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
112 typedef struct RExC_state_t {
113 U32 flags; /* are we folding, multilining? */
114 char *precomp; /* uncompiled string. */
115 REGEXP *rx_sv; /* The SV that is the regexp. */
116 regexp *rx; /* perl core regexp structure */
117 regexp_internal *rxi; /* internal data for regexp object pprivate field */
118 char *start; /* Start of input for compile */
119 char *end; /* End of input for compile */
120 char *parse; /* Input-scan pointer. */
121 I32 whilem_seen; /* number of WHILEM in this expr */
122 regnode *emit_start; /* Start of emitted-code area */
123 regnode *emit_bound; /* First regnode outside of the allocated space */
124 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
125 I32 naughty; /* How bad is this pattern? */
126 I32 sawback; /* Did we see \1, ...? */
128 I32 size; /* Code size. */
129 I32 npar; /* Capture buffer count, (OPEN). */
130 I32 cpar; /* Capture buffer count, (CLOSE). */
131 I32 nestroot; /* root parens we are in - used by accept */
135 regnode **open_parens; /* pointers to open parens */
136 regnode **close_parens; /* pointers to close parens */
137 regnode *opend; /* END node in program */
138 I32 utf8; /* whether the pattern is utf8 or not */
139 I32 orig_utf8; /* whether the pattern was originally in utf8 */
140 /* XXX use this for future optimisation of case
141 * where pattern must be upgraded to utf8. */
142 I32 uni_semantics; /* If a d charset modifier should use unicode
143 rules, even if the pattern is not in
145 HV *paren_names; /* Paren names */
147 regnode **recurse; /* Recurse regops */
148 I32 recurse_count; /* Number of recurse regops */
151 I32 override_recoding;
153 char *starttry; /* -Dr: where regtry was called. */
154 #define RExC_starttry (pRExC_state->starttry)
157 const char *lastparse;
159 AV *paren_name_list; /* idx -> name */
160 #define RExC_lastparse (pRExC_state->lastparse)
161 #define RExC_lastnum (pRExC_state->lastnum)
162 #define RExC_paren_name_list (pRExC_state->paren_name_list)
166 #define RExC_flags (pRExC_state->flags)
167 #define RExC_precomp (pRExC_state->precomp)
168 #define RExC_rx_sv (pRExC_state->rx_sv)
169 #define RExC_rx (pRExC_state->rx)
170 #define RExC_rxi (pRExC_state->rxi)
171 #define RExC_start (pRExC_state->start)
172 #define RExC_end (pRExC_state->end)
173 #define RExC_parse (pRExC_state->parse)
174 #define RExC_whilem_seen (pRExC_state->whilem_seen)
175 #ifdef RE_TRACK_PATTERN_OFFSETS
176 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
178 #define RExC_emit (pRExC_state->emit)
179 #define RExC_emit_start (pRExC_state->emit_start)
180 #define RExC_emit_bound (pRExC_state->emit_bound)
181 #define RExC_naughty (pRExC_state->naughty)
182 #define RExC_sawback (pRExC_state->sawback)
183 #define RExC_seen (pRExC_state->seen)
184 #define RExC_size (pRExC_state->size)
185 #define RExC_npar (pRExC_state->npar)
186 #define RExC_nestroot (pRExC_state->nestroot)
187 #define RExC_extralen (pRExC_state->extralen)
188 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
189 #define RExC_seen_evals (pRExC_state->seen_evals)
190 #define RExC_utf8 (pRExC_state->utf8)
191 #define RExC_uni_semantics (pRExC_state->uni_semantics)
192 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
193 #define RExC_open_parens (pRExC_state->open_parens)
194 #define RExC_close_parens (pRExC_state->close_parens)
195 #define RExC_opend (pRExC_state->opend)
196 #define RExC_paren_names (pRExC_state->paren_names)
197 #define RExC_recurse (pRExC_state->recurse)
198 #define RExC_recurse_count (pRExC_state->recurse_count)
199 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
200 #define RExC_contains_locale (pRExC_state->contains_locale)
201 #define RExC_override_recoding (pRExC_state->override_recoding)
204 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
205 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
206 ((*s) == '{' && regcurly(s)))
209 #undef SPSTART /* dratted cpp namespace... */
212 * Flags to be passed up and down.
214 #define WORST 0 /* Worst case. */
215 #define HASWIDTH 0x01 /* Known to match non-null strings. */
217 /* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
218 * character, and if utf8, must be invariant. Note that this is not the same
219 * thing as REGNODE_SIMPLE */
221 #define SPSTART 0x04 /* Starts with * or +. */
222 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
223 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
225 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
227 /* whether trie related optimizations are enabled */
228 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
229 #define TRIE_STUDY_OPT
230 #define FULL_TRIE_STUDY
236 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
237 #define PBITVAL(paren) (1 << ((paren) & 7))
238 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
239 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
240 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
242 /* If not already in utf8, do a longjmp back to the beginning */
243 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
244 #define REQUIRE_UTF8 STMT_START { \
245 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
248 /* About scan_data_t.
250 During optimisation we recurse through the regexp program performing
251 various inplace (keyhole style) optimisations. In addition study_chunk
252 and scan_commit populate this data structure with information about
253 what strings MUST appear in the pattern. We look for the longest
254 string that must appear at a fixed location, and we look for the
255 longest string that may appear at a floating location. So for instance
260 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
261 strings (because they follow a .* construct). study_chunk will identify
262 both FOO and BAR as being the longest fixed and floating strings respectively.
264 The strings can be composites, for instance
268 will result in a composite fixed substring 'foo'.
270 For each string some basic information is maintained:
272 - offset or min_offset
273 This is the position the string must appear at, or not before.
274 It also implicitly (when combined with minlenp) tells us how many
275 characters must match before the string we are searching for.
276 Likewise when combined with minlenp and the length of the string it
277 tells us how many characters must appear after the string we have
281 Only used for floating strings. This is the rightmost point that
282 the string can appear at. If set to I32 max it indicates that the
283 string can occur infinitely far to the right.
286 A pointer to the minimum length of the pattern that the string
287 was found inside. This is important as in the case of positive
288 lookahead or positive lookbehind we can have multiple patterns
293 The minimum length of the pattern overall is 3, the minimum length
294 of the lookahead part is 3, but the minimum length of the part that
295 will actually match is 1. So 'FOO's minimum length is 3, but the
296 minimum length for the F is 1. This is important as the minimum length
297 is used to determine offsets in front of and behind the string being
298 looked for. Since strings can be composites this is the length of the
299 pattern at the time it was committed with a scan_commit. Note that
300 the length is calculated by study_chunk, so that the minimum lengths
301 are not known until the full pattern has been compiled, thus the
302 pointer to the value.
306 In the case of lookbehind the string being searched for can be
307 offset past the start point of the final matching string.
308 If this value was just blithely removed from the min_offset it would
309 invalidate some of the calculations for how many chars must match
310 before or after (as they are derived from min_offset and minlen and
311 the length of the string being searched for).
312 When the final pattern is compiled and the data is moved from the
313 scan_data_t structure into the regexp structure the information
314 about lookbehind is factored in, with the information that would
315 have been lost precalculated in the end_shift field for the
318 The fields pos_min and pos_delta are used to store the minimum offset
319 and the delta to the maximum offset at the current point in the pattern.
323 typedef struct scan_data_t {
324 /*I32 len_min; unused */
325 /*I32 len_delta; unused */
329 I32 last_end; /* min value, <0 unless valid. */
332 SV **longest; /* Either &l_fixed, or &l_float. */
333 SV *longest_fixed; /* longest fixed string found in pattern */
334 I32 offset_fixed; /* offset where it starts */
335 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
336 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
337 SV *longest_float; /* longest floating string found in pattern */
338 I32 offset_float_min; /* earliest point in string it can appear */
339 I32 offset_float_max; /* latest point in string it can appear */
340 I32 *minlen_float; /* pointer to the minlen relevant to the string */
341 I32 lookbehind_float; /* is the position of the string modified by LB */
345 struct regnode_charclass_class *start_class;
349 * Forward declarations for pregcomp()'s friends.
352 static const scan_data_t zero_scan_data =
353 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
355 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
356 #define SF_BEFORE_SEOL 0x0001
357 #define SF_BEFORE_MEOL 0x0002
358 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
359 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
362 # define SF_FIX_SHIFT_EOL (0+2)
363 # define SF_FL_SHIFT_EOL (0+4)
365 # define SF_FIX_SHIFT_EOL (+2)
366 # define SF_FL_SHIFT_EOL (+4)
369 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
370 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
372 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
373 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
374 #define SF_IS_INF 0x0040
375 #define SF_HAS_PAR 0x0080
376 #define SF_IN_PAR 0x0100
377 #define SF_HAS_EVAL 0x0200
378 #define SCF_DO_SUBSTR 0x0400
379 #define SCF_DO_STCLASS_AND 0x0800
380 #define SCF_DO_STCLASS_OR 0x1000
381 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
382 #define SCF_WHILEM_VISITED_POS 0x2000
384 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
385 #define SCF_SEEN_ACCEPT 0x8000
387 #define UTF cBOOL(RExC_utf8)
389 /* The enums for all these are ordered so things work out correctly */
390 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
391 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
392 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
393 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
394 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
395 #define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
396 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
398 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
400 #define OOB_UNICODE 12345678
401 #define OOB_NAMEDCLASS -1
403 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
404 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
407 /* length of regex to show in messages that don't mark a position within */
408 #define RegexLengthToShowInErrorMessages 127
411 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
412 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
413 * op/pragma/warn/regcomp.
415 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
416 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
418 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
421 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
422 * arg. Show regex, up to a maximum length. If it's too long, chop and add
425 #define _FAIL(code) STMT_START { \
426 const char *ellipses = ""; \
427 IV len = RExC_end - RExC_precomp; \
430 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
431 if (len > RegexLengthToShowInErrorMessages) { \
432 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
433 len = RegexLengthToShowInErrorMessages - 10; \
439 #define FAIL(msg) _FAIL( \
440 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
441 msg, (int)len, RExC_precomp, ellipses))
443 #define FAIL2(msg,arg) _FAIL( \
444 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
445 arg, (int)len, RExC_precomp, ellipses))
448 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
450 #define Simple_vFAIL(m) STMT_START { \
451 const IV offset = RExC_parse - RExC_precomp; \
452 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
453 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
457 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
459 #define vFAIL(m) STMT_START { \
461 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
466 * Like Simple_vFAIL(), but accepts two arguments.
468 #define Simple_vFAIL2(m,a1) STMT_START { \
469 const IV offset = RExC_parse - RExC_precomp; \
470 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
471 (int)offset, RExC_precomp, RExC_precomp + offset); \
475 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
477 #define vFAIL2(m,a1) STMT_START { \
479 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
480 Simple_vFAIL2(m, a1); \
485 * Like Simple_vFAIL(), but accepts three arguments.
487 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
488 const IV offset = RExC_parse - RExC_precomp; \
489 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
490 (int)offset, RExC_precomp, RExC_precomp + offset); \
494 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
496 #define vFAIL3(m,a1,a2) STMT_START { \
498 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
499 Simple_vFAIL3(m, a1, a2); \
503 * Like Simple_vFAIL(), but accepts four arguments.
505 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
506 const IV offset = RExC_parse - RExC_precomp; \
507 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
508 (int)offset, RExC_precomp, RExC_precomp + offset); \
511 #define ckWARNreg(loc,m) STMT_START { \
512 const IV offset = loc - RExC_precomp; \
513 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
514 (int)offset, RExC_precomp, RExC_precomp + offset); \
517 #define ckWARNregdep(loc,m) STMT_START { \
518 const IV offset = loc - RExC_precomp; \
519 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
521 (int)offset, RExC_precomp, RExC_precomp + offset); \
524 #define ckWARN2regdep(loc,m, a1) STMT_START { \
525 const IV offset = loc - RExC_precomp; \
526 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
528 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
531 #define ckWARN2reg(loc, m, a1) STMT_START { \
532 const IV offset = loc - RExC_precomp; \
533 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
534 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
537 #define vWARN3(loc, m, a1, a2) STMT_START { \
538 const IV offset = loc - RExC_precomp; \
539 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
540 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
543 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
544 const IV offset = loc - RExC_precomp; \
545 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
546 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
549 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
550 const IV offset = loc - RExC_precomp; \
551 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
552 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
555 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
556 const IV offset = loc - RExC_precomp; \
557 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
558 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
561 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
562 const IV offset = loc - RExC_precomp; \
563 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
564 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
568 /* Allow for side effects in s */
569 #define REGC(c,s) STMT_START { \
570 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
573 /* Macros for recording node offsets. 20001227 mjd@plover.com
574 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
575 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
576 * Element 0 holds the number n.
577 * Position is 1 indexed.
579 #ifndef RE_TRACK_PATTERN_OFFSETS
580 #define Set_Node_Offset_To_R(node,byte)
581 #define Set_Node_Offset(node,byte)
582 #define Set_Cur_Node_Offset
583 #define Set_Node_Length_To_R(node,len)
584 #define Set_Node_Length(node,len)
585 #define Set_Node_Cur_Length(node)
586 #define Node_Offset(n)
587 #define Node_Length(n)
588 #define Set_Node_Offset_Length(node,offset,len)
589 #define ProgLen(ri) ri->u.proglen
590 #define SetProgLen(ri,x) ri->u.proglen = x
592 #define ProgLen(ri) ri->u.offsets[0]
593 #define SetProgLen(ri,x) ri->u.offsets[0] = x
594 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
596 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
597 __LINE__, (int)(node), (int)(byte))); \
599 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
601 RExC_offsets[2*(node)-1] = (byte); \
606 #define Set_Node_Offset(node,byte) \
607 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
608 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
610 #define Set_Node_Length_To_R(node,len) STMT_START { \
612 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
613 __LINE__, (int)(node), (int)(len))); \
615 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
617 RExC_offsets[2*(node)] = (len); \
622 #define Set_Node_Length(node,len) \
623 Set_Node_Length_To_R((node)-RExC_emit_start, len)
624 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
625 #define Set_Node_Cur_Length(node) \
626 Set_Node_Length(node, RExC_parse - parse_start)
628 /* Get offsets and lengths */
629 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
630 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
632 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
633 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
634 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
638 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
639 #define EXPERIMENTAL_INPLACESCAN
640 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
642 #define DEBUG_STUDYDATA(str,data,depth) \
643 DEBUG_OPTIMISE_MORE_r(if(data){ \
644 PerlIO_printf(Perl_debug_log, \
645 "%*s" str "Pos:%"IVdf"/%"IVdf \
646 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
647 (int)(depth)*2, "", \
648 (IV)((data)->pos_min), \
649 (IV)((data)->pos_delta), \
650 (UV)((data)->flags), \
651 (IV)((data)->whilem_c), \
652 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
653 is_inf ? "INF " : "" \
655 if ((data)->last_found) \
656 PerlIO_printf(Perl_debug_log, \
657 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
658 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
659 SvPVX_const((data)->last_found), \
660 (IV)((data)->last_end), \
661 (IV)((data)->last_start_min), \
662 (IV)((data)->last_start_max), \
663 ((data)->longest && \
664 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
665 SvPVX_const((data)->longest_fixed), \
666 (IV)((data)->offset_fixed), \
667 ((data)->longest && \
668 (data)->longest==&((data)->longest_float)) ? "*" : "", \
669 SvPVX_const((data)->longest_float), \
670 (IV)((data)->offset_float_min), \
671 (IV)((data)->offset_float_max) \
673 PerlIO_printf(Perl_debug_log,"\n"); \
676 static void clear_re(pTHX_ void *r);
678 /* Mark that we cannot extend a found fixed substring at this point.
679 Update the longest found anchored substring and the longest found
680 floating substrings if needed. */
683 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
685 const STRLEN l = CHR_SVLEN(data->last_found);
686 const STRLEN old_l = CHR_SVLEN(*data->longest);
687 GET_RE_DEBUG_FLAGS_DECL;
689 PERL_ARGS_ASSERT_SCAN_COMMIT;
691 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
692 SvSetMagicSV(*data->longest, data->last_found);
693 if (*data->longest == data->longest_fixed) {
694 data->offset_fixed = l ? data->last_start_min : data->pos_min;
695 if (data->flags & SF_BEFORE_EOL)
697 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
699 data->flags &= ~SF_FIX_BEFORE_EOL;
700 data->minlen_fixed=minlenp;
701 data->lookbehind_fixed=0;
703 else { /* *data->longest == data->longest_float */
704 data->offset_float_min = l ? data->last_start_min : data->pos_min;
705 data->offset_float_max = (l
706 ? data->last_start_max
707 : data->pos_min + data->pos_delta);
708 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
709 data->offset_float_max = I32_MAX;
710 if (data->flags & SF_BEFORE_EOL)
712 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
714 data->flags &= ~SF_FL_BEFORE_EOL;
715 data->minlen_float=minlenp;
716 data->lookbehind_float=0;
719 SvCUR_set(data->last_found, 0);
721 SV * const sv = data->last_found;
722 if (SvUTF8(sv) && SvMAGICAL(sv)) {
723 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
729 data->flags &= ~SF_BEFORE_EOL;
730 DEBUG_STUDYDATA("commit: ",data,0);
733 /* Can match anything (initialization) */
735 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
737 PERL_ARGS_ASSERT_CL_ANYTHING;
739 ANYOF_BITMAP_SETALL(cl);
740 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
741 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
743 /* If any portion of the regex is to operate under locale rules,
744 * initialization includes it. The reason this isn't done for all regexes
745 * is that the optimizer was written under the assumption that locale was
746 * all-or-nothing. Given the complexity and lack of documentation in the
747 * optimizer, and that there are inadequate test cases for locale, so many
748 * parts of it may not work properly, it is safest to avoid locale unless
750 if (RExC_contains_locale) {
751 ANYOF_CLASS_SETALL(cl); /* /l uses class */
752 cl->flags |= ANYOF_LOCALE;
755 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
759 /* Can match anything (initialization) */
761 S_cl_is_anything(const struct regnode_charclass_class *cl)
765 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
767 for (value = 0; value <= ANYOF_MAX; value += 2)
768 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
770 if (!(cl->flags & ANYOF_UNICODE_ALL))
772 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
777 /* Can match anything (initialization) */
779 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
781 PERL_ARGS_ASSERT_CL_INIT;
783 Zero(cl, 1, struct regnode_charclass_class);
785 cl_anything(pRExC_state, cl);
786 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
789 /* These two functions currently do the exact same thing */
790 #define cl_init_zero S_cl_init
792 /* 'AND' a given class with another one. Can create false positives. 'cl'
793 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
794 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
796 S_cl_and(struct regnode_charclass_class *cl,
797 const struct regnode_charclass_class *and_with)
799 PERL_ARGS_ASSERT_CL_AND;
801 assert(and_with->type == ANYOF);
803 /* I (khw) am not sure all these restrictions are necessary XXX */
804 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
805 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
806 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
807 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
808 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
811 if (and_with->flags & ANYOF_INVERT)
812 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
813 cl->bitmap[i] &= ~and_with->bitmap[i];
815 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
816 cl->bitmap[i] &= and_with->bitmap[i];
817 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
819 if (and_with->flags & ANYOF_INVERT) {
821 /* Here, the and'ed node is inverted. Get the AND of the flags that
822 * aren't affected by the inversion. Those that are affected are
823 * handled individually below */
824 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
825 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
826 cl->flags |= affected_flags;
828 /* We currently don't know how to deal with things that aren't in the
829 * bitmap, but we know that the intersection is no greater than what
830 * is already in cl, so let there be false positives that get sorted
831 * out after the synthetic start class succeeds, and the node is
832 * matched for real. */
834 /* The inversion of these two flags indicate that the resulting
835 * intersection doesn't have them */
836 if (and_with->flags & ANYOF_UNICODE_ALL) {
837 cl->flags &= ~ANYOF_UNICODE_ALL;
839 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
840 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
843 else { /* and'd node is not inverted */
844 U8 outside_bitmap_but_not_utf8; /* Temp variable */
846 if (! ANYOF_NONBITMAP(and_with)) {
848 /* Here 'and_with' doesn't match anything outside the bitmap
849 * (except possibly ANYOF_UNICODE_ALL), which means the
850 * intersection can't either, except for ANYOF_UNICODE_ALL, in
851 * which case we don't know what the intersection is, but it's no
852 * greater than what cl already has, so can just leave it alone,
853 * with possible false positives */
854 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
855 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
856 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
859 else if (! ANYOF_NONBITMAP(cl)) {
861 /* Here, 'and_with' does match something outside the bitmap, and cl
862 * doesn't have a list of things to match outside the bitmap. If
863 * cl can match all code points above 255, the intersection will
864 * be those above-255 code points that 'and_with' matches. If cl
865 * can't match all Unicode code points, it means that it can't
866 * match anything outside the bitmap (since the 'if' that got us
867 * into this block tested for that), so we leave the bitmap empty.
869 if (cl->flags & ANYOF_UNICODE_ALL) {
870 ARG_SET(cl, ARG(and_with));
872 /* and_with's ARG may match things that don't require UTF8.
873 * And now cl's will too, in spite of this being an 'and'. See
874 * the comments below about the kludge */
875 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
879 /* Here, both 'and_with' and cl match something outside the
880 * bitmap. Currently we do not do the intersection, so just match
881 * whatever cl had at the beginning. */
885 /* Take the intersection of the two sets of flags. However, the
886 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
887 * kludge around the fact that this flag is not treated like the others
888 * which are initialized in cl_anything(). The way the optimizer works
889 * is that the synthetic start class (SSC) is initialized to match
890 * anything, and then the first time a real node is encountered, its
891 * values are AND'd with the SSC's with the result being the values of
892 * the real node. However, there are paths through the optimizer where
893 * the AND never gets called, so those initialized bits are set
894 * inappropriately, which is not usually a big deal, as they just cause
895 * false positives in the SSC, which will just mean a probably
896 * imperceptible slow down in execution. However this bit has a
897 * higher false positive consequence in that it can cause utf8.pm,
898 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
899 * bigger slowdown and also causes significant extra memory to be used.
900 * In order to prevent this, the code now takes a different tack. The
901 * bit isn't set unless some part of the regular expression needs it,
902 * but once set it won't get cleared. This means that these extra
903 * modules won't get loaded unless there was some path through the
904 * pattern that would have required them anyway, and so any false
905 * positives that occur by not ANDing them out when they could be
906 * aren't as severe as they would be if we treated this bit like all
908 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
909 & ANYOF_NONBITMAP_NON_UTF8;
910 cl->flags &= and_with->flags;
911 cl->flags |= outside_bitmap_but_not_utf8;
915 /* 'OR' a given class with another one. Can create false positives. 'cl'
916 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
917 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
919 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
921 PERL_ARGS_ASSERT_CL_OR;
923 if (or_with->flags & ANYOF_INVERT) {
925 /* Here, the or'd node is to be inverted. This means we take the
926 * complement of everything not in the bitmap, but currently we don't
927 * know what that is, so give up and match anything */
928 if (ANYOF_NONBITMAP(or_with)) {
929 cl_anything(pRExC_state, cl);
932 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
933 * <= (B1 | !B2) | (CL1 | !CL2)
934 * which is wasteful if CL2 is small, but we ignore CL2:
935 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
936 * XXXX Can we handle case-fold? Unclear:
937 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
938 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
940 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
941 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
942 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
945 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
946 cl->bitmap[i] |= ~or_with->bitmap[i];
947 } /* XXXX: logic is complicated otherwise */
949 cl_anything(pRExC_state, cl);
952 /* And, we can just take the union of the flags that aren't affected
953 * by the inversion */
954 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
956 /* For the remaining flags:
957 ANYOF_UNICODE_ALL and inverted means to not match anything above
958 255, which means that the union with cl should just be
959 what cl has in it, so can ignore this flag
960 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
961 is 127-255 to match them, but then invert that, so the
962 union with cl should just be what cl has in it, so can
965 } else { /* 'or_with' is not inverted */
966 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
967 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
968 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
969 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
972 /* OR char bitmap and class bitmap separately */
973 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
974 cl->bitmap[i] |= or_with->bitmap[i];
975 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
976 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
977 cl->classflags[i] |= or_with->classflags[i];
978 cl->flags |= ANYOF_CLASS;
981 else { /* XXXX: logic is complicated, leave it along for a moment. */
982 cl_anything(pRExC_state, cl);
985 if (ANYOF_NONBITMAP(or_with)) {
987 /* Use the added node's outside-the-bit-map match if there isn't a
988 * conflict. If there is a conflict (both nodes match something
989 * outside the bitmap, but what they match outside is not the same
990 * pointer, and hence not easily compared until XXX we extend
991 * inversion lists this far), give up and allow the start class to
992 * match everything outside the bitmap. If that stuff is all above
993 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
994 if (! ANYOF_NONBITMAP(cl)) {
995 ARG_SET(cl, ARG(or_with));
997 else if (ARG(cl) != ARG(or_with)) {
999 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1000 cl_anything(pRExC_state, cl);
1003 cl->flags |= ANYOF_UNICODE_ALL;
1008 /* Take the union */
1009 cl->flags |= or_with->flags;
1013 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1014 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1015 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1016 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1021 dump_trie(trie,widecharmap,revcharmap)
1022 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1023 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1025 These routines dump out a trie in a somewhat readable format.
1026 The _interim_ variants are used for debugging the interim
1027 tables that are used to generate the final compressed
1028 representation which is what dump_trie expects.
1030 Part of the reason for their existence is to provide a form
1031 of documentation as to how the different representations function.
1036 Dumps the final compressed table form of the trie to Perl_debug_log.
1037 Used for debugging make_trie().
1041 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1042 AV *revcharmap, U32 depth)
1045 SV *sv=sv_newmortal();
1046 int colwidth= widecharmap ? 6 : 4;
1048 GET_RE_DEBUG_FLAGS_DECL;
1050 PERL_ARGS_ASSERT_DUMP_TRIE;
1052 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1053 (int)depth * 2 + 2,"",
1054 "Match","Base","Ofs" );
1056 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1057 SV ** const tmp = av_fetch( revcharmap, state, 0);
1059 PerlIO_printf( Perl_debug_log, "%*s",
1061 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1062 PL_colors[0], PL_colors[1],
1063 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1064 PERL_PV_ESCAPE_FIRSTCHAR
1069 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1070 (int)depth * 2 + 2,"");
1072 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1073 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1074 PerlIO_printf( Perl_debug_log, "\n");
1076 for( state = 1 ; state < trie->statecount ; state++ ) {
1077 const U32 base = trie->states[ state ].trans.base;
1079 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1081 if ( trie->states[ state ].wordnum ) {
1082 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1084 PerlIO_printf( Perl_debug_log, "%6s", "" );
1087 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1092 while( ( base + ofs < trie->uniquecharcount ) ||
1093 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1094 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1097 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1099 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1100 if ( ( base + ofs >= trie->uniquecharcount ) &&
1101 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1102 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1104 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1106 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1108 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1112 PerlIO_printf( Perl_debug_log, "]");
1115 PerlIO_printf( Perl_debug_log, "\n" );
1117 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1118 for (word=1; word <= trie->wordcount; word++) {
1119 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1120 (int)word, (int)(trie->wordinfo[word].prev),
1121 (int)(trie->wordinfo[word].len));
1123 PerlIO_printf(Perl_debug_log, "\n" );
1126 Dumps a fully constructed but uncompressed trie in list form.
1127 List tries normally only are used for construction when the number of
1128 possible chars (trie->uniquecharcount) is very high.
1129 Used for debugging make_trie().
1132 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1133 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1137 SV *sv=sv_newmortal();
1138 int colwidth= widecharmap ? 6 : 4;
1139 GET_RE_DEBUG_FLAGS_DECL;
1141 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1143 /* print out the table precompression. */
1144 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1145 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1146 "------:-----+-----------------\n" );
1148 for( state=1 ; state < next_alloc ; state ++ ) {
1151 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1152 (int)depth * 2 + 2,"", (UV)state );
1153 if ( ! trie->states[ state ].wordnum ) {
1154 PerlIO_printf( Perl_debug_log, "%5s| ","");
1156 PerlIO_printf( Perl_debug_log, "W%4x| ",
1157 trie->states[ state ].wordnum
1160 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1161 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1163 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1165 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1166 PL_colors[0], PL_colors[1],
1167 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1168 PERL_PV_ESCAPE_FIRSTCHAR
1170 TRIE_LIST_ITEM(state,charid).forid,
1171 (UV)TRIE_LIST_ITEM(state,charid).newstate
1174 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1175 (int)((depth * 2) + 14), "");
1178 PerlIO_printf( Perl_debug_log, "\n");
1183 Dumps a fully constructed but uncompressed trie in table form.
1184 This is the normal DFA style state transition table, with a few
1185 twists to facilitate compression later.
1186 Used for debugging make_trie().
1189 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1190 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1195 SV *sv=sv_newmortal();
1196 int colwidth= widecharmap ? 6 : 4;
1197 GET_RE_DEBUG_FLAGS_DECL;
1199 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1202 print out the table precompression so that we can do a visual check
1203 that they are identical.
1206 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1208 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1209 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1211 PerlIO_printf( Perl_debug_log, "%*s",
1213 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1214 PL_colors[0], PL_colors[1],
1215 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1216 PERL_PV_ESCAPE_FIRSTCHAR
1222 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1224 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1225 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1228 PerlIO_printf( Perl_debug_log, "\n" );
1230 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1232 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1233 (int)depth * 2 + 2,"",
1234 (UV)TRIE_NODENUM( state ) );
1236 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1237 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1239 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1241 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1243 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1244 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1246 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1247 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1255 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1256 startbranch: the first branch in the whole branch sequence
1257 first : start branch of sequence of branch-exact nodes.
1258 May be the same as startbranch
1259 last : Thing following the last branch.
1260 May be the same as tail.
1261 tail : item following the branch sequence
1262 count : words in the sequence
1263 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1264 depth : indent depth
1266 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1268 A trie is an N'ary tree where the branches are determined by digital
1269 decomposition of the key. IE, at the root node you look up the 1st character and
1270 follow that branch repeat until you find the end of the branches. Nodes can be
1271 marked as "accepting" meaning they represent a complete word. Eg:
1275 would convert into the following structure. Numbers represent states, letters
1276 following numbers represent valid transitions on the letter from that state, if
1277 the number is in square brackets it represents an accepting state, otherwise it
1278 will be in parenthesis.
1280 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1284 (1) +-i->(6)-+-s->[7]
1286 +-s->(3)-+-h->(4)-+-e->[5]
1288 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1290 This shows that when matching against the string 'hers' we will begin at state 1
1291 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1292 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1293 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1294 single traverse. We store a mapping from accepting to state to which word was
1295 matched, and then when we have multiple possibilities we try to complete the
1296 rest of the regex in the order in which they occured in the alternation.
1298 The only prior NFA like behaviour that would be changed by the TRIE support is
1299 the silent ignoring of duplicate alternations which are of the form:
1301 / (DUPE|DUPE) X? (?{ ... }) Y /x
1303 Thus EVAL blocks following a trie may be called a different number of times with
1304 and without the optimisation. With the optimisations dupes will be silently
1305 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1306 the following demonstrates:
1308 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1310 which prints out 'word' three times, but
1312 'words'=~/(word|word|word)(?{ print $1 })S/
1314 which doesnt print it out at all. This is due to other optimisations kicking in.
1316 Example of what happens on a structural level:
1318 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1320 1: CURLYM[1] {1,32767}(18)
1331 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1332 and should turn into:
1334 1: CURLYM[1] {1,32767}(18)
1336 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1344 Cases where tail != last would be like /(?foo|bar)baz/:
1354 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1355 and would end up looking like:
1358 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1365 d = uvuni_to_utf8_flags(d, uv, 0);
1367 is the recommended Unicode-aware way of saying
1372 #define TRIE_STORE_REVCHAR(val) \
1375 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1376 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1377 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1378 SvCUR_set(zlopp, kapow - flrbbbbb); \
1381 av_push(revcharmap, zlopp); \
1383 char ooooff = (char)val; \
1384 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1388 #define TRIE_READ_CHAR STMT_START { \
1391 /* if it is UTF then it is either already folded, or does not need folding */ \
1392 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1394 else if (folder == PL_fold_latin1) { \
1395 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1396 if ( foldlen > 0 ) { \
1397 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1403 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1404 skiplen = UNISKIP(uvc); \
1405 foldlen -= skiplen; \
1406 scan = foldbuf + skiplen; \
1409 /* raw data, will be folded later if needed */ \
1417 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1418 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1419 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1420 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1422 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1423 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1424 TRIE_LIST_CUR( state )++; \
1427 #define TRIE_LIST_NEW(state) STMT_START { \
1428 Newxz( trie->states[ state ].trans.list, \
1429 4, reg_trie_trans_le ); \
1430 TRIE_LIST_CUR( state ) = 1; \
1431 TRIE_LIST_LEN( state ) = 4; \
1434 #define TRIE_HANDLE_WORD(state) STMT_START { \
1435 U16 dupe= trie->states[ state ].wordnum; \
1436 regnode * const noper_next = regnext( noper ); \
1439 /* store the word for dumping */ \
1441 if (OP(noper) != NOTHING) \
1442 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1444 tmp = newSVpvn_utf8( "", 0, UTF ); \
1445 av_push( trie_words, tmp ); \
1449 trie->wordinfo[curword].prev = 0; \
1450 trie->wordinfo[curword].len = wordlen; \
1451 trie->wordinfo[curword].accept = state; \
1453 if ( noper_next < tail ) { \
1455 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1456 trie->jump[curword] = (U16)(noper_next - convert); \
1458 jumper = noper_next; \
1460 nextbranch= regnext(cur); \
1464 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1465 /* chain, so that when the bits of chain are later */\
1466 /* linked together, the dups appear in the chain */\
1467 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1468 trie->wordinfo[dupe].prev = curword; \
1470 /* we haven't inserted this word yet. */ \
1471 trie->states[ state ].wordnum = curword; \
1476 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1477 ( ( base + charid >= ucharcount \
1478 && base + charid < ubound \
1479 && state == trie->trans[ base - ucharcount + charid ].check \
1480 && trie->trans[ base - ucharcount + charid ].next ) \
1481 ? trie->trans[ base - ucharcount + charid ].next \
1482 : ( state==1 ? special : 0 ) \
1486 #define MADE_JUMP_TRIE 2
1487 #define MADE_EXACT_TRIE 4
1490 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1493 /* first pass, loop through and scan words */
1494 reg_trie_data *trie;
1495 HV *widecharmap = NULL;
1496 AV *revcharmap = newAV();
1498 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1503 regnode *jumper = NULL;
1504 regnode *nextbranch = NULL;
1505 regnode *convert = NULL;
1506 U32 *prev_states; /* temp array mapping each state to previous one */
1507 /* we just use folder as a flag in utf8 */
1508 const U8 * folder = NULL;
1511 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1512 AV *trie_words = NULL;
1513 /* along with revcharmap, this only used during construction but both are
1514 * useful during debugging so we store them in the struct when debugging.
1517 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1518 STRLEN trie_charcount=0;
1520 SV *re_trie_maxbuff;
1521 GET_RE_DEBUG_FLAGS_DECL;
1523 PERL_ARGS_ASSERT_MAKE_TRIE;
1525 PERL_UNUSED_ARG(depth);
1532 case EXACTFU_TRICKYFOLD:
1533 case EXACTFU: folder = PL_fold_latin1; break;
1534 case EXACTF: folder = PL_fold; break;
1535 case EXACTFL: folder = PL_fold_locale; break;
1536 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1539 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1541 trie->startstate = 1;
1542 trie->wordcount = word_count;
1543 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1544 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1546 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1547 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1548 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1551 trie_words = newAV();
1554 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1555 if (!SvIOK(re_trie_maxbuff)) {
1556 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1558 DEBUG_TRIE_COMPILE_r({
1559 PerlIO_printf( Perl_debug_log,
1560 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1561 (int)depth * 2 + 2, "",
1562 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1563 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1567 /* Find the node we are going to overwrite */
1568 if ( first == startbranch && OP( last ) != BRANCH ) {
1569 /* whole branch chain */
1572 /* branch sub-chain */
1573 convert = NEXTOPER( first );
1576 /* -- First loop and Setup --
1578 We first traverse the branches and scan each word to determine if it
1579 contains widechars, and how many unique chars there are, this is
1580 important as we have to build a table with at least as many columns as we
1583 We use an array of integers to represent the character codes 0..255
1584 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1585 native representation of the character value as the key and IV's for the
1588 *TODO* If we keep track of how many times each character is used we can
1589 remap the columns so that the table compression later on is more
1590 efficient in terms of memory by ensuring the most common value is in the
1591 middle and the least common are on the outside. IMO this would be better
1592 than a most to least common mapping as theres a decent chance the most
1593 common letter will share a node with the least common, meaning the node
1594 will not be compressible. With a middle is most common approach the worst
1595 case is when we have the least common nodes twice.
1599 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1600 regnode *noper = NEXTOPER( cur );
1601 const U8 *uc = (U8*)STRING( noper );
1602 const U8 *e = uc + STR_LEN( noper );
1604 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1606 const U8 *scan = (U8*)NULL;
1607 U32 wordlen = 0; /* required init */
1609 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1611 if (OP(noper) == NOTHING) {
1612 regnode *noper_next= regnext(noper);
1613 if (noper_next != tail && OP(noper_next) == flags) {
1615 uc= (U8*)STRING(noper);
1616 e= uc + STR_LEN(noper);
1617 trie->minlen= STR_LEN(noper);
1624 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1625 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1626 regardless of encoding */
1627 if (OP( noper ) == EXACTFU_SS) {
1628 /* false positives are ok, so just set this */
1629 TRIE_BITMAP_SET(trie,0xDF);
1632 for ( ; uc < e ; uc += len ) {
1633 TRIE_CHARCOUNT(trie)++;
1638 U8 folded= folder[ (U8) uvc ];
1639 if ( !trie->charmap[ folded ] ) {
1640 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1641 TRIE_STORE_REVCHAR( folded );
1644 if ( !trie->charmap[ uvc ] ) {
1645 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1646 TRIE_STORE_REVCHAR( uvc );
1649 /* store the codepoint in the bitmap, and its folded
1651 TRIE_BITMAP_SET(trie, uvc);
1653 /* store the folded codepoint */
1654 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1657 /* store first byte of utf8 representation of
1658 variant codepoints */
1659 if (! UNI_IS_INVARIANT(uvc)) {
1660 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1663 set_bit = 0; /* We've done our bit :-) */
1668 widecharmap = newHV();
1670 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1673 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1675 if ( !SvTRUE( *svpp ) ) {
1676 sv_setiv( *svpp, ++trie->uniquecharcount );
1677 TRIE_STORE_REVCHAR(uvc);
1681 if( cur == first ) {
1682 trie->minlen = chars;
1683 trie->maxlen = chars;
1684 } else if (chars < trie->minlen) {
1685 trie->minlen = chars;
1686 } else if (chars > trie->maxlen) {
1687 trie->maxlen = chars;
1689 if (OP( noper ) == EXACTFU_SS) {
1690 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1691 if (trie->minlen > 1)
1694 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1695 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1696 * - We assume that any such sequence might match a 2 byte string */
1697 if (trie->minlen > 2 )
1701 } /* end first pass */
1702 DEBUG_TRIE_COMPILE_r(
1703 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1704 (int)depth * 2 + 2,"",
1705 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1706 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1707 (int)trie->minlen, (int)trie->maxlen )
1711 We now know what we are dealing with in terms of unique chars and
1712 string sizes so we can calculate how much memory a naive
1713 representation using a flat table will take. If it's over a reasonable
1714 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1715 conservative but potentially much slower representation using an array
1718 At the end we convert both representations into the same compressed
1719 form that will be used in regexec.c for matching with. The latter
1720 is a form that cannot be used to construct with but has memory
1721 properties similar to the list form and access properties similar
1722 to the table form making it both suitable for fast searches and
1723 small enough that its feasable to store for the duration of a program.
1725 See the comment in the code where the compressed table is produced
1726 inplace from the flat tabe representation for an explanation of how
1727 the compression works.
1732 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1735 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1737 Second Pass -- Array Of Lists Representation
1739 Each state will be represented by a list of charid:state records
1740 (reg_trie_trans_le) the first such element holds the CUR and LEN
1741 points of the allocated array. (See defines above).
1743 We build the initial structure using the lists, and then convert
1744 it into the compressed table form which allows faster lookups
1745 (but cant be modified once converted).
1748 STRLEN transcount = 1;
1750 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1751 "%*sCompiling trie using list compiler\n",
1752 (int)depth * 2 + 2, ""));
1754 trie->states = (reg_trie_state *)
1755 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1756 sizeof(reg_trie_state) );
1760 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1762 regnode *noper = NEXTOPER( cur );
1763 U8 *uc = (U8*)STRING( noper );
1764 const U8 *e = uc + STR_LEN( noper );
1765 U32 state = 1; /* required init */
1766 U16 charid = 0; /* sanity init */
1767 U8 *scan = (U8*)NULL; /* sanity init */
1768 STRLEN foldlen = 0; /* required init */
1769 U32 wordlen = 0; /* required init */
1770 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1773 if (OP(noper) == NOTHING) {
1774 regnode *noper_next= regnext(noper);
1775 if (noper_next != tail && OP(noper_next) == flags) {
1777 uc= (U8*)STRING(noper);
1778 e= uc + STR_LEN(noper);
1782 if (OP(noper) != NOTHING) {
1783 for ( ; uc < e ; uc += len ) {
1788 charid = trie->charmap[ uvc ];
1790 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1794 charid=(U16)SvIV( *svpp );
1797 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1804 if ( !trie->states[ state ].trans.list ) {
1805 TRIE_LIST_NEW( state );
1807 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1808 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1809 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1814 newstate = next_alloc++;
1815 prev_states[newstate] = state;
1816 TRIE_LIST_PUSH( state, charid, newstate );
1821 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1825 TRIE_HANDLE_WORD(state);
1827 } /* end second pass */
1829 /* next alloc is the NEXT state to be allocated */
1830 trie->statecount = next_alloc;
1831 trie->states = (reg_trie_state *)
1832 PerlMemShared_realloc( trie->states,
1834 * sizeof(reg_trie_state) );
1836 /* and now dump it out before we compress it */
1837 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1838 revcharmap, next_alloc,
1842 trie->trans = (reg_trie_trans *)
1843 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1850 for( state=1 ; state < next_alloc ; state ++ ) {
1854 DEBUG_TRIE_COMPILE_MORE_r(
1855 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1859 if (trie->states[state].trans.list) {
1860 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1864 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1865 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1866 if ( forid < minid ) {
1868 } else if ( forid > maxid ) {
1872 if ( transcount < tp + maxid - minid + 1) {
1874 trie->trans = (reg_trie_trans *)
1875 PerlMemShared_realloc( trie->trans,
1877 * sizeof(reg_trie_trans) );
1878 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1880 base = trie->uniquecharcount + tp - minid;
1881 if ( maxid == minid ) {
1883 for ( ; zp < tp ; zp++ ) {
1884 if ( ! trie->trans[ zp ].next ) {
1885 base = trie->uniquecharcount + zp - minid;
1886 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1887 trie->trans[ zp ].check = state;
1893 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1894 trie->trans[ tp ].check = state;
1899 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1900 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1901 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1902 trie->trans[ tid ].check = state;
1904 tp += ( maxid - minid + 1 );
1906 Safefree(trie->states[ state ].trans.list);
1909 DEBUG_TRIE_COMPILE_MORE_r(
1910 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1913 trie->states[ state ].trans.base=base;
1915 trie->lasttrans = tp + 1;
1919 Second Pass -- Flat Table Representation.
1921 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1922 We know that we will need Charcount+1 trans at most to store the data
1923 (one row per char at worst case) So we preallocate both structures
1924 assuming worst case.
1926 We then construct the trie using only the .next slots of the entry
1929 We use the .check field of the first entry of the node temporarily to
1930 make compression both faster and easier by keeping track of how many non
1931 zero fields are in the node.
1933 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1936 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1937 number representing the first entry of the node, and state as a
1938 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1939 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1940 are 2 entrys per node. eg:
1948 The table is internally in the right hand, idx form. However as we also
1949 have to deal with the states array which is indexed by nodenum we have to
1950 use TRIE_NODENUM() to convert.
1953 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1954 "%*sCompiling trie using table compiler\n",
1955 (int)depth * 2 + 2, ""));
1957 trie->trans = (reg_trie_trans *)
1958 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1959 * trie->uniquecharcount + 1,
1960 sizeof(reg_trie_trans) );
1961 trie->states = (reg_trie_state *)
1962 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1963 sizeof(reg_trie_state) );
1964 next_alloc = trie->uniquecharcount + 1;
1967 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1969 regnode *noper = NEXTOPER( cur );
1970 const U8 *uc = (U8*)STRING( noper );
1971 const U8 *e = uc + STR_LEN( noper );
1973 U32 state = 1; /* required init */
1975 U16 charid = 0; /* sanity init */
1976 U32 accept_state = 0; /* sanity init */
1977 U8 *scan = (U8*)NULL; /* sanity init */
1979 STRLEN foldlen = 0; /* required init */
1980 U32 wordlen = 0; /* required init */
1982 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1984 if (OP(noper) == NOTHING) {
1985 regnode *noper_next= regnext(noper);
1986 if (noper_next != tail && OP(noper_next) == flags) {
1988 uc= (U8*)STRING(noper);
1989 e= uc + STR_LEN(noper);
1993 if ( OP(noper) != NOTHING ) {
1994 for ( ; uc < e ; uc += len ) {
1999 charid = trie->charmap[ uvc ];
2001 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2002 charid = svpp ? (U16)SvIV(*svpp) : 0;
2006 if ( !trie->trans[ state + charid ].next ) {
2007 trie->trans[ state + charid ].next = next_alloc;
2008 trie->trans[ state ].check++;
2009 prev_states[TRIE_NODENUM(next_alloc)]
2010 = TRIE_NODENUM(state);
2011 next_alloc += trie->uniquecharcount;
2013 state = trie->trans[ state + charid ].next;
2015 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2017 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2020 accept_state = TRIE_NODENUM( state );
2021 TRIE_HANDLE_WORD(accept_state);
2023 } /* end second pass */
2025 /* and now dump it out before we compress it */
2026 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2028 next_alloc, depth+1));
2032 * Inplace compress the table.*
2034 For sparse data sets the table constructed by the trie algorithm will
2035 be mostly 0/FAIL transitions or to put it another way mostly empty.
2036 (Note that leaf nodes will not contain any transitions.)
2038 This algorithm compresses the tables by eliminating most such
2039 transitions, at the cost of a modest bit of extra work during lookup:
2041 - Each states[] entry contains a .base field which indicates the
2042 index in the state[] array wheres its transition data is stored.
2044 - If .base is 0 there are no valid transitions from that node.
2046 - If .base is nonzero then charid is added to it to find an entry in
2049 -If trans[states[state].base+charid].check!=state then the
2050 transition is taken to be a 0/Fail transition. Thus if there are fail
2051 transitions at the front of the node then the .base offset will point
2052 somewhere inside the previous nodes data (or maybe even into a node
2053 even earlier), but the .check field determines if the transition is
2057 The following process inplace converts the table to the compressed
2058 table: We first do not compress the root node 1,and mark all its
2059 .check pointers as 1 and set its .base pointer as 1 as well. This
2060 allows us to do a DFA construction from the compressed table later,
2061 and ensures that any .base pointers we calculate later are greater
2064 - We set 'pos' to indicate the first entry of the second node.
2066 - We then iterate over the columns of the node, finding the first and
2067 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2068 and set the .check pointers accordingly, and advance pos
2069 appropriately and repreat for the next node. Note that when we copy
2070 the next pointers we have to convert them from the original
2071 NODEIDX form to NODENUM form as the former is not valid post
2074 - If a node has no transitions used we mark its base as 0 and do not
2075 advance the pos pointer.
2077 - If a node only has one transition we use a second pointer into the
2078 structure to fill in allocated fail transitions from other states.
2079 This pointer is independent of the main pointer and scans forward
2080 looking for null transitions that are allocated to a state. When it
2081 finds one it writes the single transition into the "hole". If the
2082 pointer doesnt find one the single transition is appended as normal.
2084 - Once compressed we can Renew/realloc the structures to release the
2087 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2088 specifically Fig 3.47 and the associated pseudocode.
2092 const U32 laststate = TRIE_NODENUM( next_alloc );
2095 trie->statecount = laststate;
2097 for ( state = 1 ; state < laststate ; state++ ) {
2099 const U32 stateidx = TRIE_NODEIDX( state );
2100 const U32 o_used = trie->trans[ stateidx ].check;
2101 U32 used = trie->trans[ stateidx ].check;
2102 trie->trans[ stateidx ].check = 0;
2104 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2105 if ( flag || trie->trans[ stateidx + charid ].next ) {
2106 if ( trie->trans[ stateidx + charid ].next ) {
2108 for ( ; zp < pos ; zp++ ) {
2109 if ( ! trie->trans[ zp ].next ) {
2113 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2114 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2115 trie->trans[ zp ].check = state;
2116 if ( ++zp > pos ) pos = zp;
2123 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2125 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2126 trie->trans[ pos ].check = state;
2131 trie->lasttrans = pos + 1;
2132 trie->states = (reg_trie_state *)
2133 PerlMemShared_realloc( trie->states, laststate
2134 * sizeof(reg_trie_state) );
2135 DEBUG_TRIE_COMPILE_MORE_r(
2136 PerlIO_printf( Perl_debug_log,
2137 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2138 (int)depth * 2 + 2,"",
2139 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2142 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2145 } /* end table compress */
2147 DEBUG_TRIE_COMPILE_MORE_r(
2148 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2149 (int)depth * 2 + 2, "",
2150 (UV)trie->statecount,
2151 (UV)trie->lasttrans)
2153 /* resize the trans array to remove unused space */
2154 trie->trans = (reg_trie_trans *)
2155 PerlMemShared_realloc( trie->trans, trie->lasttrans
2156 * sizeof(reg_trie_trans) );
2158 { /* Modify the program and insert the new TRIE node */
2159 U8 nodetype =(U8)(flags & 0xFF);
2163 regnode *optimize = NULL;
2164 #ifdef RE_TRACK_PATTERN_OFFSETS
2167 U32 mjd_nodelen = 0;
2168 #endif /* RE_TRACK_PATTERN_OFFSETS */
2169 #endif /* DEBUGGING */
2171 This means we convert either the first branch or the first Exact,
2172 depending on whether the thing following (in 'last') is a branch
2173 or not and whther first is the startbranch (ie is it a sub part of
2174 the alternation or is it the whole thing.)
2175 Assuming its a sub part we convert the EXACT otherwise we convert
2176 the whole branch sequence, including the first.
2178 /* Find the node we are going to overwrite */
2179 if ( first != startbranch || OP( last ) == BRANCH ) {
2180 /* branch sub-chain */
2181 NEXT_OFF( first ) = (U16)(last - first);
2182 #ifdef RE_TRACK_PATTERN_OFFSETS
2184 mjd_offset= Node_Offset((convert));
2185 mjd_nodelen= Node_Length((convert));
2188 /* whole branch chain */
2190 #ifdef RE_TRACK_PATTERN_OFFSETS
2193 const regnode *nop = NEXTOPER( convert );
2194 mjd_offset= Node_Offset((nop));
2195 mjd_nodelen= Node_Length((nop));
2199 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2200 (int)depth * 2 + 2, "",
2201 (UV)mjd_offset, (UV)mjd_nodelen)
2204 /* But first we check to see if there is a common prefix we can
2205 split out as an EXACT and put in front of the TRIE node. */
2206 trie->startstate= 1;
2207 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2209 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2213 const U32 base = trie->states[ state ].trans.base;
2215 if ( trie->states[state].wordnum )
2218 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2219 if ( ( base + ofs >= trie->uniquecharcount ) &&
2220 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2221 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2223 if ( ++count > 1 ) {
2224 SV **tmp = av_fetch( revcharmap, ofs, 0);
2225 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2226 if ( state == 1 ) break;
2228 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2230 PerlIO_printf(Perl_debug_log,
2231 "%*sNew Start State=%"UVuf" Class: [",
2232 (int)depth * 2 + 2, "",
2235 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2236 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2238 TRIE_BITMAP_SET(trie,*ch);
2240 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2242 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2246 TRIE_BITMAP_SET(trie,*ch);
2248 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2249 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2255 SV **tmp = av_fetch( revcharmap, idx, 0);
2257 char *ch = SvPV( *tmp, len );
2259 SV *sv=sv_newmortal();
2260 PerlIO_printf( Perl_debug_log,
2261 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2262 (int)depth * 2 + 2, "",
2264 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2265 PL_colors[0], PL_colors[1],
2266 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2267 PERL_PV_ESCAPE_FIRSTCHAR
2272 OP( convert ) = nodetype;
2273 str=STRING(convert);
2276 STR_LEN(convert) += len;
2282 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2287 trie->prefixlen = (state-1);
2289 regnode *n = convert+NODE_SZ_STR(convert);
2290 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2291 trie->startstate = state;
2292 trie->minlen -= (state - 1);
2293 trie->maxlen -= (state - 1);
2295 /* At least the UNICOS C compiler choked on this
2296 * being argument to DEBUG_r(), so let's just have
2299 #ifdef PERL_EXT_RE_BUILD
2305 regnode *fix = convert;
2306 U32 word = trie->wordcount;
2308 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2309 while( ++fix < n ) {
2310 Set_Node_Offset_Length(fix, 0, 0);
2313 SV ** const tmp = av_fetch( trie_words, word, 0 );
2315 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2316 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2318 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2326 NEXT_OFF(convert) = (U16)(tail - convert);
2327 DEBUG_r(optimize= n);
2333 if ( trie->maxlen ) {
2334 NEXT_OFF( convert ) = (U16)(tail - convert);
2335 ARG_SET( convert, data_slot );
2336 /* Store the offset to the first unabsorbed branch in
2337 jump[0], which is otherwise unused by the jump logic.
2338 We use this when dumping a trie and during optimisation. */
2340 trie->jump[0] = (U16)(nextbranch - convert);
2342 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2343 * and there is a bitmap
2344 * and the first "jump target" node we found leaves enough room
2345 * then convert the TRIE node into a TRIEC node, with the bitmap
2346 * embedded inline in the opcode - this is hypothetically faster.
2348 if ( !trie->states[trie->startstate].wordnum
2350 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2352 OP( convert ) = TRIEC;
2353 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2354 PerlMemShared_free(trie->bitmap);
2357 OP( convert ) = TRIE;
2359 /* store the type in the flags */
2360 convert->flags = nodetype;
2364 + regarglen[ OP( convert ) ];
2366 /* XXX We really should free up the resource in trie now,
2367 as we won't use them - (which resources?) dmq */
2369 /* needed for dumping*/
2370 DEBUG_r(if (optimize) {
2371 regnode *opt = convert;
2373 while ( ++opt < optimize) {
2374 Set_Node_Offset_Length(opt,0,0);
2377 Try to clean up some of the debris left after the
2380 while( optimize < jumper ) {
2381 mjd_nodelen += Node_Length((optimize));
2382 OP( optimize ) = OPTIMIZED;
2383 Set_Node_Offset_Length(optimize,0,0);
2386 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2388 } /* end node insert */
2390 /* Finish populating the prev field of the wordinfo array. Walk back
2391 * from each accept state until we find another accept state, and if
2392 * so, point the first word's .prev field at the second word. If the
2393 * second already has a .prev field set, stop now. This will be the
2394 * case either if we've already processed that word's accept state,
2395 * or that state had multiple words, and the overspill words were
2396 * already linked up earlier.
2403 for (word=1; word <= trie->wordcount; word++) {
2405 if (trie->wordinfo[word].prev)
2407 state = trie->wordinfo[word].accept;
2409 state = prev_states[state];
2412 prev = trie->states[state].wordnum;
2416 trie->wordinfo[word].prev = prev;
2418 Safefree(prev_states);
2422 /* and now dump out the compressed format */
2423 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2425 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2427 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2428 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2430 SvREFCNT_dec(revcharmap);
2434 : trie->startstate>1
2440 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2442 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2444 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2445 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2448 We find the fail state for each state in the trie, this state is the longest proper
2449 suffix of the current state's 'word' that is also a proper prefix of another word in our
2450 trie. State 1 represents the word '' and is thus the default fail state. This allows
2451 the DFA not to have to restart after its tried and failed a word at a given point, it
2452 simply continues as though it had been matching the other word in the first place.
2454 'abcdgu'=~/abcdefg|cdgu/
2455 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2456 fail, which would bring us to the state representing 'd' in the second word where we would
2457 try 'g' and succeed, proceeding to match 'cdgu'.
2459 /* add a fail transition */
2460 const U32 trie_offset = ARG(source);
2461 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2463 const U32 ucharcount = trie->uniquecharcount;
2464 const U32 numstates = trie->statecount;
2465 const U32 ubound = trie->lasttrans + ucharcount;
2469 U32 base = trie->states[ 1 ].trans.base;
2472 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2473 GET_RE_DEBUG_FLAGS_DECL;
2475 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2477 PERL_UNUSED_ARG(depth);
2481 ARG_SET( stclass, data_slot );
2482 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2483 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2484 aho->trie=trie_offset;
2485 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2486 Copy( trie->states, aho->states, numstates, reg_trie_state );
2487 Newxz( q, numstates, U32);
2488 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2491 /* initialize fail[0..1] to be 1 so that we always have
2492 a valid final fail state */
2493 fail[ 0 ] = fail[ 1 ] = 1;
2495 for ( charid = 0; charid < ucharcount ; charid++ ) {
2496 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2498 q[ q_write ] = newstate;
2499 /* set to point at the root */
2500 fail[ q[ q_write++ ] ]=1;
2503 while ( q_read < q_write) {
2504 const U32 cur = q[ q_read++ % numstates ];
2505 base = trie->states[ cur ].trans.base;
2507 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2508 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2510 U32 fail_state = cur;
2513 fail_state = fail[ fail_state ];
2514 fail_base = aho->states[ fail_state ].trans.base;
2515 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2517 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2518 fail[ ch_state ] = fail_state;
2519 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2521 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2523 q[ q_write++ % numstates] = ch_state;
2527 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2528 when we fail in state 1, this allows us to use the
2529 charclass scan to find a valid start char. This is based on the principle
2530 that theres a good chance the string being searched contains lots of stuff
2531 that cant be a start char.
2533 fail[ 0 ] = fail[ 1 ] = 0;
2534 DEBUG_TRIE_COMPILE_r({
2535 PerlIO_printf(Perl_debug_log,
2536 "%*sStclass Failtable (%"UVuf" states): 0",
2537 (int)(depth * 2), "", (UV)numstates
2539 for( q_read=1; q_read<numstates; q_read++ ) {
2540 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2542 PerlIO_printf(Perl_debug_log, "\n");
2545 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2550 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2551 * These need to be revisited when a newer toolchain becomes available.
2553 #if defined(__sparc64__) && defined(__GNUC__)
2554 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2555 # undef SPARC64_GCC_WORKAROUND
2556 # define SPARC64_GCC_WORKAROUND 1
2560 #define DEBUG_PEEP(str,scan,depth) \
2561 DEBUG_OPTIMISE_r({if (scan){ \
2562 SV * const mysv=sv_newmortal(); \
2563 regnode *Next = regnext(scan); \
2564 regprop(RExC_rx, mysv, scan); \
2565 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2566 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2567 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2571 /* The below joins as many adjacent EXACTish nodes as possible into a single
2572 * one, and looks for problematic sequences of characters whose folds vs.
2573 * non-folds have sufficiently different lengths, that the optimizer would be
2574 * fooled into rejecting legitimate matches of them, and the trie construction
2575 * code can't cope with them. The joining is only done if:
2576 * 1) there is room in the current conglomerated node to entirely contain the
2578 * 2) they are the exact same node type
2580 * The adjacent nodes actually may be separated by NOTHING kind nodes, and
2581 * these get optimized out
2583 * If there are problematic code sequences, *min_subtract is set to the delta
2584 * that the minimum size of the node can be less than its actual size. And,
2585 * the node type of the result is changed to reflect that it contains these
2588 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2589 * and contains LATIN SMALL LETTER SHARP S
2591 * This is as good a place as any to discuss the design of handling these
2592 * problematic sequences. It's been wrong in Perl for a very long time. There
2593 * are three code points in Unicode whose folded lengths differ so much from
2594 * the un-folded lengths that it causes problems for the optimizer and trie
2595 * construction. Why only these are problematic, and not others where lengths
2596 * also differ is something I (khw) do not understand. New versions of Unicode
2597 * might add more such code points. Hopefully the logic in fold_grind.t that
2598 * figures out what to test (in part by verifying that each size-combination
2599 * gets tested) will catch any that do come along, so they can be added to the
2600 * special handling below. The chances of new ones are actually rather small,
2601 * as most, if not all, of the world's scripts that have casefolding have
2602 * already been encoded by Unicode. Also, a number of Unicode's decisions were
2603 * made to allow compatibility with pre-existing standards, and almost all of
2604 * those have already been dealt with. These would otherwise be the most
2605 * likely candidates for generating further tricky sequences. In other words,
2606 * Unicode by itself is unlikely to add new ones unless it is for compatibility
2607 * with pre-existing standards, and there aren't many of those left.
2609 * The previous designs for dealing with these involved assigning a special
2610 * node for them. This approach doesn't work, as evidenced by this example:
2611 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2612 * Both these fold to "sss", but if the pattern is parsed to create a node of
2613 * that would match just the \xDF, it won't be able to handle the case where a
2614 * successful match would have to cross the node's boundary. The new approach
2615 * that hopefully generally solves the problem generates an EXACTFU_SS node
2618 * There are a number of components to the approach (a lot of work for just
2619 * three code points!):
2620 * 1) This routine examines each EXACTFish node that could contain the
2621 * problematic sequences. It returns in *min_subtract how much to
2622 * subtract from the the actual length of the string to get a real minimum
2623 * for one that could match it. This number is usually 0 except for the
2624 * problematic sequences. This delta is used by the caller to adjust the
2625 * min length of the match, and the delta between min and max, so that the
2626 * optimizer doesn't reject these possibilities based on size constraints.
2627 * 2) These sequences are not currently correctly handled by the trie code
2628 * either, so it changes the joined node type to ops that are not handled
2629 * by trie's, those new ops being EXACTFU_SS and EXACTFU_TRICKYFOLD.
2630 * 3) This is sufficient for the two Greek sequences (described below), but
2631 * the one involving the Sharp s (\xDF) needs more. The node type
2632 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2633 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2634 * case where there is a possible fold length change. That means that a
2635 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2636 * itself with length changes, and so can be processed faster. regexec.c
2637 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2638 * is pre-folded by regcomp.c. This saves effort in regex matching.
2639 * However, probably mostly for historical reasons, the pre-folding isn't
2640 * done for non-UTF8 patterns (and it can't be for EXACTF and EXACTFL
2641 * nodes, as what they fold to isn't known until runtime.) The fold
2642 * possibilities for the non-UTF8 patterns are quite simple, except for
2643 * the sharp s. All the ones that don't involve a UTF-8 target string
2644 * are members of a fold-pair, and arrays are set up for all of them
2645 * that quickly find the other member of the pair. It might actually
2646 * be faster to pre-fold these, but it isn't currently done, except for
2647 * the sharp s. Code elsewhere in this file makes sure that it gets
2648 * folded to 'ss', even if the pattern isn't UTF-8. This avoids the
2649 * issues described in the next item.
2650 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2651 * 'ss' or not is not knowable at compile time. It will match iff the
2652 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2653 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2654 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2655 * described in item 3). An assumption that the optimizer part of
2656 * regexec.c (probably unwittingly) makes is that a character in the
2657 * pattern corresponds to at most a single character in the target string.
2658 * (And I do mean character, and not byte here, unlike other parts of the
2659 * documentation that have never been updated to account for multibyte
2660 * Unicode.) This assumption is wrong only in this case, as all other
2661 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2662 * virtue of having this file pre-fold UTF-8 patterns. I'm
2663 * reluctant to try to change this assumption, so instead the code punts.
2664 * This routine examines EXACTF nodes for the sharp s, and returns a
2665 * boolean indicating whether or not the node is an EXACTF node that
2666 * contains a sharp s. When it is true, the caller sets a flag that later
2667 * causes the optimizer in this file to not set values for the floating
2668 * and fixed string lengths, and thus avoids the optimizer code in
2669 * regexec.c that makes the invalid assumption. Thus, there is no
2670 * optimization based on string lengths for EXACTF nodes that contain the
2671 * sharp s. This only happens for /id rules (which means the pattern
2675 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2676 if (PL_regkind[OP(scan)] == EXACT) \
2677 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2680 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) {
2681 /* Merge several consecutive EXACTish nodes into one. */
2682 regnode *n = regnext(scan);
2684 regnode *next = scan + NODE_SZ_STR(scan);
2688 regnode *stop = scan;
2689 GET_RE_DEBUG_FLAGS_DECL;
2691 PERL_UNUSED_ARG(depth);
2694 PERL_ARGS_ASSERT_JOIN_EXACT;
2695 #ifndef EXPERIMENTAL_INPLACESCAN
2696 PERL_UNUSED_ARG(flags);
2697 PERL_UNUSED_ARG(val);
2699 DEBUG_PEEP("join",scan,depth);
2701 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2702 * EXACT ones that are mergeable to the current one. */
2704 && (PL_regkind[OP(n)] == NOTHING
2705 || (stringok && OP(n) == OP(scan)))
2707 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2710 if (OP(n) == TAIL || n > next)
2712 if (PL_regkind[OP(n)] == NOTHING) {
2713 DEBUG_PEEP("skip:",n,depth);
2714 NEXT_OFF(scan) += NEXT_OFF(n);
2715 next = n + NODE_STEP_REGNODE;
2722 else if (stringok) {
2723 const unsigned int oldl = STR_LEN(scan);
2724 regnode * const nnext = regnext(n);
2726 if (oldl + STR_LEN(n) > U8_MAX)
2729 DEBUG_PEEP("merg",n,depth);
2732 NEXT_OFF(scan) += NEXT_OFF(n);
2733 STR_LEN(scan) += STR_LEN(n);
2734 next = n + NODE_SZ_STR(n);
2735 /* Now we can overwrite *n : */
2736 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2744 #ifdef EXPERIMENTAL_INPLACESCAN
2745 if (flags && !NEXT_OFF(n)) {
2746 DEBUG_PEEP("atch", val, depth);
2747 if (reg_off_by_arg[OP(n)]) {
2748 ARG_SET(n, val - n);
2751 NEXT_OFF(n) = val - n;
2759 *has_exactf_sharp_s = FALSE;
2761 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2762 * can now analyze for sequences of problematic code points. (Prior to
2763 * this final joining, sequences could have been split over boundaries, and
2764 * hence missed). The sequences only happen in folding, hence for any
2765 * non-EXACT EXACTish node */
2766 if (OP(scan) != EXACT) {
2768 U8 * s0 = (U8*) STRING(scan);
2769 U8 * const s_end = s0 + STR_LEN(scan);
2771 /* The below is perhaps overboard, but this allows us to save a test
2772 * each time through the loop at the expense of a mask. This is
2773 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2774 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2775 * This uses an exclusive 'or' to find that bit and then inverts it to
2776 * form a mask, with just a single 0, in the bit position where 'S' and
2778 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2779 const U8 s_masked = 's' & S_or_s_mask;
2781 /* One pass is made over the node's string looking for all the
2782 * possibilities. to avoid some tests in the loop, there are two main
2783 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2787 /* There are two problematic Greek code points in Unicode
2790 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2791 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2797 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2798 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2800 * This means that in case-insensitive matching (or "loose
2801 * matching", as Unicode calls it), an EXACTF of length six (the
2802 * UTF-8 encoded byte length of the above casefolded versions) can
2803 * match a target string of length two (the byte length of UTF-8
2804 * encoded U+0390 or U+03B0). This would rather mess up the
2805 * minimum length computation. (there are other code points that
2806 * also fold to these two sequences, but the delta is smaller)
2808 * If these sequences are found, the minimum length is decreased by
2809 * four (six minus two).
2811 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2812 * LETTER SHARP S. We decrease the min length by 1 for each
2813 * occurrence of 'ss' found */
2815 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2816 # define U390_first_byte 0xb4
2817 const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
2818 # define U3B0_first_byte 0xb5
2819 const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
2821 # define U390_first_byte 0xce
2822 const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
2823 # define U3B0_first_byte 0xcf
2824 const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
2826 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2827 yields a net of 0 */
2828 /* Examine the string for one of the problematic sequences */
2830 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2831 * sequence we are looking for is 2 */
2835 /* Look for the first byte in each problematic sequence */
2837 /* We don't have to worry about other things that fold to
2838 * 's' (such as the long s, U+017F), as all above-latin1
2839 * code points have been pre-folded */
2843 /* Current character is an 's' or 'S'. If next one is
2844 * as well, we have the dreaded sequence */
2845 if (((*(s+1) & S_or_s_mask) == s_masked)
2846 /* These two node types don't have special handling
2848 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2851 OP(scan) = EXACTFU_SS;
2852 s++; /* No need to look at this character again */
2856 case U390_first_byte:
2857 if (s_end - s >= len
2859 /* The 1's are because are skipping comparing the
2861 && memEQ(s + 1, U390_tail, len - 1))
2863 goto greek_sequence;
2867 case U3B0_first_byte:
2868 if (! (s_end - s >= len
2869 && memEQ(s + 1, U3B0_tail, len - 1)))
2876 /* This can't currently be handled by trie's, so change
2877 * the node type to indicate this. If EXACTFA and
2878 * EXACTFL were ever to be handled by trie's, this
2879 * would have to be changed. If this node has already
2880 * been changed to EXACTFU_SS in this loop, leave it as
2881 * is. (I (khw) think it doesn't matter in regexec.c
2882 * for UTF patterns, but no need to change it */
2883 if (OP(scan) == EXACTFU) {
2884 OP(scan) = EXACTFU_TRICKYFOLD;
2886 s += 6; /* We already know what this sequence is. Skip
2892 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2894 /* Here, the pattern is not UTF-8. We need to look only for the
2895 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2896 * in the final position. Otherwise we can stop looking 1 byte
2897 * earlier because have to find both the first and second 's' */
2898 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2900 for (s = s0; s < upper; s++) {
2905 && ((*(s+1) & S_or_s_mask) == s_masked))
2909 /* EXACTF nodes need to know that the minimum
2910 * length changed so that a sharp s in the string
2911 * can match this ss in the pattern, but they
2912 * remain EXACTF nodes, as they are not trie'able,
2913 * so don't have to invent a new node type to
2914 * exclude them from the trie code */
2915 if (OP(scan) != EXACTF) {
2916 OP(scan) = EXACTFU_SS;
2921 case LATIN_SMALL_LETTER_SHARP_S:
2922 if (OP(scan) == EXACTF) {
2923 *has_exactf_sharp_s = TRUE;
2932 /* Allow dumping but overwriting the collection of skipped
2933 * ops and/or strings with fake optimized ops */
2934 n = scan + NODE_SZ_STR(scan);
2942 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2946 /* REx optimizer. Converts nodes into quicker variants "in place".
2947 Finds fixed substrings. */
2949 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2950 to the position after last scanned or to NULL. */
2952 #define INIT_AND_WITHP \
2953 assert(!and_withp); \
2954 Newx(and_withp,1,struct regnode_charclass_class); \
2955 SAVEFREEPV(and_withp)
2957 /* this is a chain of data about sub patterns we are processing that
2958 need to be handled separately/specially in study_chunk. Its so
2959 we can simulate recursion without losing state. */
2961 typedef struct scan_frame {
2962 regnode *last; /* last node to process in this frame */
2963 regnode *next; /* next node to process when last is reached */
2964 struct scan_frame *prev; /*previous frame*/
2965 I32 stop; /* what stopparen do we use */
2969 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2971 #define CASE_SYNST_FNC(nAmE) \
2973 if (flags & SCF_DO_STCLASS_AND) { \
2974 for (value = 0; value < 256; value++) \
2975 if (!is_ ## nAmE ## _cp(value)) \
2976 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2979 for (value = 0; value < 256; value++) \
2980 if (is_ ## nAmE ## _cp(value)) \
2981 ANYOF_BITMAP_SET(data->start_class, value); \
2985 if (flags & SCF_DO_STCLASS_AND) { \
2986 for (value = 0; value < 256; value++) \
2987 if (is_ ## nAmE ## _cp(value)) \
2988 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2991 for (value = 0; value < 256; value++) \
2992 if (!is_ ## nAmE ## _cp(value)) \
2993 ANYOF_BITMAP_SET(data->start_class, value); \
3000 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3001 I32 *minlenp, I32 *deltap,
3006 struct regnode_charclass_class *and_withp,
3007 U32 flags, U32 depth)
3008 /* scanp: Start here (read-write). */
3009 /* deltap: Write maxlen-minlen here. */
3010 /* last: Stop before this one. */
3011 /* data: string data about the pattern */
3012 /* stopparen: treat close N as END */
3013 /* recursed: which subroutines have we recursed into */
3014 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3017 I32 min = 0, pars = 0, code;
3018 regnode *scan = *scanp, *next;
3020 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3021 int is_inf_internal = 0; /* The studied chunk is infinite */
3022 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3023 scan_data_t data_fake;
3024 SV *re_trie_maxbuff = NULL;
3025 regnode *first_non_open = scan;
3026 I32 stopmin = I32_MAX;
3027 scan_frame *frame = NULL;
3028 GET_RE_DEBUG_FLAGS_DECL;
3030 PERL_ARGS_ASSERT_STUDY_CHUNK;
3033 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3037 while (first_non_open && OP(first_non_open) == OPEN)
3038 first_non_open=regnext(first_non_open);
3043 while ( scan && OP(scan) != END && scan < last ){
3044 UV min_subtract = 0; /* How much to subtract from the minimum node
3045 length to get a real minimum (because the
3046 folded version may be shorter) */
3047 bool has_exactf_sharp_s = FALSE;
3048 /* Peephole optimizer: */
3049 DEBUG_STUDYDATA("Peep:", data,depth);
3050 DEBUG_PEEP("Peep",scan,depth);
3052 /* Its not clear to khw or hv why this is done here, and not in the
3053 * clauses that deal with EXACT nodes. khw's guess is that it's
3054 * because of a previous design */
3055 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3057 /* Follow the next-chain of the current node and optimize
3058 away all the NOTHINGs from it. */
3059 if (OP(scan) != CURLYX) {
3060 const int max = (reg_off_by_arg[OP(scan)]
3062 /* I32 may be smaller than U16 on CRAYs! */
3063 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3064 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3068 /* Skip NOTHING and LONGJMP. */
3069 while ((n = regnext(n))
3070 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3071 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3072 && off + noff < max)
3074 if (reg_off_by_arg[OP(scan)])
3077 NEXT_OFF(scan) = off;
3082 /* The principal pseudo-switch. Cannot be a switch, since we
3083 look into several different things. */
3084 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3085 || OP(scan) == IFTHEN) {
3086 next = regnext(scan);
3088 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3090 if (OP(next) == code || code == IFTHEN) {
3091 /* NOTE - There is similar code to this block below for handling
3092 TRIE nodes on a re-study. If you change stuff here check there
3094 I32 max1 = 0, min1 = I32_MAX, num = 0;
3095 struct regnode_charclass_class accum;
3096 regnode * const startbranch=scan;
3098 if (flags & SCF_DO_SUBSTR)
3099 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3100 if (flags & SCF_DO_STCLASS)
3101 cl_init_zero(pRExC_state, &accum);
3103 while (OP(scan) == code) {
3104 I32 deltanext, minnext, f = 0, fake;
3105 struct regnode_charclass_class this_class;
3108 data_fake.flags = 0;
3110 data_fake.whilem_c = data->whilem_c;
3111 data_fake.last_closep = data->last_closep;
3114 data_fake.last_closep = &fake;
3116 data_fake.pos_delta = delta;
3117 next = regnext(scan);
3118 scan = NEXTOPER(scan);
3120 scan = NEXTOPER(scan);
3121 if (flags & SCF_DO_STCLASS) {
3122 cl_init(pRExC_state, &this_class);
3123 data_fake.start_class = &this_class;
3124 f = SCF_DO_STCLASS_AND;
3126 if (flags & SCF_WHILEM_VISITED_POS)
3127 f |= SCF_WHILEM_VISITED_POS;
3129 /* we suppose the run is continuous, last=next...*/
3130 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3132 stopparen, recursed, NULL, f,depth+1);
3135 if (max1 < minnext + deltanext)
3136 max1 = minnext + deltanext;
3137 if (deltanext == I32_MAX)
3138 is_inf = is_inf_internal = 1;
3140 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3142 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3143 if ( stopmin > minnext)
3144 stopmin = min + min1;
3145 flags &= ~SCF_DO_SUBSTR;
3147 data->flags |= SCF_SEEN_ACCEPT;
3150 if (data_fake.flags & SF_HAS_EVAL)
3151 data->flags |= SF_HAS_EVAL;
3152 data->whilem_c = data_fake.whilem_c;
3154 if (flags & SCF_DO_STCLASS)
3155 cl_or(pRExC_state, &accum, &this_class);
3157 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3159 if (flags & SCF_DO_SUBSTR) {
3160 data->pos_min += min1;
3161 data->pos_delta += max1 - min1;
3162 if (max1 != min1 || is_inf)
3163 data->longest = &(data->longest_float);
3166 delta += max1 - min1;
3167 if (flags & SCF_DO_STCLASS_OR) {
3168 cl_or(pRExC_state, data->start_class, &accum);
3170 cl_and(data->start_class, and_withp);
3171 flags &= ~SCF_DO_STCLASS;
3174 else if (flags & SCF_DO_STCLASS_AND) {
3176 cl_and(data->start_class, &accum);
3177 flags &= ~SCF_DO_STCLASS;
3180 /* Switch to OR mode: cache the old value of
3181 * data->start_class */
3183 StructCopy(data->start_class, and_withp,
3184 struct regnode_charclass_class);
3185 flags &= ~SCF_DO_STCLASS_AND;
3186 StructCopy(&accum, data->start_class,
3187 struct regnode_charclass_class);
3188 flags |= SCF_DO_STCLASS_OR;
3189 data->start_class->flags |= ANYOF_EOS;
3193 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3196 Assuming this was/is a branch we are dealing with: 'scan' now
3197 points at the item that follows the branch sequence, whatever
3198 it is. We now start at the beginning of the sequence and look
3205 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3207 If we can find such a subsequence we need to turn the first
3208 element into a trie and then add the subsequent branch exact
3209 strings to the trie.
3213 1. patterns where the whole set of branches can be converted.
3215 2. patterns where only a subset can be converted.
3217 In case 1 we can replace the whole set with a single regop
3218 for the trie. In case 2 we need to keep the start and end
3221 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3222 becomes BRANCH TRIE; BRANCH X;
3224 There is an additional case, that being where there is a
3225 common prefix, which gets split out into an EXACT like node
3226 preceding the TRIE node.
3228 If x(1..n)==tail then we can do a simple trie, if not we make
3229 a "jump" trie, such that when we match the appropriate word
3230 we "jump" to the appropriate tail node. Essentially we turn
3231 a nested if into a case structure of sorts.
3236 if (!re_trie_maxbuff) {
3237 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3238 if (!SvIOK(re_trie_maxbuff))
3239 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3241 if ( SvIV(re_trie_maxbuff)>=0 ) {
3243 regnode *first = (regnode *)NULL;
3244 regnode *last = (regnode *)NULL;
3245 regnode *tail = scan;
3250 SV * const mysv = sv_newmortal(); /* for dumping */
3252 /* var tail is used because there may be a TAIL
3253 regop in the way. Ie, the exacts will point to the
3254 thing following the TAIL, but the last branch will
3255 point at the TAIL. So we advance tail. If we
3256 have nested (?:) we may have to move through several
3260 while ( OP( tail ) == TAIL ) {
3261 /* this is the TAIL generated by (?:) */
3262 tail = regnext( tail );
3266 DEBUG_TRIE_COMPILE_r({
3267 regprop(RExC_rx, mysv, tail );
3268 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3269 (int)depth * 2 + 2, "",
3270 "Looking for TRIE'able sequences. Tail node is: ",
3271 SvPV_nolen_const( mysv )
3277 Step through the branches
3278 cur represents each branch,
3279 noper is the first thing to be matched as part of that branch
3280 noper_next is the regnext() of that node.
3282 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3283 via a "jump trie" but we also support building with NOJUMPTRIE,
3284 which restricts the trie logic to structures like /FOO|BAR/.
3286 If noper is a trieable nodetype then the branch is a possible optimization
3287 target. If we are building under NOJUMPTRIE then we require that noper_next
3288 is the same as scan (our current position in the regex program).
3290 Once we have two or more consecutive such branches we can create a
3291 trie of the EXACT's contents and stitch it in place into the program.
3293 If the sequence represents all of the branches in the alternation we
3294 replace the entire thing with a single TRIE node.
3296 Otherwise when it is a subsequence we need to stitch it in place and
3297 replace only the relevant branches. This means the first branch has
3298 to remain as it is used by the alternation logic, and its next pointer,
3299 and needs to be repointed at the item on the branch chain following
3300 the last branch we have optimized away.
3302 This could be either a BRANCH, in which case the subsequence is internal,
3303 or it could be the item following the branch sequence in which case the
3304 subsequence is at the end (which does not necessarily mean the first node
3305 is the start of the alternation).
3307 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3310 ----------------+-----------
3314 EXACTFU_SS | EXACTFU
3315 EXACTFU_TRICKYFOLD | EXACTFU
3320 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3321 ( EXACT == (X) ) ? EXACT : \
3322 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3325 /* dont use tail as the end marker for this traverse */
3326 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3327 regnode * const noper = NEXTOPER( cur );
3328 U8 noper_type = OP( noper );
3329 U8 noper_trietype = TRIE_TYPE( noper_type );
3330 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3331 regnode * const noper_next = regnext( noper );
3332 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3333 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3336 DEBUG_TRIE_COMPILE_r({
3337 regprop(RExC_rx, mysv, cur);
3338 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3339 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3341 regprop(RExC_rx, mysv, noper);
3342 PerlIO_printf( Perl_debug_log, " -> %s",
3343 SvPV_nolen_const(mysv));
3346 regprop(RExC_rx, mysv, noper_next );
3347 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3348 SvPV_nolen_const(mysv));
3350 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3351 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3352 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3356 /* Is noper a trieable nodetype that can be merged with the
3357 * current trie (if there is one)? */
3361 ( noper_trietype == NOTHING)
3362 || ( trietype == NOTHING )
3363 || ( trietype == noper_trietype )
3366 && noper_next == tail
3370 /* Handle mergable triable node
3371 * Either we are the first node in a new trieable sequence,
3372 * in which case we do some bookkeeping, otherwise we update
3373 * the end pointer. */
3376 trietype = noper_trietype;
3377 if ( noper_trietype == NOTHING ) {
3378 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3379 regnode * const noper_next = regnext( noper );
3380 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3381 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3384 if ( noper_next_trietype )
3385 trietype = noper_next_trietype;
3388 if ( trietype == NOTHING )
3389 trietype = noper_trietype;
3394 } /* end handle mergable triable node */
3396 /* handle unmergable node -
3397 * noper may either be a triable node which can not be tried
3398 * together with the current trie, or a non triable node */
3400 /* If last is set and trietype is not NOTHING then we have found
3401 * at least two triable branch sequences in a row of a similar
3402 * trietype so we can turn them into a trie. If/when we
3403 * allow NOTHING to start a trie sequence this condition will be
3404 * required, and it isn't expensive so we leave it in for now. */
3405 if ( trietype != NOTHING )
3406 make_trie( pRExC_state,
3407 startbranch, first, cur, tail, count,
3408 trietype, depth+1 );
3409 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3413 && noper_next == tail
3416 /* noper is triable, so we can start a new trie sequence */
3419 trietype = noper_trietype;
3421 /* if we already saw a first but the current node is not triable then we have
3422 * to reset the first information. */
3427 } /* end handle unmergable node */
3428 } /* loop over branches */
3429 DEBUG_TRIE_COMPILE_r({
3430 regprop(RExC_rx, mysv, cur);
3431 PerlIO_printf( Perl_debug_log,
3432 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3433 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3437 if ( trietype != NOTHING ) {
3438 /* the last branch of the sequence was part of a trie,
3439 * so we have to construct it here outside of the loop
3441 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3442 #ifdef TRIE_STUDY_OPT
3443 if ( ((made == MADE_EXACT_TRIE &&
3444 startbranch == first)
3445 || ( first_non_open == first )) &&
3447 flags |= SCF_TRIE_RESTUDY;
3448 if ( startbranch == first
3451 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3456 /* at this point we know whatever we have is a NOTHING sequence/branch
3457 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3459 if ( startbranch == first ) {
3461 /* the entire thing is a NOTHING sequence, something like this:
3462 * (?:|) So we can turn it into a plain NOTHING op. */
3463 DEBUG_TRIE_COMPILE_r({
3464 regprop(RExC_rx, mysv, cur);
3465 PerlIO_printf( Perl_debug_log,
3466 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3467 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3470 OP(startbranch)= NOTHING;
3471 NEXT_OFF(startbranch)= tail - startbranch;
3472 for ( opt= startbranch + 1; opt < tail ; opt++ )
3476 } /* end if ( last) */
3477 } /* TRIE_MAXBUF is non zero */
3482 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3483 scan = NEXTOPER(NEXTOPER(scan));
3484 } else /* single branch is optimized. */
3485 scan = NEXTOPER(scan);
3487 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3488 scan_frame *newframe = NULL;
3493 if (OP(scan) != SUSPEND) {
3494 /* set the pointer */
3495 if (OP(scan) == GOSUB) {
3497 RExC_recurse[ARG2L(scan)] = scan;
3498 start = RExC_open_parens[paren-1];
3499 end = RExC_close_parens[paren-1];
3502 start = RExC_rxi->program + 1;
3506 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3507 SAVEFREEPV(recursed);
3509 if (!PAREN_TEST(recursed,paren+1)) {
3510 PAREN_SET(recursed,paren+1);
3511 Newx(newframe,1,scan_frame);
3513 if (flags & SCF_DO_SUBSTR) {
3514 SCAN_COMMIT(pRExC_state,data,minlenp);
3515 data->longest = &(data->longest_float);
3517 is_inf = is_inf_internal = 1;
3518 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3519 cl_anything(pRExC_state, data->start_class);
3520 flags &= ~SCF_DO_STCLASS;
3523 Newx(newframe,1,scan_frame);
3526 end = regnext(scan);
3531 SAVEFREEPV(newframe);
3532 newframe->next = regnext(scan);
3533 newframe->last = last;
3534 newframe->stop = stopparen;
3535 newframe->prev = frame;
3545 else if (OP(scan) == EXACT) {
3546 I32 l = STR_LEN(scan);
3549 const U8 * const s = (U8*)STRING(scan);
3550 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3551 l = utf8_length(s, s + l);
3553 uc = *((U8*)STRING(scan));
3556 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3557 /* The code below prefers earlier match for fixed
3558 offset, later match for variable offset. */
3559 if (data->last_end == -1) { /* Update the start info. */
3560 data->last_start_min = data->pos_min;
3561 data->last_start_max = is_inf
3562 ? I32_MAX : data->pos_min + data->pos_delta;
3564 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3566 SvUTF8_on(data->last_found);
3568 SV * const sv = data->last_found;
3569 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3570 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3571 if (mg && mg->mg_len >= 0)
3572 mg->mg_len += utf8_length((U8*)STRING(scan),
3573 (U8*)STRING(scan)+STR_LEN(scan));
3575 data->last_end = data->pos_min + l;
3576 data->pos_min += l; /* As in the first entry. */
3577 data->flags &= ~SF_BEFORE_EOL;
3579 if (flags & SCF_DO_STCLASS_AND) {
3580 /* Check whether it is compatible with what we know already! */
3584 /* If compatible, we or it in below. It is compatible if is
3585 * in the bitmp and either 1) its bit or its fold is set, or 2)
3586 * it's for a locale. Even if there isn't unicode semantics
3587 * here, at runtime there may be because of matching against a
3588 * utf8 string, so accept a possible false positive for
3589 * latin1-range folds */
3591 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3592 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3593 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
3594 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3599 ANYOF_CLASS_ZERO(data->start_class);
3600 ANYOF_BITMAP_ZERO(data->start_class);
3602 ANYOF_BITMAP_SET(data->start_class, uc);
3603 else if (uc >= 0x100) {
3606 /* Some Unicode code points fold to the Latin1 range; as
3607 * XXX temporary code, instead of figuring out if this is
3608 * one, just assume it is and set all the start class bits
3609 * that could be some such above 255 code point's fold
3610 * which will generate fals positives. As the code
3611 * elsewhere that does compute the fold settles down, it
3612 * can be extracted out and re-used here */
3613 for (i = 0; i < 256; i++){
3614 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3615 ANYOF_BITMAP_SET(data->start_class, i);
3619 data->start_class->flags &= ~ANYOF_EOS;
3621 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3623 else if (flags & SCF_DO_STCLASS_OR) {
3624 /* false positive possible if the class is case-folded */
3626 ANYOF_BITMAP_SET(data->start_class, uc);
3628 data->start_class->flags |= ANYOF_UNICODE_ALL;
3629 data->start_class->flags &= ~ANYOF_EOS;
3630 cl_and(data->start_class, and_withp);
3632 flags &= ~SCF_DO_STCLASS;
3634 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3635 I32 l = STR_LEN(scan);
3636 UV uc = *((U8*)STRING(scan));
3638 /* Search for fixed substrings supports EXACT only. */
3639 if (flags & SCF_DO_SUBSTR) {
3641 SCAN_COMMIT(pRExC_state, data, minlenp);
3644 const U8 * const s = (U8 *)STRING(scan);
3645 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3646 l = utf8_length(s, s + l);
3648 else if (has_exactf_sharp_s) {
3649 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3651 min += l - min_subtract;
3655 delta += min_subtract;
3656 if (flags & SCF_DO_SUBSTR) {
3657 data->pos_min += l - min_subtract;
3658 if (data->pos_min < 0) {
3661 data->pos_delta += min_subtract;
3663 data->longest = &(data->longest_float);
3666 if (flags & SCF_DO_STCLASS_AND) {
3667 /* Check whether it is compatible with what we know already! */
3670 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3671 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3672 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3676 ANYOF_CLASS_ZERO(data->start_class);
3677 ANYOF_BITMAP_ZERO(data->start_class);
3679 ANYOF_BITMAP_SET(data->start_class, uc);
3680 data->start_class->flags &= ~ANYOF_EOS;
3681 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
3682 if (OP(scan) == EXACTFL) {
3683 /* XXX This set is probably no longer necessary, and
3684 * probably wrong as LOCALE now is on in the initial
3686 data->start_class->flags |= ANYOF_LOCALE;
3690 /* Also set the other member of the fold pair. In case
3691 * that unicode semantics is called for at runtime, use
3692 * the full latin1 fold. (Can't do this for locale,
3693 * because not known until runtime) */
3694 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3696 /* All other (EXACTFL handled above) folds except under
3697 * /iaa that include s, S, and sharp_s also may include
3699 if (OP(scan) != EXACTFA) {
3700 if (uc == 's' || uc == 'S') {
3701 ANYOF_BITMAP_SET(data->start_class,
3702 LATIN_SMALL_LETTER_SHARP_S);
3704 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3705 ANYOF_BITMAP_SET(data->start_class, 's');
3706 ANYOF_BITMAP_SET(data->start_class, 'S');
3711 else if (uc >= 0x100) {
3713 for (i = 0; i < 256; i++){
3714 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3715 ANYOF_BITMAP_SET(data->start_class, i);
3720 else if (flags & SCF_DO_STCLASS_OR) {
3721 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
3722 /* false positive possible if the class is case-folded.
3723 Assume that the locale settings are the same... */
3725 ANYOF_BITMAP_SET(data->start_class, uc);
3726 if (OP(scan) != EXACTFL) {
3728 /* And set the other member of the fold pair, but
3729 * can't do that in locale because not known until
3731 ANYOF_BITMAP_SET(data->start_class,
3732 PL_fold_latin1[uc]);
3734 /* All folds except under /iaa that include s, S,
3735 * and sharp_s also may include the others */
3736 if (OP(scan) != EXACTFA) {
3737 if (uc == 's' || uc == 'S') {
3738 ANYOF_BITMAP_SET(data->start_class,
3739 LATIN_SMALL_LETTER_SHARP_S);
3741 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3742 ANYOF_BITMAP_SET(data->start_class, 's');
3743 ANYOF_BITMAP_SET(data->start_class, 'S');
3748 data->start_class->flags &= ~ANYOF_EOS;
3750 cl_and(data->start_class, and_withp);
3752 flags &= ~SCF_DO_STCLASS;
3754 else if (REGNODE_VARIES(OP(scan))) {
3755 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3756 I32 f = flags, pos_before = 0;
3757 regnode * const oscan = scan;
3758 struct regnode_charclass_class this_class;
3759 struct regnode_charclass_class *oclass = NULL;
3760 I32 next_is_eval = 0;
3762 switch (PL_regkind[OP(scan)]) {
3763 case WHILEM: /* End of (?:...)* . */
3764 scan = NEXTOPER(scan);
3767 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3768 next = NEXTOPER(scan);
3769 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3771 maxcount = REG_INFTY;
3772 next = regnext(scan);
3773 scan = NEXTOPER(scan);
3777 if (flags & SCF_DO_SUBSTR)
3782 if (flags & SCF_DO_STCLASS) {
3784 maxcount = REG_INFTY;
3785 next = regnext(scan);
3786 scan = NEXTOPER(scan);
3789 is_inf = is_inf_internal = 1;
3790 scan = regnext(scan);
3791 if (flags & SCF_DO_SUBSTR) {
3792 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3793 data->longest = &(data->longest_float);
3795 goto optimize_curly_tail;
3797 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3798 && (scan->flags == stopparen))
3803 mincount = ARG1(scan);
3804 maxcount = ARG2(scan);
3806 next = regnext(scan);
3807 if (OP(scan) == CURLYX) {
3808 I32 lp = (data ? *(data->last_closep) : 0);
3809 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3811 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3812 next_is_eval = (OP(scan) == EVAL);
3814 if (flags & SCF_DO_SUBSTR) {
3815 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3816 pos_before = data->pos_min;
3820 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3822 data->flags |= SF_IS_INF;
3824 if (flags & SCF_DO_STCLASS) {
<
https://perl5.git.perl.org / perl5.git / blob