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"
93 #define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
100 # if defined(BUGGY_MSC6)
101 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
102 # pragma optimize("a",off)
103 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
104 # pragma optimize("w",on )
105 # endif /* BUGGY_MSC6 */
109 #define STATIC static
113 typedef struct RExC_state_t {
114 U32 flags; /* are we folding, multilining? */
115 char *precomp; /* uncompiled string. */
116 REGEXP *rx_sv; /* The SV that is the regexp. */
117 regexp *rx; /* perl core regexp structure */
118 regexp_internal *rxi; /* internal data for regexp object pprivate field */
119 char *start; /* Start of input for compile */
120 char *end; /* End of input for compile */
121 char *parse; /* Input-scan pointer. */
122 I32 whilem_seen; /* number of WHILEM in this expr */
123 regnode *emit_start; /* Start of emitted-code area */
124 regnode *emit_bound; /* First regnode outside of the allocated space */
125 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
126 I32 naughty; /* How bad is this pattern? */
127 I32 sawback; /* Did we see \1, ...? */
129 I32 size; /* Code size. */
130 I32 npar; /* Capture buffer count, (OPEN). */
131 I32 cpar; /* Capture buffer count, (CLOSE). */
132 I32 nestroot; /* root parens we are in - used by accept */
136 regnode **open_parens; /* pointers to open parens */
137 regnode **close_parens; /* pointers to close parens */
138 regnode *opend; /* END node in program */
139 I32 utf8; /* whether the pattern is utf8 or not */
140 I32 orig_utf8; /* whether the pattern was originally in utf8 */
141 /* XXX use this for future optimisation of case
142 * where pattern must be upgraded to utf8. */
143 I32 uni_semantics; /* If a d charset modifier should use unicode
144 rules, even if the pattern is not in
146 HV *paren_names; /* Paren names */
148 regnode **recurse; /* Recurse regops */
149 I32 recurse_count; /* Number of recurse regops */
152 I32 override_recoding;
153 struct reg_code_block *code_blocks; /* positions of literal (?{})
155 int num_code_blocks; /* size of code_blocks[] */
156 int code_index; /* next code_blocks[] slot */
158 char *starttry; /* -Dr: where regtry was called. */
159 #define RExC_starttry (pRExC_state->starttry)
162 const char *lastparse;
164 AV *paren_name_list; /* idx -> name */
165 #define RExC_lastparse (pRExC_state->lastparse)
166 #define RExC_lastnum (pRExC_state->lastnum)
167 #define RExC_paren_name_list (pRExC_state->paren_name_list)
171 #define RExC_flags (pRExC_state->flags)
172 #define RExC_precomp (pRExC_state->precomp)
173 #define RExC_rx_sv (pRExC_state->rx_sv)
174 #define RExC_rx (pRExC_state->rx)
175 #define RExC_rxi (pRExC_state->rxi)
176 #define RExC_start (pRExC_state->start)
177 #define RExC_end (pRExC_state->end)
178 #define RExC_parse (pRExC_state->parse)
179 #define RExC_whilem_seen (pRExC_state->whilem_seen)
180 #ifdef RE_TRACK_PATTERN_OFFSETS
181 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
183 #define RExC_emit (pRExC_state->emit)
184 #define RExC_emit_start (pRExC_state->emit_start)
185 #define RExC_emit_bound (pRExC_state->emit_bound)
186 #define RExC_naughty (pRExC_state->naughty)
187 #define RExC_sawback (pRExC_state->sawback)
188 #define RExC_seen (pRExC_state->seen)
189 #define RExC_size (pRExC_state->size)
190 #define RExC_npar (pRExC_state->npar)
191 #define RExC_nestroot (pRExC_state->nestroot)
192 #define RExC_extralen (pRExC_state->extralen)
193 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
194 #define RExC_seen_evals (pRExC_state->seen_evals)
195 #define RExC_utf8 (pRExC_state->utf8)
196 #define RExC_uni_semantics (pRExC_state->uni_semantics)
197 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
198 #define RExC_open_parens (pRExC_state->open_parens)
199 #define RExC_close_parens (pRExC_state->close_parens)
200 #define RExC_opend (pRExC_state->opend)
201 #define RExC_paren_names (pRExC_state->paren_names)
202 #define RExC_recurse (pRExC_state->recurse)
203 #define RExC_recurse_count (pRExC_state->recurse_count)
204 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
205 #define RExC_contains_locale (pRExC_state->contains_locale)
206 #define RExC_override_recoding (pRExC_state->override_recoding)
209 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
210 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
211 ((*s) == '{' && regcurly(s)))
214 #undef SPSTART /* dratted cpp namespace... */
217 * Flags to be passed up and down.
219 #define WORST 0 /* Worst case. */
220 #define HASWIDTH 0x01 /* Known to match non-null strings. */
222 /* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
223 * character, and if utf8, must be invariant. Note that this is not the same
224 * thing as REGNODE_SIMPLE */
226 #define SPSTART 0x04 /* Starts with * or +. */
227 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
228 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
230 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
232 /* whether trie related optimizations are enabled */
233 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
234 #define TRIE_STUDY_OPT
235 #define FULL_TRIE_STUDY
241 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
242 #define PBITVAL(paren) (1 << ((paren) & 7))
243 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
244 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
245 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
247 /* If not already in utf8, do a longjmp back to the beginning */
248 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
249 #define REQUIRE_UTF8 STMT_START { \
250 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
253 /* About scan_data_t.
255 During optimisation we recurse through the regexp program performing
256 various inplace (keyhole style) optimisations. In addition study_chunk
257 and scan_commit populate this data structure with information about
258 what strings MUST appear in the pattern. We look for the longest
259 string that must appear at a fixed location, and we look for the
260 longest string that may appear at a floating location. So for instance
265 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
266 strings (because they follow a .* construct). study_chunk will identify
267 both FOO and BAR as being the longest fixed and floating strings respectively.
269 The strings can be composites, for instance
273 will result in a composite fixed substring 'foo'.
275 For each string some basic information is maintained:
277 - offset or min_offset
278 This is the position the string must appear at, or not before.
279 It also implicitly (when combined with minlenp) tells us how many
280 characters must match before the string we are searching for.
281 Likewise when combined with minlenp and the length of the string it
282 tells us how many characters must appear after the string we have
286 Only used for floating strings. This is the rightmost point that
287 the string can appear at. If set to I32 max it indicates that the
288 string can occur infinitely far to the right.
291 A pointer to the minimum length of the pattern that the string
292 was found inside. This is important as in the case of positive
293 lookahead or positive lookbehind we can have multiple patterns
298 The minimum length of the pattern overall is 3, the minimum length
299 of the lookahead part is 3, but the minimum length of the part that
300 will actually match is 1. So 'FOO's minimum length is 3, but the
301 minimum length for the F is 1. This is important as the minimum length
302 is used to determine offsets in front of and behind the string being
303 looked for. Since strings can be composites this is the length of the
304 pattern at the time it was committed with a scan_commit. Note that
305 the length is calculated by study_chunk, so that the minimum lengths
306 are not known until the full pattern has been compiled, thus the
307 pointer to the value.
311 In the case of lookbehind the string being searched for can be
312 offset past the start point of the final matching string.
313 If this value was just blithely removed from the min_offset it would
314 invalidate some of the calculations for how many chars must match
315 before or after (as they are derived from min_offset and minlen and
316 the length of the string being searched for).
317 When the final pattern is compiled and the data is moved from the
318 scan_data_t structure into the regexp structure the information
319 about lookbehind is factored in, with the information that would
320 have been lost precalculated in the end_shift field for the
323 The fields pos_min and pos_delta are used to store the minimum offset
324 and the delta to the maximum offset at the current point in the pattern.
328 typedef struct scan_data_t {
329 /*I32 len_min; unused */
330 /*I32 len_delta; unused */
334 I32 last_end; /* min value, <0 unless valid. */
337 SV **longest; /* Either &l_fixed, or &l_float. */
338 SV *longest_fixed; /* longest fixed string found in pattern */
339 I32 offset_fixed; /* offset where it starts */
340 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
341 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
342 SV *longest_float; /* longest floating string found in pattern */
343 I32 offset_float_min; /* earliest point in string it can appear */
344 I32 offset_float_max; /* latest point in string it can appear */
345 I32 *minlen_float; /* pointer to the minlen relevant to the string */
346 I32 lookbehind_float; /* is the position of the string modified by LB */
350 struct regnode_charclass_class *start_class;
354 * Forward declarations for pregcomp()'s friends.
357 static const scan_data_t zero_scan_data =
358 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
360 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
361 #define SF_BEFORE_SEOL 0x0001
362 #define SF_BEFORE_MEOL 0x0002
363 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
364 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
367 # define SF_FIX_SHIFT_EOL (0+2)
368 # define SF_FL_SHIFT_EOL (0+4)
370 # define SF_FIX_SHIFT_EOL (+2)
371 # define SF_FL_SHIFT_EOL (+4)
374 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
375 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
377 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
378 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
379 #define SF_IS_INF 0x0040
380 #define SF_HAS_PAR 0x0080
381 #define SF_IN_PAR 0x0100
382 #define SF_HAS_EVAL 0x0200
383 #define SCF_DO_SUBSTR 0x0400
384 #define SCF_DO_STCLASS_AND 0x0800
385 #define SCF_DO_STCLASS_OR 0x1000
386 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
387 #define SCF_WHILEM_VISITED_POS 0x2000
389 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
390 #define SCF_SEEN_ACCEPT 0x8000
392 #define UTF cBOOL(RExC_utf8)
394 /* The enums for all these are ordered so things work out correctly */
395 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
396 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
397 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
398 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
399 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
400 #define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
401 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
403 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
405 #define OOB_UNICODE 12345678
406 #define OOB_NAMEDCLASS -1
408 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
409 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
412 /* length of regex to show in messages that don't mark a position within */
413 #define RegexLengthToShowInErrorMessages 127
416 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
417 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
418 * op/pragma/warn/regcomp.
420 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
421 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
423 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
426 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
427 * arg. Show regex, up to a maximum length. If it's too long, chop and add
430 #define _FAIL(code) STMT_START { \
431 const char *ellipses = ""; \
432 IV len = RExC_end - RExC_precomp; \
435 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
436 if (len > RegexLengthToShowInErrorMessages) { \
437 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
438 len = RegexLengthToShowInErrorMessages - 10; \
444 #define FAIL(msg) _FAIL( \
445 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
446 msg, (int)len, RExC_precomp, ellipses))
448 #define FAIL2(msg,arg) _FAIL( \
449 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
450 arg, (int)len, RExC_precomp, ellipses))
453 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
455 #define Simple_vFAIL(m) STMT_START { \
456 const IV offset = RExC_parse - RExC_precomp; \
457 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
458 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
462 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
464 #define vFAIL(m) STMT_START { \
466 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
471 * Like Simple_vFAIL(), but accepts two arguments.
473 #define Simple_vFAIL2(m,a1) STMT_START { \
474 const IV offset = RExC_parse - RExC_precomp; \
475 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
476 (int)offset, RExC_precomp, RExC_precomp + offset); \
480 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
482 #define vFAIL2(m,a1) STMT_START { \
484 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
485 Simple_vFAIL2(m, a1); \
490 * Like Simple_vFAIL(), but accepts three arguments.
492 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
493 const IV offset = RExC_parse - RExC_precomp; \
494 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
495 (int)offset, RExC_precomp, RExC_precomp + offset); \
499 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
501 #define vFAIL3(m,a1,a2) STMT_START { \
503 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
504 Simple_vFAIL3(m, a1, a2); \
508 * Like Simple_vFAIL(), but accepts four arguments.
510 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
511 const IV offset = RExC_parse - RExC_precomp; \
512 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
513 (int)offset, RExC_precomp, RExC_precomp + offset); \
516 #define ckWARNreg(loc,m) STMT_START { \
517 const IV offset = loc - RExC_precomp; \
518 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
519 (int)offset, RExC_precomp, RExC_precomp + offset); \
522 #define ckWARNregdep(loc,m) STMT_START { \
523 const IV offset = loc - RExC_precomp; \
524 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
526 (int)offset, RExC_precomp, RExC_precomp + offset); \
529 #define ckWARN2regdep(loc,m, a1) STMT_START { \
530 const IV offset = loc - RExC_precomp; \
531 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
533 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
536 #define ckWARN2reg(loc, m, a1) STMT_START { \
537 const IV offset = loc - RExC_precomp; \
538 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
539 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
542 #define vWARN3(loc, m, a1, a2) STMT_START { \
543 const IV offset = loc - RExC_precomp; \
544 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
545 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
548 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
549 const IV offset = loc - RExC_precomp; \
550 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
551 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
554 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
555 const IV offset = loc - RExC_precomp; \
556 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
557 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
560 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
561 const IV offset = loc - RExC_precomp; \
562 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
563 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
566 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
567 const IV offset = loc - RExC_precomp; \
568 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
569 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
573 /* Allow for side effects in s */
574 #define REGC(c,s) STMT_START { \
575 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
578 /* Macros for recording node offsets. 20001227 mjd@plover.com
579 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
580 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
581 * Element 0 holds the number n.
582 * Position is 1 indexed.
584 #ifndef RE_TRACK_PATTERN_OFFSETS
585 #define Set_Node_Offset_To_R(node,byte)
586 #define Set_Node_Offset(node,byte)
587 #define Set_Cur_Node_Offset
588 #define Set_Node_Length_To_R(node,len)
589 #define Set_Node_Length(node,len)
590 #define Set_Node_Cur_Length(node)
591 #define Node_Offset(n)
592 #define Node_Length(n)
593 #define Set_Node_Offset_Length(node,offset,len)
594 #define ProgLen(ri) ri->u.proglen
595 #define SetProgLen(ri,x) ri->u.proglen = x
597 #define ProgLen(ri) ri->u.offsets[0]
598 #define SetProgLen(ri,x) ri->u.offsets[0] = x
599 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
601 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
602 __LINE__, (int)(node), (int)(byte))); \
604 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
606 RExC_offsets[2*(node)-1] = (byte); \
611 #define Set_Node_Offset(node,byte) \
612 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
613 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
615 #define Set_Node_Length_To_R(node,len) STMT_START { \
617 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
618 __LINE__, (int)(node), (int)(len))); \
620 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
622 RExC_offsets[2*(node)] = (len); \
627 #define Set_Node_Length(node,len) \
628 Set_Node_Length_To_R((node)-RExC_emit_start, len)
629 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
630 #define Set_Node_Cur_Length(node) \
631 Set_Node_Length(node, RExC_parse - parse_start)
633 /* Get offsets and lengths */
634 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
635 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
637 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
638 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
639 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
643 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
644 #define EXPERIMENTAL_INPLACESCAN
645 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
647 #define DEBUG_STUDYDATA(str,data,depth) \
648 DEBUG_OPTIMISE_MORE_r(if(data){ \
649 PerlIO_printf(Perl_debug_log, \
650 "%*s" str "Pos:%"IVdf"/%"IVdf \
651 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
652 (int)(depth)*2, "", \
653 (IV)((data)->pos_min), \
654 (IV)((data)->pos_delta), \
655 (UV)((data)->flags), \
656 (IV)((data)->whilem_c), \
657 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
658 is_inf ? "INF " : "" \
660 if ((data)->last_found) \
661 PerlIO_printf(Perl_debug_log, \
662 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
663 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
664 SvPVX_const((data)->last_found), \
665 (IV)((data)->last_end), \
666 (IV)((data)->last_start_min), \
667 (IV)((data)->last_start_max), \
668 ((data)->longest && \
669 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
670 SvPVX_const((data)->longest_fixed), \
671 (IV)((data)->offset_fixed), \
672 ((data)->longest && \
673 (data)->longest==&((data)->longest_float)) ? "*" : "", \
674 SvPVX_const((data)->longest_float), \
675 (IV)((data)->offset_float_min), \
676 (IV)((data)->offset_float_max) \
678 PerlIO_printf(Perl_debug_log,"\n"); \
681 static void clear_re(pTHX_ void *r);
683 /* Mark that we cannot extend a found fixed substring at this point.
684 Update the longest found anchored substring and the longest found
685 floating substrings if needed. */
688 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
690 const STRLEN l = CHR_SVLEN(data->last_found);
691 const STRLEN old_l = CHR_SVLEN(*data->longest);
692 GET_RE_DEBUG_FLAGS_DECL;
694 PERL_ARGS_ASSERT_SCAN_COMMIT;
696 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
697 SvSetMagicSV(*data->longest, data->last_found);
698 if (*data->longest == data->longest_fixed) {
699 data->offset_fixed = l ? data->last_start_min : data->pos_min;
700 if (data->flags & SF_BEFORE_EOL)
702 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
704 data->flags &= ~SF_FIX_BEFORE_EOL;
705 data->minlen_fixed=minlenp;
706 data->lookbehind_fixed=0;
708 else { /* *data->longest == data->longest_float */
709 data->offset_float_min = l ? data->last_start_min : data->pos_min;
710 data->offset_float_max = (l
711 ? data->last_start_max
712 : data->pos_min + data->pos_delta);
713 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
714 data->offset_float_max = I32_MAX;
715 if (data->flags & SF_BEFORE_EOL)
717 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
719 data->flags &= ~SF_FL_BEFORE_EOL;
720 data->minlen_float=minlenp;
721 data->lookbehind_float=0;
724 SvCUR_set(data->last_found, 0);
726 SV * const sv = data->last_found;
727 if (SvUTF8(sv) && SvMAGICAL(sv)) {
728 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
734 data->flags &= ~SF_BEFORE_EOL;
735 DEBUG_STUDYDATA("commit: ",data,0);
738 /* Can match anything (initialization) */
740 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
742 PERL_ARGS_ASSERT_CL_ANYTHING;
744 ANYOF_BITMAP_SETALL(cl);
745 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
746 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
748 /* If any portion of the regex is to operate under locale rules,
749 * initialization includes it. The reason this isn't done for all regexes
750 * is that the optimizer was written under the assumption that locale was
751 * all-or-nothing. Given the complexity and lack of documentation in the
752 * optimizer, and that there are inadequate test cases for locale, so many
753 * parts of it may not work properly, it is safest to avoid locale unless
755 if (RExC_contains_locale) {
756 ANYOF_CLASS_SETALL(cl); /* /l uses class */
757 cl->flags |= ANYOF_LOCALE;
760 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
764 /* Can match anything (initialization) */
766 S_cl_is_anything(const struct regnode_charclass_class *cl)
770 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
772 for (value = 0; value <= ANYOF_MAX; value += 2)
773 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
775 if (!(cl->flags & ANYOF_UNICODE_ALL))
777 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
782 /* Can match anything (initialization) */
784 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
786 PERL_ARGS_ASSERT_CL_INIT;
788 Zero(cl, 1, struct regnode_charclass_class);
790 cl_anything(pRExC_state, cl);
791 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
794 /* These two functions currently do the exact same thing */
795 #define cl_init_zero S_cl_init
797 /* 'AND' a given class with another one. Can create false positives. 'cl'
798 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
799 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
801 S_cl_and(struct regnode_charclass_class *cl,
802 const struct regnode_charclass_class *and_with)
804 PERL_ARGS_ASSERT_CL_AND;
806 assert(and_with->type == ANYOF);
808 /* I (khw) am not sure all these restrictions are necessary XXX */
809 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
810 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
811 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
812 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
813 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
816 if (and_with->flags & ANYOF_INVERT)
817 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
818 cl->bitmap[i] &= ~and_with->bitmap[i];
820 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
821 cl->bitmap[i] &= and_with->bitmap[i];
822 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
824 if (and_with->flags & ANYOF_INVERT) {
826 /* Here, the and'ed node is inverted. Get the AND of the flags that
827 * aren't affected by the inversion. Those that are affected are
828 * handled individually below */
829 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
830 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
831 cl->flags |= affected_flags;
833 /* We currently don't know how to deal with things that aren't in the
834 * bitmap, but we know that the intersection is no greater than what
835 * is already in cl, so let there be false positives that get sorted
836 * out after the synthetic start class succeeds, and the node is
837 * matched for real. */
839 /* The inversion of these two flags indicate that the resulting
840 * intersection doesn't have them */
841 if (and_with->flags & ANYOF_UNICODE_ALL) {
842 cl->flags &= ~ANYOF_UNICODE_ALL;
844 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
845 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
848 else { /* and'd node is not inverted */
849 U8 outside_bitmap_but_not_utf8; /* Temp variable */
851 if (! ANYOF_NONBITMAP(and_with)) {
853 /* Here 'and_with' doesn't match anything outside the bitmap
854 * (except possibly ANYOF_UNICODE_ALL), which means the
855 * intersection can't either, except for ANYOF_UNICODE_ALL, in
856 * which case we don't know what the intersection is, but it's no
857 * greater than what cl already has, so can just leave it alone,
858 * with possible false positives */
859 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
860 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
861 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
864 else if (! ANYOF_NONBITMAP(cl)) {
866 /* Here, 'and_with' does match something outside the bitmap, and cl
867 * doesn't have a list of things to match outside the bitmap. If
868 * cl can match all code points above 255, the intersection will
869 * be those above-255 code points that 'and_with' matches. If cl
870 * can't match all Unicode code points, it means that it can't
871 * match anything outside the bitmap (since the 'if' that got us
872 * into this block tested for that), so we leave the bitmap empty.
874 if (cl->flags & ANYOF_UNICODE_ALL) {
875 ARG_SET(cl, ARG(and_with));
877 /* and_with's ARG may match things that don't require UTF8.
878 * And now cl's will too, in spite of this being an 'and'. See
879 * the comments below about the kludge */
880 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
884 /* Here, both 'and_with' and cl match something outside the
885 * bitmap. Currently we do not do the intersection, so just match
886 * whatever cl had at the beginning. */
890 /* Take the intersection of the two sets of flags. However, the
891 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
892 * kludge around the fact that this flag is not treated like the others
893 * which are initialized in cl_anything(). The way the optimizer works
894 * is that the synthetic start class (SSC) is initialized to match
895 * anything, and then the first time a real node is encountered, its
896 * values are AND'd with the SSC's with the result being the values of
897 * the real node. However, there are paths through the optimizer where
898 * the AND never gets called, so those initialized bits are set
899 * inappropriately, which is not usually a big deal, as they just cause
900 * false positives in the SSC, which will just mean a probably
901 * imperceptible slow down in execution. However this bit has a
902 * higher false positive consequence in that it can cause utf8.pm,
903 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
904 * bigger slowdown and also causes significant extra memory to be used.
905 * In order to prevent this, the code now takes a different tack. The
906 * bit isn't set unless some part of the regular expression needs it,
907 * but once set it won't get cleared. This means that these extra
908 * modules won't get loaded unless there was some path through the
909 * pattern that would have required them anyway, and so any false
910 * positives that occur by not ANDing them out when they could be
911 * aren't as severe as they would be if we treated this bit like all
913 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
914 & ANYOF_NONBITMAP_NON_UTF8;
915 cl->flags &= and_with->flags;
916 cl->flags |= outside_bitmap_but_not_utf8;
920 /* 'OR' a given class with another one. Can create false positives. 'cl'
921 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
922 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
924 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
926 PERL_ARGS_ASSERT_CL_OR;
928 if (or_with->flags & ANYOF_INVERT) {
930 /* Here, the or'd node is to be inverted. This means we take the
931 * complement of everything not in the bitmap, but currently we don't
932 * know what that is, so give up and match anything */
933 if (ANYOF_NONBITMAP(or_with)) {
934 cl_anything(pRExC_state, cl);
937 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
938 * <= (B1 | !B2) | (CL1 | !CL2)
939 * which is wasteful if CL2 is small, but we ignore CL2:
940 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
941 * XXXX Can we handle case-fold? Unclear:
942 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
943 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
945 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
946 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
947 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
950 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
951 cl->bitmap[i] |= ~or_with->bitmap[i];
952 } /* XXXX: logic is complicated otherwise */
954 cl_anything(pRExC_state, cl);
957 /* And, we can just take the union of the flags that aren't affected
958 * by the inversion */
959 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
961 /* For the remaining flags:
962 ANYOF_UNICODE_ALL and inverted means to not match anything above
963 255, which means that the union with cl should just be
964 what cl has in it, so can ignore this flag
965 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
966 is 127-255 to match them, but then invert that, so the
967 union with cl should just be what cl has in it, so can
970 } else { /* 'or_with' is not inverted */
971 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
972 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
973 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
974 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
977 /* OR char bitmap and class bitmap separately */
978 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
979 cl->bitmap[i] |= or_with->bitmap[i];
980 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
981 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
982 cl->classflags[i] |= or_with->classflags[i];
983 cl->flags |= ANYOF_CLASS;
986 else { /* XXXX: logic is complicated, leave it along for a moment. */
987 cl_anything(pRExC_state, cl);
990 if (ANYOF_NONBITMAP(or_with)) {
992 /* Use the added node's outside-the-bit-map match if there isn't a
993 * conflict. If there is a conflict (both nodes match something
994 * outside the bitmap, but what they match outside is not the same
995 * pointer, and hence not easily compared until XXX we extend
996 * inversion lists this far), give up and allow the start class to
997 * match everything outside the bitmap. If that stuff is all above
998 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
999 if (! ANYOF_NONBITMAP(cl)) {
1000 ARG_SET(cl, ARG(or_with));
1002 else if (ARG(cl) != ARG(or_with)) {
1004 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1005 cl_anything(pRExC_state, cl);
1008 cl->flags |= ANYOF_UNICODE_ALL;
1013 /* Take the union */
1014 cl->flags |= or_with->flags;
1018 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1019 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1020 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1021 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1026 dump_trie(trie,widecharmap,revcharmap)
1027 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1028 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1030 These routines dump out a trie in a somewhat readable format.
1031 The _interim_ variants are used for debugging the interim
1032 tables that are used to generate the final compressed
1033 representation which is what dump_trie expects.
1035 Part of the reason for their existence is to provide a form
1036 of documentation as to how the different representations function.
1041 Dumps the final compressed table form of the trie to Perl_debug_log.
1042 Used for debugging make_trie().
1046 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1047 AV *revcharmap, U32 depth)
1050 SV *sv=sv_newmortal();
1051 int colwidth= widecharmap ? 6 : 4;
1053 GET_RE_DEBUG_FLAGS_DECL;
1055 PERL_ARGS_ASSERT_DUMP_TRIE;
1057 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1058 (int)depth * 2 + 2,"",
1059 "Match","Base","Ofs" );
1061 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1062 SV ** const tmp = av_fetch( revcharmap, state, 0);
1064 PerlIO_printf( Perl_debug_log, "%*s",
1066 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1067 PL_colors[0], PL_colors[1],
1068 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1069 PERL_PV_ESCAPE_FIRSTCHAR
1074 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1075 (int)depth * 2 + 2,"");
1077 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1078 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1079 PerlIO_printf( Perl_debug_log, "\n");
1081 for( state = 1 ; state < trie->statecount ; state++ ) {
1082 const U32 base = trie->states[ state ].trans.base;
1084 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1086 if ( trie->states[ state ].wordnum ) {
1087 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1089 PerlIO_printf( Perl_debug_log, "%6s", "" );
1092 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1097 while( ( base + ofs < trie->uniquecharcount ) ||
1098 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1099 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1102 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1104 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1105 if ( ( base + ofs >= trie->uniquecharcount ) &&
1106 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1107 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1109 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1111 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1113 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1117 PerlIO_printf( Perl_debug_log, "]");
1120 PerlIO_printf( Perl_debug_log, "\n" );
1122 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1123 for (word=1; word <= trie->wordcount; word++) {
1124 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1125 (int)word, (int)(trie->wordinfo[word].prev),
1126 (int)(trie->wordinfo[word].len));
1128 PerlIO_printf(Perl_debug_log, "\n" );
1131 Dumps a fully constructed but uncompressed trie in list form.
1132 List tries normally only are used for construction when the number of
1133 possible chars (trie->uniquecharcount) is very high.
1134 Used for debugging make_trie().
1137 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1138 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1142 SV *sv=sv_newmortal();
1143 int colwidth= widecharmap ? 6 : 4;
1144 GET_RE_DEBUG_FLAGS_DECL;
1146 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1148 /* print out the table precompression. */
1149 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1150 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1151 "------:-----+-----------------\n" );
1153 for( state=1 ; state < next_alloc ; state ++ ) {
1156 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1157 (int)depth * 2 + 2,"", (UV)state );
1158 if ( ! trie->states[ state ].wordnum ) {
1159 PerlIO_printf( Perl_debug_log, "%5s| ","");
1161 PerlIO_printf( Perl_debug_log, "W%4x| ",
1162 trie->states[ state ].wordnum
1165 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1166 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1168 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1170 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1171 PL_colors[0], PL_colors[1],
1172 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1173 PERL_PV_ESCAPE_FIRSTCHAR
1175 TRIE_LIST_ITEM(state,charid).forid,
1176 (UV)TRIE_LIST_ITEM(state,charid).newstate
1179 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1180 (int)((depth * 2) + 14), "");
1183 PerlIO_printf( Perl_debug_log, "\n");
1188 Dumps a fully constructed but uncompressed trie in table form.
1189 This is the normal DFA style state transition table, with a few
1190 twists to facilitate compression later.
1191 Used for debugging make_trie().
1194 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1195 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1200 SV *sv=sv_newmortal();
1201 int colwidth= widecharmap ? 6 : 4;
1202 GET_RE_DEBUG_FLAGS_DECL;
1204 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1207 print out the table precompression so that we can do a visual check
1208 that they are identical.
1211 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1213 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1214 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1216 PerlIO_printf( Perl_debug_log, "%*s",
1218 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1219 PL_colors[0], PL_colors[1],
1220 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1221 PERL_PV_ESCAPE_FIRSTCHAR
1227 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1229 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1230 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1233 PerlIO_printf( Perl_debug_log, "\n" );
1235 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1237 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1238 (int)depth * 2 + 2,"",
1239 (UV)TRIE_NODENUM( state ) );
1241 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1242 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1244 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1246 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1248 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1249 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1251 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1252 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1260 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1261 startbranch: the first branch in the whole branch sequence
1262 first : start branch of sequence of branch-exact nodes.
1263 May be the same as startbranch
1264 last : Thing following the last branch.
1265 May be the same as tail.
1266 tail : item following the branch sequence
1267 count : words in the sequence
1268 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1269 depth : indent depth
1271 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1273 A trie is an N'ary tree where the branches are determined by digital
1274 decomposition of the key. IE, at the root node you look up the 1st character and
1275 follow that branch repeat until you find the end of the branches. Nodes can be
1276 marked as "accepting" meaning they represent a complete word. Eg:
1280 would convert into the following structure. Numbers represent states, letters
1281 following numbers represent valid transitions on the letter from that state, if
1282 the number is in square brackets it represents an accepting state, otherwise it
1283 will be in parenthesis.
1285 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1289 (1) +-i->(6)-+-s->[7]
1291 +-s->(3)-+-h->(4)-+-e->[5]
1293 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1295 This shows that when matching against the string 'hers' we will begin at state 1
1296 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1297 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1298 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1299 single traverse. We store a mapping from accepting to state to which word was
1300 matched, and then when we have multiple possibilities we try to complete the
1301 rest of the regex in the order in which they occured in the alternation.
1303 The only prior NFA like behaviour that would be changed by the TRIE support is
1304 the silent ignoring of duplicate alternations which are of the form:
1306 / (DUPE|DUPE) X? (?{ ... }) Y /x
1308 Thus EVAL blocks following a trie may be called a different number of times with
1309 and without the optimisation. With the optimisations dupes will be silently
1310 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1311 the following demonstrates:
1313 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1315 which prints out 'word' three times, but
1317 'words'=~/(word|word|word)(?{ print $1 })S/
1319 which doesnt print it out at all. This is due to other optimisations kicking in.
1321 Example of what happens on a structural level:
1323 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1325 1: CURLYM[1] {1,32767}(18)
1336 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1337 and should turn into:
1339 1: CURLYM[1] {1,32767}(18)
1341 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1349 Cases where tail != last would be like /(?foo|bar)baz/:
1359 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1360 and would end up looking like:
1363 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1370 d = uvuni_to_utf8_flags(d, uv, 0);
1372 is the recommended Unicode-aware way of saying
1377 #define TRIE_STORE_REVCHAR(val) \
1380 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1381 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1382 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1383 SvCUR_set(zlopp, kapow - flrbbbbb); \
1386 av_push(revcharmap, zlopp); \
1388 char ooooff = (char)val; \
1389 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1393 #define TRIE_READ_CHAR STMT_START { \
1396 /* if it is UTF then it is either already folded, or does not need folding */ \
1397 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1399 else if (folder == PL_fold_latin1) { \
1400 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1401 if ( foldlen > 0 ) { \
1402 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1408 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1409 skiplen = UNISKIP(uvc); \
1410 foldlen -= skiplen; \
1411 scan = foldbuf + skiplen; \
1414 /* raw data, will be folded later if needed */ \
1422 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1423 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1424 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1425 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1427 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1428 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1429 TRIE_LIST_CUR( state )++; \
1432 #define TRIE_LIST_NEW(state) STMT_START { \
1433 Newxz( trie->states[ state ].trans.list, \
1434 4, reg_trie_trans_le ); \
1435 TRIE_LIST_CUR( state ) = 1; \
1436 TRIE_LIST_LEN( state ) = 4; \
1439 #define TRIE_HANDLE_WORD(state) STMT_START { \
1440 U16 dupe= trie->states[ state ].wordnum; \
1441 regnode * const noper_next = regnext( noper ); \
1444 /* store the word for dumping */ \
1446 if (OP(noper) != NOTHING) \
1447 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1449 tmp = newSVpvn_utf8( "", 0, UTF ); \
1450 av_push( trie_words, tmp ); \
1454 trie->wordinfo[curword].prev = 0; \
1455 trie->wordinfo[curword].len = wordlen; \
1456 trie->wordinfo[curword].accept = state; \
1458 if ( noper_next < tail ) { \
1460 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1461 trie->jump[curword] = (U16)(noper_next - convert); \
1463 jumper = noper_next; \
1465 nextbranch= regnext(cur); \
1469 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1470 /* chain, so that when the bits of chain are later */\
1471 /* linked together, the dups appear in the chain */\
1472 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1473 trie->wordinfo[dupe].prev = curword; \
1475 /* we haven't inserted this word yet. */ \
1476 trie->states[ state ].wordnum = curword; \
1481 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1482 ( ( base + charid >= ucharcount \
1483 && base + charid < ubound \
1484 && state == trie->trans[ base - ucharcount + charid ].check \
1485 && trie->trans[ base - ucharcount + charid ].next ) \
1486 ? trie->trans[ base - ucharcount + charid ].next \
1487 : ( state==1 ? special : 0 ) \
1491 #define MADE_JUMP_TRIE 2
1492 #define MADE_EXACT_TRIE 4
1495 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1498 /* first pass, loop through and scan words */
1499 reg_trie_data *trie;
1500 HV *widecharmap = NULL;
1501 AV *revcharmap = newAV();
1503 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1508 regnode *jumper = NULL;
1509 regnode *nextbranch = NULL;
1510 regnode *convert = NULL;
1511 U32 *prev_states; /* temp array mapping each state to previous one */
1512 /* we just use folder as a flag in utf8 */
1513 const U8 * folder = NULL;
1516 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1517 AV *trie_words = NULL;
1518 /* along with revcharmap, this only used during construction but both are
1519 * useful during debugging so we store them in the struct when debugging.
1522 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1523 STRLEN trie_charcount=0;
1525 SV *re_trie_maxbuff;
1526 GET_RE_DEBUG_FLAGS_DECL;
1528 PERL_ARGS_ASSERT_MAKE_TRIE;
1530 PERL_UNUSED_ARG(depth);
1537 case EXACTFU_TRICKYFOLD:
1538 case EXACTFU: folder = PL_fold_latin1; break;
1539 case EXACTF: folder = PL_fold; break;
1540 case EXACTFL: folder = PL_fold_locale; break;
1541 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1544 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1546 trie->startstate = 1;
1547 trie->wordcount = word_count;
1548 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1549 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1551 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1552 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1553 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1556 trie_words = newAV();
1559 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1560 if (!SvIOK(re_trie_maxbuff)) {
1561 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1563 DEBUG_TRIE_COMPILE_r({
1564 PerlIO_printf( Perl_debug_log,
1565 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1566 (int)depth * 2 + 2, "",
1567 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1568 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1572 /* Find the node we are going to overwrite */
1573 if ( first == startbranch && OP( last ) != BRANCH ) {
1574 /* whole branch chain */
1577 /* branch sub-chain */
1578 convert = NEXTOPER( first );
1581 /* -- First loop and Setup --
1583 We first traverse the branches and scan each word to determine if it
1584 contains widechars, and how many unique chars there are, this is
1585 important as we have to build a table with at least as many columns as we
1588 We use an array of integers to represent the character codes 0..255
1589 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1590 native representation of the character value as the key and IV's for the
1593 *TODO* If we keep track of how many times each character is used we can
1594 remap the columns so that the table compression later on is more
1595 efficient in terms of memory by ensuring the most common value is in the
1596 middle and the least common are on the outside. IMO this would be better
1597 than a most to least common mapping as theres a decent chance the most
1598 common letter will share a node with the least common, meaning the node
1599 will not be compressible. With a middle is most common approach the worst
1600 case is when we have the least common nodes twice.
1604 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1605 regnode *noper = NEXTOPER( cur );
1606 const U8 *uc = (U8*)STRING( noper );
1607 const U8 *e = uc + STR_LEN( noper );
1609 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1611 const U8 *scan = (U8*)NULL;
1612 U32 wordlen = 0; /* required init */
1614 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1616 if (OP(noper) == NOTHING) {
1617 regnode *noper_next= regnext(noper);
1618 if (noper_next != tail && OP(noper_next) == flags) {
1620 uc= (U8*)STRING(noper);
1621 e= uc + STR_LEN(noper);
1622 trie->minlen= STR_LEN(noper);
1629 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1630 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1631 regardless of encoding */
1632 if (OP( noper ) == EXACTFU_SS) {
1633 /* false positives are ok, so just set this */
1634 TRIE_BITMAP_SET(trie,0xDF);
1637 for ( ; uc < e ; uc += len ) {
1638 TRIE_CHARCOUNT(trie)++;
1643 U8 folded= folder[ (U8) uvc ];
1644 if ( !trie->charmap[ folded ] ) {
1645 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1646 TRIE_STORE_REVCHAR( folded );
1649 if ( !trie->charmap[ uvc ] ) {
1650 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1651 TRIE_STORE_REVCHAR( uvc );
1654 /* store the codepoint in the bitmap, and its folded
1656 TRIE_BITMAP_SET(trie, uvc);
1658 /* store the folded codepoint */
1659 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1662 /* store first byte of utf8 representation of
1663 variant codepoints */
1664 if (! UNI_IS_INVARIANT(uvc)) {
1665 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1668 set_bit = 0; /* We've done our bit :-) */
1673 widecharmap = newHV();
1675 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1678 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1680 if ( !SvTRUE( *svpp ) ) {
1681 sv_setiv( *svpp, ++trie->uniquecharcount );
1682 TRIE_STORE_REVCHAR(uvc);
1686 if( cur == first ) {
1687 trie->minlen = chars;
1688 trie->maxlen = chars;
1689 } else if (chars < trie->minlen) {
1690 trie->minlen = chars;
1691 } else if (chars > trie->maxlen) {
1692 trie->maxlen = chars;
1694 if (OP( noper ) == EXACTFU_SS) {
1695 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1696 if (trie->minlen > 1)
1699 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1700 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1701 * - We assume that any such sequence might match a 2 byte string */
1702 if (trie->minlen > 2 )
1706 } /* end first pass */
1707 DEBUG_TRIE_COMPILE_r(
1708 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1709 (int)depth * 2 + 2,"",
1710 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1711 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1712 (int)trie->minlen, (int)trie->maxlen )
1716 We now know what we are dealing with in terms of unique chars and
1717 string sizes so we can calculate how much memory a naive
1718 representation using a flat table will take. If it's over a reasonable
1719 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1720 conservative but potentially much slower representation using an array
1723 At the end we convert both representations into the same compressed
1724 form that will be used in regexec.c for matching with. The latter
1725 is a form that cannot be used to construct with but has memory
1726 properties similar to the list form and access properties similar
1727 to the table form making it both suitable for fast searches and
1728 small enough that its feasable to store for the duration of a program.
1730 See the comment in the code where the compressed table is produced
1731 inplace from the flat tabe representation for an explanation of how
1732 the compression works.
1737 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1740 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1742 Second Pass -- Array Of Lists Representation
1744 Each state will be represented by a list of charid:state records
1745 (reg_trie_trans_le) the first such element holds the CUR and LEN
1746 points of the allocated array. (See defines above).
1748 We build the initial structure using the lists, and then convert
1749 it into the compressed table form which allows faster lookups
1750 (but cant be modified once converted).
1753 STRLEN transcount = 1;
1755 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1756 "%*sCompiling trie using list compiler\n",
1757 (int)depth * 2 + 2, ""));
1759 trie->states = (reg_trie_state *)
1760 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1761 sizeof(reg_trie_state) );
1765 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1767 regnode *noper = NEXTOPER( cur );
1768 U8 *uc = (U8*)STRING( noper );
1769 const U8 *e = uc + STR_LEN( noper );
1770 U32 state = 1; /* required init */
1771 U16 charid = 0; /* sanity init */
1772 U8 *scan = (U8*)NULL; /* sanity init */
1773 STRLEN foldlen = 0; /* required init */
1774 U32 wordlen = 0; /* required init */
1775 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1778 if (OP(noper) == NOTHING) {
1779 regnode *noper_next= regnext(noper);
1780 if (noper_next != tail && OP(noper_next) == flags) {
1782 uc= (U8*)STRING(noper);
1783 e= uc + STR_LEN(noper);
1787 if (OP(noper) != NOTHING) {
1788 for ( ; uc < e ; uc += len ) {
1793 charid = trie->charmap[ uvc ];
1795 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1799 charid=(U16)SvIV( *svpp );
1802 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1809 if ( !trie->states[ state ].trans.list ) {
1810 TRIE_LIST_NEW( state );
1812 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1813 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1814 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1819 newstate = next_alloc++;
1820 prev_states[newstate] = state;
1821 TRIE_LIST_PUSH( state, charid, newstate );
1826 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1830 TRIE_HANDLE_WORD(state);
1832 } /* end second pass */
1834 /* next alloc is the NEXT state to be allocated */
1835 trie->statecount = next_alloc;
1836 trie->states = (reg_trie_state *)
1837 PerlMemShared_realloc( trie->states,
1839 * sizeof(reg_trie_state) );
1841 /* and now dump it out before we compress it */
1842 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1843 revcharmap, next_alloc,
1847 trie->trans = (reg_trie_trans *)
1848 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1855 for( state=1 ; state < next_alloc ; state ++ ) {
1859 DEBUG_TRIE_COMPILE_MORE_r(
1860 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1864 if (trie->states[state].trans.list) {
1865 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1869 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1870 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1871 if ( forid < minid ) {
1873 } else if ( forid > maxid ) {
1877 if ( transcount < tp + maxid - minid + 1) {
1879 trie->trans = (reg_trie_trans *)
1880 PerlMemShared_realloc( trie->trans,
1882 * sizeof(reg_trie_trans) );
1883 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1885 base = trie->uniquecharcount + tp - minid;
1886 if ( maxid == minid ) {
1888 for ( ; zp < tp ; zp++ ) {
1889 if ( ! trie->trans[ zp ].next ) {
1890 base = trie->uniquecharcount + zp - minid;
1891 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1892 trie->trans[ zp ].check = state;
1898 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1899 trie->trans[ tp ].check = state;
1904 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1905 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1906 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1907 trie->trans[ tid ].check = state;
1909 tp += ( maxid - minid + 1 );
1911 Safefree(trie->states[ state ].trans.list);
1914 DEBUG_TRIE_COMPILE_MORE_r(
1915 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1918 trie->states[ state ].trans.base=base;
1920 trie->lasttrans = tp + 1;
1924 Second Pass -- Flat Table Representation.
1926 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1927 We know that we will need Charcount+1 trans at most to store the data
1928 (one row per char at worst case) So we preallocate both structures
1929 assuming worst case.
1931 We then construct the trie using only the .next slots of the entry
1934 We use the .check field of the first entry of the node temporarily to
1935 make compression both faster and easier by keeping track of how many non
1936 zero fields are in the node.
1938 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1941 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1942 number representing the first entry of the node, and state as a
1943 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1944 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1945 are 2 entrys per node. eg:
1953 The table is internally in the right hand, idx form. However as we also
1954 have to deal with the states array which is indexed by nodenum we have to
1955 use TRIE_NODENUM() to convert.
1958 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1959 "%*sCompiling trie using table compiler\n",
1960 (int)depth * 2 + 2, ""));
1962 trie->trans = (reg_trie_trans *)
1963 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1964 * trie->uniquecharcount + 1,
1965 sizeof(reg_trie_trans) );
1966 trie->states = (reg_trie_state *)
1967 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1968 sizeof(reg_trie_state) );
1969 next_alloc = trie->uniquecharcount + 1;
1972 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1974 regnode *noper = NEXTOPER( cur );
1975 const U8 *uc = (U8*)STRING( noper );
1976 const U8 *e = uc + STR_LEN( noper );
1978 U32 state = 1; /* required init */
1980 U16 charid = 0; /* sanity init */
1981 U32 accept_state = 0; /* sanity init */
1982 U8 *scan = (U8*)NULL; /* sanity init */
1984 STRLEN foldlen = 0; /* required init */
1985 U32 wordlen = 0; /* required init */
1987 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1989 if (OP(noper) == NOTHING) {
1990 regnode *noper_next= regnext(noper);
1991 if (noper_next != tail && OP(noper_next) == flags) {
1993 uc= (U8*)STRING(noper);
1994 e= uc + STR_LEN(noper);
1998 if ( OP(noper) != NOTHING ) {
1999 for ( ; uc < e ; uc += len ) {
2004 charid = trie->charmap[ uvc ];
2006 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2007 charid = svpp ? (U16)SvIV(*svpp) : 0;
2011 if ( !trie->trans[ state + charid ].next ) {
2012 trie->trans[ state + charid ].next = next_alloc;
2013 trie->trans[ state ].check++;
2014 prev_states[TRIE_NODENUM(next_alloc)]
2015 = TRIE_NODENUM(state);
2016 next_alloc += trie->uniquecharcount;
2018 state = trie->trans[ state + charid ].next;
2020 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2022 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2025 accept_state = TRIE_NODENUM( state );
2026 TRIE_HANDLE_WORD(accept_state);
2028 } /* end second pass */
2030 /* and now dump it out before we compress it */
2031 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2033 next_alloc, depth+1));
2037 * Inplace compress the table.*
2039 For sparse data sets the table constructed by the trie algorithm will
2040 be mostly 0/FAIL transitions or to put it another way mostly empty.
2041 (Note that leaf nodes will not contain any transitions.)
2043 This algorithm compresses the tables by eliminating most such
2044 transitions, at the cost of a modest bit of extra work during lookup:
2046 - Each states[] entry contains a .base field which indicates the
2047 index in the state[] array wheres its transition data is stored.
2049 - If .base is 0 there are no valid transitions from that node.
2051 - If .base is nonzero then charid is added to it to find an entry in
2054 -If trans[states[state].base+charid].check!=state then the
2055 transition is taken to be a 0/Fail transition. Thus if there are fail
2056 transitions at the front of the node then the .base offset will point
2057 somewhere inside the previous nodes data (or maybe even into a node
2058 even earlier), but the .check field determines if the transition is
2062 The following process inplace converts the table to the compressed
2063 table: We first do not compress the root node 1,and mark all its
2064 .check pointers as 1 and set its .base pointer as 1 as well. This
2065 allows us to do a DFA construction from the compressed table later,
2066 and ensures that any .base pointers we calculate later are greater
2069 - We set 'pos' to indicate the first entry of the second node.
2071 - We then iterate over the columns of the node, finding the first and
2072 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2073 and set the .check pointers accordingly, and advance pos
2074 appropriately and repreat for the next node. Note that when we copy
2075 the next pointers we have to convert them from the original
2076 NODEIDX form to NODENUM form as the former is not valid post
2079 - If a node has no transitions used we mark its base as 0 and do not
2080 advance the pos pointer.
2082 - If a node only has one transition we use a second pointer into the
2083 structure to fill in allocated fail transitions from other states.
2084 This pointer is independent of the main pointer and scans forward
2085 looking for null transitions that are allocated to a state. When it
2086 finds one it writes the single transition into the "hole". If the
2087 pointer doesnt find one the single transition is appended as normal.
2089 - Once compressed we can Renew/realloc the structures to release the
2092 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2093 specifically Fig 3.47 and the associated pseudocode.
2097 const U32 laststate = TRIE_NODENUM( next_alloc );
2100 trie->statecount = laststate;
2102 for ( state = 1 ; state < laststate ; state++ ) {
2104 const U32 stateidx = TRIE_NODEIDX( state );
2105 const U32 o_used = trie->trans[ stateidx ].check;
2106 U32 used = trie->trans[ stateidx ].check;
2107 trie->trans[ stateidx ].check = 0;
2109 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2110 if ( flag || trie->trans[ stateidx + charid ].next ) {
2111 if ( trie->trans[ stateidx + charid ].next ) {
2113 for ( ; zp < pos ; zp++ ) {
2114 if ( ! trie->trans[ zp ].next ) {
2118 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2119 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2120 trie->trans[ zp ].check = state;
2121 if ( ++zp > pos ) pos = zp;
2128 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2130 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2131 trie->trans[ pos ].check = state;
2136 trie->lasttrans = pos + 1;
2137 trie->states = (reg_trie_state *)
2138 PerlMemShared_realloc( trie->states, laststate
2139 * sizeof(reg_trie_state) );
2140 DEBUG_TRIE_COMPILE_MORE_r(
2141 PerlIO_printf( Perl_debug_log,
2142 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2143 (int)depth * 2 + 2,"",
2144 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2147 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2150 } /* end table compress */
2152 DEBUG_TRIE_COMPILE_MORE_r(
2153 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2154 (int)depth * 2 + 2, "",
2155 (UV)trie->statecount,
2156 (UV)trie->lasttrans)
2158 /* resize the trans array to remove unused space */
2159 trie->trans = (reg_trie_trans *)
2160 PerlMemShared_realloc( trie->trans, trie->lasttrans
2161 * sizeof(reg_trie_trans) );
2163 { /* Modify the program and insert the new TRIE node */
2164 U8 nodetype =(U8)(flags & 0xFF);
2168 regnode *optimize = NULL;
2169 #ifdef RE_TRACK_PATTERN_OFFSETS
2172 U32 mjd_nodelen = 0;
2173 #endif /* RE_TRACK_PATTERN_OFFSETS */
2174 #endif /* DEBUGGING */
2176 This means we convert either the first branch or the first Exact,
2177 depending on whether the thing following (in 'last') is a branch
2178 or not and whther first is the startbranch (ie is it a sub part of
2179 the alternation or is it the whole thing.)
2180 Assuming its a sub part we convert the EXACT otherwise we convert
2181 the whole branch sequence, including the first.
2183 /* Find the node we are going to overwrite */
2184 if ( first != startbranch || OP( last ) == BRANCH ) {
2185 /* branch sub-chain */
2186 NEXT_OFF( first ) = (U16)(last - first);
2187 #ifdef RE_TRACK_PATTERN_OFFSETS
2189 mjd_offset= Node_Offset((convert));
2190 mjd_nodelen= Node_Length((convert));
2193 /* whole branch chain */
2195 #ifdef RE_TRACK_PATTERN_OFFSETS
2198 const regnode *nop = NEXTOPER( convert );
2199 mjd_offset= Node_Offset((nop));
2200 mjd_nodelen= Node_Length((nop));
2204 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2205 (int)depth * 2 + 2, "",
2206 (UV)mjd_offset, (UV)mjd_nodelen)
2209 /* But first we check to see if there is a common prefix we can
2210 split out as an EXACT and put in front of the TRIE node. */
2211 trie->startstate= 1;
2212 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2214 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2218 const U32 base = trie->states[ state ].trans.base;
2220 if ( trie->states[state].wordnum )
2223 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2224 if ( ( base + ofs >= trie->uniquecharcount ) &&
2225 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2226 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2228 if ( ++count > 1 ) {
2229 SV **tmp = av_fetch( revcharmap, ofs, 0);
2230 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2231 if ( state == 1 ) break;
2233 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2235 PerlIO_printf(Perl_debug_log,
2236 "%*sNew Start State=%"UVuf" Class: [",
2237 (int)depth * 2 + 2, "",
2240 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2241 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2243 TRIE_BITMAP_SET(trie,*ch);
2245 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2247 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2251 TRIE_BITMAP_SET(trie,*ch);
2253 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2254 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2260 SV **tmp = av_fetch( revcharmap, idx, 0);
2262 char *ch = SvPV( *tmp, len );
2264 SV *sv=sv_newmortal();
2265 PerlIO_printf( Perl_debug_log,
2266 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2267 (int)depth * 2 + 2, "",
2269 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2270 PL_colors[0], PL_colors[1],
2271 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2272 PERL_PV_ESCAPE_FIRSTCHAR
2277 OP( convert ) = nodetype;
2278 str=STRING(convert);
2281 STR_LEN(convert) += len;
2287 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2292 trie->prefixlen = (state-1);
2294 regnode *n = convert+NODE_SZ_STR(convert);
2295 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2296 trie->startstate = state;
2297 trie->minlen -= (state - 1);
2298 trie->maxlen -= (state - 1);
2300 /* At least the UNICOS C compiler choked on this
2301 * being argument to DEBUG_r(), so let's just have
2304 #ifdef PERL_EXT_RE_BUILD
2310 regnode *fix = convert;
2311 U32 word = trie->wordcount;
2313 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2314 while( ++fix < n ) {
2315 Set_Node_Offset_Length(fix, 0, 0);
2318 SV ** const tmp = av_fetch( trie_words, word, 0 );
2320 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2321 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2323 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2331 NEXT_OFF(convert) = (U16)(tail - convert);
2332 DEBUG_r(optimize= n);
2338 if ( trie->maxlen ) {
2339 NEXT_OFF( convert ) = (U16)(tail - convert);
2340 ARG_SET( convert, data_slot );
2341 /* Store the offset to the first unabsorbed branch in
2342 jump[0], which is otherwise unused by the jump logic.
2343 We use this when dumping a trie and during optimisation. */
2345 trie->jump[0] = (U16)(nextbranch - convert);
2347 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2348 * and there is a bitmap
2349 * and the first "jump target" node we found leaves enough room
2350 * then convert the TRIE node into a TRIEC node, with the bitmap
2351 * embedded inline in the opcode - this is hypothetically faster.
2353 if ( !trie->states[trie->startstate].wordnum
2355 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2357 OP( convert ) = TRIEC;
2358 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2359 PerlMemShared_free(trie->bitmap);
2362 OP( convert ) = TRIE;
2364 /* store the type in the flags */
2365 convert->flags = nodetype;
2369 + regarglen[ OP( convert ) ];
2371 /* XXX We really should free up the resource in trie now,
2372 as we won't use them - (which resources?) dmq */
2374 /* needed for dumping*/
2375 DEBUG_r(if (optimize) {
2376 regnode *opt = convert;
2378 while ( ++opt < optimize) {
2379 Set_Node_Offset_Length(opt,0,0);
2382 Try to clean up some of the debris left after the
2385 while( optimize < jumper ) {
2386 mjd_nodelen += Node_Length((optimize));
2387 OP( optimize ) = OPTIMIZED;
2388 Set_Node_Offset_Length(optimize,0,0);
2391 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2393 } /* end node insert */
2395 /* Finish populating the prev field of the wordinfo array. Walk back
2396 * from each accept state until we find another accept state, and if
2397 * so, point the first word's .prev field at the second word. If the
2398 * second already has a .prev field set, stop now. This will be the
2399 * case either if we've already processed that word's accept state,
2400 * or that state had multiple words, and the overspill words were
2401 * already linked up earlier.
2408 for (word=1; word <= trie->wordcount; word++) {
2410 if (trie->wordinfo[word].prev)
2412 state = trie->wordinfo[word].accept;
2414 state = prev_states[state];
2417 prev = trie->states[state].wordnum;
2421 trie->wordinfo[word].prev = prev;
2423 Safefree(prev_states);
2427 /* and now dump out the compressed format */
2428 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2430 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2432 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2433 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2435 SvREFCNT_dec(revcharmap);
2439 : trie->startstate>1
2445 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2447 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2449 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2450 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2453 We find the fail state for each state in the trie, this state is the longest proper
2454 suffix of the current state's 'word' that is also a proper prefix of another word in our
2455 trie. State 1 represents the word '' and is thus the default fail state. This allows
2456 the DFA not to have to restart after its tried and failed a word at a given point, it
2457 simply continues as though it had been matching the other word in the first place.
2459 'abcdgu'=~/abcdefg|cdgu/
2460 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2461 fail, which would bring us to the state representing 'd' in the second word where we would
2462 try 'g' and succeed, proceeding to match 'cdgu'.
2464 /* add a fail transition */
2465 const U32 trie_offset = ARG(source);
2466 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2468 const U32 ucharcount = trie->uniquecharcount;
2469 const U32 numstates = trie->statecount;
2470 const U32 ubound = trie->lasttrans + ucharcount;
2474 U32 base = trie->states[ 1 ].trans.base;
2477 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2478 GET_RE_DEBUG_FLAGS_DECL;
2480 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2482 PERL_UNUSED_ARG(depth);
2486 ARG_SET( stclass, data_slot );
2487 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2488 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2489 aho->trie=trie_offset;
2490 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2491 Copy( trie->states, aho->states, numstates, reg_trie_state );
2492 Newxz( q, numstates, U32);
2493 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2496 /* initialize fail[0..1] to be 1 so that we always have
2497 a valid final fail state */
2498 fail[ 0 ] = fail[ 1 ] = 1;
2500 for ( charid = 0; charid < ucharcount ; charid++ ) {
2501 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2503 q[ q_write ] = newstate;
2504 /* set to point at the root */
2505 fail[ q[ q_write++ ] ]=1;
2508 while ( q_read < q_write) {
2509 const U32 cur = q[ q_read++ % numstates ];
2510 base = trie->states[ cur ].trans.base;
2512 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2513 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2515 U32 fail_state = cur;
2518 fail_state = fail[ fail_state ];
2519 fail_base = aho->states[ fail_state ].trans.base;
2520 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2522 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2523 fail[ ch_state ] = fail_state;
2524 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2526 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2528 q[ q_write++ % numstates] = ch_state;
2532 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2533 when we fail in state 1, this allows us to use the
2534 charclass scan to find a valid start char. This is based on the principle
2535 that theres a good chance the string being searched contains lots of stuff
2536 that cant be a start char.
2538 fail[ 0 ] = fail[ 1 ] = 0;
2539 DEBUG_TRIE_COMPILE_r({
2540 PerlIO_printf(Perl_debug_log,
2541 "%*sStclass Failtable (%"UVuf" states): 0",
2542 (int)(depth * 2), "", (UV)numstates
2544 for( q_read=1; q_read<numstates; q_read++ ) {
2545 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2547 PerlIO_printf(Perl_debug_log, "\n");
2550 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2555 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2556 * These need to be revisited when a newer toolchain becomes available.
2558 #if defined(__sparc64__) && defined(__GNUC__)
2559 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2560 # undef SPARC64_GCC_WORKAROUND
2561 # define SPARC64_GCC_WORKAROUND 1
2565 #define DEBUG_PEEP(str,scan,depth) \
2566 DEBUG_OPTIMISE_r({if (scan){ \
2567 SV * const mysv=sv_newmortal(); \
2568 regnode *Next = regnext(scan); \
2569 regprop(RExC_rx, mysv, scan); \
2570 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2571 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2572 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2576 /* The below joins as many adjacent EXACTish nodes as possible into a single
2577 * one, and looks for problematic sequences of characters whose folds vs.
2578 * non-folds have sufficiently different lengths, that the optimizer would be
2579 * fooled into rejecting legitimate matches of them, and the trie construction
2580 * code can't cope with them. The joining is only done if:
2581 * 1) there is room in the current conglomerated node to entirely contain the
2583 * 2) they are the exact same node type
2585 * The adjacent nodes actually may be separated by NOTHING kind nodes, and
2586 * these get optimized out
2588 * If there are problematic code sequences, *min_subtract is set to the delta
2589 * that the minimum size of the node can be less than its actual size. And,
2590 * the node type of the result is changed to reflect that it contains these
2593 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2594 * and contains LATIN SMALL LETTER SHARP S
2596 * This is as good a place as any to discuss the design of handling these
2597 * problematic sequences. It's been wrong in Perl for a very long time. There
2598 * are three code points in Unicode whose folded lengths differ so much from
2599 * the un-folded lengths that it causes problems for the optimizer and trie
2600 * construction. Why only these are problematic, and not others where lengths
2601 * also differ is something I (khw) do not understand. New versions of Unicode
2602 * might add more such code points. Hopefully the logic in fold_grind.t that
2603 * figures out what to test (in part by verifying that each size-combination
2604 * gets tested) will catch any that do come along, so they can be added to the
2605 * special handling below. The chances of new ones are actually rather small,
2606 * as most, if not all, of the world's scripts that have casefolding have
2607 * already been encoded by Unicode. Also, a number of Unicode's decisions were
2608 * made to allow compatibility with pre-existing standards, and almost all of
2609 * those have already been dealt with. These would otherwise be the most
2610 * likely candidates for generating further tricky sequences. In other words,
2611 * Unicode by itself is unlikely to add new ones unless it is for compatibility
2612 * with pre-existing standards, and there aren't many of those left.
2614 * The previous designs for dealing with these involved assigning a special
2615 * node for them. This approach doesn't work, as evidenced by this example:
2616 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2617 * Both these fold to "sss", but if the pattern is parsed to create a node of
2618 * that would match just the \xDF, it won't be able to handle the case where a
2619 * successful match would have to cross the node's boundary. The new approach
2620 * that hopefully generally solves the problem generates an EXACTFU_SS node
2623 * There are a number of components to the approach (a lot of work for just
2624 * three code points!):
2625 * 1) This routine examines each EXACTFish node that could contain the
2626 * problematic sequences. It returns in *min_subtract how much to
2627 * subtract from the the actual length of the string to get a real minimum
2628 * for one that could match it. This number is usually 0 except for the
2629 * problematic sequences. This delta is used by the caller to adjust the
2630 * min length of the match, and the delta between min and max, so that the
2631 * optimizer doesn't reject these possibilities based on size constraints.
2632 * 2) These sequences are not currently correctly handled by the trie code
2633 * either, so it changes the joined node type to ops that are not handled
2634 * by trie's, those new ops being EXACTFU_SS and EXACTFU_TRICKYFOLD.
2635 * 3) This is sufficient for the two Greek sequences (described below), but
2636 * the one involving the Sharp s (\xDF) needs more. The node type
2637 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2638 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2639 * case where there is a possible fold length change. That means that a
2640 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2641 * itself with length changes, and so can be processed faster. regexec.c
2642 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2643 * is pre-folded by regcomp.c. This saves effort in regex matching.
2644 * However, probably mostly for historical reasons, the pre-folding isn't
2645 * done for non-UTF8 patterns (and it can't be for EXACTF and EXACTFL
2646 * nodes, as what they fold to isn't known until runtime.) The fold
2647 * possibilities for the non-UTF8 patterns are quite simple, except for
2648 * the sharp s. All the ones that don't involve a UTF-8 target string
2649 * are members of a fold-pair, and arrays are set up for all of them
2650 * that quickly find the other member of the pair. It might actually
2651 * be faster to pre-fold these, but it isn't currently done, except for
2652 * the sharp s. Code elsewhere in this file makes sure that it gets
2653 * folded to 'ss', even if the pattern isn't UTF-8. This avoids the
2654 * issues described in the next item.
2655 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2656 * 'ss' or not is not knowable at compile time. It will match iff the
2657 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2658 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2659 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2660 * described in item 3). An assumption that the optimizer part of
2661 * regexec.c (probably unwittingly) makes is that a character in the
2662 * pattern corresponds to at most a single character in the target string.
2663 * (And I do mean character, and not byte here, unlike other parts of the
2664 * documentation that have never been updated to account for multibyte
2665 * Unicode.) This assumption is wrong only in this case, as all other
2666 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2667 * virtue of having this file pre-fold UTF-8 patterns. I'm
2668 * reluctant to try to change this assumption, so instead the code punts.
2669 * This routine examines EXACTF nodes for the sharp s, and returns a
2670 * boolean indicating whether or not the node is an EXACTF node that
2671 * contains a sharp s. When it is true, the caller sets a flag that later
2672 * causes the optimizer in this file to not set values for the floating
2673 * and fixed string lengths, and thus avoids the optimizer code in
2674 * regexec.c that makes the invalid assumption. Thus, there is no
2675 * optimization based on string lengths for EXACTF nodes that contain the
2676 * sharp s. This only happens for /id rules (which means the pattern
2680 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2681 if (PL_regkind[OP(scan)] == EXACT) \
2682 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2685 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) {
2686 /* Merge several consecutive EXACTish nodes into one. */
2687 regnode *n = regnext(scan);
2689 regnode *next = scan + NODE_SZ_STR(scan);
2693 regnode *stop = scan;
2694 GET_RE_DEBUG_FLAGS_DECL;
2696 PERL_UNUSED_ARG(depth);
2699 PERL_ARGS_ASSERT_JOIN_EXACT;
2700 #ifndef EXPERIMENTAL_INPLACESCAN
2701 PERL_UNUSED_ARG(flags);
2702 PERL_UNUSED_ARG(val);
2704 DEBUG_PEEP("join",scan,depth);
2706 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2707 * EXACT ones that are mergeable to the current one. */
2709 && (PL_regkind[OP(n)] == NOTHING
2710 || (stringok && OP(n) == OP(scan)))
2712 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2715 if (OP(n) == TAIL || n > next)
2717 if (PL_regkind[OP(n)] == NOTHING) {
2718 DEBUG_PEEP("skip:",n,depth);
2719 NEXT_OFF(scan) += NEXT_OFF(n);
2720 next = n + NODE_STEP_REGNODE;
2727 else if (stringok) {
2728 const unsigned int oldl = STR_LEN(scan);
2729 regnode * const nnext = regnext(n);
2731 if (oldl + STR_LEN(n) > U8_MAX)
2734 DEBUG_PEEP("merg",n,depth);
2737 NEXT_OFF(scan) += NEXT_OFF(n);
2738 STR_LEN(scan) += STR_LEN(n);
2739 next = n + NODE_SZ_STR(n);
2740 /* Now we can overwrite *n : */
2741 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2749 #ifdef EXPERIMENTAL_INPLACESCAN
2750 if (flags && !NEXT_OFF(n)) {
2751 DEBUG_PEEP("atch", val, depth);
2752 if (reg_off_by_arg[OP(n)]) {
2753 ARG_SET(n, val - n);
2756 NEXT_OFF(n) = val - n;
2764 *has_exactf_sharp_s = FALSE;
2766 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2767 * can now analyze for sequences of problematic code points. (Prior to
2768 * this final joining, sequences could have been split over boundaries, and
2769 * hence missed). The sequences only happen in folding, hence for any
2770 * non-EXACT EXACTish node */
2771 if (OP(scan) != EXACT) {
2773 U8 * s0 = (U8*) STRING(scan);
2774 U8 * const s_end = s0 + STR_LEN(scan);
2776 /* The below is perhaps overboard, but this allows us to save a test
2777 * each time through the loop at the expense of a mask. This is
2778 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2779 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2780 * This uses an exclusive 'or' to find that bit and then inverts it to
2781 * form a mask, with just a single 0, in the bit position where 'S' and
2783 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2784 const U8 s_masked = 's' & S_or_s_mask;
2786 /* One pass is made over the node's string looking for all the
2787 * possibilities. to avoid some tests in the loop, there are two main
2788 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2792 /* There are two problematic Greek code points in Unicode
2795 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2796 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2802 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2803 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2805 * This means that in case-insensitive matching (or "loose
2806 * matching", as Unicode calls it), an EXACTF of length six (the
2807 * UTF-8 encoded byte length of the above casefolded versions) can
2808 * match a target string of length two (the byte length of UTF-8
2809 * encoded U+0390 or U+03B0). This would rather mess up the
2810 * minimum length computation. (there are other code points that
2811 * also fold to these two sequences, but the delta is smaller)
2813 * If these sequences are found, the minimum length is decreased by
2814 * four (six minus two).
2816 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2817 * LETTER SHARP S. We decrease the min length by 1 for each
2818 * occurrence of 'ss' found */
2820 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2821 # define U390_first_byte 0xb4
2822 const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
2823 # define U3B0_first_byte 0xb5
2824 const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
2826 # define U390_first_byte 0xce
2827 const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
2828 # define U3B0_first_byte 0xcf
2829 const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
2831 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2832 yields a net of 0 */
2833 /* Examine the string for one of the problematic sequences */
2835 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2836 * sequence we are looking for is 2 */
2840 /* Look for the first byte in each problematic sequence */
2842 /* We don't have to worry about other things that fold to
2843 * 's' (such as the long s, U+017F), as all above-latin1
2844 * code points have been pre-folded */
2848 /* Current character is an 's' or 'S'. If next one is
2849 * as well, we have the dreaded sequence */
2850 if (((*(s+1) & S_or_s_mask) == s_masked)
2851 /* These two node types don't have special handling
2853 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2856 OP(scan) = EXACTFU_SS;
2857 s++; /* No need to look at this character again */
2861 case U390_first_byte:
2862 if (s_end - s >= len
2864 /* The 1's are because are skipping comparing the
2866 && memEQ(s + 1, U390_tail, len - 1))
2868 goto greek_sequence;
2872 case U3B0_first_byte:
2873 if (! (s_end - s >= len
2874 && memEQ(s + 1, U3B0_tail, len - 1)))
2881 /* This can't currently be handled by trie's, so change
2882 * the node type to indicate this. If EXACTFA and
2883 * EXACTFL were ever to be handled by trie's, this
2884 * would have to be changed. If this node has already
2885 * been changed to EXACTFU_SS in this loop, leave it as
2886 * is. (I (khw) think it doesn't matter in regexec.c
2887 * for UTF patterns, but no need to change it */
2888 if (OP(scan) == EXACTFU) {
2889 OP(scan) = EXACTFU_TRICKYFOLD;
2891 s += 6; /* We already know what this sequence is. Skip
2897 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2899 /* Here, the pattern is not UTF-8. We need to look only for the
2900 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2901 * in the final position. Otherwise we can stop looking 1 byte
2902 * earlier because have to find both the first and second 's' */
2903 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2905 for (s = s0; s < upper; s++) {
2910 && ((*(s+1) & S_or_s_mask) == s_masked))
2914 /* EXACTF nodes need to know that the minimum
2915 * length changed so that a sharp s in the string
2916 * can match this ss in the pattern, but they
2917 * remain EXACTF nodes, as they are not trie'able,
2918 * so don't have to invent a new node type to
2919 * exclude them from the trie code */
2920 if (OP(scan) != EXACTF) {
2921 OP(scan) = EXACTFU_SS;
2926 case LATIN_SMALL_LETTER_SHARP_S:
2927 if (OP(scan) == EXACTF) {
2928 *has_exactf_sharp_s = TRUE;
2937 /* Allow dumping but overwriting the collection of skipped
2938 * ops and/or strings with fake optimized ops */
2939 n = scan + NODE_SZ_STR(scan);
2947 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2951 /* REx optimizer. Converts nodes into quicker variants "in place".
2952 Finds fixed substrings. */
2954 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2955 to the position after last scanned or to NULL. */
2957 #define INIT_AND_WITHP \
2958 assert(!and_withp); \
2959 Newx(and_withp,1,struct regnode_charclass_class); \
2960 SAVEFREEPV(and_withp)
2962 /* this is a chain of data about sub patterns we are processing that
2963 need to be handled separately/specially in study_chunk. Its so
2964 we can simulate recursion without losing state. */
2966 typedef struct scan_frame {
2967 regnode *last; /* last node to process in this frame */
2968 regnode *next; /* next node to process when last is reached */
2969 struct scan_frame *prev; /*previous frame*/
2970 I32 stop; /* what stopparen do we use */
2974 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2976 #define CASE_SYNST_FNC(nAmE) \
2978 if (flags & SCF_DO_STCLASS_AND) { \
2979 for (value = 0; value < 256; value++) \
2980 if (!is_ ## nAmE ## _cp(value)) \
2981 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2984 for (value = 0; value < 256; value++) \
2985 if (is_ ## nAmE ## _cp(value)) \
2986 ANYOF_BITMAP_SET(data->start_class, value); \
2990 if (flags & SCF_DO_STCLASS_AND) { \
2991 for (value = 0; value < 256; value++) \
2992 if (is_ ## nAmE ## _cp(value)) \
2993 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2996 for (value = 0; value < 256; value++) \
2997 if (!is_ ## nAmE ## _cp(value)) \
2998 ANYOF_BITMAP_SET(data->start_class, value); \
3005 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3006 I32 *minlenp, I32 *deltap,
3011 struct regnode_charclass_class *and_withp,
3012 U32 flags, U32 depth)
3013 /* scanp: Start here (read-write). */
3014 /* deltap: Write maxlen-minlen here. */
3015 /* last: Stop before this one. */
3016 /* data: string data about the pattern */
3017 /* stopparen: treat close N as END */
3018 /* recursed: which subroutines have we recursed into */
3019 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3022 I32 min = 0, pars = 0, code;
3023 regnode *scan = *scanp, *next;
3025 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3026 int is_inf_internal = 0; /* The studied chunk is infinite */
3027 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3028 scan_data_t data_fake;
3029 SV *re_trie_maxbuff = NULL;
3030 regnode *first_non_open = scan;
3031 I32 stopmin = I32_MAX;
3032 scan_frame *frame = NULL;
3033 GET_RE_DEBUG_FLAGS_DECL;
3035 PERL_ARGS_ASSERT_STUDY_CHUNK;
3038 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3042 while (first_non_open && OP(first_non_open) == OPEN)
3043 first_non_open=regnext(first_non_open);
3048 while ( scan && OP(scan) != END && scan < last ){
3049 UV min_subtract = 0; /* How much to subtract from the minimum node
3050 length to get a real minimum (because the
3051 folded version may be shorter) */
3052 bool has_exactf_sharp_s = FALSE;
3053 /* Peephole optimizer: */
3054 DEBUG_STUDYDATA("Peep:", data,depth);
3055 DEBUG_PEEP("Peep",scan,depth);
3057 /* Its not clear to khw or hv why this is done here, and not in the
3058 * clauses that deal with EXACT nodes. khw's guess is that it's
3059 * because of a previous design */
3060 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3062 /* Follow the next-chain of the current node and optimize
3063 away all the NOTHINGs from it. */
3064 if (OP(scan) != CURLYX) {
3065 const int max = (reg_off_by_arg[OP(scan)]
3067 /* I32 may be smaller than U16 on CRAYs! */
3068 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3069 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3073 /* Skip NOTHING and LONGJMP. */
3074 while ((n = regnext(n))
3075 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3076 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3077 && off + noff < max)
3079 if (reg_off_by_arg[OP(scan)])
3082 NEXT_OFF(scan) = off;
3087 /* The principal pseudo-switch. Cannot be a switch, since we
3088 look into several different things. */
3089 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3090 || OP(scan) == IFTHEN) {
3091 next = regnext(scan);
3093 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3095 if (OP(next) == code || code == IFTHEN) {
3096 /* NOTE - There is similar code to this block below for handling
3097 TRIE nodes on a re-study. If you change stuff here check there
3099 I32 max1 = 0, min1 = I32_MAX, num = 0;
3100 struct regnode_charclass_class accum;
3101 regnode * const startbranch=scan;
3103 if (flags & SCF_DO_SUBSTR)
3104 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3105 if (flags & SCF_DO_STCLASS)
3106 cl_init_zero(pRExC_state, &accum);
3108 while (OP(scan) == code) {
3109 I32 deltanext, minnext, f = 0, fake;
3110 struct regnode_charclass_class this_class;
3113 data_fake.flags = 0;
3115 data_fake.whilem_c = data->whilem_c;
3116 data_fake.last_closep = data->last_closep;
3119 data_fake.last_closep = &fake;
3121 data_fake.pos_delta = delta;
3122 next = regnext(scan);
3123 scan = NEXTOPER(scan);
3125 scan = NEXTOPER(scan);
3126 if (flags & SCF_DO_STCLASS) {
3127 cl_init(pRExC_state, &this_class);
3128 data_fake.start_class = &this_class;
3129 f = SCF_DO_STCLASS_AND;
3131 if (flags & SCF_WHILEM_VISITED_POS)
3132 f |= SCF_WHILEM_VISITED_POS;
3134 /* we suppose the run is continuous, last=next...*/
3135 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3137 stopparen, recursed, NULL, f,depth+1);
3140 if (max1 < minnext + deltanext)
3141 max1 = minnext + deltanext;
3142 if (deltanext == I32_MAX)
3143 is_inf = is_inf_internal = 1;
3145 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3147 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3148 if ( stopmin > minnext)
3149 stopmin = min + min1;
3150 flags &= ~SCF_DO_SUBSTR;
3152 data->flags |= SCF_SEEN_ACCEPT;
3155 if (data_fake.flags & SF_HAS_EVAL)
3156 data->flags |= SF_HAS_EVAL;
3157 data->whilem_c = data_fake.whilem_c;
3159 if (flags & SCF_DO_STCLASS)
3160 cl_or(pRExC_state, &accum, &this_class);
3162 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3164 if (flags & SCF_DO_SUBSTR) {
3165 data->pos_min += min1;
3166 data->pos_delta += max1 - min1;
3167 if (max1 != min1 || is_inf)
3168 data->longest = &(data->longest_float);
3171 delta += max1 - min1;
3172 if (flags & SCF_DO_STCLASS_OR) {
3173 cl_or(pRExC_state, data->start_class, &accum);
3175 cl_and(data->start_class, and_withp);
3176 flags &= ~SCF_DO_STCLASS;
3179 else if (flags & SCF_DO_STCLASS_AND) {
3181 cl_and(data->start_class, &accum);
3182 flags &= ~SCF_DO_STCLASS;
3185 /* Switch to OR mode: cache the old value of
3186 * data->start_class */
3188 StructCopy(data->start_class, and_withp,
3189 struct regnode_charclass_class);
3190 flags &= ~SCF_DO_STCLASS_AND;
3191 StructCopy(&accum, data->start_class,
3192 struct regnode_charclass_class);
3193 flags |= SCF_DO_STCLASS_OR;
3194 data->start_class->flags |= ANYOF_EOS;
3198 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3201 Assuming this was/is a branch we are dealing with: 'scan' now
3202 points at the item that follows the branch sequence, whatever
3203 it is. We now start at the beginning of the sequence and look
3210 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3212 If we can find such a subsequence we need to turn the first
3213 element into a trie and then add the subsequent branch exact
3214 strings to the trie.
3218 1. patterns where the whole set of branches can be converted.
3220 2. patterns where only a subset can be converted.
3222 In case 1 we can replace the whole set with a single regop
3223 for the trie. In case 2 we need to keep the start and end
3226 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3227 becomes BRANCH TRIE; BRANCH X;
3229 There is an additional case, that being where there is a
3230 common prefix, which gets split out into an EXACT like node
3231 preceding the TRIE node.
3233 If x(1..n)==tail then we can do a simple trie, if not we make
3234 a "jump" trie, such that when we match the appropriate word
3235 we "jump" to the appropriate tail node. Essentially we turn
3236 a nested if into a case structure of sorts.
3241 if (!re_trie_maxbuff) {
3242 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3243 if (!SvIOK(re_trie_maxbuff))
3244 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3246 if ( SvIV(re_trie_maxbuff)>=0 ) {
3248 regnode *first = (regnode *)NULL;
3249 regnode *last = (regnode *)NULL;
3250 regnode *tail = scan;
3255 SV * const mysv = sv_newmortal(); /* for dumping */
3257 /* var tail is used because there may be a TAIL
3258 regop in the way. Ie, the exacts will point to the
3259 thing following the TAIL, but the last branch will
3260 point at the TAIL. So we advance tail. If we
3261 have nested (?:) we may have to move through several
3265 while ( OP( tail ) == TAIL ) {
3266 /* this is the TAIL generated by (?:) */
3267 tail = regnext( tail );
3271 DEBUG_TRIE_COMPILE_r({
3272 regprop(RExC_rx, mysv, tail );
3273 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3274 (int)depth * 2 + 2, "",
3275 "Looking for TRIE'able sequences. Tail node is: ",
3276 SvPV_nolen_const( mysv )
3282 Step through the branches
3283 cur represents each branch,
3284 noper is the first thing to be matched as part of that branch
3285 noper_next is the regnext() of that node.
3287 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3288 via a "jump trie" but we also support building with NOJUMPTRIE,
3289 which restricts the trie logic to structures like /FOO|BAR/.
3291 If noper is a trieable nodetype then the branch is a possible optimization
3292 target. If we are building under NOJUMPTRIE then we require that noper_next
3293 is the same as scan (our current position in the regex program).
3295 Once we have two or more consecutive such branches we can create a
3296 trie of the EXACT's contents and stitch it in place into the program.
3298 If the sequence represents all of the branches in the alternation we
3299 replace the entire thing with a single TRIE node.
3301 Otherwise when it is a subsequence we need to stitch it in place and
3302 replace only the relevant branches. This means the first branch has
3303 to remain as it is used by the alternation logic, and its next pointer,
3304 and needs to be repointed at the item on the branch chain following
3305 the last branch we have optimized away.
3307 This could be either a BRANCH, in which case the subsequence is internal,
3308 or it could be the item following the branch sequence in which case the
3309 subsequence is at the end (which does not necessarily mean the first node
3310 is the start of the alternation).
3312 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3315 ----------------+-----------
3319 EXACTFU_SS | EXACTFU
3320 EXACTFU_TRICKYFOLD | EXACTFU
3325 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3326 ( EXACT == (X) ) ? EXACT : \
3327 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3330 /* dont use tail as the end marker for this traverse */
3331 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3332 regnode * const noper = NEXTOPER( cur );
3333 U8 noper_type = OP( noper );
3334 U8 noper_trietype = TRIE_TYPE( noper_type );
3335 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3336 regnode * const noper_next = regnext( noper );
3337 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3338 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3341 DEBUG_TRIE_COMPILE_r({
3342 regprop(RExC_rx, mysv, cur);
3343 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3344 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3346 regprop(RExC_rx, mysv, noper);
3347 PerlIO_printf( Perl_debug_log, " -> %s",
3348 SvPV_nolen_const(mysv));
3351 regprop(RExC_rx, mysv, noper_next );
3352 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3353 SvPV_nolen_const(mysv));
3355 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3356 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3357 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3361 /* Is noper a trieable nodetype that can be merged with the
3362 * current trie (if there is one)? */
3366 ( noper_trietype == NOTHING)
3367 || ( trietype == NOTHING )
3368 || ( trietype == noper_trietype )
3371 && noper_next == tail
3375 /* Handle mergable triable node
3376 * Either we are the first node in a new trieable sequence,
3377 * in which case we do some bookkeeping, otherwise we update
3378 * the end pointer. */
3381 trietype = noper_trietype;
3382 if ( noper_trietype == NOTHING ) {
3383 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3384 regnode * const noper_next = regnext( noper );
3385 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3386 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3389 if ( noper_next_trietype )
3390 trietype = noper_next_trietype;
3393 if ( trietype == NOTHING )
3394 trietype = noper_trietype;
3399 } /* end handle mergable triable node */
3401 /* handle unmergable node -
3402 * noper may either be a triable node which can not be tried
3403 * together with the current trie, or a non triable node */
3405 /* If last is set and trietype is not NOTHING then we have found
3406 * at least two triable branch sequences in a row of a similar
3407 * trietype so we can turn them into a trie. If/when we
3408 * allow NOTHING to start a trie sequence this condition will be
3409 * required, and it isn't expensive so we leave it in for now. */
3410 if ( trietype != NOTHING )
3411 make_trie( pRExC_state,
3412 startbranch, first, cur, tail, count,
3413 trietype, depth+1 );
3414 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3418 && noper_next == tail
3421 /* noper is triable, so we can start a new trie sequence */
3424 trietype = noper_trietype;
3426 /* if we already saw a first but the current node is not triable then we have
3427 * to reset the first information. */
3432 } /* end handle unmergable node */
3433 } /* loop over branches */
3434 DEBUG_TRIE_COMPILE_r({
3435 regprop(RExC_rx, mysv, cur);
3436 PerlIO_printf( Perl_debug_log,
3437 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3438 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3442 if ( trietype != NOTHING ) {
3443 /* the last branch of the sequence was part of a trie,
3444 * so we have to construct it here outside of the loop
3446 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3447 #ifdef TRIE_STUDY_OPT
3448 if ( ((made == MADE_EXACT_TRIE &&
3449 startbranch == first)
3450 || ( first_non_open == first )) &&
3452 flags |= SCF_TRIE_RESTUDY;
3453 if ( startbranch == first
3456 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3461 /* at this point we know whatever we have is a NOTHING sequence/branch
3462 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3464 if ( startbranch == first ) {
3466 /* the entire thing is a NOTHING sequence, something like this:
3467 * (?:|) So we can turn it into a plain NOTHING op. */
3468 DEBUG_TRIE_COMPILE_r({
3469 regprop(RExC_rx, mysv, cur);
3470 PerlIO_printf( Perl_debug_log,
3471 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3472 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3475 OP(startbranch)= NOTHING;
3476 NEXT_OFF(startbranch)= tail - startbranch;
3477 for ( opt= startbranch + 1; opt < tail ; opt++ )
3481 } /* end if ( last) */
3482 } /* TRIE_MAXBUF is non zero */
3487 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3488 scan = NEXTOPER(NEXTOPER(scan));
3489 } else /* single branch is optimized. */
3490 scan = NEXTOPER(scan);
3492 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3493 scan_frame *newframe = NULL;
3498 if (OP(scan) != SUSPEND) {
3499 /* set the pointer */
3500 if (OP(scan) == GOSUB) {
3502 RExC_recurse[ARG2L(scan)] = scan;
3503 start = RExC_open_parens[paren-1];
3504 end = RExC_close_parens[paren-1];
3507 start = RExC_rxi->program + 1;
3511 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3512 SAVEFREEPV(recursed);
3514 if (!PAREN_TEST(recursed,paren+1)) {
3515 PAREN_SET(recursed,paren+1);
3516 Newx(newframe,1,scan_frame);
3518 if (flags & SCF_DO_SUBSTR) {
3519 SCAN_COMMIT(pRExC_state,data,minlenp);
3520 data->longest = &(data->longest_float);
3522 is_inf = is_inf_internal = 1;
3523 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3524 cl_anything(pRExC_state, data->start_class);
3525 flags &= ~SCF_DO_STCLASS;
3528 Newx(newframe,1,scan_frame);
3531 end = regnext(scan);
3536 SAVEFREEPV(newframe);
3537 newframe->next = regnext(scan);
3538 newframe->last = last;
3539 newframe->stop = stopparen;
3540 newframe->prev = frame;
3550 else if (OP(scan) == EXACT) {
3551 I32 l = STR_LEN(scan);
3554 const U8 * const s = (U8*)STRING(scan);
3555 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3556 l = utf8_length(s, s + l);
3558 uc = *((U8*)STRING(scan));
3561 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3562 /* The code below prefers earlier match for fixed
3563 offset, later match for variable offset. */
3564 if (data->last_end == -1) { /* Update the start info. */
3565 data->last_start_min = data->pos_min;
3566 data->last_start_max = is_inf
3567 ? I32_MAX : data->pos_min + data->pos_delta;
3569 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3571 SvUTF8_on(data->last_found);
3573 SV * const sv = data->last_found;
3574 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3575 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3576 if (mg && mg->mg_len >= 0)
3577 mg->mg_len += utf8_length((U8*)STRING(scan),
3578 (U8*)STRING(scan)+STR_LEN(scan));
3580 data->last_end = data->pos_min + l;
3581 data->pos_min += l; /* As in the first entry. */
3582 data->flags &= ~SF_BEFORE_EOL;
3584 if (flags & SCF_DO_STCLASS_AND) {
3585 /* Check whether it is compatible with what we know already! */
3589 /* If compatible, we or it in below. It is compatible if is
3590 * in the bitmp and either 1) its bit or its fold is set, or 2)
3591 * it's for a locale. Even if there isn't unicode semantics
3592 * here, at runtime there may be because of matching against a
3593 * utf8 string, so accept a possible false positive for
3594 * latin1-range folds */
3596 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3597 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3598 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
3599 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3604 ANYOF_CLASS_ZERO(data->start_class);
3605 ANYOF_BITMAP_ZERO(data->start_class);
3607 ANYOF_BITMAP_SET(data->start_class, uc);
3608 else if (uc >= 0x100) {
3611 /* Some Unicode code points fold to the Latin1 range; as
3612 * XXX temporary code, instead of figuring out if this is
3613 * one, just assume it is and set all the start class bits
3614 * that could be some such above 255 code point's fold
3615 * which will generate fals positives. As the code
3616 * elsewhere that does compute the fold settles down, it
3617 * can be extracted out and re-used here */
3618 for (i = 0; i < 256; i++){
3619 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3620 ANYOF_BITMAP_SET(data->start_class, i);
3624 data->start_class->flags &= ~ANYOF_EOS;
3626 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3628 else if (flags & SCF_DO_STCLASS_OR) {
3629 /* false positive possible if the class is case-folded */
3631 ANYOF_BITMAP_SET(data->start_class, uc);
3633 data->start_class->flags |= ANYOF_UNICODE_ALL;
3634 data->start_class->flags &= ~ANYOF_EOS;
3635 cl_and(data->start_class, and_withp);
3637 flags &= ~SCF_DO_STCLASS;
3639 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3640 I32 l = STR_LEN(scan);
3641 UV uc = *((U8*)STRING(scan));
3643 /* Search for fixed substrings supports EXACT only. */
3644 if (flags & SCF_DO_SUBSTR) {
3646 SCAN_COMMIT(pRExC_state, data, minlenp);
3649 const U8 * const s = (U8 *)STRING(scan);
3650 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3651 l = utf8_length(s, s + l);
3653 else if (has_exactf_sharp_s) {
3654 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3656 min += l - min_subtract;
3660 delta += min_subtract;
3661 if (flags & SCF_DO_SUBSTR) {
3662 data->pos_min += l - min_subtract;
3663 if (data->pos_min < 0) {
3666 data->pos_delta += min_subtract;
3668 data->longest = &(data->longest_float);
3671 if (flags & SCF_DO_STCLASS_AND) {
3672 /* Check whether it is compatible with what we know already! */
3675 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3676 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3677 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3681 ANYOF_CLASS_ZERO(data->start_class);
3682 ANYOF_BITMAP_ZERO(data->start_class);
3684 ANYOF_BITMAP_SET(data->start_class, uc);
3685 data->start_class->flags &= ~ANYOF_EOS;
3686 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
3687 if (OP(scan) == EXACTFL) {
3688 /* XXX This set is probably no longer necessary, and
3689 * probably wrong as LOCALE now is on in the initial
3691 data->start_class->flags |= ANYOF_LOCALE;
3695 /* Also set the other member of the fold pair. In case
3696 * that unicode semantics is called for at runtime, use
3697 * the full latin1 fold. (Can't do this for locale,
3698 * because not known until runtime) */
3699 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3701 /* All other (EXACTFL handled above) folds except under
3702 * /iaa that include s, S, and sharp_s also may include
3704 if (OP(scan) != EXACTFA) {
3705 if (uc == 's' || uc == 'S') {
3706 ANYOF_BITMAP_SET(data->start_class,
3707 LATIN_SMALL_LETTER_SHARP_S);
3709 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3710 ANYOF_BITMAP_SET(data->start_class, 's');
3711 ANYOF_BITMAP_SET(data->start_class, 'S');
3716 else if (uc >= 0x100) {
3718 for (i = 0; i < 256; i++){
3719 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3720 ANYOF_BITMAP_SET(data->start_class, i);
3725 else if (flags & SCF_DO_STCLASS_OR) {
3726 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
3727 /* false positive possible if the class is case-folded.
3728 Assume that the locale settings are the same... */
3730 ANYOF_BITMAP_SET(data->start_class, uc);
3731 if (OP(scan) != EXACTFL) {
3733 /* And set the other member of the fold pair, but
3734 * can't do that in locale because not known until
3736 ANYOF_BITMAP_SET(data->start_class,
3737 PL_fold_latin1[uc]);
3739 /* All folds except under /iaa that include s, S,
3740 * and sharp_s also may include the others */
3741 if (OP(scan) != EXACTFA) {
3742 if (uc == 's' || uc == 'S') {
3743 ANYOF_BITMAP_SET(data->start_class,
3744 LATIN_SMALL_LETTER_SHARP_S);
3746 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3747 ANYOF_BITMAP_SET(data->start_class, 's');
3748 ANYOF_BITMAP_SET(data->start_class, 'S');
3753 data->start_class->flags &= ~ANYOF_EOS;
3755 cl_and(data->start_class, and_withp);
3757 flags &= ~SCF_DO_STCLASS;
3759 else if (REGNODE_VARIES(OP(scan))) {
3760 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3761 I32 f = flags, pos_before = 0;
3762 regnode * const oscan = scan;
3763 struct regnode_charclass_class this_class;
3764 struct regnode_charclass_class *oclass = NULL;
3765 I32 next_is_eval = 0;
3767 switch (PL_regkind[OP(scan)]) {
3768 case WHILEM: /* End of (?:...)* . */
3769 scan = NEXTOPER(scan);
3772 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3773 next = NEXTOPER(scan);
3774 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3776 maxcount = REG_INFTY;
3777 next = regnext(scan);
3778 scan = NEXTOPER(scan);
3782 if (flags & SCF_DO_SUBSTR)
3787 if (flags & SCF_DO_STCLASS) {
3789 maxcount = REG_INFTY;
3790 next = regnext(scan);
3791 scan = NEXTOPER(scan);
3794 is_inf = is_inf_internal = 1;
3795 scan = regnext(scan);
3796 if (flags & SCF_DO_SUBSTR) {
3797 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3798 data->longest = &(data->longest_float);
3800 goto optimize_curly_tail;
3802 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3803 && (scan->flags == stopparen))
3808 mincount = ARG1(scan);
3809 maxcount = ARG2(scan);
3811 next = regnext(scan);
3812 if (OP(scan) == CURLYX) {
3813 I32 lp = (data ? *(data->last_closep) : 0);
3814 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3816 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3817 next_is_eval = (OP(scan) == EVAL);
3819 if (flags & SCF_DO_SUBSTR) {
3820 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3821 pos_before = data->pos_min;
3825 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3827 data->flags |= SF_IS_INF;
3829 if (flags & SCF_DO_STCLASS) {
3830 cl_init(pRExC_state, &this_class);
3831 oclass = data->start_class;
3832 data->start_class = &this_class;
3833 f |= SCF_DO_STCLASS_AND;
https://perl5.git.perl.org / perl5.git / blob