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; /* RXf_* are we folding, multilining? */
115 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
116 char *precomp; /* uncompiled string. */
117 REGEXP *rx_sv; /* The SV that is the regexp. */
118 regexp *rx; /* perl core regexp structure */
119 regexp_internal *rxi; /* internal data for regexp object pprivate field */
120 char *start; /* Start of input for compile */
121 char *end; /* End of input for compile */
122 char *parse; /* Input-scan pointer. */
123 I32 whilem_seen; /* number of WHILEM in this expr */
124 regnode *emit_start; /* Start of emitted-code area */
125 regnode *emit_bound; /* First regnode outside of the allocated space */
126 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
127 I32 naughty; /* How bad is this pattern? */
128 I32 sawback; /* Did we see \1, ...? */
130 I32 size; /* Code size. */
131 I32 npar; /* Capture buffer count, (OPEN). */
132 I32 cpar; /* Capture buffer count, (CLOSE). */
133 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)
161 SV *runtime_code_qr; /* qr with the runtime code blocks */
163 const char *lastparse;
165 AV *paren_name_list; /* idx -> name */
166 #define RExC_lastparse (pRExC_state->lastparse)
167 #define RExC_lastnum (pRExC_state->lastnum)
168 #define RExC_paren_name_list (pRExC_state->paren_name_list)
172 #define RExC_flags (pRExC_state->flags)
173 #define RExC_pm_flags (pRExC_state->pm_flags)
174 #define RExC_precomp (pRExC_state->precomp)
175 #define RExC_rx_sv (pRExC_state->rx_sv)
176 #define RExC_rx (pRExC_state->rx)
177 #define RExC_rxi (pRExC_state->rxi)
178 #define RExC_start (pRExC_state->start)
179 #define RExC_end (pRExC_state->end)
180 #define RExC_parse (pRExC_state->parse)
181 #define RExC_whilem_seen (pRExC_state->whilem_seen)
182 #ifdef RE_TRACK_PATTERN_OFFSETS
183 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
185 #define RExC_emit (pRExC_state->emit)
186 #define RExC_emit_start (pRExC_state->emit_start)
187 #define RExC_emit_bound (pRExC_state->emit_bound)
188 #define RExC_naughty (pRExC_state->naughty)
189 #define RExC_sawback (pRExC_state->sawback)
190 #define RExC_seen (pRExC_state->seen)
191 #define RExC_size (pRExC_state->size)
192 #define RExC_npar (pRExC_state->npar)
193 #define RExC_nestroot (pRExC_state->nestroot)
194 #define RExC_extralen (pRExC_state->extralen)
195 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
196 #define RExC_utf8 (pRExC_state->utf8)
197 #define RExC_uni_semantics (pRExC_state->uni_semantics)
198 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
199 #define RExC_open_parens (pRExC_state->open_parens)
200 #define RExC_close_parens (pRExC_state->close_parens)
201 #define RExC_opend (pRExC_state->opend)
202 #define RExC_paren_names (pRExC_state->paren_names)
203 #define RExC_recurse (pRExC_state->recurse)
204 #define RExC_recurse_count (pRExC_state->recurse_count)
205 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
206 #define RExC_contains_locale (pRExC_state->contains_locale)
207 #define RExC_override_recoding (pRExC_state->override_recoding)
210 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
211 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
212 ((*s) == '{' && regcurly(s)))
215 #undef SPSTART /* dratted cpp namespace... */
218 * Flags to be passed up and down.
220 #define WORST 0 /* Worst case. */
221 #define HASWIDTH 0x01 /* Known to match non-null strings. */
223 /* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
224 * character, and if utf8, must be invariant. Note that this is not the same
225 * thing as REGNODE_SIMPLE */
227 #define SPSTART 0x04 /* Starts with * or +. */
228 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
229 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
231 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
233 /* whether trie related optimizations are enabled */
234 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
235 #define TRIE_STUDY_OPT
236 #define FULL_TRIE_STUDY
242 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
243 #define PBITVAL(paren) (1 << ((paren) & 7))
244 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
245 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
246 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
248 /* If not already in utf8, do a longjmp back to the beginning */
249 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
250 #define REQUIRE_UTF8 STMT_START { \
251 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
254 /* About scan_data_t.
256 During optimisation we recurse through the regexp program performing
257 various inplace (keyhole style) optimisations. In addition study_chunk
258 and scan_commit populate this data structure with information about
259 what strings MUST appear in the pattern. We look for the longest
260 string that must appear at a fixed location, and we look for the
261 longest string that may appear at a floating location. So for instance
266 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
267 strings (because they follow a .* construct). study_chunk will identify
268 both FOO and BAR as being the longest fixed and floating strings respectively.
270 The strings can be composites, for instance
274 will result in a composite fixed substring 'foo'.
276 For each string some basic information is maintained:
278 - offset or min_offset
279 This is the position the string must appear at, or not before.
280 It also implicitly (when combined with minlenp) tells us how many
281 characters must match before the string we are searching for.
282 Likewise when combined with minlenp and the length of the string it
283 tells us how many characters must appear after the string we have
287 Only used for floating strings. This is the rightmost point that
288 the string can appear at. If set to I32 max it indicates that the
289 string can occur infinitely far to the right.
292 A pointer to the minimum length of the pattern that the string
293 was found inside. This is important as in the case of positive
294 lookahead or positive lookbehind we can have multiple patterns
299 The minimum length of the pattern overall is 3, the minimum length
300 of the lookahead part is 3, but the minimum length of the part that
301 will actually match is 1. So 'FOO's minimum length is 3, but the
302 minimum length for the F is 1. This is important as the minimum length
303 is used to determine offsets in front of and behind the string being
304 looked for. Since strings can be composites this is the length of the
305 pattern at the time it was committed with a scan_commit. Note that
306 the length is calculated by study_chunk, so that the minimum lengths
307 are not known until the full pattern has been compiled, thus the
308 pointer to the value.
312 In the case of lookbehind the string being searched for can be
313 offset past the start point of the final matching string.
314 If this value was just blithely removed from the min_offset it would
315 invalidate some of the calculations for how many chars must match
316 before or after (as they are derived from min_offset and minlen and
317 the length of the string being searched for).
318 When the final pattern is compiled and the data is moved from the
319 scan_data_t structure into the regexp structure the information
320 about lookbehind is factored in, with the information that would
321 have been lost precalculated in the end_shift field for the
324 The fields pos_min and pos_delta are used to store the minimum offset
325 and the delta to the maximum offset at the current point in the pattern.
329 typedef struct scan_data_t {
330 /*I32 len_min; unused */
331 /*I32 len_delta; unused */
335 I32 last_end; /* min value, <0 unless valid. */
338 SV **longest; /* Either &l_fixed, or &l_float. */
339 SV *longest_fixed; /* longest fixed string found in pattern */
340 I32 offset_fixed; /* offset where it starts */
341 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
342 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
343 SV *longest_float; /* longest floating string found in pattern */
344 I32 offset_float_min; /* earliest point in string it can appear */
345 I32 offset_float_max; /* latest point in string it can appear */
346 I32 *minlen_float; /* pointer to the minlen relevant to the string */
347 I32 lookbehind_float; /* is the position of the string modified by LB */
351 struct regnode_charclass_class *start_class;
355 * Forward declarations for pregcomp()'s friends.
358 static const scan_data_t zero_scan_data =
359 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
361 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
362 #define SF_BEFORE_SEOL 0x0001
363 #define SF_BEFORE_MEOL 0x0002
364 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
365 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
368 # define SF_FIX_SHIFT_EOL (0+2)
369 # define SF_FL_SHIFT_EOL (0+4)
371 # define SF_FIX_SHIFT_EOL (+2)
372 # define SF_FL_SHIFT_EOL (+4)
375 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
376 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
378 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
379 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
380 #define SF_IS_INF 0x0040
381 #define SF_HAS_PAR 0x0080
382 #define SF_IN_PAR 0x0100
383 #define SF_HAS_EVAL 0x0200
384 #define SCF_DO_SUBSTR 0x0400
385 #define SCF_DO_STCLASS_AND 0x0800
386 #define SCF_DO_STCLASS_OR 0x1000
387 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
388 #define SCF_WHILEM_VISITED_POS 0x2000
390 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
391 #define SCF_SEEN_ACCEPT 0x8000
393 #define UTF cBOOL(RExC_utf8)
395 /* The enums for all these are ordered so things work out correctly */
396 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
397 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
398 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
399 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
400 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
401 #define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
402 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
404 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
406 #define OOB_UNICODE 12345678
407 #define OOB_NAMEDCLASS -1
409 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
410 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
413 /* length of regex to show in messages that don't mark a position within */
414 #define RegexLengthToShowInErrorMessages 127
417 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
418 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
419 * op/pragma/warn/regcomp.
421 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
422 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
424 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
427 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
428 * arg. Show regex, up to a maximum length. If it's too long, chop and add
431 #define _FAIL(code) STMT_START { \
432 const char *ellipses = ""; \
433 IV len = RExC_end - RExC_precomp; \
436 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
437 if (len > RegexLengthToShowInErrorMessages) { \
438 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
439 len = RegexLengthToShowInErrorMessages - 10; \
445 #define FAIL(msg) _FAIL( \
446 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
447 msg, (int)len, RExC_precomp, ellipses))
449 #define FAIL2(msg,arg) _FAIL( \
450 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
451 arg, (int)len, RExC_precomp, ellipses))
454 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
456 #define Simple_vFAIL(m) STMT_START { \
457 const IV offset = RExC_parse - RExC_precomp; \
458 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
459 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
463 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
465 #define vFAIL(m) STMT_START { \
467 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
472 * Like Simple_vFAIL(), but accepts two arguments.
474 #define Simple_vFAIL2(m,a1) STMT_START { \
475 const IV offset = RExC_parse - RExC_precomp; \
476 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
477 (int)offset, RExC_precomp, RExC_precomp + offset); \
481 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
483 #define vFAIL2(m,a1) STMT_START { \
485 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
486 Simple_vFAIL2(m, a1); \
491 * Like Simple_vFAIL(), but accepts three arguments.
493 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
494 const IV offset = RExC_parse - RExC_precomp; \
495 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
496 (int)offset, RExC_precomp, RExC_precomp + offset); \
500 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
502 #define vFAIL3(m,a1,a2) STMT_START { \
504 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
505 Simple_vFAIL3(m, a1, a2); \
509 * Like Simple_vFAIL(), but accepts four arguments.
511 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
512 const IV offset = RExC_parse - RExC_precomp; \
513 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
514 (int)offset, RExC_precomp, RExC_precomp + offset); \
517 #define ckWARNreg(loc,m) STMT_START { \
518 const IV offset = loc - RExC_precomp; \
519 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
520 (int)offset, RExC_precomp, RExC_precomp + offset); \
523 #define ckWARNregdep(loc,m) STMT_START { \
524 const IV offset = loc - RExC_precomp; \
525 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
527 (int)offset, RExC_precomp, RExC_precomp + offset); \
530 #define ckWARN2regdep(loc,m, a1) STMT_START { \
531 const IV offset = loc - RExC_precomp; \
532 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
534 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
537 #define ckWARN2reg(loc, m, a1) STMT_START { \
538 const IV offset = loc - RExC_precomp; \
539 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
540 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
543 #define vWARN3(loc, m, a1, a2) STMT_START { \
544 const IV offset = loc - RExC_precomp; \
545 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
546 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
549 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
550 const IV offset = loc - RExC_precomp; \
551 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
552 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
555 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
556 const IV offset = loc - RExC_precomp; \
557 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
558 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
561 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
562 const IV offset = loc - RExC_precomp; \
563 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
564 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
567 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
568 const IV offset = loc - RExC_precomp; \
569 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
570 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
574 /* Allow for side effects in s */
575 #define REGC(c,s) STMT_START { \
576 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
579 /* Macros for recording node offsets. 20001227 mjd@plover.com
580 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
581 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
582 * Element 0 holds the number n.
583 * Position is 1 indexed.
585 #ifndef RE_TRACK_PATTERN_OFFSETS
586 #define Set_Node_Offset_To_R(node,byte)
587 #define Set_Node_Offset(node,byte)
588 #define Set_Cur_Node_Offset
589 #define Set_Node_Length_To_R(node,len)
590 #define Set_Node_Length(node,len)
591 #define Set_Node_Cur_Length(node)
592 #define Node_Offset(n)
593 #define Node_Length(n)
594 #define Set_Node_Offset_Length(node,offset,len)
595 #define ProgLen(ri) ri->u.proglen
596 #define SetProgLen(ri,x) ri->u.proglen = x
598 #define ProgLen(ri) ri->u.offsets[0]
599 #define SetProgLen(ri,x) ri->u.offsets[0] = x
600 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
602 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
603 __LINE__, (int)(node), (int)(byte))); \
605 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
607 RExC_offsets[2*(node)-1] = (byte); \
612 #define Set_Node_Offset(node,byte) \
613 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
614 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
616 #define Set_Node_Length_To_R(node,len) STMT_START { \
618 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
619 __LINE__, (int)(node), (int)(len))); \
621 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
623 RExC_offsets[2*(node)] = (len); \
628 #define Set_Node_Length(node,len) \
629 Set_Node_Length_To_R((node)-RExC_emit_start, len)
630 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
631 #define Set_Node_Cur_Length(node) \
632 Set_Node_Length(node, RExC_parse - parse_start)
634 /* Get offsets and lengths */
635 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
636 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
638 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
639 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
640 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
644 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
645 #define EXPERIMENTAL_INPLACESCAN
646 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
648 #define DEBUG_STUDYDATA(str,data,depth) \
649 DEBUG_OPTIMISE_MORE_r(if(data){ \
650 PerlIO_printf(Perl_debug_log, \
651 "%*s" str "Pos:%"IVdf"/%"IVdf \
652 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
653 (int)(depth)*2, "", \
654 (IV)((data)->pos_min), \
655 (IV)((data)->pos_delta), \
656 (UV)((data)->flags), \
657 (IV)((data)->whilem_c), \
658 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
659 is_inf ? "INF " : "" \
661 if ((data)->last_found) \
662 PerlIO_printf(Perl_debug_log, \
663 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
664 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
665 SvPVX_const((data)->last_found), \
666 (IV)((data)->last_end), \
667 (IV)((data)->last_start_min), \
668 (IV)((data)->last_start_max), \
669 ((data)->longest && \
670 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
671 SvPVX_const((data)->longest_fixed), \
672 (IV)((data)->offset_fixed), \
673 ((data)->longest && \
674 (data)->longest==&((data)->longest_float)) ? "*" : "", \
675 SvPVX_const((data)->longest_float), \
676 (IV)((data)->offset_float_min), \
677 (IV)((data)->offset_float_max) \
679 PerlIO_printf(Perl_debug_log,"\n"); \
682 static void clear_re(pTHX_ void *r);
684 /* Mark that we cannot extend a found fixed substring at this point.
685 Update the longest found anchored substring and the longest found
686 floating substrings if needed. */
689 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
691 const STRLEN l = CHR_SVLEN(data->last_found);
692 const STRLEN old_l = CHR_SVLEN(*data->longest);
693 GET_RE_DEBUG_FLAGS_DECL;
695 PERL_ARGS_ASSERT_SCAN_COMMIT;
697 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
698 SvSetMagicSV(*data->longest, data->last_found);
699 if (*data->longest == data->longest_fixed) {
700 data->offset_fixed = l ? data->last_start_min : data->pos_min;
701 if (data->flags & SF_BEFORE_EOL)
703 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
705 data->flags &= ~SF_FIX_BEFORE_EOL;
706 data->minlen_fixed=minlenp;
707 data->lookbehind_fixed=0;
709 else { /* *data->longest == data->longest_float */
710 data->offset_float_min = l ? data->last_start_min : data->pos_min;
711 data->offset_float_max = (l
712 ? data->last_start_max
713 : data->pos_min + data->pos_delta);
714 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
715 data->offset_float_max = I32_MAX;
716 if (data->flags & SF_BEFORE_EOL)
718 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
720 data->flags &= ~SF_FL_BEFORE_EOL;
721 data->minlen_float=minlenp;
722 data->lookbehind_float=0;
725 SvCUR_set(data->last_found, 0);
727 SV * const sv = data->last_found;
728 if (SvUTF8(sv) && SvMAGICAL(sv)) {
729 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
735 data->flags &= ~SF_BEFORE_EOL;
736 DEBUG_STUDYDATA("commit: ",data,0);
739 /* Can match anything (initialization) */
741 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
743 PERL_ARGS_ASSERT_CL_ANYTHING;
745 ANYOF_BITMAP_SETALL(cl);
746 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
747 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
749 /* If any portion of the regex is to operate under locale rules,
750 * initialization includes it. The reason this isn't done for all regexes
751 * is that the optimizer was written under the assumption that locale was
752 * all-or-nothing. Given the complexity and lack of documentation in the
753 * optimizer, and that there are inadequate test cases for locale, so many
754 * parts of it may not work properly, it is safest to avoid locale unless
756 if (RExC_contains_locale) {
757 ANYOF_CLASS_SETALL(cl); /* /l uses class */
758 cl->flags |= ANYOF_LOCALE;
761 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
765 /* Can match anything (initialization) */
767 S_cl_is_anything(const struct regnode_charclass_class *cl)
771 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
773 for (value = 0; value <= ANYOF_MAX; value += 2)
774 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
776 if (!(cl->flags & ANYOF_UNICODE_ALL))
778 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
783 /* Can match anything (initialization) */
785 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
787 PERL_ARGS_ASSERT_CL_INIT;
789 Zero(cl, 1, struct regnode_charclass_class);
791 cl_anything(pRExC_state, cl);
792 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
795 /* These two functions currently do the exact same thing */
796 #define cl_init_zero S_cl_init
798 /* 'AND' a given class with another one. Can create false positives. 'cl'
799 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
800 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
802 S_cl_and(struct regnode_charclass_class *cl,
803 const struct regnode_charclass_class *and_with)
805 PERL_ARGS_ASSERT_CL_AND;
807 assert(and_with->type == ANYOF);
809 /* I (khw) am not sure all these restrictions are necessary XXX */
810 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
811 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
812 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
813 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
814 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
817 if (and_with->flags & ANYOF_INVERT)
818 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
819 cl->bitmap[i] &= ~and_with->bitmap[i];
821 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
822 cl->bitmap[i] &= and_with->bitmap[i];
823 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
825 if (and_with->flags & ANYOF_INVERT) {
827 /* Here, the and'ed node is inverted. Get the AND of the flags that
828 * aren't affected by the inversion. Those that are affected are
829 * handled individually below */
830 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
831 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
832 cl->flags |= affected_flags;
834 /* We currently don't know how to deal with things that aren't in the
835 * bitmap, but we know that the intersection is no greater than what
836 * is already in cl, so let there be false positives that get sorted
837 * out after the synthetic start class succeeds, and the node is
838 * matched for real. */
840 /* The inversion of these two flags indicate that the resulting
841 * intersection doesn't have them */
842 if (and_with->flags & ANYOF_UNICODE_ALL) {
843 cl->flags &= ~ANYOF_UNICODE_ALL;
845 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
846 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
849 else { /* and'd node is not inverted */
850 U8 outside_bitmap_but_not_utf8; /* Temp variable */
852 if (! ANYOF_NONBITMAP(and_with)) {
854 /* Here 'and_with' doesn't match anything outside the bitmap
855 * (except possibly ANYOF_UNICODE_ALL), which means the
856 * intersection can't either, except for ANYOF_UNICODE_ALL, in
857 * which case we don't know what the intersection is, but it's no
858 * greater than what cl already has, so can just leave it alone,
859 * with possible false positives */
860 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
861 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
862 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
865 else if (! ANYOF_NONBITMAP(cl)) {
867 /* Here, 'and_with' does match something outside the bitmap, and cl
868 * doesn't have a list of things to match outside the bitmap. If
869 * cl can match all code points above 255, the intersection will
870 * be those above-255 code points that 'and_with' matches. If cl
871 * can't match all Unicode code points, it means that it can't
872 * match anything outside the bitmap (since the 'if' that got us
873 * into this block tested for that), so we leave the bitmap empty.
875 if (cl->flags & ANYOF_UNICODE_ALL) {
876 ARG_SET(cl, ARG(and_with));
878 /* and_with's ARG may match things that don't require UTF8.
879 * And now cl's will too, in spite of this being an 'and'. See
880 * the comments below about the kludge */
881 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
885 /* Here, both 'and_with' and cl match something outside the
886 * bitmap. Currently we do not do the intersection, so just match
887 * whatever cl had at the beginning. */
891 /* Take the intersection of the two sets of flags. However, the
892 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
893 * kludge around the fact that this flag is not treated like the others
894 * which are initialized in cl_anything(). The way the optimizer works
895 * is that the synthetic start class (SSC) is initialized to match
896 * anything, and then the first time a real node is encountered, its
897 * values are AND'd with the SSC's with the result being the values of
898 * the real node. However, there are paths through the optimizer where
899 * the AND never gets called, so those initialized bits are set
900 * inappropriately, which is not usually a big deal, as they just cause
901 * false positives in the SSC, which will just mean a probably
902 * imperceptible slow down in execution. However this bit has a
903 * higher false positive consequence in that it can cause utf8.pm,
904 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
905 * bigger slowdown and also causes significant extra memory to be used.
906 * In order to prevent this, the code now takes a different tack. The
907 * bit isn't set unless some part of the regular expression needs it,
908 * but once set it won't get cleared. This means that these extra
909 * modules won't get loaded unless there was some path through the
910 * pattern that would have required them anyway, and so any false
911 * positives that occur by not ANDing them out when they could be
912 * aren't as severe as they would be if we treated this bit like all
914 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
915 & ANYOF_NONBITMAP_NON_UTF8;
916 cl->flags &= and_with->flags;
917 cl->flags |= outside_bitmap_but_not_utf8;
921 /* 'OR' a given class with another one. Can create false positives. 'cl'
922 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
923 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
925 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
927 PERL_ARGS_ASSERT_CL_OR;
929 if (or_with->flags & ANYOF_INVERT) {
931 /* Here, the or'd node is to be inverted. This means we take the
932 * complement of everything not in the bitmap, but currently we don't
933 * know what that is, so give up and match anything */
934 if (ANYOF_NONBITMAP(or_with)) {
935 cl_anything(pRExC_state, cl);
938 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
939 * <= (B1 | !B2) | (CL1 | !CL2)
940 * which is wasteful if CL2 is small, but we ignore CL2:
941 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
942 * XXXX Can we handle case-fold? Unclear:
943 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
944 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
946 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
947 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
948 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
951 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
952 cl->bitmap[i] |= ~or_with->bitmap[i];
953 } /* XXXX: logic is complicated otherwise */
955 cl_anything(pRExC_state, cl);
958 /* And, we can just take the union of the flags that aren't affected
959 * by the inversion */
960 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
962 /* For the remaining flags:
963 ANYOF_UNICODE_ALL and inverted means to not match anything above
964 255, which means that the union with cl should just be
965 what cl has in it, so can ignore this flag
966 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
967 is 127-255 to match them, but then invert that, so the
968 union with cl should just be what cl has in it, so can
971 } else { /* 'or_with' is not inverted */
972 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
973 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
974 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
975 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
978 /* OR char bitmap and class bitmap separately */
979 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
980 cl->bitmap[i] |= or_with->bitmap[i];
981 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
982 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
983 cl->classflags[i] |= or_with->classflags[i];
984 cl->flags |= ANYOF_CLASS;
987 else { /* XXXX: logic is complicated, leave it along for a moment. */
988 cl_anything(pRExC_state, cl);
991 if (ANYOF_NONBITMAP(or_with)) {
993 /* Use the added node's outside-the-bit-map match if there isn't a
994 * conflict. If there is a conflict (both nodes match something
995 * outside the bitmap, but what they match outside is not the same
996 * pointer, and hence not easily compared until XXX we extend
997 * inversion lists this far), give up and allow the start class to
998 * match everything outside the bitmap. If that stuff is all above
999 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1000 if (! ANYOF_NONBITMAP(cl)) {
1001 ARG_SET(cl, ARG(or_with));
1003 else if (ARG(cl) != ARG(or_with)) {
1005 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1006 cl_anything(pRExC_state, cl);
1009 cl->flags |= ANYOF_UNICODE_ALL;
1014 /* Take the union */
1015 cl->flags |= or_with->flags;
1019 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1020 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1021 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1022 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1027 dump_trie(trie,widecharmap,revcharmap)
1028 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1029 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1031 These routines dump out a trie in a somewhat readable format.
1032 The _interim_ variants are used for debugging the interim
1033 tables that are used to generate the final compressed
1034 representation which is what dump_trie expects.
1036 Part of the reason for their existence is to provide a form
1037 of documentation as to how the different representations function.
1042 Dumps the final compressed table form of the trie to Perl_debug_log.
1043 Used for debugging make_trie().
1047 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1048 AV *revcharmap, U32 depth)
1051 SV *sv=sv_newmortal();
1052 int colwidth= widecharmap ? 6 : 4;
1054 GET_RE_DEBUG_FLAGS_DECL;
1056 PERL_ARGS_ASSERT_DUMP_TRIE;
1058 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1059 (int)depth * 2 + 2,"",
1060 "Match","Base","Ofs" );
1062 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1063 SV ** const tmp = av_fetch( revcharmap, state, 0);
1065 PerlIO_printf( Perl_debug_log, "%*s",
1067 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1068 PL_colors[0], PL_colors[1],
1069 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1070 PERL_PV_ESCAPE_FIRSTCHAR
1075 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1076 (int)depth * 2 + 2,"");
1078 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1079 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1080 PerlIO_printf( Perl_debug_log, "\n");
1082 for( state = 1 ; state < trie->statecount ; state++ ) {
1083 const U32 base = trie->states[ state ].trans.base;
1085 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1087 if ( trie->states[ state ].wordnum ) {
1088 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1090 PerlIO_printf( Perl_debug_log, "%6s", "" );
1093 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1098 while( ( base + ofs < trie->uniquecharcount ) ||
1099 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1100 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1103 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1105 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1106 if ( ( base + ofs >= trie->uniquecharcount ) &&
1107 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1108 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1110 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1112 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1114 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1118 PerlIO_printf( Perl_debug_log, "]");
1121 PerlIO_printf( Perl_debug_log, "\n" );
1123 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1124 for (word=1; word <= trie->wordcount; word++) {
1125 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1126 (int)word, (int)(trie->wordinfo[word].prev),
1127 (int)(trie->wordinfo[word].len));
1129 PerlIO_printf(Perl_debug_log, "\n" );
1132 Dumps a fully constructed but uncompressed trie in list form.
1133 List tries normally only are used for construction when the number of
1134 possible chars (trie->uniquecharcount) is very high.
1135 Used for debugging make_trie().
1138 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1139 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1143 SV *sv=sv_newmortal();
1144 int colwidth= widecharmap ? 6 : 4;
1145 GET_RE_DEBUG_FLAGS_DECL;
1147 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1149 /* print out the table precompression. */
1150 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1151 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1152 "------:-----+-----------------\n" );
1154 for( state=1 ; state < next_alloc ; state ++ ) {
1157 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1158 (int)depth * 2 + 2,"", (UV)state );
1159 if ( ! trie->states[ state ].wordnum ) {
1160 PerlIO_printf( Perl_debug_log, "%5s| ","");
1162 PerlIO_printf( Perl_debug_log, "W%4x| ",
1163 trie->states[ state ].wordnum
1166 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1167 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1169 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1171 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1172 PL_colors[0], PL_colors[1],
1173 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1174 PERL_PV_ESCAPE_FIRSTCHAR
1176 TRIE_LIST_ITEM(state,charid).forid,
1177 (UV)TRIE_LIST_ITEM(state,charid).newstate
1180 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1181 (int)((depth * 2) + 14), "");
1184 PerlIO_printf( Perl_debug_log, "\n");
1189 Dumps a fully constructed but uncompressed trie in table form.
1190 This is the normal DFA style state transition table, with a few
1191 twists to facilitate compression later.
1192 Used for debugging make_trie().
1195 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1196 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1201 SV *sv=sv_newmortal();
1202 int colwidth= widecharmap ? 6 : 4;
1203 GET_RE_DEBUG_FLAGS_DECL;
1205 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1208 print out the table precompression so that we can do a visual check
1209 that they are identical.
1212 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1214 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1215 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1217 PerlIO_printf( Perl_debug_log, "%*s",
1219 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1220 PL_colors[0], PL_colors[1],
1221 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1222 PERL_PV_ESCAPE_FIRSTCHAR
1228 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1230 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1231 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1234 PerlIO_printf( Perl_debug_log, "\n" );
1236 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1238 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1239 (int)depth * 2 + 2,"",
1240 (UV)TRIE_NODENUM( state ) );
1242 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1243 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1245 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1247 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1249 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1250 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1252 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1253 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1261 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1262 startbranch: the first branch in the whole branch sequence
1263 first : start branch of sequence of branch-exact nodes.
1264 May be the same as startbranch
1265 last : Thing following the last branch.
1266 May be the same as tail.
1267 tail : item following the branch sequence
1268 count : words in the sequence
1269 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1270 depth : indent depth
1272 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1274 A trie is an N'ary tree where the branches are determined by digital
1275 decomposition of the key. IE, at the root node you look up the 1st character and
1276 follow that branch repeat until you find the end of the branches. Nodes can be
1277 marked as "accepting" meaning they represent a complete word. Eg:
1281 would convert into the following structure. Numbers represent states, letters
1282 following numbers represent valid transitions on the letter from that state, if
1283 the number is in square brackets it represents an accepting state, otherwise it
1284 will be in parenthesis.
1286 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1290 (1) +-i->(6)-+-s->[7]
1292 +-s->(3)-+-h->(4)-+-e->[5]
1294 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1296 This shows that when matching against the string 'hers' we will begin at state 1
1297 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1298 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1299 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1300 single traverse. We store a mapping from accepting to state to which word was
1301 matched, and then when we have multiple possibilities we try to complete the
1302 rest of the regex in the order in which they occured in the alternation.
1304 The only prior NFA like behaviour that would be changed by the TRIE support is
1305 the silent ignoring of duplicate alternations which are of the form:
1307 / (DUPE|DUPE) X? (?{ ... }) Y /x
1309 Thus EVAL blocks following a trie may be called a different number of times with
1310 and without the optimisation. With the optimisations dupes will be silently
1311 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1312 the following demonstrates:
1314 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1316 which prints out 'word' three times, but
1318 'words'=~/(word|word|word)(?{ print $1 })S/
1320 which doesnt print it out at all. This is due to other optimisations kicking in.
1322 Example of what happens on a structural level:
1324 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1326 1: CURLYM[1] {1,32767}(18)
1337 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1338 and should turn into:
1340 1: CURLYM[1] {1,32767}(18)
1342 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1350 Cases where tail != last would be like /(?foo|bar)baz/:
1360 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1361 and would end up looking like:
1364 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1371 d = uvuni_to_utf8_flags(d, uv, 0);
1373 is the recommended Unicode-aware way of saying
1378 #define TRIE_STORE_REVCHAR(val) \
1381 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1382 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1383 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1384 SvCUR_set(zlopp, kapow - flrbbbbb); \
1387 av_push(revcharmap, zlopp); \
1389 char ooooff = (char)val; \
1390 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1394 #define TRIE_READ_CHAR STMT_START { \
1397 /* if it is UTF then it is either already folded, or does not need folding */ \
1398 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1400 else if (folder == PL_fold_latin1) { \
1401 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1402 if ( foldlen > 0 ) { \
1403 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1409 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1410 skiplen = UNISKIP(uvc); \
1411 foldlen -= skiplen; \
1412 scan = foldbuf + skiplen; \
1415 /* raw data, will be folded later if needed */ \
1423 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1424 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1425 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1426 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1428 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1429 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1430 TRIE_LIST_CUR( state )++; \
1433 #define TRIE_LIST_NEW(state) STMT_START { \
1434 Newxz( trie->states[ state ].trans.list, \
1435 4, reg_trie_trans_le ); \
1436 TRIE_LIST_CUR( state ) = 1; \
1437 TRIE_LIST_LEN( state ) = 4; \
1440 #define TRIE_HANDLE_WORD(state) STMT_START { \
1441 U16 dupe= trie->states[ state ].wordnum; \
1442 regnode * const noper_next = regnext( noper ); \
1445 /* store the word for dumping */ \
1447 if (OP(noper) != NOTHING) \
1448 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1450 tmp = newSVpvn_utf8( "", 0, UTF ); \
1451 av_push( trie_words, tmp ); \
1455 trie->wordinfo[curword].prev = 0; \
1456 trie->wordinfo[curword].len = wordlen; \
1457 trie->wordinfo[curword].accept = state; \
1459 if ( noper_next < tail ) { \
1461 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1462 trie->jump[curword] = (U16)(noper_next - convert); \
1464 jumper = noper_next; \
1466 nextbranch= regnext(cur); \
1470 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1471 /* chain, so that when the bits of chain are later */\
1472 /* linked together, the dups appear in the chain */\
1473 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1474 trie->wordinfo[dupe].prev = curword; \
1476 /* we haven't inserted this word yet. */ \
1477 trie->states[ state ].wordnum = curword; \
1482 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1483 ( ( base + charid >= ucharcount \
1484 && base + charid < ubound \
1485 && state == trie->trans[ base - ucharcount + charid ].check \
1486 && trie->trans[ base - ucharcount + charid ].next ) \
1487 ? trie->trans[ base - ucharcount + charid ].next \
1488 : ( state==1 ? special : 0 ) \
1492 #define MADE_JUMP_TRIE 2
1493 #define MADE_EXACT_TRIE 4
1496 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1499 /* first pass, loop through and scan words */
1500 reg_trie_data *trie;
1501 HV *widecharmap = NULL;
1502 AV *revcharmap = newAV();
1504 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1509 regnode *jumper = NULL;
1510 regnode *nextbranch = NULL;
1511 regnode *convert = NULL;
1512 U32 *prev_states; /* temp array mapping each state to previous one */
1513 /* we just use folder as a flag in utf8 */
1514 const U8 * folder = NULL;
1517 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1518 AV *trie_words = NULL;
1519 /* along with revcharmap, this only used during construction but both are
1520 * useful during debugging so we store them in the struct when debugging.
1523 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1524 STRLEN trie_charcount=0;
1526 SV *re_trie_maxbuff;
1527 GET_RE_DEBUG_FLAGS_DECL;
1529 PERL_ARGS_ASSERT_MAKE_TRIE;
1531 PERL_UNUSED_ARG(depth);
1538 case EXACTFU_TRICKYFOLD:
1539 case EXACTFU: folder = PL_fold_latin1; break;
1540 case EXACTF: folder = PL_fold; break;
1541 case EXACTFL: folder = PL_fold_locale; break;
1542 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1545 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1547 trie->startstate = 1;
1548 trie->wordcount = word_count;
1549 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1550 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1552 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1553 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1554 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1557 trie_words = newAV();
1560 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1561 if (!SvIOK(re_trie_maxbuff)) {
1562 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1564 DEBUG_TRIE_COMPILE_r({
1565 PerlIO_printf( Perl_debug_log,
1566 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1567 (int)depth * 2 + 2, "",
1568 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1569 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1573 /* Find the node we are going to overwrite */
1574 if ( first == startbranch && OP( last ) != BRANCH ) {
1575 /* whole branch chain */
1578 /* branch sub-chain */
1579 convert = NEXTOPER( first );
1582 /* -- First loop and Setup --
1584 We first traverse the branches and scan each word to determine if it
1585 contains widechars, and how many unique chars there are, this is
1586 important as we have to build a table with at least as many columns as we
1589 We use an array of integers to represent the character codes 0..255
1590 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1591 native representation of the character value as the key and IV's for the
1594 *TODO* If we keep track of how many times each character is used we can
1595 remap the columns so that the table compression later on is more
1596 efficient in terms of memory by ensuring the most common value is in the
1597 middle and the least common are on the outside. IMO this would be better
1598 than a most to least common mapping as theres a decent chance the most
1599 common letter will share a node with the least common, meaning the node
1600 will not be compressible. With a middle is most common approach the worst
1601 case is when we have the least common nodes twice.
1605 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1606 regnode *noper = NEXTOPER( cur );
1607 const U8 *uc = (U8*)STRING( noper );
1608 const U8 *e = uc + STR_LEN( noper );
1610 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1612 const U8 *scan = (U8*)NULL;
1613 U32 wordlen = 0; /* required init */
1615 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1617 if (OP(noper) == NOTHING) {
1618 regnode *noper_next= regnext(noper);
1619 if (noper_next != tail && OP(noper_next) == flags) {
1621 uc= (U8*)STRING(noper);
1622 e= uc + STR_LEN(noper);
1623 trie->minlen= STR_LEN(noper);
1630 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1631 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1632 regardless of encoding */
1633 if (OP( noper ) == EXACTFU_SS) {
1634 /* false positives are ok, so just set this */
1635 TRIE_BITMAP_SET(trie,0xDF);
1638 for ( ; uc < e ; uc += len ) {
1639 TRIE_CHARCOUNT(trie)++;
1644 U8 folded= folder[ (U8) uvc ];
1645 if ( !trie->charmap[ folded ] ) {
1646 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1647 TRIE_STORE_REVCHAR( folded );
1650 if ( !trie->charmap[ uvc ] ) {
1651 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1652 TRIE_STORE_REVCHAR( uvc );
1655 /* store the codepoint in the bitmap, and its folded
1657 TRIE_BITMAP_SET(trie, uvc);
1659 /* store the folded codepoint */
1660 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1663 /* store first byte of utf8 representation of
1664 variant codepoints */
1665 if (! UNI_IS_INVARIANT(uvc)) {
1666 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1669 set_bit = 0; /* We've done our bit :-) */
1674 widecharmap = newHV();
1676 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1679 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1681 if ( !SvTRUE( *svpp ) ) {
1682 sv_setiv( *svpp, ++trie->uniquecharcount );
1683 TRIE_STORE_REVCHAR(uvc);
1687 if( cur == first ) {
1688 trie->minlen = chars;
1689 trie->maxlen = chars;
1690 } else if (chars < trie->minlen) {
1691 trie->minlen = chars;
1692 } else if (chars > trie->maxlen) {
1693 trie->maxlen = chars;
1695 if (OP( noper ) == EXACTFU_SS) {
1696 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1697 if (trie->minlen > 1)
1700 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1701 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1702 * - We assume that any such sequence might match a 2 byte string */
1703 if (trie->minlen > 2 )
1707 } /* end first pass */
1708 DEBUG_TRIE_COMPILE_r(
1709 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1710 (int)depth * 2 + 2,"",
1711 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1712 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1713 (int)trie->minlen, (int)trie->maxlen )
1717 We now know what we are dealing with in terms of unique chars and
1718 string sizes so we can calculate how much memory a naive
1719 representation using a flat table will take. If it's over a reasonable
1720 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1721 conservative but potentially much slower representation using an array
1724 At the end we convert both representations into the same compressed
1725 form that will be used in regexec.c for matching with. The latter
1726 is a form that cannot be used to construct with but has memory
1727 properties similar to the list form and access properties similar
1728 to the table form making it both suitable for fast searches and
1729 small enough that its feasable to store for the duration of a program.
1731 See the comment in the code where the compressed table is produced
1732 inplace from the flat tabe representation for an explanation of how
1733 the compression works.
1738 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1741 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1743 Second Pass -- Array Of Lists Representation
1745 Each state will be represented by a list of charid:state records
1746 (reg_trie_trans_le) the first such element holds the CUR and LEN
1747 points of the allocated array. (See defines above).
1749 We build the initial structure using the lists, and then convert
1750 it into the compressed table form which allows faster lookups
1751 (but cant be modified once converted).
1754 STRLEN transcount = 1;
1756 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1757 "%*sCompiling trie using list compiler\n",
1758 (int)depth * 2 + 2, ""));
1760 trie->states = (reg_trie_state *)
1761 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1762 sizeof(reg_trie_state) );
1766 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1768 regnode *noper = NEXTOPER( cur );
1769 U8 *uc = (U8*)STRING( noper );
1770 const U8 *e = uc + STR_LEN( noper );
1771 U32 state = 1; /* required init */
1772 U16 charid = 0; /* sanity init */
1773 U8 *scan = (U8*)NULL; /* sanity init */
1774 STRLEN foldlen = 0; /* required init */
1775 U32 wordlen = 0; /* required init */
1776 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1779 if (OP(noper) == NOTHING) {
1780 regnode *noper_next= regnext(noper);
1781 if (noper_next != tail && OP(noper_next) == flags) {
1783 uc= (U8*)STRING(noper);
1784 e= uc + STR_LEN(noper);
1788 if (OP(noper) != NOTHING) {
1789 for ( ; uc < e ; uc += len ) {
1794 charid = trie->charmap[ uvc ];
1796 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1800 charid=(U16)SvIV( *svpp );
1803 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1810 if ( !trie->states[ state ].trans.list ) {
1811 TRIE_LIST_NEW( state );
1813 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1814 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1815 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1820 newstate = next_alloc++;
1821 prev_states[newstate] = state;
1822 TRIE_LIST_PUSH( state, charid, newstate );
1827 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1831 TRIE_HANDLE_WORD(state);
1833 } /* end second pass */
1835 /* next alloc is the NEXT state to be allocated */
1836 trie->statecount = next_alloc;
1837 trie->states = (reg_trie_state *)
1838 PerlMemShared_realloc( trie->states,
1840 * sizeof(reg_trie_state) );
1842 /* and now dump it out before we compress it */
1843 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1844 revcharmap, next_alloc,
1848 trie->trans = (reg_trie_trans *)
1849 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1856 for( state=1 ; state < next_alloc ; state ++ ) {
1860 DEBUG_TRIE_COMPILE_MORE_r(
1861 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1865 if (trie->states[state].trans.list) {
1866 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1870 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1871 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1872 if ( forid < minid ) {
1874 } else if ( forid > maxid ) {
1878 if ( transcount < tp + maxid - minid + 1) {
1880 trie->trans = (reg_trie_trans *)
1881 PerlMemShared_realloc( trie->trans,
1883 * sizeof(reg_trie_trans) );
1884 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1886 base = trie->uniquecharcount + tp - minid;
1887 if ( maxid == minid ) {
1889 for ( ; zp < tp ; zp++ ) {
1890 if ( ! trie->trans[ zp ].next ) {
1891 base = trie->uniquecharcount + zp - minid;
1892 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1893 trie->trans[ zp ].check = state;
1899 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1900 trie->trans[ tp ].check = state;
1905 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1906 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1907 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1908 trie->trans[ tid ].check = state;
1910 tp += ( maxid - minid + 1 );
1912 Safefree(trie->states[ state ].trans.list);
1915 DEBUG_TRIE_COMPILE_MORE_r(
1916 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1919 trie->states[ state ].trans.base=base;
1921 trie->lasttrans = tp + 1;
1925 Second Pass -- Flat Table Representation.
1927 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1928 We know that we will need Charcount+1 trans at most to store the data
1929 (one row per char at worst case) So we preallocate both structures
1930 assuming worst case.
1932 We then construct the trie using only the .next slots of the entry
1935 We use the .check field of the first entry of the node temporarily to
1936 make compression both faster and easier by keeping track of how many non
1937 zero fields are in the node.
1939 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1942 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1943 number representing the first entry of the node, and state as a
1944 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1945 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1946 are 2 entrys per node. eg:
1954 The table is internally in the right hand, idx form. However as we also
1955 have to deal with the states array which is indexed by nodenum we have to
1956 use TRIE_NODENUM() to convert.
1959 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1960 "%*sCompiling trie using table compiler\n",
1961 (int)depth * 2 + 2, ""));
1963 trie->trans = (reg_trie_trans *)
1964 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1965 * trie->uniquecharcount + 1,
1966 sizeof(reg_trie_trans) );
1967 trie->states = (reg_trie_state *)
1968 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1969 sizeof(reg_trie_state) );
1970 next_alloc = trie->uniquecharcount + 1;
1973 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1975 regnode *noper = NEXTOPER( cur );
1976 const U8 *uc = (U8*)STRING( noper );
1977 const U8 *e = uc + STR_LEN( noper );
1979 U32 state = 1; /* required init */
1981 U16 charid = 0; /* sanity init */
1982 U32 accept_state = 0; /* sanity init */
1983 U8 *scan = (U8*)NULL; /* sanity init */
1985 STRLEN foldlen = 0; /* required init */
1986 U32 wordlen = 0; /* required init */
1988 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1990 if (OP(noper) == NOTHING) {
1991 regnode *noper_next= regnext(noper);
1992 if (noper_next != tail && OP(noper_next) == flags) {
1994 uc= (U8*)STRING(noper);
1995 e= uc + STR_LEN(noper);
1999 if ( OP(noper) != NOTHING ) {
2000 for ( ; uc < e ; uc += len ) {
2005 charid = trie->charmap[ uvc ];
2007 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2008 charid = svpp ? (U16)SvIV(*svpp) : 0;
2012 if ( !trie->trans[ state + charid ].next ) {
2013 trie->trans[ state + charid ].next = next_alloc;
2014 trie->trans[ state ].check++;
2015 prev_states[TRIE_NODENUM(next_alloc)]
2016 = TRIE_NODENUM(state);
2017 next_alloc += trie->uniquecharcount;
2019 state = trie->trans[ state + charid ].next;
2021 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2023 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2026 accept_state = TRIE_NODENUM( state );
2027 TRIE_HANDLE_WORD(accept_state);
2029 } /* end second pass */
2031 /* and now dump it out before we compress it */
2032 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2034 next_alloc, depth+1));
2038 * Inplace compress the table.*
2040 For sparse data sets the table constructed by the trie algorithm will
2041 be mostly 0/FAIL transitions or to put it another way mostly empty.
2042 (Note that leaf nodes will not contain any transitions.)
2044 This algorithm compresses the tables by eliminating most such
2045 transitions, at the cost of a modest bit of extra work during lookup:
2047 - Each states[] entry contains a .base field which indicates the
2048 index in the state[] array wheres its transition data is stored.
2050 - If .base is 0 there are no valid transitions from that node.
2052 - If .base is nonzero then charid is added to it to find an entry in
2055 -If trans[states[state].base+charid].check!=state then the
2056 transition is taken to be a 0/Fail transition. Thus if there are fail
2057 transitions at the front of the node then the .base offset will point
2058 somewhere inside the previous nodes data (or maybe even into a node
2059 even earlier), but the .check field determines if the transition is
2063 The following process inplace converts the table to the compressed
2064 table: We first do not compress the root node 1,and mark all its
2065 .check pointers as 1 and set its .base pointer as 1 as well. This
2066 allows us to do a DFA construction from the compressed table later,
2067 and ensures that any .base pointers we calculate later are greater
2070 - We set 'pos' to indicate the first entry of the second node.
2072 - We then iterate over the columns of the node, finding the first and
2073 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2074 and set the .check pointers accordingly, and advance pos
2075 appropriately and repreat for the next node. Note that when we copy
2076 the next pointers we have to convert them from the original
2077 NODEIDX form to NODENUM form as the former is not valid post
2080 - If a node has no transitions used we mark its base as 0 and do not
2081 advance the pos pointer.
2083 - If a node only has one transition we use a second pointer into the
2084 structure to fill in allocated fail transitions from other states.
2085 This pointer is independent of the main pointer and scans forward
2086 looking for null transitions that are allocated to a state. When it
2087 finds one it writes the single transition into the "hole". If the
2088 pointer doesnt find one the single transition is appended as normal.
2090 - Once compressed we can Renew/realloc the structures to release the
2093 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2094 specifically Fig 3.47 and the associated pseudocode.
2098 const U32 laststate = TRIE_NODENUM( next_alloc );
2101 trie->statecount = laststate;
2103 for ( state = 1 ; state < laststate ; state++ ) {
2105 const U32 stateidx = TRIE_NODEIDX( state );
2106 const U32 o_used = trie->trans[ stateidx ].check;
2107 U32 used = trie->trans[ stateidx ].check;
2108 trie->trans[ stateidx ].check = 0;
2110 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2111 if ( flag || trie->trans[ stateidx + charid ].next ) {
2112 if ( trie->trans[ stateidx + charid ].next ) {
2114 for ( ; zp < pos ; zp++ ) {
2115 if ( ! trie->trans[ zp ].next ) {
2119 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2120 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2121 trie->trans[ zp ].check = state;
2122 if ( ++zp > pos ) pos = zp;
2129 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2131 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2132 trie->trans[ pos ].check = state;
2137 trie->lasttrans = pos + 1;
2138 trie->states = (reg_trie_state *)
2139 PerlMemShared_realloc( trie->states, laststate
2140 * sizeof(reg_trie_state) );
2141 DEBUG_TRIE_COMPILE_MORE_r(
2142 PerlIO_printf( Perl_debug_log,
2143 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2144 (int)depth * 2 + 2,"",
2145 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2148 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2151 } /* end table compress */
2153 DEBUG_TRIE_COMPILE_MORE_r(
2154 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2155 (int)depth * 2 + 2, "",
2156 (UV)trie->statecount,
2157 (UV)trie->lasttrans)
2159 /* resize the trans array to remove unused space */
2160 trie->trans = (reg_trie_trans *)
2161 PerlMemShared_realloc( trie->trans, trie->lasttrans
2162 * sizeof(reg_trie_trans) );
2164 { /* Modify the program and insert the new TRIE node */
2165 U8 nodetype =(U8)(flags & 0xFF);
2169 regnode *optimize = NULL;
2170 #ifdef RE_TRACK_PATTERN_OFFSETS
2173 U32 mjd_nodelen = 0;
2174 #endif /* RE_TRACK_PATTERN_OFFSETS */
2175 #endif /* DEBUGGING */
2177 This means we convert either the first branch or the first Exact,
2178 depending on whether the thing following (in 'last') is a branch
2179 or not and whther first is the startbranch (ie is it a sub part of
2180 the alternation or is it the whole thing.)
2181 Assuming its a sub part we convert the EXACT otherwise we convert
2182 the whole branch sequence, including the first.
2184 /* Find the node we are going to overwrite */
2185 if ( first != startbranch || OP( last ) == BRANCH ) {
2186 /* branch sub-chain */
2187 NEXT_OFF( first ) = (U16)(last - first);
2188 #ifdef RE_TRACK_PATTERN_OFFSETS
2190 mjd_offset= Node_Offset((convert));
2191 mjd_nodelen= Node_Length((convert));
2194 /* whole branch chain */
2196 #ifdef RE_TRACK_PATTERN_OFFSETS
2199 const regnode *nop = NEXTOPER( convert );
2200 mjd_offset= Node_Offset((nop));
2201 mjd_nodelen= Node_Length((nop));
2205 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2206 (int)depth * 2 + 2, "",
2207 (UV)mjd_offset, (UV)mjd_nodelen)
2210 /* But first we check to see if there is a common prefix we can
2211 split out as an EXACT and put in front of the TRIE node. */
2212 trie->startstate= 1;
2213 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2215 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2219 const U32 base = trie->states[ state ].trans.base;
2221 if ( trie->states[state].wordnum )
2224 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2225 if ( ( base + ofs >= trie->uniquecharcount ) &&
2226 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2227 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2229 if ( ++count > 1 ) {
2230 SV **tmp = av_fetch( revcharmap, ofs, 0);
2231 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2232 if ( state == 1 ) break;
2234 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2236 PerlIO_printf(Perl_debug_log,
2237 "%*sNew Start State=%"UVuf" Class: [",
2238 (int)depth * 2 + 2, "",
2241 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2242 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2244 TRIE_BITMAP_SET(trie,*ch);
2246 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2248 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2252 TRIE_BITMAP_SET(trie,*ch);
2254 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2255 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2261 SV **tmp = av_fetch( revcharmap, idx, 0);
2263 char *ch = SvPV( *tmp, len );
2265 SV *sv=sv_newmortal();
2266 PerlIO_printf( Perl_debug_log,
2267 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2268 (int)depth * 2 + 2, "",
2270 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2271 PL_colors[0], PL_colors[1],
2272 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2273 PERL_PV_ESCAPE_FIRSTCHAR
2278 OP( convert ) = nodetype;
2279 str=STRING(convert);
2282 STR_LEN(convert) += len;
2288 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2293 trie->prefixlen = (state-1);
2295 regnode *n = convert+NODE_SZ_STR(convert);
2296 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2297 trie->startstate = state;
2298 trie->minlen -= (state - 1);
2299 trie->maxlen -= (state - 1);
2301 /* At least the UNICOS C compiler choked on this
2302 * being argument to DEBUG_r(), so let's just have
2305 #ifdef PERL_EXT_RE_BUILD
2311 regnode *fix = convert;
2312 U32 word = trie->wordcount;
2314 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2315 while( ++fix < n ) {
2316 Set_Node_Offset_Length(fix, 0, 0);
2319 SV ** const tmp = av_fetch( trie_words, word, 0 );
2321 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2322 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2324 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2332 NEXT_OFF(convert) = (U16)(tail - convert);
2333 DEBUG_r(optimize= n);
2339 if ( trie->maxlen ) {
2340 NEXT_OFF( convert ) = (U16)(tail - convert);
2341 ARG_SET( convert, data_slot );
2342 /* Store the offset to the first unabsorbed branch in
2343 jump[0], which is otherwise unused by the jump logic.
2344 We use this when dumping a trie and during optimisation. */
2346 trie->jump[0] = (U16)(nextbranch - convert);
2348 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2349 * and there is a bitmap
2350 * and the first "jump target" node we found leaves enough room
2351 * then convert the TRIE node into a TRIEC node, with the bitmap
2352 * embedded inline in the opcode - this is hypothetically faster.
2354 if ( !trie->states[trie->startstate].wordnum
2356 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2358 OP( convert ) = TRIEC;
2359 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2360 PerlMemShared_free(trie->bitmap);
2363 OP( convert ) = TRIE;
2365 /* store the type in the flags */
2366 convert->flags = nodetype;
2370 + regarglen[ OP( convert ) ];
2372 /* XXX We really should free up the resource in trie now,
2373 as we won't use them - (which resources?) dmq */
2375 /* needed for dumping*/
2376 DEBUG_r(if (optimize) {
2377 regnode *opt = convert;
2379 while ( ++opt < optimize) {
2380 Set_Node_Offset_Length(opt,0,0);
2383 Try to clean up some of the debris left after the
2386 while( optimize < jumper ) {
2387 mjd_nodelen += Node_Length((optimize));
2388 OP( optimize ) = OPTIMIZED;
2389 Set_Node_Offset_Length(optimize,0,0);
2392 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2394 } /* end node insert */
2396 /* Finish populating the prev field of the wordinfo array. Walk back
2397 * from each accept state until we find another accept state, and if
2398 * so, point the first word's .prev field at the second word. If the
2399 * second already has a .prev field set, stop now. This will be the
2400 * case either if we've already processed that word's accept state,
2401 * or that state had multiple words, and the overspill words were
2402 * already linked up earlier.
2409 for (word=1; word <= trie->wordcount; word++) {
2411 if (trie->wordinfo[word].prev)
2413 state = trie->wordinfo[word].accept;
2415 state = prev_states[state];
2418 prev = trie->states[state].wordnum;
2422 trie->wordinfo[word].prev = prev;
2424 Safefree(prev_states);
2428 /* and now dump out the compressed format */
2429 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2431 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2433 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2434 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2436 SvREFCNT_dec(revcharmap);
2440 : trie->startstate>1
2446 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2448 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2450 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2451 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2454 We find the fail state for each state in the trie, this state is the longest proper
2455 suffix of the current state's 'word' that is also a proper prefix of another word in our
2456 trie. State 1 represents the word '' and is thus the default fail state. This allows
2457 the DFA not to have to restart after its tried and failed a word at a given point, it
2458 simply continues as though it had been matching the other word in the first place.
2460 'abcdgu'=~/abcdefg|cdgu/
2461 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2462 fail, which would bring us to the state representing 'd' in the second word where we would
2463 try 'g' and succeed, proceeding to match 'cdgu'.
2465 /* add a fail transition */
2466 const U32 trie_offset = ARG(source);
2467 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2469 const U32 ucharcount = trie->uniquecharcount;
2470 const U32 numstates = trie->statecount;
2471 const U32 ubound = trie->lasttrans + ucharcount;
2475 U32 base = trie->states[ 1 ].trans.base;
2478 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2479 GET_RE_DEBUG_FLAGS_DECL;
2481 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2483 PERL_UNUSED_ARG(depth);
2487 ARG_SET( stclass, data_slot );
2488 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2489 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2490 aho->trie=trie_offset;
2491 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2492 Copy( trie->states, aho->states, numstates, reg_trie_state );
2493 Newxz( q, numstates, U32);
2494 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2497 /* initialize fail[0..1] to be 1 so that we always have
2498 a valid final fail state */
2499 fail[ 0 ] = fail[ 1 ] = 1;
2501 for ( charid = 0; charid < ucharcount ; charid++ ) {
2502 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2504 q[ q_write ] = newstate;
2505 /* set to point at the root */
2506 fail[ q[ q_write++ ] ]=1;
2509 while ( q_read < q_write) {
2510 const U32 cur = q[ q_read++ % numstates ];
2511 base = trie->states[ cur ].trans.base;
2513 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2514 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2516 U32 fail_state = cur;
2519 fail_state = fail[ fail_state ];
2520 fail_base = aho->states[ fail_state ].trans.base;
2521 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2523 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2524 fail[ ch_state ] = fail_state;
2525 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2527 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2529 q[ q_write++ % numstates] = ch_state;
2533 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2534 when we fail in state 1, this allows us to use the
2535 charclass scan to find a valid start char. This is based on the principle
2536 that theres a good chance the string being searched contains lots of stuff
2537 that cant be a start char.
2539 fail[ 0 ] = fail[ 1 ] = 0;
2540 DEBUG_TRIE_COMPILE_r({
2541 PerlIO_printf(Perl_debug_log,
2542 "%*sStclass Failtable (%"UVuf" states): 0",
2543 (int)(depth * 2), "", (UV)numstates
2545 for( q_read=1; q_read<numstates; q_read++ ) {
2546 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2548 PerlIO_printf(Perl_debug_log, "\n");
2551 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2556 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2557 * These need to be revisited when a newer toolchain becomes available.
2559 #if defined(__sparc64__) && defined(__GNUC__)
2560 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2561 # undef SPARC64_GCC_WORKAROUND
2562 # define SPARC64_GCC_WORKAROUND 1
2566 #define DEBUG_PEEP(str,scan,depth) \
2567 DEBUG_OPTIMISE_r({if (scan){ \
2568 SV * const mysv=sv_newmortal(); \
2569 regnode *Next = regnext(scan); \
2570 regprop(RExC_rx, mysv, scan); \
2571 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2572 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2573 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2577 /* The below joins as many adjacent EXACTish nodes as possible into a single
2578 * one, and looks for problematic sequences of characters whose folds vs.
2579 * non-folds have sufficiently different lengths, that the optimizer would be
2580 * fooled into rejecting legitimate matches of them, and the trie construction
2581 * code can't cope with them. The joining is only done if:
2582 * 1) there is room in the current conglomerated node to entirely contain the
2584 * 2) they are the exact same node type
2586 * The adjacent nodes actually may be separated by NOTHING kind nodes, and
2587 * these get optimized out
2589 * If there are problematic code sequences, *min_subtract is set to the delta
2590 * that the minimum size of the node can be less than its actual size. And,
2591 * the node type of the result is changed to reflect that it contains these
2594 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2595 * and contains LATIN SMALL LETTER SHARP S
2597 * This is as good a place as any to discuss the design of handling these
2598 * problematic sequences. It's been wrong in Perl for a very long time. There
2599 * are three code points in Unicode whose folded lengths differ so much from
2600 * the un-folded lengths that it causes problems for the optimizer and trie
2601 * construction. Why only these are problematic, and not others where lengths
2602 * also differ is something I (khw) do not understand. New versions of Unicode
2603 * might add more such code points. Hopefully the logic in fold_grind.t that
2604 * figures out what to test (in part by verifying that each size-combination
2605 * gets tested) will catch any that do come along, so they can be added to the
2606 * special handling below. The chances of new ones are actually rather small,
2607 * as most, if not all, of the world's scripts that have casefolding have
2608 * already been encoded by Unicode. Also, a number of Unicode's decisions were
2609 * made to allow compatibility with pre-existing standards, and almost all of
2610 * those have already been dealt with. These would otherwise be the most
2611 * likely candidates for generating further tricky sequences. In other words,
2612 * Unicode by itself is unlikely to add new ones unless it is for compatibility
2613 * with pre-existing standards, and there aren't many of those left.
2615 * The previous designs for dealing with these involved assigning a special
2616 * node for them. This approach doesn't work, as evidenced by this example:
2617 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2618 * Both these fold to "sss", but if the pattern is parsed to create a node of
2619 * that would match just the \xDF, it won't be able to handle the case where a
2620 * successful match would have to cross the node's boundary. The new approach
2621 * that hopefully generally solves the problem generates an EXACTFU_SS node
2624 * There are a number of components to the approach (a lot of work for just
2625 * three code points!):
2626 * 1) This routine examines each EXACTFish node that could contain the
2627 * problematic sequences. It returns in *min_subtract how much to
2628 * subtract from the the actual length of the string to get a real minimum
2629 * for one that could match it. This number is usually 0 except for the
2630 * problematic sequences. This delta is used by the caller to adjust the
2631 * min length of the match, and the delta between min and max, so that the
2632 * optimizer doesn't reject these possibilities based on size constraints.
2633 * 2) These sequences are not currently correctly handled by the trie code
2634 * either, so it changes the joined node type to ops that are not handled
2635 * by trie's, those new ops being EXACTFU_SS and EXACTFU_TRICKYFOLD.
2636 * 3) This is sufficient for the two Greek sequences (described below), but
2637 * the one involving the Sharp s (\xDF) needs more. The node type
2638 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2639 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2640 * case where there is a possible fold length change. That means that a
2641 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2642 * itself with length changes, and so can be processed faster. regexec.c
2643 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2644 * is pre-folded by regcomp.c. This saves effort in regex matching.
2645 * However, probably mostly for historical reasons, the pre-folding isn't
2646 * done for non-UTF8 patterns (and it can't be for EXACTF and EXACTFL
2647 * nodes, as what they fold to isn't known until runtime.) The fold
2648 * possibilities for the non-UTF8 patterns are quite simple, except for
2649 * the sharp s. All the ones that don't involve a UTF-8 target string
2650 * are members of a fold-pair, and arrays are set up for all of them
2651 * that quickly find the other member of the pair. It might actually
2652 * be faster to pre-fold these, but it isn't currently done, except for
2653 * the sharp s. Code elsewhere in this file makes sure that it gets
2654 * folded to 'ss', even if the pattern isn't UTF-8. This avoids the
2655 * issues described in the next item.
2656 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2657 * 'ss' or not is not knowable at compile time. It will match iff the
2658 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2659 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2660 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2661 * described in item 3). An assumption that the optimizer part of
2662 * regexec.c (probably unwittingly) makes is that a character in the
2663 * pattern corresponds to at most a single character in the target string.
2664 * (And I do mean character, and not byte here, unlike other parts of the
2665 * documentation that have never been updated to account for multibyte
2666 * Unicode.) This assumption is wrong only in this case, as all other
2667 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2668 * virtue of having this file pre-fold UTF-8 patterns. I'm
2669 * reluctant to try to change this assumption, so instead the code punts.
2670 * This routine examines EXACTF nodes for the sharp s, and returns a
2671 * boolean indicating whether or not the node is an EXACTF node that
2672 * contains a sharp s. When it is true, the caller sets a flag that later
2673 * causes the optimizer in this file to not set values for the floating
2674 * and fixed string lengths, and thus avoids the optimizer code in
2675 * regexec.c that makes the invalid assumption. Thus, there is no
2676 * optimization based on string lengths for EXACTF nodes that contain the
2677 * sharp s. This only happens for /id rules (which means the pattern
2681 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2682 if (PL_regkind[OP(scan)] == EXACT) \
2683 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2686 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) {
2687 /* Merge several consecutive EXACTish nodes into one. */
2688 regnode *n = regnext(scan);
2690 regnode *next = scan + NODE_SZ_STR(scan);
2694 regnode *stop = scan;
2695 GET_RE_DEBUG_FLAGS_DECL;
2697 PERL_UNUSED_ARG(depth);
2700 PERL_ARGS_ASSERT_JOIN_EXACT;
2701 #ifndef EXPERIMENTAL_INPLACESCAN
2702 PERL_UNUSED_ARG(flags);
2703 PERL_UNUSED_ARG(val);
2705 DEBUG_PEEP("join",scan,depth);
2707 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2708 * EXACT ones that are mergeable to the current one. */
2710 && (PL_regkind[OP(n)] == NOTHING
2711 || (stringok && OP(n) == OP(scan)))
2713 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2716 if (OP(n) == TAIL || n > next)
2718 if (PL_regkind[OP(n)] == NOTHING) {
2719 DEBUG_PEEP("skip:",n,depth);
2720 NEXT_OFF(scan) += NEXT_OFF(n);
2721 next = n + NODE_STEP_REGNODE;
2728 else if (stringok) {
2729 const unsigned int oldl = STR_LEN(scan);
2730 regnode * const nnext = regnext(n);
2732 if (oldl + STR_LEN(n) > U8_MAX)
2735 DEBUG_PEEP("merg",n,depth);
2738 NEXT_OFF(scan) += NEXT_OFF(n);
2739 STR_LEN(scan) += STR_LEN(n);
2740 next = n + NODE_SZ_STR(n);
2741 /* Now we can overwrite *n : */
2742 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2750 #ifdef EXPERIMENTAL_INPLACESCAN
2751 if (flags && !NEXT_OFF(n)) {
2752 DEBUG_PEEP("atch", val, depth);
2753 if (reg_off_by_arg[OP(n)]) {
2754 ARG_SET(n, val - n);
2757 NEXT_OFF(n) = val - n;
2765 *has_exactf_sharp_s = FALSE;
2767 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2768 * can now analyze for sequences of problematic code points. (Prior to
2769 * this final joining, sequences could have been split over boundaries, and
2770 * hence missed). The sequences only happen in folding, hence for any
2771 * non-EXACT EXACTish node */
2772 if (OP(scan) != EXACT) {
2774 U8 * s0 = (U8*) STRING(scan);
2775 U8 * const s_end = s0 + STR_LEN(scan);
2777 /* The below is perhaps overboard, but this allows us to save a test
2778 * each time through the loop at the expense of a mask. This is
2779 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2780 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2781 * This uses an exclusive 'or' to find that bit and then inverts it to
2782 * form a mask, with just a single 0, in the bit position where 'S' and
2784 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2785 const U8 s_masked = 's' & S_or_s_mask;
2787 /* One pass is made over the node's string looking for all the
2788 * possibilities. to avoid some tests in the loop, there are two main
2789 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2793 /* There are two problematic Greek code points in Unicode
2796 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2797 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2803 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2804 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2806 * This means that in case-insensitive matching (or "loose
2807 * matching", as Unicode calls it), an EXACTF of length six (the
2808 * UTF-8 encoded byte length of the above casefolded versions) can
2809 * match a target string of length two (the byte length of UTF-8
2810 * encoded U+0390 or U+03B0). This would rather mess up the
2811 * minimum length computation. (there are other code points that
2812 * also fold to these two sequences, but the delta is smaller)
2814 * If these sequences are found, the minimum length is decreased by
2815 * four (six minus two).
2817 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2818 * LETTER SHARP S. We decrease the min length by 1 for each
2819 * occurrence of 'ss' found */
2821 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2822 # define U390_first_byte 0xb4
2823 const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
2824 # define U3B0_first_byte 0xb5
2825 const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
2827 # define U390_first_byte 0xce
2828 const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
2829 # define U3B0_first_byte 0xcf
2830 const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
2832 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2833 yields a net of 0 */
2834 /* Examine the string for one of the problematic sequences */
2836 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2837 * sequence we are looking for is 2 */
2841 /* Look for the first byte in each problematic sequence */
2843 /* We don't have to worry about other things that fold to
2844 * 's' (such as the long s, U+017F), as all above-latin1
2845 * code points have been pre-folded */
2849 /* Current character is an 's' or 'S'. If next one is
2850 * as well, we have the dreaded sequence */
2851 if (((*(s+1) & S_or_s_mask) == s_masked)
2852 /* These two node types don't have special handling
2854 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2857 OP(scan) = EXACTFU_SS;
2858 s++; /* No need to look at this character again */
2862 case U390_first_byte:
2863 if (s_end - s >= len
2865 /* The 1's are because are skipping comparing the
2867 && memEQ(s + 1, U390_tail, len - 1))
2869 goto greek_sequence;
2873 case U3B0_first_byte:
2874 if (! (s_end - s >= len
2875 && memEQ(s + 1, U3B0_tail, len - 1)))
2882 /* This can't currently be handled by trie's, so change
2883 * the node type to indicate this. If EXACTFA and
2884 * EXACTFL were ever to be handled by trie's, this
2885 * would have to be changed. If this node has already
2886 * been changed to EXACTFU_SS in this loop, leave it as
2887 * is. (I (khw) think it doesn't matter in regexec.c
2888 * for UTF patterns, but no need to change it */
2889 if (OP(scan) == EXACTFU) {
2890 OP(scan) = EXACTFU_TRICKYFOLD;
2892 s += 6; /* We already know what this sequence is. Skip
2898 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2900 /* Here, the pattern is not UTF-8. We need to look only for the
2901 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2902 * in the final position. Otherwise we can stop looking 1 byte
2903 * earlier because have to find both the first and second 's' */
2904 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2906 for (s = s0; s < upper; s++) {
2911 && ((*(s+1) & S_or_s_mask) == s_masked))
2915 /* EXACTF nodes need to know that the minimum
2916 * length changed so that a sharp s in the string
2917 * can match this ss in the pattern, but they
2918 * remain EXACTF nodes, as they are not trie'able,
2919 * so don't have to invent a new node type to
2920 * exclude them from the trie code */
2921 if (OP(scan) != EXACTF) {
2922 OP(scan) = EXACTFU_SS;
2927 case LATIN_SMALL_LETTER_SHARP_S:
2928 if (OP(scan) == EXACTF) {
2929 *has_exactf_sharp_s = TRUE;
2938 /* Allow dumping but overwriting the collection of skipped
2939 * ops and/or strings with fake optimized ops */
2940 n = scan + NODE_SZ_STR(scan);
2948 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2952 /* REx optimizer. Converts nodes into quicker variants "in place".
2953 Finds fixed substrings. */
2955 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2956 to the position after last scanned or to NULL. */
2958 #define INIT_AND_WITHP \
2959 assert(!and_withp); \
2960 Newx(and_withp,1,struct regnode_charclass_class); \
2961 SAVEFREEPV(and_withp)
2963 /* this is a chain of data about sub patterns we are processing that
2964 need to be handled separately/specially in study_chunk. Its so
2965 we can simulate recursion without losing state. */
2967 typedef struct scan_frame {
2968 regnode *last; /* last node to process in this frame */
2969 regnode *next; /* next node to process when last is reached */
2970 struct scan_frame *prev; /*previous frame*/
2971 I32 stop; /* what stopparen do we use */
2975 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2977 #define CASE_SYNST_FNC(nAmE) \
2979 if (flags & SCF_DO_STCLASS_AND) { \
2980 for (value = 0; value < 256; value++) \
2981 if (!is_ ## nAmE ## _cp(value)) \
2982 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2985 for (value = 0; value < 256; value++) \
2986 if (is_ ## nAmE ## _cp(value)) \
2987 ANYOF_BITMAP_SET(data->start_class, value); \
2991 if (flags & SCF_DO_STCLASS_AND) { \
2992 for (value = 0; value < 256; value++) \
2993 if (is_ ## nAmE ## _cp(value)) \
2994 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2997 for (value = 0; value < 256; value++) \
2998 if (!is_ ## nAmE ## _cp(value)) \
2999 ANYOF_BITMAP_SET(data->start_class, value); \
3006 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3007 I32 *minlenp, I32 *deltap,
3012 struct regnode_charclass_class *and_withp,
3013 U32 flags, U32 depth)
3014 /* scanp: Start here (read-write). */
3015 /* deltap: Write maxlen-minlen here. */
3016 /* last: Stop before this one. */
3017 /* data: string data about the pattern */
3018 /* stopparen: treat close N as END */
3019 /* recursed: which subroutines have we recursed into */
3020 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3023 I32 min = 0, pars = 0, code;
3024 regnode *scan = *scanp, *next;
3026 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3027 int is_inf_internal = 0; /* The studied chunk is infinite */
3028 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3029 scan_data_t data_fake;
3030 SV *re_trie_maxbuff = NULL;
3031 regnode *first_non_open = scan;
3032 I32 stopmin = I32_MAX;
3033 scan_frame *frame = NULL;
3034 GET_RE_DEBUG_FLAGS_DECL;
3036 PERL_ARGS_ASSERT_STUDY_CHUNK;
3039 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3043 while (first_non_open && OP(first_non_open) == OPEN)
3044 first_non_open=regnext(first_non_open);
3049 while ( scan && OP(scan) != END && scan < last ){
3050 UV min_subtract = 0; /* How much to subtract from the minimum node
3051 length to get a real minimum (because the
3052 folded version may be shorter) */
3053 bool has_exactf_sharp_s = FALSE;
3054 /* Peephole optimizer: */
3055 DEBUG_STUDYDATA("Peep:", data,depth);
3056 DEBUG_PEEP("Peep",scan,depth);
3058 /* Its not clear to khw or hv why this is done here, and not in the
3059 * clauses that deal with EXACT nodes. khw's guess is that it's
3060 * because of a previous design */
3061 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3063 /* Follow the next-chain of the current node and optimize
3064 away all the NOTHINGs from it. */
3065 if (OP(scan) != CURLYX) {
3066 const int max = (reg_off_by_arg[OP(scan)]
3068 /* I32 may be smaller than U16 on CRAYs! */
3069 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3070 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3074 /* Skip NOTHING and LONGJMP. */
3075 while ((n = regnext(n))
3076 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3077 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3078 && off + noff < max)
3080 if (reg_off_by_arg[OP(scan)])
3083 NEXT_OFF(scan) = off;
3088 /* The principal pseudo-switch. Cannot be a switch, since we
3089 look into several different things. */
3090 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3091 || OP(scan) == IFTHEN) {
3092 next = regnext(scan);
3094 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3096 if (OP(next) == code || code == IFTHEN) {
3097 /* NOTE - There is similar code to this block below for handling
3098 TRIE nodes on a re-study. If you change stuff here check there
3100 I32 max1 = 0, min1 = I32_MAX, num = 0;
3101 struct regnode_charclass_class accum;
3102 regnode * const startbranch=scan;
3104 if (flags & SCF_DO_SUBSTR)
3105 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3106 if (flags & SCF_DO_STCLASS)
3107 cl_init_zero(pRExC_state, &accum);
3109 while (OP(scan) == code) {
3110 I32 deltanext, minnext, f = 0, fake;
3111 struct regnode_charclass_class this_class;
3114 data_fake.flags = 0;
3116 data_fake.whilem_c = data->whilem_c;
3117 data_fake.last_closep = data->last_closep;
3120 data_fake.last_closep = &fake;
3122 data_fake.pos_delta = delta;
3123 next = regnext(scan);
3124 scan = NEXTOPER(scan);
3126 scan = NEXTOPER(scan);
3127 if (flags & SCF_DO_STCLASS) {
3128 cl_init(pRExC_state, &this_class);
3129 data_fake.start_class = &this_class;
3130 f = SCF_DO_STCLASS_AND;
3132 if (flags & SCF_WHILEM_VISITED_POS)
3133 f |= SCF_WHILEM_VISITED_POS;
3135 /* we suppose the run is continuous, last=next...*/
3136 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3138 stopparen, recursed, NULL, f,depth+1);
3141 if (max1 < minnext + deltanext)
3142 max1 = minnext + deltanext;
3143 if (deltanext == I32_MAX)
3144 is_inf = is_inf_internal = 1;
3146 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3148 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3149 if ( stopmin > minnext)
3150 stopmin = min + min1;
3151 flags &= ~SCF_DO_SUBSTR;
3153 data->flags |= SCF_SEEN_ACCEPT;
3156 if (data_fake.flags & SF_HAS_EVAL)
3157 data->flags |= SF_HAS_EVAL;
3158 data->whilem_c = data_fake.whilem_c;
3160 if (flags & SCF_DO_STCLASS)
3161 cl_or(pRExC_state, &accum, &this_class);
3163 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3165 if (flags & SCF_DO_SUBSTR) {
3166 data->pos_min += min1;
3167 data->pos_delta += max1 - min1;
3168 if (max1 != min1 || is_inf)
3169 data->longest = &(data->longest_float);
3172 delta += max1 - min1;
3173 if (flags & SCF_DO_STCLASS_OR) {
3174 cl_or(pRExC_state, data->start_class, &accum);
3176 cl_and(data->start_class, and_withp);
3177 flags &= ~SCF_DO_STCLASS;
3180 else if (flags & SCF_DO_STCLASS_AND) {
3182 cl_and(data->start_class, &accum);
3183 flags &= ~SCF_DO_STCLASS;
3186 /* Switch to OR mode: cache the old value of
3187 * data->start_class */
3189 StructCopy(data->start_class, and_withp,
3190 struct regnode_charclass_class);
3191 flags &= ~SCF_DO_STCLASS_AND;
3192 StructCopy(&accum, data->start_class,
3193 struct regnode_charclass_class);
3194 flags |= SCF_DO_STCLASS_OR;
3195 data->start_class->flags |= ANYOF_EOS;
3199 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3202 Assuming this was/is a branch we are dealing with: 'scan' now
3203 points at the item that follows the branch sequence, whatever
3204 it is. We now start at the beginning of the sequence and look
3211 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3213 If we can find such a subsequence we need to turn the first
3214 element into a trie and then add the subsequent branch exact
3215 strings to the trie.
3219 1. patterns where the whole set of branches can be converted.
3221 2. patterns where only a subset can be converted.
3223 In case 1 we can replace the whole set with a single regop
3224 for the trie. In case 2 we need to keep the start and end
3227 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3228 becomes BRANCH TRIE; BRANCH X;
3230 There is an additional case, that being where there is a
3231 common prefix, which gets split out into an EXACT like node
3232 preceding the TRIE node.
3234 If x(1..n)==tail then we can do a simple trie, if not we make
3235 a "jump" trie, such that when we match the appropriate word
3236 we "jump" to the appropriate tail node. Essentially we turn
3237 a nested if into a case structure of sorts.
3242 if (!re_trie_maxbuff) {
3243 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3244 if (!SvIOK(re_trie_maxbuff))
3245 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3247 if ( SvIV(re_trie_maxbuff)>=0 ) {
3249 regnode *first = (regnode *)NULL;
3250 regnode *last = (regnode *)NULL;
3251 regnode *tail = scan;
3256 SV * const mysv = sv_newmortal(); /* for dumping */
3258 /* var tail is used because there may be a TAIL
3259 regop in the way. Ie, the exacts will point to the
3260 thing following the TAIL, but the last branch will
3261 point at the TAIL. So we advance tail. If we
3262 have nested (?:) we may have to move through several
3266 while ( OP( tail ) == TAIL ) {
3267 /* this is the TAIL generated by (?:) */
3268 tail = regnext( tail );
3272 DEBUG_TRIE_COMPILE_r({
3273 regprop(RExC_rx, mysv, tail );
3274 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3275 (int)depth * 2 + 2, "",
3276 "Looking for TRIE'able sequences. Tail node is: ",
3277 SvPV_nolen_const( mysv )
3283 Step through the branches
3284 cur represents each branch,
3285 noper is the first thing to be matched as part of that branch
3286 noper_next is the regnext() of that node.
3288 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3289 via a "jump trie" but we also support building with NOJUMPTRIE,
3290 which restricts the trie logic to structures like /FOO|BAR/.
3292 If noper is a trieable nodetype then the branch is a possible optimization
3293 target. If we are building under NOJUMPTRIE then we require that noper_next
3294 is the same as scan (our current position in the regex program).
3296 Once we have two or more consecutive such branches we can create a
3297 trie of the EXACT's contents and stitch it in place into the program.
3299 If the sequence represents all of the branches in the alternation we
3300 replace the entire thing with a single TRIE node.
3302 Otherwise when it is a subsequence we need to stitch it in place and
3303 replace only the relevant branches. This means the first branch has
3304 to remain as it is used by the alternation logic, and its next pointer,
3305 and needs to be repointed at the item on the branch chain following
3306 the last branch we have optimized away.
3308 This could be either a BRANCH, in which case the subsequence is internal,
3309 or it could be the item following the branch sequence in which case the
3310 subsequence is at the end (which does not necessarily mean the first node
3311 is the start of the alternation).
3313 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3316 ----------------+-----------
3320 EXACTFU_SS | EXACTFU
3321 EXACTFU_TRICKYFOLD | EXACTFU
3326 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3327 ( EXACT == (X) ) ? EXACT : \
3328 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3331 /* dont use tail as the end marker for this traverse */
3332 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3333 regnode * const noper = NEXTOPER( cur );
3334 U8 noper_type = OP( noper );
3335 U8 noper_trietype = TRIE_TYPE( noper_type );
3336 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3337 regnode * const noper_next = regnext( noper );
3338 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3339 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3342 DEBUG_TRIE_COMPILE_r({
3343 regprop(RExC_rx, mysv, cur);
3344 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3345 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3347 regprop(RExC_rx, mysv, noper);
3348 PerlIO_printf( Perl_debug_log, " -> %s",
3349 SvPV_nolen_const(mysv));
3352 regprop(RExC_rx, mysv, noper_next );
3353 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3354 SvPV_nolen_const(mysv));
3356 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3357 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3358 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3362 /* Is noper a trieable nodetype that can be merged with the
3363 * current trie (if there is one)? */
3367 ( noper_trietype == NOTHING)
3368 || ( trietype == NOTHING )
3369 || ( trietype == noper_trietype )
3372 && noper_next == tail
3376 /* Handle mergable triable node
3377 * Either we are the first node in a new trieable sequence,
3378 * in which case we do some bookkeeping, otherwise we update
3379 * the end pointer. */
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;
3391 } else if (noper_next_type) {
3392 /* a NOTHING regop is 1 regop wide. We need at least two
3393 * for a trie so we can't merge this in */
3397 trietype = noper_trietype;
3400 if ( trietype == NOTHING )
3401 trietype = noper_trietype;
3406 } /* end handle mergable triable node */
3408 /* handle unmergable node -
3409 * noper may either be a triable node which can not be tried
3410 * together with the current trie, or a non triable node */
3412 /* If last is set and trietype is not NOTHING then we have found
3413 * at least two triable branch sequences in a row of a similar
3414 * trietype so we can turn them into a trie. If/when we
3415 * allow NOTHING to start a trie sequence this condition will be
3416 * required, and it isn't expensive so we leave it in for now. */
3417 if ( trietype != NOTHING )
3418 make_trie( pRExC_state,
3419 startbranch, first, cur, tail, count,
3420 trietype, depth+1 );
3421 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3425 && noper_next == tail
3428 /* noper is triable, so we can start a new trie sequence */
3431 trietype = noper_trietype;
3433 /* if we already saw a first but the current node is not triable then we have
3434 * to reset the first information. */
3439 } /* end handle unmergable node */
3440 } /* loop over branches */
3441 DEBUG_TRIE_COMPILE_r({
3442 regprop(RExC_rx, mysv, cur);
3443 PerlIO_printf( Perl_debug_log,
3444 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3445 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3449 if ( trietype != NOTHING ) {
3450 /* the last branch of the sequence was part of a trie,
3451 * so we have to construct it here outside of the loop
3453 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3454 #ifdef TRIE_STUDY_OPT
3455 if ( ((made == MADE_EXACT_TRIE &&
3456 startbranch == first)
3457 || ( first_non_open == first )) &&
3459 flags |= SCF_TRIE_RESTUDY;
3460 if ( startbranch == first
3463 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3468 /* at this point we know whatever we have is a NOTHING sequence/branch
3469 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3471 if ( startbranch == first ) {
3473 /* the entire thing is a NOTHING sequence, something like this:
3474 * (?:|) So we can turn it into a plain NOTHING op. */
3475 DEBUG_TRIE_COMPILE_r({
3476 regprop(RExC_rx, mysv, cur);
3477 PerlIO_printf( Perl_debug_log,
3478 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3479 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3482 OP(startbranch)= NOTHING;
3483 NEXT_OFF(startbranch)= tail - startbranch;
3484 for ( opt= startbranch + 1; opt < tail ; opt++ )
3488 } /* end if ( last) */
3489 } /* TRIE_MAXBUF is non zero */
3494 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3495 scan = NEXTOPER(NEXTOPER(scan));
3496 } else /* single branch is optimized. */
3497 scan = NEXTOPER(scan);
3499 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3500 scan_frame *newframe = NULL;
3505 if (OP(scan) != SUSPEND) {
3506 /* set the pointer */
3507 if (OP(scan) == GOSUB) {
3509 RExC_recurse[ARG2L(scan)] = scan;
3510 start = RExC_open_parens[paren-1];
3511 end = RExC_close_parens[paren-1];
3514 start = RExC_rxi->program + 1;
3518 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3519 SAVEFREEPV(recursed);
3521 if (!PAREN_TEST(recursed,paren+1)) {
3522 PAREN_SET(recursed,paren+1);
3523 Newx(newframe,1,scan_frame);
3525 if (flags & SCF_DO_SUBSTR) {
3526 SCAN_COMMIT(pRExC_state,data,minlenp);
3527 data->longest = &(data->longest_float);
3529 is_inf = is_inf_internal = 1;
3530 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3531 cl_anything(pRExC_state, data->start_class);
3532 flags &= ~SCF_DO_STCLASS;
3535 Newx(newframe,1,scan_frame);
3538 end = regnext(scan);
3543 SAVEFREEPV(newframe);
3544 newframe->next = regnext(scan);
3545 newframe->last = last;
3546 newframe->stop = stopparen;
3547 newframe->prev = frame;
3557 else if (OP(scan) == EXACT) {
3558 I32 l = STR_LEN(scan);
3561 const U8 * const s = (U8*)STRING(scan);
3562 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3563 l = utf8_length(s, s + l);
3565 uc = *((U8*)STRING(scan));
3568 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3569 /* The code below prefers earlier match for fixed
3570 offset, later match for variable offset. */
3571 if (data->last_end == -1) { /* Update the start info. */
3572 data->last_start_min = data->pos_min;
3573 data->last_start_max = is_inf
3574 ? I32_MAX : data->pos_min + data->pos_delta;
3576 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3578 SvUTF8_on(data->last_found);
3580 SV * const sv = data->last_found;
3581 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3582 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3583 if (mg && mg->mg_len >= 0)
3584 mg->mg_len += utf8_length((U8*)STRING(scan),
3585 (U8*)STRING(scan)+STR_LEN(scan));
3587 data->last_end = data->pos_min + l;
3588 data->pos_min += l; /* As in the first entry. */
3589 data->flags &= ~SF_BEFORE_EOL;
3591 if (flags & SCF_DO_STCLASS_AND) {
3592 /* Check whether it is compatible with what we know already! */
3596 /* If compatible, we or it in below. It is compatible if is
3597 * in the bitmp and either 1) its bit or its fold is set, or 2)
3598 * it's for a locale. Even if there isn't unicode semantics
3599 * here, at runtime there may be because of matching against a
3600 * utf8 string, so accept a possible false positive for
3601 * latin1-range folds */
3603 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3604 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3605 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
3606 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3611 ANYOF_CLASS_ZERO(data->start_class);
3612 ANYOF_BITMAP_ZERO(data->start_class);
3614 ANYOF_BITMAP_SET(data->start_class, uc);
3615 else if (uc >= 0x100) {
3618 /* Some Unicode code points fold to the Latin1 range; as
3619 * XXX temporary code, instead of figuring out if this is
3620 * one, just assume it is and set all the start class bits
3621 * that could be some such above 255 code point's fold
3622 * which will generate fals positives. As the code
3623 * elsewhere that does compute the fold settles down, it
3624 * can be extracted out and re-used here */
3625 for (i = 0; i < 256; i++){
3626 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3627 ANYOF_BITMAP_SET(data->start_class, i);
3631 data->start_class->flags &= ~ANYOF_EOS;
3633 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3635 else if (flags & SCF_DO_STCLASS_OR) {
3636 /* false positive possible if the class is case-folded */
3638 ANYOF_BITMAP_SET(data->start_class, uc);
3640 data->start_class->flags |= ANYOF_UNICODE_ALL;
3641 data->start_class->flags &= ~ANYOF_EOS;
3642 cl_and(data->start_class, and_withp);
3644 flags &= ~SCF_DO_STCLASS;
3646 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3647 I32 l = STR_LEN(scan);
3648 UV uc = *((U8*)STRING(scan));
3650 /* Search for fixed substrings supports EXACT only. */
3651 if (flags & SCF_DO_SUBSTR) {
3653 SCAN_COMMIT(pRExC_state, data, minlenp);
3656 const U8 * const s = (U8 *)STRING(scan);
3657 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3658 l = utf8_length(s, s + l);
3660 else if (has_exactf_sharp_s) {
3661 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3663 min += l - min_subtract;
3667 delta += min_subtract;
3668 if (flags & SCF_DO_SUBSTR) {
3669 data->pos_min += l - min_subtract;
3670 if (data->pos_min < 0) {
3673 data->pos_delta += min_subtract;
3675 data->longest = &(data->longest_float);
3678 if (flags & SCF_DO_STCLASS_AND) {
3679 /* Check whether it is compatible with what we know already! */
3682 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3683 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3684 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3688 ANYOF_CLASS_ZERO(data->start_class);
3689 ANYOF_BITMAP_ZERO(data->start_class);
3691 ANYOF_BITMAP_SET(data->start_class, uc);
3692 data->start_class->flags &= ~ANYOF_EOS;
3693 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
3694 if (OP(scan) == EXACTFL) {
3695 /* XXX This set is probably no longer necessary, and
3696 * probably wrong as LOCALE now is on in the initial
3698 data->start_class->flags |= ANYOF_LOCALE;
3702 /* Also set the other member of the fold pair. In case
3703 * that unicode semantics is called for at runtime, use
3704 * the full latin1 fold. (Can't do this for locale,
3705 * because not known until runtime) */
3706 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3708 /* All other (EXACTFL handled above) folds except under
3709 * /iaa that include s, S, and sharp_s also may include
3711 if (OP(scan) != EXACTFA) {
3712 if (uc == 's' || uc == 'S') {
3713 ANYOF_BITMAP_SET(data->start_class,
3714 LATIN_SMALL_LETTER_SHARP_S);
3716 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3717 ANYOF_BITMAP_SET(data->start_class, 's');
3718 ANYOF_BITMAP_SET(data->start_class, 'S');
3723 else if (uc >= 0x100) {
3725 for (i = 0; i < 256; i++){
3726 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3727 ANYOF_BITMAP_SET(data->start_class, i);
3732 else if (flags & SCF_DO_STCLASS_OR) {
3733 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
3734 /* false positive possible if the class is case-folded.
3735 Assume that the locale settings are the same... */
3737 ANYOF_BITMAP_SET(data->start_class, uc);
3738 if (OP(scan) != EXACTFL) {
3740 /* And set the other member of the fold pair, but
3741 * can't do that in locale because not known until
3743 ANYOF_BITMAP_SET(data->start_class,
3744 PL_fold_latin1[uc]);
3746 /* All folds except under /iaa that include s, S,
3747 * and sharp_s also may include the others */
3748 if (OP(scan) != EXACTFA) {
3749 if (uc == 's' || uc == 'S') {
3750 ANYOF_BITMAP_SET(data->start_class,
3751 LATIN_SMALL_LETTER_SHARP_S);
3753 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3754 ANYOF_BITMAP_SET(data->start_class, 's');
3755 ANYOF_BITMAP_SET(data->start_class, 'S');
3760 data->start_class->flags &= ~ANYOF_EOS;
3762 cl_and(data->start_class, and_withp);
3764 flags &= ~SCF_DO_STCLASS;
3766 else if (REGNODE_VARIES(OP(scan))) {
3767 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3768 I32 f = flags, pos_before = 0;
3769 regnode * const oscan = scan;
3770 struct regnode_charclass_class this_class;
3771 struct regnode_charclass_class *oclass = NULL;
3772 I32 next_is_eval = 0;
3774 switch (PL_regkind[OP(scan)]) {
3775 case WHILEM: /* End of (?:...)* . */
3776 scan = NEXTOPER(scan);
3779 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3780 next = NEXTOPER(scan);
3781 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3783 maxcount = REG_INFTY;
3784 next = regnext(scan);
3785 scan = NEXTOPER(scan);
3789 if (flags & SCF_DO_SUBSTR)
3794 if (flags & SCF_DO_STCLASS) {
3796 maxcount = REG_INFTY;
3797 next = regnext(scan);
3798 scan = NEXTOPER(scan);
3801 is_inf = is_inf_internal = 1;
3802 scan = regnext(scan);
3803 if (flags & SCF_DO_SUBSTR) {
3804 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3805 data->longest = &(data->longest_float);
3807 goto optimize_curly_tail;
3809 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3810 && (scan->flags == stopparen))
3815 mincount = ARG1(scan);
3816 maxcount = ARG2(scan);
3818 next = regnext(scan);
3819 if (OP(scan) == CURLYX) {
3820 I32 lp = (data ? *(data->last_closep) : 0);
3821 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3823 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3824 next_is_eval = (OP(scan) == EVAL);
3826 if (flags & SCF_DO_SUBSTR) {
3827 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3828 pos_before = data->pos_min;
https://perl5.git.perl.org / perl5.git / blob