev;
2358 Safefree(prev_states);
2362 /* and now dump out the compressed format */
2363 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2365 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2367 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2368 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2370 SvREFCNT_dec(revcharmap);
2374 : trie->startstate>1
2380 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2382 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2384 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2385 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2388 We find the fail state for each state in the trie, this state is the longest proper
2389 suffix of the current state's 'word' that is also a proper prefix of another word in our
2390 trie. State 1 represents the word '' and is thus the default fail state. This allows
2391 the DFA not to have to restart after its tried and failed a word at a given point, it
2392 simply continues as though it had been matching the other word in the first place.
2394 'abcdgu'=~/abcdefg|cdgu/
2395 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2396 fail, which would bring us to the state representing 'd' in the second word where we would
2397 try 'g' and succeed, proceeding to match 'cdgu'.
2399 /* add a fail transition */
2400 const U32 trie_offset = ARG(source);
2401 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2403 const U32 ucharcount = trie->uniquecharcount;
2404 const U32 numstates = trie->statecount;
2405 const U32 ubound = trie->lasttrans + ucharcount;
2409 U32 base = trie->states[ 1 ].trans.base;
2412 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2413 GET_RE_DEBUG_FLAGS_DECL;
2415 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2417 PERL_UNUSED_ARG(depth);
2421 ARG_SET( stclass, data_slot );
2422 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2423 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2424 aho->trie=trie_offset;
2425 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2426 Copy( trie->states, aho->states, numstates, reg_trie_state );
2427 Newxz( q, numstates, U32);
2428 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2431 /* initialize fail[0..1] to be 1 so that we always have
2432 a valid final fail state */
2433 fail[ 0 ] = fail[ 1 ] = 1;
2435 for ( charid = 0; charid < ucharcount ; charid++ ) {
2436 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2438 q[ q_write ] = newstate;
2439 /* set to point at the root */
2440 fail[ q[ q_write++ ] ]=1;
2443 while ( q_read < q_write) {
2444 const U32 cur = q[ q_read++ % numstates ];
2445 base = trie->states[ cur ].trans.base;
2447 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2448 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2450 U32 fail_state = cur;
2453 fail_state = fail[ fail_state ];
2454 fail_base = aho->states[ fail_state ].trans.base;
2455 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2457 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2458 fail[ ch_state ] = fail_state;
2459 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2461 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2463 q[ q_write++ % numstates] = ch_state;
2467 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2468 when we fail in state 1, this allows us to use the
2469 charclass scan to find a valid start char. This is based on the principle
2470 that theres a good chance the string being searched contains lots of stuff
2471 that cant be a start char.
2473 fail[ 0 ] = fail[ 1 ] = 0;
2474 DEBUG_TRIE_COMPILE_r({
2475 PerlIO_printf(Perl_debug_log,
2476 "%*sStclass Failtable (%"UVuf" states): 0",
2477 (int)(depth * 2), "", (UV)numstates
2479 for( q_read=1; q_read<numstates; q_read++ ) {
2480 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2482 PerlIO_printf(Perl_debug_log, "\n");
2485 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2490 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2491 * These need to be revisited when a newer toolchain becomes available.
2493 #if defined(__sparc64__) && defined(__GNUC__)
2494 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2495 # undef SPARC64_GCC_WORKAROUND
2496 # define SPARC64_GCC_WORKAROUND 1
2500 #define DEBUG_PEEP(str,scan,depth) \
2501 DEBUG_OPTIMISE_r({if (scan){ \
2502 SV * const mysv=sv_newmortal(); \
2503 regnode *Next = regnext(scan); \
2504 regprop(RExC_rx, mysv, scan); \
2505 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2506 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2507 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2511 /* The below joins as many adjacent EXACTish nodes as possible into a single
2512 * one, and looks for problematic sequences of characters whose folds vs.
2513 * non-folds have sufficiently different lengths, that the optimizer would be
2514 * fooled into rejecting legitimate matches of them, and the trie construction
2515 * code can't cope with them. The joining is only done if:
2516 * 1) there is room in the current conglomerated node to entirely contain the
2518 * 2) they are the exact same node type
2520 * The adjacent nodes actually may be separated by NOTHING kind nodes, and
2521 * these get optimized out
2523 * If there are problematic code sequences, *min_subtract is set to the delta
2524 * that the minimum size of the node can be less than its actual size. And,
2525 * the node type of the result is changed to reflect that it contains these
2528 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2529 * and contains LATIN SMALL LETTER SHARP S
2531 * This is as good a place as any to discuss the design of handling these
2532 * problematic sequences. It's been wrong in Perl for a very long time. There
2533 * are three code points in Unicode whose folded lengths differ so much from
2534 * the un-folded lengths that it causes problems for the optimizer and trie
2535 * construction. Why only these are problematic, and not others where lengths
2536 * also differ is something I (khw) do not understand. New versions of Unicode
2537 * might add more such code points. Hopefully the logic in fold_grind.t that
2538 * figures out what to test (in part by verifying that each size-combination
2539 * gets tested) will catch any that do come along, so they can be added to the
2540 * special handling below. The chances of new ones are actually rather small,
2541 * as most, if not all, of the world's scripts that have casefolding have
2542 * already been encoded by Unicode. Also, a number of Unicode's decisions were
2543 * made to allow compatibility with pre-existing standards, and almost all of
2544 * those have already been dealt with. These would otherwise be the most
2545 * likely candidates for generating further tricky sequences. In other words,
2546 * Unicode by itself is unlikely to add new ones unless it is for compatibility
2547 * with pre-existing standards, and there aren't many of those left.
2549 * The previous designs for dealing with these involved assigning a special
2550 * node for them. This approach doesn't work, as evidenced by this example:
2551 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2552 * Both these fold to "sss", but if the pattern is parsed to create a node of
2553 * that would match just the \xDF, it won't be able to handle the case where a
2554 * successful match would have to cross the node's boundary. The new approach
2555 * that hopefully generally solves the problem generates an EXACTFU_SS node
2558 * There are a number of components to the approach (a lot of work for just
2559 * three code points!):
2560 * 1) This routine examines each EXACTFish node that could contain the
2561 * problematic sequences. It returns in *min_subtract how much to
2562 * subtract from the the actual length of the string to get a real minimum
2563 * for one that could match it. This number is usually 0 except for the
2564 * problematic sequences. This delta is used by the caller to adjust the
2565 * min length of the match, and the delta between min and max, so that the
2566 * optimizer doesn't reject these possibilities based on size constraints.
2567 * 2) These sequences are not currently correctly handled by the trie code
2568 * either, so it changes the joined node type to ops that are not handled
2569 * by trie's, those new ops being EXACTFU_SS and EXACTFU_NO_TRIE.
2570 * 3) This is sufficient for the two Greek sequences (described below), but
2571 * the one involving the Sharp s (\xDF) needs more. The node type
2572 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2573 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2574 * case where there is a possible fold length change. That means that a
2575 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2576 * itself with length changes, and so can be processed faster. regexec.c
2577 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2578 * is pre-folded by regcomp.c. This saves effort in regex matching.
2579 * However, probably mostly for historical reasons, the pre-folding isn't
2580 * done for non-UTF8 patterns (and it can't be for EXACTF and EXACTFL
2581 * nodes, as what they fold to isn't known until runtime.) The fold
2582 * possibilities for the non-UTF8 patterns are quite simple, except for
2583 * the sharp s. All the ones that don't involve a UTF-8 target string
2584 * are members of a fold-pair, and arrays are set up for all of them
2585 * that quickly find the other member of the pair. It might actually
2586 * be faster to pre-fold these, but it isn't currently done, except for
2587 * the sharp s. Code elsewhere in this file makes sure that it gets
2588 * folded to 'ss', even if the pattern isn't UTF-8. This avoids the
2589 * issues described in the next item.
2590 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2591 * 'ss' or not is not knowable at compile time. It will match iff the
2592 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2593 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2594 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2595 * described in item 3). An assumption that the optimizer part of
2596 * regexec.c (probably unwittingly) makes is that a character in the
2597 * pattern corresponds to at most a single character in the target string.
2598 * (And I do mean character, and not byte here, unlike other parts of the
2599 * documentation that have never been updated to account for multibyte
2600 * Unicode.) This assumption is wrong only in this case, as all other
2601 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2602 * virtue of having this file pre-fold UTF-8 patterns. I'm
2603 * reluctant to try to change this assumption, so instead the code punts.
2604 * This routine examines EXACTF nodes for the sharp s, and returns a
2605 * boolean indicating whether or not the node is an EXACTF node that
2606 * contains a sharp s. When it is true, the caller sets a flag that later
2607 * causes the optimizer in this file to not set values for the floating
2608 * and fixed string lengths, and thus avoids the optimizer code in
2609 * regexec.c that makes the invalid assumption. Thus, there is no
2610 * optimization based on string lengths for EXACTF nodes that contain the
2611 * sharp s. This only happens for /id rules (which means the pattern
2615 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2616 if (PL_regkind[OP(scan)] == EXACT) \
2617 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2620 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) {
2621 /* Merge several consecutive EXACTish nodes into one. */
2622 regnode *n = regnext(scan);
2624 regnode *next = scan + NODE_SZ_STR(scan);
2628 regnode *stop = scan;
2629 GET_RE_DEBUG_FLAGS_DECL;
2631 PERL_UNUSED_ARG(depth);
2634 PERL_ARGS_ASSERT_JOIN_EXACT;
2635 #ifndef EXPERIMENTAL_INPLACESCAN
2636 PERL_UNUSED_ARG(flags);
2637 PERL_UNUSED_ARG(val);
2639 DEBUG_PEEP("join",scan,depth);
2641 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2642 * EXACT ones that are mergeable to the current one. */
2644 && (PL_regkind[OP(n)] == NOTHING
2645 || (stringok && OP(n) == OP(scan)))
2647 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2650 if (OP(n) == TAIL || n > next)
2652 if (PL_regkind[OP(n)] == NOTHING) {
2653 DEBUG_PEEP("skip:",n,depth);
2654 NEXT_OFF(scan) += NEXT_OFF(n);
2655 next = n + NODE_STEP_REGNODE;
2662 else if (stringok) {
2663 const unsigned int oldl = STR_LEN(scan);
2664 regnode * const nnext = regnext(n);
2666 if (oldl + STR_LEN(n) > U8_MAX)
2669 DEBUG_PEEP("merg",n,depth);
2672 NEXT_OFF(scan) += NEXT_OFF(n);
2673 STR_LEN(scan) += STR_LEN(n);
2674 next = n + NODE_SZ_STR(n);
2675 /* Now we can overwrite *n : */
2676 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2684 #ifdef EXPERIMENTAL_INPLACESCAN
2685 if (flags && !NEXT_OFF(n)) {
2686 DEBUG_PEEP("atch", val, depth);
2687 if (reg_off_by_arg[OP(n)]) {
2688 ARG_SET(n, val - n);
2691 NEXT_OFF(n) = val - n;
2699 *has_exactf_sharp_s = FALSE;
2701 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2702 * can now analyze for sequences of problematic code points. (Prior to
2703 * this final joining, sequences could have been split over boundaries, and
2704 * hence missed). The sequences only happen in folding, hence for any
2705 * non-EXACT EXACTish node */
2706 if (OP(scan) != EXACT) {
2708 U8 * s0 = (U8*) STRING(scan);
2709 U8 * const s_end = s0 + STR_LEN(scan);
2711 /* The below is perhaps overboard, but this allows us to save a test
2712 * each time through the loop at the expense of a mask. This is
2713 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2714 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2715 * This uses an exclusive 'or' to find that bit and then inverts it to
2716 * form a mask, with just a single 0, in the bit position where 'S' and
2718 const U8 S_or_s_mask = ~ ('S' ^ 's');
2719 const U8 s_masked = 's' & S_or_s_mask;
2721 /* One pass is made over the node's string looking for all the
2722 * possibilities. to avoid some tests in the loop, there are two main
2723 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2727 /* There are two problematic Greek code points in Unicode
2730 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2731 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2737 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2738 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2740 * This means that in case-insensitive matching (or "loose
2741 * matching", as Unicode calls it), an EXACTF of length six (the
2742 * UTF-8 encoded byte length of the above casefolded versions) can
2743 * match a target string of length two (the byte length of UTF-8
2744 * encoded U+0390 or U+03B0). This would rather mess up the
2745 * minimum length computation. (there are other code points that
2746 * also fold to these two sequences, but the delta is smaller)
2748 * If these sequences are found, the minimum length is decreased by
2749 * four (six minus two).
2751 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2752 * LETTER SHARP S. We decrease the min length by 1 for each
2753 * occurrence of 'ss' found */
2755 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2756 # define U390_first_byte 0xb4
2757 const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
2758 # define U3B0_first_byte 0xb5
2759 const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
2761 # define U390_first_byte 0xce
2762 const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
2763 # define U3B0_first_byte 0xcf
2764 const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
2766 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2767 yields a net of 0 */
2768 /* Examine the string for one of the problematic sequences */
2770 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2771 * sequence we are looking for is 2 */
2775 /* Look for the first byte in each problematic sequence */
2777 /* We don't have to worry about other things that fold to
2778 * 's' (such as the long s, U+017F), as all above-latin1
2779 * code points have been pre-folded */
2783 /* Current character is an 's' or 'S'. If next one is
2784 * as well, we have the dreaded sequence */
2785 if (((*(s+1) & S_or_s_mask) == s_masked)
2786 /* These two node types don't have special handling
2788 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2791 OP(scan) = EXACTFU_SS;
2792 s++; /* No need to look at this character again */
2796 case U390_first_byte:
2797 if (s_end - s >= len
2799 /* The 1's are because are skipping comparing the
2801 && memEQ(s + 1, U390_tail, len - 1))
2803 goto greek_sequence;
2807 case U3B0_first_byte:
2808 if (! (s_end - s >= len
2809 && memEQ(s + 1, U3B0_tail, len - 1)))
2816 /* This can't currently be handled by trie's, so change
2817 * the node type to indicate this. If EXACTFA and
2818 * EXACTFL were ever to be handled by trie's, this
2819 * would have to be changed. If this node has already
2820 * been changed to EXACTFU_SS in this loop, leave it as
2821 * is. (I (khw) think it doesn't matter in regexec.c
2822 * for UTF patterns, but no need to change it */
2823 if (OP(scan) == EXACTFU) {
2824 OP(scan) = EXACTFU_NO_TRIE;
2826 s += 6; /* We already know what this sequence is. Skip
2832 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2834 /* Here, the pattern is not UTF-8. We need to look only for the
2835 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2836 * in the final position. Otherwise we can stop looking 1 byte
2837 * earlier because have to find both the first and second 's' */
2838 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2840 for (s = s0; s < upper; s++) {
2845 && ((*(s+1) & S_or_s_mask) == s_masked))
2849 /* EXACTF nodes need to know that the minimum
2850 * length changed so that a sharp s in the string
2851 * can match this ss in the pattern, but they
2852 * remain EXACTF nodes, as they are not trie'able,
2853 * so don't have to invent a new node type to
2854 * exclude them from the trie code */
2855 if (OP(scan) != EXACTF) {
2856 OP(scan) = EXACTFU_SS;
2861 case LATIN_SMALL_LETTER_SHARP_S:
2862 if (OP(scan) == EXACTF) {
2863 *has_exactf_sharp_s = TRUE;
2872 /* Allow dumping but overwriting the collection of skipped
2873 * ops and/or strings with fake optimized ops */
2874 n = scan + NODE_SZ_STR(scan);
2882 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2886 /* REx optimizer. Converts nodes into quicker variants "in place".
2887 Finds fixed substrings. */
2889 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2890 to the position after last scanned or to NULL. */
2892 #define INIT_AND_WITHP \
2893 assert(!and_withp); \
2894 Newx(and_withp,1,struct regnode_charclass_class); \
2895 SAVEFREEPV(and_withp)
2897 /* this is a chain of data about sub patterns we are processing that
2898 need to be handled separately/specially in study_chunk. Its so
2899 we can simulate recursion without losing state. */
2901 typedef struct scan_frame {
2902 regnode *last; /* last node to process in this frame */
2903 regnode *next; /* next node to process when last is reached */
2904 struct scan_frame *prev; /*previous frame*/
2905 I32 stop; /* what stopparen do we use */
2909 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2911 #define CASE_SYNST_FNC(nAmE) \
2913 if (flags & SCF_DO_STCLASS_AND) { \
2914 for (value = 0; value < 256; value++) \
2915 if (!is_ ## nAmE ## _cp(value)) \
2916 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2919 for (value = 0; value < 256; value++) \
2920 if (is_ ## nAmE ## _cp(value)) \
2921 ANYOF_BITMAP_SET(data->start_class, value); \
2925 if (flags & SCF_DO_STCLASS_AND) { \
2926 for (value = 0; value < 256; value++) \
2927 if (is_ ## nAmE ## _cp(value)) \
2928 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2931 for (value = 0; value < 256; value++) \
2932 if (!is_ ## nAmE ## _cp(value)) \
2933 ANYOF_BITMAP_SET(data->start_class, value); \
2940 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2941 I32 *minlenp, I32 *deltap,
2946 struct regnode_charclass_class *and_withp,
2947 U32 flags, U32 depth)
2948 /* scanp: Start here (read-write). */
2949 /* deltap: Write maxlen-minlen here. */
2950 /* last: Stop before this one. */
2951 /* data: string data about the pattern */
2952 /* stopparen: treat close N as END */
2953 /* recursed: which subroutines have we recursed into */
2954 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
2957 I32 min = 0, pars = 0, code;
2958 regnode *scan = *scanp, *next;
2960 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2961 int is_inf_internal = 0; /* The studied chunk is infinite */
2962 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2963 scan_data_t data_fake;
2964 SV *re_trie_maxbuff = NULL;
2965 regnode *first_non_open = scan;
2966 I32 stopmin = I32_MAX;
2967 scan_frame *frame = NULL;
2968 GET_RE_DEBUG_FLAGS_DECL;
2970 PERL_ARGS_ASSERT_STUDY_CHUNK;
2973 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
2977 while (first_non_open && OP(first_non_open) == OPEN)
2978 first_non_open=regnext(first_non_open);
2983 while ( scan && OP(scan) != END && scan < last ){
2984 UV min_subtract = 0; /* How much to subtract from the minimum node
2985 length to get a real minimum (because the
2986 folded version may be shorter) */
2987 bool has_exactf_sharp_s = FALSE;
2988 /* Peephole optimizer: */
2989 DEBUG_STUDYDATA("Peep:", data,depth);
2990 DEBUG_PEEP("Peep",scan,depth);
2992 /* Its not clear to khw or hv why this is done here, and not in the
2993 * clauses that deal with EXACT nodes. khw's guess is that it's
2994 * because of a previous design */
2995 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
2997 /* Follow the next-chain of the current node and optimize
2998 away all the NOTHINGs from it. */
2999 if (OP(scan) != CURLYX) {
3000 const int max = (reg_off_by_arg[OP(scan)]
3002 /* I32 may be smaller than U16 on CRAYs! */
3003 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3004 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3008 /* Skip NOTHING and LONGJMP. */
3009 while ((n = regnext(n))
3010 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3011 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3012 && off + noff < max)
3014 if (reg_off_by_arg[OP(scan)])
3017 NEXT_OFF(scan) = off;
3022 /* The principal pseudo-switch. Cannot be a switch, since we
3023 look into several different things. */
3024 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3025 || OP(scan) == IFTHEN) {
3026 next = regnext(scan);
3028 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3030 if (OP(next) == code || code == IFTHEN) {
3031 /* NOTE - There is similar code to this block below for handling
3032 TRIE nodes on a re-study. If you change stuff here check there
3034 I32 max1 = 0, min1 = I32_MAX, num = 0;
3035 struct regnode_charclass_class accum;
3036 regnode * const startbranch=scan;
3038 if (flags & SCF_DO_SUBSTR)
3039 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3040 if (flags & SCF_DO_STCLASS)
3041 cl_init_zero(pRExC_state, &accum);
3043 while (OP(scan) == code) {
3044 I32 deltanext, minnext, f = 0, fake;
3045 struct regnode_charclass_class this_class;
3048 data_fake.flags = 0;
3050 data_fake.whilem_c = data->whilem_c;
3051 data_fake.last_closep = data->last_closep;
3054 data_fake.last_closep = &fake;
3056 data_fake.pos_delta = delta;
3057 next = regnext(scan);
3058 scan = NEXTOPER(scan);
3060 scan = NEXTOPER(scan);
3061 if (flags & SCF_DO_STCLASS) {
3062 cl_init(pRExC_state, &this_class);
3063 data_fake.start_class = &this_class;
3064 f = SCF_DO_STCLASS_AND;
3066 if (flags & SCF_WHILEM_VISITED_POS)
3067 f |= SCF_WHILEM_VISITED_POS;
3069 /* we suppose the run is continuous, last=next...*/
3070 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3072 stopparen, recursed, NULL, f,depth+1);
3075 if (max1 < minnext + deltanext)
3076 max1 = minnext + deltanext;
3077 if (deltanext == I32_MAX)
3078 is_inf = is_inf_internal = 1;
3080 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3082 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3083 if ( stopmin > minnext)
3084 stopmin = min + min1;
3085 flags &= ~SCF_DO_SUBSTR;
3087 data->flags |= SCF_SEEN_ACCEPT;
3090 if (data_fake.flags & SF_HAS_EVAL)
3091 data->flags |= SF_HAS_EVAL;
3092 data->whilem_c = data_fake.whilem_c;
3094 if (flags & SCF_DO_STCLASS)
3095 cl_or(pRExC_state, &accum, &this_class);
3097 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3099 if (flags & SCF_DO_SUBSTR) {
3100 data->pos_min += min1;
3101 data->pos_delta += max1 - min1;
3102 if (max1 != min1 || is_inf)
3103 data->longest = &(data->longest_float);
3106 delta += max1 - min1;
3107 if (flags & SCF_DO_STCLASS_OR) {
3108 cl_or(pRExC_state, data->start_class, &accum);
3110 cl_and(data->start_class, and_withp);
3111 flags &= ~SCF_DO_STCLASS;
3114 else if (flags & SCF_DO_STCLASS_AND) {
3116 cl_and(data->start_class, &accum);
3117 flags &= ~SCF_DO_STCLASS;
3120 /* Switch to OR mode: cache the old value of
3121 * data->start_class */
3123 StructCopy(data->start_class, and_withp,
3124 struct regnode_charclass_class);
3125 flags &= ~SCF_DO_STCLASS_AND;
3126 StructCopy(&accum, data->start_class,
3127 struct regnode_charclass_class);
3128 flags |= SCF_DO_STCLASS_OR;
3129 data->start_class->flags |= ANYOF_EOS;
3133 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3136 Assuming this was/is a branch we are dealing with: 'scan' now
3137 points at the item that follows the branch sequence, whatever
3138 it is. We now start at the beginning of the sequence and look
3145 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3147 If we can find such a subsequence we need to turn the first
3148 element into a trie and then add the subsequent branch exact
3149 strings to the trie.
3153 1. patterns where the whole set of branches can be converted.
3155 2. patterns where only a subset can be converted.
3157 In case 1 we can replace the whole set with a single regop
3158 for the trie. In case 2 we need to keep the start and end
3161 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3162 becomes BRANCH TRIE; BRANCH X;
3164 There is an additional case, that being where there is a
3165 common prefix, which gets split out into an EXACT like node
3166 preceding the TRIE node.
3168 If x(1..n)==tail then we can do a simple trie, if not we make
3169 a "jump" trie, such that when we match the appropriate word
3170 we "jump" to the appropriate tail node. Essentially we turn
3171 a nested if into a case structure of sorts.
3176 if (!re_trie_maxbuff) {
3177 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3178 if (!SvIOK(re_trie_maxbuff))
3179 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3181 if ( SvIV(re_trie_maxbuff)>=0 ) {
3183 regnode *first = (regnode *)NULL;
3184 regnode *last = (regnode *)NULL;
3185 regnode *tail = scan;
3190 SV * const mysv = sv_newmortal(); /* for dumping */
3192 /* var tail is used because there may be a TAIL
3193 regop in the way. Ie, the exacts will point to the
3194 thing following the TAIL, but the last branch will
3195 point at the TAIL. So we advance tail. If we
3196 have nested (?:) we may have to move through several
3200 while ( OP( tail ) == TAIL ) {
3201 /* this is the TAIL generated by (?:) */
3202 tail = regnext( tail );
3207 regprop(RExC_rx, mysv, tail );
3208 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3209 (int)depth * 2 + 2, "",
3210 "Looking for TRIE'able sequences. Tail node is: ",
3211 SvPV_nolen_const( mysv )
3217 step through the branches, cur represents each
3218 branch, noper is the first thing to be matched
3219 as part of that branch and noper_next is the
3220 regnext() of that node. if noper is an EXACT
3221 and noper_next is the same as scan (our current
3222 position in the regex) then the EXACT branch is
3223 a possible optimization target. Once we have
3224 two or more consecutive such branches we can
3225 create a trie of the EXACT's contents and stich
3226 it in place. If the sequence represents all of
3227 the branches we eliminate the whole thing and
3228 replace it with a single TRIE. If it is a
3229 subsequence then we need to stitch it in. This
3230 means the first branch has to remain, and needs
3231 to be repointed at the item on the branch chain
3232 following the last branch optimized. This could
3233 be either a BRANCH, in which case the
3234 subsequence is internal, or it could be the
3235 item following the branch sequence in which
3236 case the subsequence is at the end.
3240 /* dont use tail as the end marker for this traverse */
3241 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3242 regnode * const noper = NEXTOPER( cur );
3243 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3244 regnode * const noper_next = regnext( noper );
3248 regprop(RExC_rx, mysv, cur);
3249 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3250 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3252 regprop(RExC_rx, mysv, noper);
3253 PerlIO_printf( Perl_debug_log, " -> %s",
3254 SvPV_nolen_const(mysv));
3257 regprop(RExC_rx, mysv, noper_next );
3258 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3259 SvPV_nolen_const(mysv));
3261 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
3262 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
3264 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
3265 : PL_regkind[ OP( noper ) ] == EXACT )
3266 || OP(noper) == NOTHING )
3268 && noper_next == tail
3273 if ( !first || optype == NOTHING ) {
3274 if (!first) first = cur;
3275 optype = OP( noper );
3281 Currently the trie logic handles case insensitive matching properly only
3282 when the pattern is UTF-8 and the node is EXACTFU (thus forcing unicode
3285 If/when this is fixed the following define can be swapped
3286 in below to fully enable trie logic.
3288 #define TRIE_TYPE_IS_SAFE 1
3290 Note that join_exact() assumes that the other types of EXACTFish nodes are not
3291 used in tries, so that would have to be updated if this changed
3294 #define TRIE_TYPE_IS_SAFE ((UTF && optype == EXACTFU) || optype==EXACT)
3296 if ( last && TRIE_TYPE_IS_SAFE ) {
3297 make_trie( pRExC_state,
3298 startbranch, first, cur, tail, count,
3301 if ( PL_regkind[ OP( noper ) ] == EXACT
3303 && noper_next == tail
3308 optype = OP( noper );
3318 regprop(RExC_rx, mysv, cur);
3319 PerlIO_printf( Perl_debug_log,
3320 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3321 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3325 if ( last && TRIE_TYPE_IS_SAFE ) {
3326 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
3327 #ifdef TRIE_STUDY_OPT
3328 if ( ((made == MADE_EXACT_TRIE &&
3329 startbranch == first)
3330 || ( first_non_open == first )) &&
3332 flags |= SCF_TRIE_RESTUDY;
3333 if ( startbranch == first
3336 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3346 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3347 scan = NEXTOPER(NEXTOPER(scan));
3348 } else /* single branch is optimized. */
3349 scan = NEXTOPER(scan);
3351 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3352 scan_frame *newframe = NULL;
3357 if (OP(scan) != SUSPEND) {
3358 /* set the pointer */
3359 if (OP(scan) == GOSUB) {
3361 RExC_recurse[ARG2L(scan)] = scan;
3362 start = RExC_open_parens[paren-1];
3363 end = RExC_close_parens[paren-1];
3366 start = RExC_rxi->program + 1;
3370 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3371 SAVEFREEPV(recursed);
3373 if (!PAREN_TEST(recursed,paren+1)) {
3374 PAREN_SET(recursed,paren+1);
3375 Newx(newframe,1,scan_frame);
3377 if (flags & SCF_DO_SUBSTR) {
3378 SCAN_COMMIT(pRExC_state,data,minlenp);
3379 data->longest = &(data->longest_float);
3381 is_inf = is_inf_internal = 1;
3382 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3383 cl_anything(pRExC_state, data->start_class);
3384 flags &= ~SCF_DO_STCLASS;
3387 Newx(newframe,1,scan_frame);
3390 end = regnext(scan);
3395 SAVEFREEPV(newframe);
3396 newframe->next = regnext(scan);
3397 newframe->last = last;
3398 newframe->stop = stopparen;
3399 newframe->prev = frame;
3409 else if (OP(scan) == EXACT) {
3410 I32 l = STR_LEN(scan);
3413 const U8 * const s = (U8*)STRING(scan);
3414 l = utf8_length(s, s + l);
3415 uc = utf8_to_uvchr(s, NULL);
3417 uc = *((U8*)STRING(scan));
3420 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3421 /* The code below prefers earlier match for fixed
3422 offset, later match for variable offset. */
3423 if (data->last_end == -1) { /* Update the start info. */
3424 data->last_start_min = data->pos_min;
3425 data->last_start_max = is_inf
3426 ? I32_MAX : data->pos_min + data->pos_delta;
3428 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3430 SvUTF8_on(data->last_found);
3432 SV * const sv = data->last_found;
3433 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3434 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3435 if (mg && mg->mg_len >= 0)
3436 mg->mg_len += utf8_length((U8*)STRING(scan),
3437 (U8*)STRING(scan)+STR_LEN(scan));
3439 data->last_end = data->pos_min + l;
3440 data->pos_min += l; /* As in the first entry. */
3441 data->flags &= ~SF_BEFORE_EOL;
3443 if (flags & SCF_DO_STCLASS_AND) {
3444 /* Check whether it is compatible with what we know already! */
3448 /* If compatible, we or it in below. It is compatible if is
3449 * in the bitmp and either 1) its bit or its fold is set, or 2)
3450 * it's for a locale. Even if there isn't unicode semantics
3451 * here, at runtime there may be because of matching against a
3452 * utf8 string, so accept a possible false positive for
3453 * latin1-range folds */
3455 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3456 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3457 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
3458 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3463 ANYOF_CLASS_ZERO(data->start_class);
3464 ANYOF_BITMAP_ZERO(data->start_class);
3466 ANYOF_BITMAP_SET(data->start_class, uc);
3467 else if (uc >= 0x100) {
3470 /* Some Unicode code points fold to the Latin1 range; as
3471 * XXX temporary code, instead of figuring out if this is
3472 * one, just assume it is and set all the start class bits
3473 * that could be some such above 255 code point's fold
3474 * which will generate fals positives. As the code
3475 * elsewhere that does compute the fold settles down, it
3476 * can be extracted out and re-used here */
3477 for (i = 0; i < 256; i++){
3478 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3479 ANYOF_BITMAP_SET(data->start_class, i);
3483 data->start_class->flags &= ~ANYOF_EOS;
3485 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3487 else if (flags & SCF_DO_STCLASS_OR) {
3488 /* false positive possible if the class is case-folded */
3490 ANYOF_BITMAP_SET(data->start_class, uc);
3492 data->start_class->flags |= ANYOF_UNICODE_ALL;
3493 data->start_class->flags &= ~ANYOF_EOS;
3494 cl_and(data->start_class, and_withp);
3496 flags &= ~SCF_DO_STCLASS;
3498 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3499 I32 l = STR_LEN(scan);
3500 UV uc = *((U8*)STRING(scan));
3502 /* Search for fixed substrings supports EXACT only. */
3503 if (flags & SCF_DO_SUBSTR) {
3505 SCAN_COMMIT(pRExC_state, data, minlenp);
3508 const U8 * const s = (U8 *)STRING(scan);
3509 l = utf8_length(s, s + l);
3510 uc = utf8_to_uvchr(s, NULL);
3512 else if (has_exactf_sharp_s) {
3513 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3515 min += l - min_subtract;
3519 delta += min_subtract;
3520 if (flags & SCF_DO_SUBSTR) {
3521 data->pos_min += l - min_subtract;
3522 if (data->pos_min < 0) {
3525 data->pos_delta += min_subtract;
3527 data->longest = &(data->longest_float);
3530 if (flags & SCF_DO_STCLASS_AND) {
3531 /* Check whether it is compatible with what we know already! */
3534 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3535 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3536 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3540 ANYOF_CLASS_ZERO(data->start_class);
3541 ANYOF_BITMAP_ZERO(data->start_class);
3543 ANYOF_BITMAP_SET(data->start_class, uc);
3544 data->start_class->flags &= ~ANYOF_EOS;
3545 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
3546 if (OP(scan) == EXACTFL) {
3547 /* XXX This set is probably no longer necessary, and
3548 * probably wrong as LOCALE now is on in the initial
3550 data->start_class->flags |= ANYOF_LOCALE;
3554 /* Also set the other member of the fold pair. In case
3555 * that unicode semantics is called for at runtime, use
3556 * the full latin1 fold. (Can't do this for locale,
3557 * because not known until runtime) */
3558 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3560 /* All other (EXACTFL handled above) folds except under
3561 * /iaa that include s, S, and sharp_s also may include
3563 if (OP(scan) != EXACTFA) {
3564 if (uc == 's' || uc == 'S') {
3565 ANYOF_BITMAP_SET(data->start_class,
3566 LATIN_SMALL_LETTER_SHARP_S);
3568 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3569 ANYOF_BITMAP_SET(data->start_class, 's');
3570 ANYOF_BITMAP_SET(data->start_class, 'S');
3575 else if (uc >= 0x100) {
3577 for (i = 0; i < 256; i++){
3578 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3579 ANYOF_BITMAP_SET(data->start_class, i);
3584 else if (flags & SCF_DO_STCLASS_OR) {
3585 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
3586 /* false positive possible if the class is case-folded.
3587 Assume that the locale settings are the same... */
3589 ANYOF_BITMAP_SET(data->start_class, uc);
3590 if (OP(scan) != EXACTFL) {
3592 /* And set the other member of the fold pair, but
3593 * can't do that in locale because not known until
3595 ANYOF_BITMAP_SET(data->start_class,
3596 PL_fold_latin1[uc]);
3598 /* All folds except under /iaa that include s, S,
3599 * and sharp_s also may include the others */
3600 if (OP(scan) != EXACTFA) {
3601 if (uc == 's' || uc == 'S') {
3602 ANYOF_BITMAP_SET(data->start_class,
3603 LATIN_SMALL_LETTER_SHARP_S);
3605 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3606 ANYOF_BITMAP_SET(data->start_class, 's');
3607 ANYOF_BITMAP_SET(data->start_class, 'S');
3612 data->start_class->flags &= ~ANYOF_EOS;
3614 cl_and(data->start_class, and_withp);
3616 flags &= ~SCF_DO_STCLASS;
3618 else if (REGNODE_VARIES(OP(scan))) {
3619 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3620 I32 f = flags, pos_before = 0;
3621 regnode * const oscan = scan;
3622 struct regnode_charclass_class this_class;
3623 struct regnode_charclass_class *oclass = NULL;
3624 I32 next_is_eval = 0;
3626 switch (PL_regkind[OP(scan)]) {
3627 case WHILEM: /* End of (?:...)* . */
3628 scan = NEXTOPER(scan);
3631 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3632 next = NEXTOPER(scan);
3633 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3635 maxcount = REG_INFTY;
3636 next = regnext(scan);
3637 scan = NEXTOPER(scan);
3641 if (flags & SCF_DO_SUBSTR)
3646 if (flags & SCF_DO_STCLASS) {
3648 maxcount = REG_INFTY;
3649 next = regnext(scan);
3650 scan = NEXTOPER(scan);
3653 is_inf = is_inf_internal = 1;
3654 scan = regnext(scan);
3655 if (flags & SCF_DO_SUBSTR) {
3656 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3657 data->longest = &(data->longest_float);
3659 goto optimize_curly_tail;
3661 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3662 && (scan->flags == stopparen))
3667 mincount = ARG1(scan);
3668 maxcount = ARG2(scan);
3670 next = regnext(scan);
3671 if (OP(scan) == CURLYX) {
3672 I32 lp = (data ? *(data->last_closep) : 0);
3673 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3675 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3676 next_is_eval = (OP(scan) == EVAL);
3678 if (flags & SCF_DO_SUBSTR) {
3679 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3680 pos_before = data->pos_min;
3684 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3686 data->flags |= SF_IS_INF;
3688 if (flags & SCF_DO_STCLASS) {
3689 cl_init(pRExC_state, &this_class);
3690 oclass = data->start_class;
3691 data->start_class = &this_class;
3692 f |= SCF_DO_STCLASS_AND;
3693 f &= ~SCF_DO_STCLASS_OR;
3695 /* Exclude from super-linear cache processing any {n,m}
3696 regops for which the combination of input pos and regex
3697 pos is not enough information to determine if a match
3700 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3701 regex pos at the \s*, the prospects for a match depend not
3702 only on the input position but also on how many (bar\s*)
3703 repeats into the {4,8} we are. */
3704 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3705 f &= ~SCF_WHILEM_VISITED_POS;
3707 /* This will finish on WHILEM, setting scan, or on NULL: */
3708 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3709 last, data, stopparen, recursed, NULL,
3711 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3713 if (flags & SCF_DO_STCLASS)
3714 data->start_class = oclass;
3715 if (mincount == 0 || minnext == 0) {
3716 if (flags & SCF_DO_STCLASS_OR) {
3717 cl_or(pRExC_state, data->start_class, &this_class);
3719 else if (flags & SCF_DO_STCLASS_AND) {
3720 /* Switch to OR mode: cache the old value of
3721 * data->start_class */
3723 StructCopy(data->start_class, and_withp,
3724 struct regnode_charclass_class);
3725 flags &= ~SCF_DO_STCLASS_AND;
3726 StructCopy(&this_class, data->start_class,
3727 struct regnode_charclass_class);
3728 flags |= SCF_DO_STCLASS_OR;
3729 data->start_class->flags |= ANYOF_EOS;
3731 } else { /* Non-zero len */
3732 if (flags & SCF_DO_STCLASS_OR) {
3733 cl_or(pRExC_state, data->start_class, &this_class);
3734 cl_and(data->start_class, and_withp);
3736 else if (flags & SCF_DO_STCLASS_AND)
3737 cl_and(data->start_class, &this_class);
3738 flags &= ~SCF_DO_STCLASS;
3740 if (!scan) /* It was not CURLYX, but CURLY. */
3742 if ( /* ? quantifier ok, except for (?{ ... }) */
3743 (next_is_eval || !(mincount == 0 && maxcount == 1))
3744 && (minnext == 0) && (deltanext == 0)
3745 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3746 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3748 ckWARNreg(RExC_parse,
3749 "Quantifier unexpected on zero-length expression");
3752 min += minnext * mincount;
3753 is_inf_internal |= ((maxcount == REG_INFTY
3754 && (minnext + deltanext) > 0)
3755 || deltanext == I32_MAX);
3756 is_inf |= is_inf_internal;
3757 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3759 /* Try powerful optimization CURLYX => CURLYN. */
3760 if ( OP(oscan) == CURLYX && data
3761 && data->flags & SF_IN_PAR
3762 && !(data->flags & SF_HAS_EVAL)
3763 && !deltanext && minnext == 1 ) {
3764 /* Try to optimize to CURLYN. */
3765 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3766 regnode * const nxt1 = nxt;
3773 if (!REGNODE_SIMPLE(OP(nxt))
3774 && !(PL_regkind[OP(nxt)] == EXACT
3775 && STR_LEN(nxt) == 1))
3781 if (OP(nxt) != CLOSE)
3783 if (RExC_open_parens) {
3784 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3785 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3787 /* Now we know that nxt2 is the only contents: */
3788 oscan->flags = (U8)ARG(nxt);
3790 OP(nxt1) = NOTHING; /* was OPEN. */
3793 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3794 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3795 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3796 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3797 OP(nxt + 1) = OPTIMIZED; /* was count. */
3798 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3803 /* Try optimization CURLYX => CURLYM. */
3804 if ( OP(oscan) == CURLYX && data
3805 && !(data->flags & SF_HAS_PAR)
3806 && !(data->flags & SF_HAS_EVAL)
3807 && !deltanext /* atom is fixed width */
3808 && minnext != 0 /* CURLYM can't handle zero width */
3810 /* XXXX How to optimize if data == 0? */
3811 /* Optimize to a simpler form. */
3812 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3816 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3817 && (OP(nxt2) != WHILEM))
3819 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3820 /* Need to optimize away parenths. */
3821 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
3822 /* Set the parenth number. */
3823 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3825 oscan->flags = (U8)ARG(nxt);
3826 if (RExC_open_parens) {
3827 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3828 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3830 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3831 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3834 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3835 OP(nxt + 1) = OPTIMIZED; /* was count. */
3836 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
3837 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
3840 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3841 regnode *nnxt = regnext(nxt1);
3843 if (reg_off_by_arg[OP(nxt1)])
3844 ARG_SET(nxt1, nxt2 - nxt1);
3845 else if (nxt2 - nxt1 < U16_MAX)
3846 NEXT_OFF(nxt1) = nxt2 - nxt1;
3848 OP(nxt) = NOTHING; /* Cannot beautify */
3853 /* Optimize again: */
3854 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
3855 NULL, stopparen, recursed, NULL, 0,depth+1);