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
84 extern const struct regexp_engine my_reg_engine;
89 #include "dquote_static.c"
90 #ifndef PERL_IN_XSUB_RE
91 # include "charclass_invlists.h"
94 #define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
101 # if defined(BUGGY_MSC6)
102 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
103 # pragma optimize("a",off)
104 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
105 # pragma optimize("w",on )
106 # endif /* BUGGY_MSC6 */
110 #define STATIC static
114 typedef struct RExC_state_t {
115 U32 flags; /* RXf_* are we folding, multilining? */
116 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
117 char *precomp; /* uncompiled string. */
118 REGEXP *rx_sv; /* The SV that is the regexp. */
119 regexp *rx; /* perl core regexp structure */
120 regexp_internal *rxi; /* internal data for regexp object pprivate field */
121 char *start; /* Start of input for compile */
122 char *end; /* End of input for compile */
123 char *parse; /* Input-scan pointer. */
124 I32 whilem_seen; /* number of WHILEM in this expr */
125 regnode *emit_start; /* Start of emitted-code area */
126 regnode *emit_bound; /* First regnode outside of the allocated space */
127 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
128 I32 naughty; /* How bad is this pattern? */
129 I32 sawback; /* Did we see \1, ...? */
131 I32 size; /* Code size. */
132 I32 npar; /* Capture buffer count, (OPEN). */
133 I32 cpar; /* Capture buffer count, (CLOSE). */
134 I32 nestroot; /* root parens we are in - used by accept */
137 regnode **open_parens; /* pointers to open parens */
138 regnode **close_parens; /* pointers to close parens */
139 regnode *opend; /* END node in program */
140 I32 utf8; /* whether the pattern is utf8 or not */
141 I32 orig_utf8; /* whether the pattern was originally in utf8 */
142 /* XXX use this for future optimisation of case
143 * where pattern must be upgraded to utf8. */
144 I32 uni_semantics; /* If a d charset modifier should use unicode
145 rules, even if the pattern is not in
147 HV *paren_names; /* Paren names */
149 regnode **recurse; /* Recurse regops */
150 I32 recurse_count; /* Number of recurse regops */
153 I32 override_recoding;
154 struct reg_code_block *code_blocks; /* positions of literal (?{})
156 int num_code_blocks; /* size of code_blocks[] */
157 int code_index; /* next code_blocks[] slot */
159 char *starttry; /* -Dr: where regtry was called. */
160 #define RExC_starttry (pRExC_state->starttry)
162 SV *runtime_code_qr; /* qr with the runtime code blocks */
164 const char *lastparse;
166 AV *paren_name_list; /* idx -> name */
167 #define RExC_lastparse (pRExC_state->lastparse)
168 #define RExC_lastnum (pRExC_state->lastnum)
169 #define RExC_paren_name_list (pRExC_state->paren_name_list)
173 #define RExC_flags (pRExC_state->flags)
174 #define RExC_pm_flags (pRExC_state->pm_flags)
175 #define RExC_precomp (pRExC_state->precomp)
176 #define RExC_rx_sv (pRExC_state->rx_sv)
177 #define RExC_rx (pRExC_state->rx)
178 #define RExC_rxi (pRExC_state->rxi)
179 #define RExC_start (pRExC_state->start)
180 #define RExC_end (pRExC_state->end)
181 #define RExC_parse (pRExC_state->parse)
182 #define RExC_whilem_seen (pRExC_state->whilem_seen)
183 #ifdef RE_TRACK_PATTERN_OFFSETS
184 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
186 #define RExC_emit (pRExC_state->emit)
187 #define RExC_emit_start (pRExC_state->emit_start)
188 #define RExC_emit_bound (pRExC_state->emit_bound)
189 #define RExC_naughty (pRExC_state->naughty)
190 #define RExC_sawback (pRExC_state->sawback)
191 #define RExC_seen (pRExC_state->seen)
192 #define RExC_size (pRExC_state->size)
193 #define RExC_npar (pRExC_state->npar)
194 #define RExC_nestroot (pRExC_state->nestroot)
195 #define RExC_extralen (pRExC_state->extralen)
196 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
197 #define RExC_utf8 (pRExC_state->utf8)
198 #define RExC_uni_semantics (pRExC_state->uni_semantics)
199 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
200 #define RExC_open_parens (pRExC_state->open_parens)
201 #define RExC_close_parens (pRExC_state->close_parens)
202 #define RExC_opend (pRExC_state->opend)
203 #define RExC_paren_names (pRExC_state->paren_names)
204 #define RExC_recurse (pRExC_state->recurse)
205 #define RExC_recurse_count (pRExC_state->recurse_count)
206 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
207 #define RExC_contains_locale (pRExC_state->contains_locale)
208 #define RExC_override_recoding (pRExC_state->override_recoding)
211 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
212 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
213 ((*s) == '{' && regcurly(s)))
216 #undef SPSTART /* dratted cpp namespace... */
219 * Flags to be passed up and down.
221 #define WORST 0 /* Worst case. */
222 #define HASWIDTH 0x01 /* Known to match non-null strings. */
224 /* Simple enough to be STAR/PLUS operand; in an EXACT node must be a single
225 * character, and if utf8, must be invariant. Note that this is not the same
226 * thing as REGNODE_SIMPLE */
228 #define SPSTART 0x04 /* Starts with * or +. */
229 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
230 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
232 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
234 /* whether trie related optimizations are enabled */
235 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
236 #define TRIE_STUDY_OPT
237 #define FULL_TRIE_STUDY
243 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
244 #define PBITVAL(paren) (1 << ((paren) & 7))
245 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
246 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
247 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
249 /* If not already in utf8, do a longjmp back to the beginning */
250 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
251 #define REQUIRE_UTF8 STMT_START { \
252 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
255 /* About scan_data_t.
257 During optimisation we recurse through the regexp program performing
258 various inplace (keyhole style) optimisations. In addition study_chunk
259 and scan_commit populate this data structure with information about
260 what strings MUST appear in the pattern. We look for the longest
261 string that must appear at a fixed location, and we look for the
262 longest string that may appear at a floating location. So for instance
267 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
268 strings (because they follow a .* construct). study_chunk will identify
269 both FOO and BAR as being the longest fixed and floating strings respectively.
271 The strings can be composites, for instance
275 will result in a composite fixed substring 'foo'.
277 For each string some basic information is maintained:
279 - offset or min_offset
280 This is the position the string must appear at, or not before.
281 It also implicitly (when combined with minlenp) tells us how many
282 characters must match before the string we are searching for.
283 Likewise when combined with minlenp and the length of the string it
284 tells us how many characters must appear after the string we have
288 Only used for floating strings. This is the rightmost point that
289 the string can appear at. If set to I32 max it indicates that the
290 string can occur infinitely far to the right.
293 A pointer to the minimum length of the pattern that the string
294 was found inside. This is important as in the case of positive
295 lookahead or positive lookbehind we can have multiple patterns
300 The minimum length of the pattern overall is 3, the minimum length
301 of the lookahead part is 3, but the minimum length of the part that
302 will actually match is 1. So 'FOO's minimum length is 3, but the
303 minimum length for the F is 1. This is important as the minimum length
304 is used to determine offsets in front of and behind the string being
305 looked for. Since strings can be composites this is the length of the
306 pattern at the time it was committed with a scan_commit. Note that
307 the length is calculated by study_chunk, so that the minimum lengths
308 are not known until the full pattern has been compiled, thus the
309 pointer to the value.
313 In the case of lookbehind the string being searched for can be
314 offset past the start point of the final matching string.
315 If this value was just blithely removed from the min_offset it would
316 invalidate some of the calculations for how many chars must match
317 before or after (as they are derived from min_offset and minlen and
318 the length of the string being searched for).
319 When the final pattern is compiled and the data is moved from the
320 scan_data_t structure into the regexp structure the information
321 about lookbehind is factored in, with the information that would
322 have been lost precalculated in the end_shift field for the
325 The fields pos_min and pos_delta are used to store the minimum offset
326 and the delta to the maximum offset at the current point in the pattern.
330 typedef struct scan_data_t {
331 /*I32 len_min; unused */
332 /*I32 len_delta; unused */
336 I32 last_end; /* min value, <0 unless valid. */
339 SV **longest; /* Either &l_fixed, or &l_float. */
340 SV *longest_fixed; /* longest fixed string found in pattern */
341 I32 offset_fixed; /* offset where it starts */
342 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
343 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
344 SV *longest_float; /* longest floating string found in pattern */
345 I32 offset_float_min; /* earliest point in string it can appear */
346 I32 offset_float_max; /* latest point in string it can appear */
347 I32 *minlen_float; /* pointer to the minlen relevant to the string */
348 I32 lookbehind_float; /* is the position of the string modified by LB */
352 struct regnode_charclass_class *start_class;
356 * Forward declarations for pregcomp()'s friends.
359 static const scan_data_t zero_scan_data =
360 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
362 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
363 #define SF_BEFORE_SEOL 0x0001
364 #define SF_BEFORE_MEOL 0x0002
365 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
366 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
369 # define SF_FIX_SHIFT_EOL (0+2)
370 # define SF_FL_SHIFT_EOL (0+4)
372 # define SF_FIX_SHIFT_EOL (+2)
373 # define SF_FL_SHIFT_EOL (+4)
376 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
377 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
379 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
380 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
381 #define SF_IS_INF 0x0040
382 #define SF_HAS_PAR 0x0080
383 #define SF_IN_PAR 0x0100
384 #define SF_HAS_EVAL 0x0200
385 #define SCF_DO_SUBSTR 0x0400
386 #define SCF_DO_STCLASS_AND 0x0800
387 #define SCF_DO_STCLASS_OR 0x1000
388 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
389 #define SCF_WHILEM_VISITED_POS 0x2000
391 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
392 #define SCF_SEEN_ACCEPT 0x8000
394 #define UTF cBOOL(RExC_utf8)
396 /* The enums for all these are ordered so things work out correctly */
397 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
398 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
399 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
400 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
401 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
402 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
403 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
405 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
407 #define OOB_NAMEDCLASS -1
409 /* There is no code point that is out-of-bounds, so this is problematic. But
410 * its only current use is to initialize a variable that is always set before
412 #define OOB_UNICODE 0xDEADBEEF
414 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
415 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
418 /* length of regex to show in messages that don't mark a position within */
419 #define RegexLengthToShowInErrorMessages 127
422 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
423 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
424 * op/pragma/warn/regcomp.
426 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
427 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
429 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
432 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
433 * arg. Show regex, up to a maximum length. If it's too long, chop and add
436 #define _FAIL(code) STMT_START { \
437 const char *ellipses = ""; \
438 IV len = RExC_end - RExC_precomp; \
441 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
442 if (len > RegexLengthToShowInErrorMessages) { \
443 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
444 len = RegexLengthToShowInErrorMessages - 10; \
450 #define FAIL(msg) _FAIL( \
451 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
452 msg, (int)len, RExC_precomp, ellipses))
454 #define FAIL2(msg,arg) _FAIL( \
455 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
456 arg, (int)len, RExC_precomp, ellipses))
459 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
461 #define Simple_vFAIL(m) STMT_START { \
462 const IV offset = RExC_parse - RExC_precomp; \
463 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
464 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
468 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
470 #define vFAIL(m) STMT_START { \
472 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
477 * Like Simple_vFAIL(), but accepts two arguments.
479 #define Simple_vFAIL2(m,a1) STMT_START { \
480 const IV offset = RExC_parse - RExC_precomp; \
481 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
482 (int)offset, RExC_precomp, RExC_precomp + offset); \
486 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
488 #define vFAIL2(m,a1) STMT_START { \
490 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
491 Simple_vFAIL2(m, a1); \
496 * Like Simple_vFAIL(), but accepts three arguments.
498 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
499 const IV offset = RExC_parse - RExC_precomp; \
500 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
501 (int)offset, RExC_precomp, RExC_precomp + offset); \
505 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
507 #define vFAIL3(m,a1,a2) STMT_START { \
509 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
510 Simple_vFAIL3(m, a1, a2); \
514 * Like Simple_vFAIL(), but accepts four arguments.
516 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
517 const IV offset = RExC_parse - RExC_precomp; \
518 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
519 (int)offset, RExC_precomp, RExC_precomp + offset); \
522 #define ckWARNreg(loc,m) STMT_START { \
523 const IV offset = loc - RExC_precomp; \
524 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
525 (int)offset, RExC_precomp, RExC_precomp + offset); \
528 #define ckWARNregdep(loc,m) STMT_START { \
529 const IV offset = loc - RExC_precomp; \
530 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
532 (int)offset, RExC_precomp, RExC_precomp + offset); \
535 #define ckWARN2regdep(loc,m, a1) STMT_START { \
536 const IV offset = loc - RExC_precomp; \
537 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
539 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
542 #define ckWARN2reg(loc, m, a1) STMT_START { \
543 const IV offset = loc - RExC_precomp; \
544 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
545 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
548 #define vWARN3(loc, m, a1, a2) STMT_START { \
549 const IV offset = loc - RExC_precomp; \
550 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
551 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
554 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
555 const IV offset = loc - RExC_precomp; \
556 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
557 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
560 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
561 const IV offset = loc - RExC_precomp; \
562 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
563 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
566 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
567 const IV offset = loc - RExC_precomp; \
568 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
569 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
572 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
573 const IV offset = loc - RExC_precomp; \
574 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
575 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
579 /* Allow for side effects in s */
580 #define REGC(c,s) STMT_START { \
581 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
584 /* Macros for recording node offsets. 20001227 mjd@plover.com
585 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
586 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
587 * Element 0 holds the number n.
588 * Position is 1 indexed.
590 #ifndef RE_TRACK_PATTERN_OFFSETS
591 #define Set_Node_Offset_To_R(node,byte)
592 #define Set_Node_Offset(node,byte)
593 #define Set_Cur_Node_Offset
594 #define Set_Node_Length_To_R(node,len)
595 #define Set_Node_Length(node,len)
596 #define Set_Node_Cur_Length(node)
597 #define Node_Offset(n)
598 #define Node_Length(n)
599 #define Set_Node_Offset_Length(node,offset,len)
600 #define ProgLen(ri) ri->u.proglen
601 #define SetProgLen(ri,x) ri->u.proglen = x
603 #define ProgLen(ri) ri->u.offsets[0]
604 #define SetProgLen(ri,x) ri->u.offsets[0] = x
605 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
607 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
608 __LINE__, (int)(node), (int)(byte))); \
610 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
612 RExC_offsets[2*(node)-1] = (byte); \
617 #define Set_Node_Offset(node,byte) \
618 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
619 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
621 #define Set_Node_Length_To_R(node,len) STMT_START { \
623 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
624 __LINE__, (int)(node), (int)(len))); \
626 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
628 RExC_offsets[2*(node)] = (len); \
633 #define Set_Node_Length(node,len) \
634 Set_Node_Length_To_R((node)-RExC_emit_start, len)
635 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
636 #define Set_Node_Cur_Length(node) \
637 Set_Node_Length(node, RExC_parse - parse_start)
639 /* Get offsets and lengths */
640 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
641 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
643 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
644 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
645 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
649 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
650 #define EXPERIMENTAL_INPLACESCAN
651 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
653 #define DEBUG_STUDYDATA(str,data,depth) \
654 DEBUG_OPTIMISE_MORE_r(if(data){ \
655 PerlIO_printf(Perl_debug_log, \
656 "%*s" str "Pos:%"IVdf"/%"IVdf \
657 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
658 (int)(depth)*2, "", \
659 (IV)((data)->pos_min), \
660 (IV)((data)->pos_delta), \
661 (UV)((data)->flags), \
662 (IV)((data)->whilem_c), \
663 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
664 is_inf ? "INF " : "" \
666 if ((data)->last_found) \
667 PerlIO_printf(Perl_debug_log, \
668 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
669 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
670 SvPVX_const((data)->last_found), \
671 (IV)((data)->last_end), \
672 (IV)((data)->last_start_min), \
673 (IV)((data)->last_start_max), \
674 ((data)->longest && \
675 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
676 SvPVX_const((data)->longest_fixed), \
677 (IV)((data)->offset_fixed), \
678 ((data)->longest && \
679 (data)->longest==&((data)->longest_float)) ? "*" : "", \
680 SvPVX_const((data)->longest_float), \
681 (IV)((data)->offset_float_min), \
682 (IV)((data)->offset_float_max) \
684 PerlIO_printf(Perl_debug_log,"\n"); \
687 static void clear_re(pTHX_ void *r);
689 /* Mark that we cannot extend a found fixed substring at this point.
690 Update the longest found anchored substring and the longest found
691 floating substrings if needed. */
694 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
696 const STRLEN l = CHR_SVLEN(data->last_found);
697 const STRLEN old_l = CHR_SVLEN(*data->longest);
698 GET_RE_DEBUG_FLAGS_DECL;
700 PERL_ARGS_ASSERT_SCAN_COMMIT;
702 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
703 SvSetMagicSV(*data->longest, data->last_found);
704 if (*data->longest == data->longest_fixed) {
705 data->offset_fixed = l ? data->last_start_min : data->pos_min;
706 if (data->flags & SF_BEFORE_EOL)
708 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
710 data->flags &= ~SF_FIX_BEFORE_EOL;
711 data->minlen_fixed=minlenp;
712 data->lookbehind_fixed=0;
714 else { /* *data->longest == data->longest_float */
715 data->offset_float_min = l ? data->last_start_min : data->pos_min;
716 data->offset_float_max = (l
717 ? data->last_start_max
718 : data->pos_min + data->pos_delta);
719 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
720 data->offset_float_max = I32_MAX;
721 if (data->flags & SF_BEFORE_EOL)
723 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
725 data->flags &= ~SF_FL_BEFORE_EOL;
726 data->minlen_float=minlenp;
727 data->lookbehind_float=0;
730 SvCUR_set(data->last_found, 0);
732 SV * const sv = data->last_found;
733 if (SvUTF8(sv) && SvMAGICAL(sv)) {
734 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
740 data->flags &= ~SF_BEFORE_EOL;
741 DEBUG_STUDYDATA("commit: ",data,0);
744 /* Can match anything (initialization) */
746 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
748 PERL_ARGS_ASSERT_CL_ANYTHING;
750 ANYOF_BITMAP_SETALL(cl);
751 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
752 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
754 /* If any portion of the regex is to operate under locale rules,
755 * initialization includes it. The reason this isn't done for all regexes
756 * is that the optimizer was written under the assumption that locale was
757 * all-or-nothing. Given the complexity and lack of documentation in the
758 * optimizer, and that there are inadequate test cases for locale, so many
759 * parts of it may not work properly, it is safest to avoid locale unless
761 if (RExC_contains_locale) {
762 ANYOF_CLASS_SETALL(cl); /* /l uses class */
763 cl->flags |= ANYOF_LOCALE;
766 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
770 /* Can match anything (initialization) */
772 S_cl_is_anything(const struct regnode_charclass_class *cl)
776 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
778 for (value = 0; value <= ANYOF_MAX; value += 2)
779 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
781 if (!(cl->flags & ANYOF_UNICODE_ALL))
783 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
788 /* Can match anything (initialization) */
790 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
792 PERL_ARGS_ASSERT_CL_INIT;
794 Zero(cl, 1, struct regnode_charclass_class);
796 cl_anything(pRExC_state, cl);
797 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
800 /* These two functions currently do the exact same thing */
801 #define cl_init_zero S_cl_init
803 /* 'AND' a given class with another one. Can create false positives. 'cl'
804 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
805 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
807 S_cl_and(struct regnode_charclass_class *cl,
808 const struct regnode_charclass_class *and_with)
810 PERL_ARGS_ASSERT_CL_AND;
812 assert(and_with->type == ANYOF);
814 /* I (khw) am not sure all these restrictions are necessary XXX */
815 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
816 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
817 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
818 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
819 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
822 if (and_with->flags & ANYOF_INVERT)
823 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
824 cl->bitmap[i] &= ~and_with->bitmap[i];
826 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
827 cl->bitmap[i] &= and_with->bitmap[i];
828 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
830 if (and_with->flags & ANYOF_INVERT) {
832 /* Here, the and'ed node is inverted. Get the AND of the flags that
833 * aren't affected by the inversion. Those that are affected are
834 * handled individually below */
835 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
836 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
837 cl->flags |= affected_flags;
839 /* We currently don't know how to deal with things that aren't in the
840 * bitmap, but we know that the intersection is no greater than what
841 * is already in cl, so let there be false positives that get sorted
842 * out after the synthetic start class succeeds, and the node is
843 * matched for real. */
845 /* The inversion of these two flags indicate that the resulting
846 * intersection doesn't have them */
847 if (and_with->flags & ANYOF_UNICODE_ALL) {
848 cl->flags &= ~ANYOF_UNICODE_ALL;
850 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
851 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
854 else { /* and'd node is not inverted */
855 U8 outside_bitmap_but_not_utf8; /* Temp variable */
857 if (! ANYOF_NONBITMAP(and_with)) {
859 /* Here 'and_with' doesn't match anything outside the bitmap
860 * (except possibly ANYOF_UNICODE_ALL), which means the
861 * intersection can't either, except for ANYOF_UNICODE_ALL, in
862 * which case we don't know what the intersection is, but it's no
863 * greater than what cl already has, so can just leave it alone,
864 * with possible false positives */
865 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
866 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
867 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
870 else if (! ANYOF_NONBITMAP(cl)) {
872 /* Here, 'and_with' does match something outside the bitmap, and cl
873 * doesn't have a list of things to match outside the bitmap. If
874 * cl can match all code points above 255, the intersection will
875 * be those above-255 code points that 'and_with' matches. If cl
876 * can't match all Unicode code points, it means that it can't
877 * match anything outside the bitmap (since the 'if' that got us
878 * into this block tested for that), so we leave the bitmap empty.
880 if (cl->flags & ANYOF_UNICODE_ALL) {
881 ARG_SET(cl, ARG(and_with));
883 /* and_with's ARG may match things that don't require UTF8.
884 * And now cl's will too, in spite of this being an 'and'. See
885 * the comments below about the kludge */
886 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
890 /* Here, both 'and_with' and cl match something outside the
891 * bitmap. Currently we do not do the intersection, so just match
892 * whatever cl had at the beginning. */
896 /* Take the intersection of the two sets of flags. However, the
897 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
898 * kludge around the fact that this flag is not treated like the others
899 * which are initialized in cl_anything(). The way the optimizer works
900 * is that the synthetic start class (SSC) is initialized to match
901 * anything, and then the first time a real node is encountered, its
902 * values are AND'd with the SSC's with the result being the values of
903 * the real node. However, there are paths through the optimizer where
904 * the AND never gets called, so those initialized bits are set
905 * inappropriately, which is not usually a big deal, as they just cause
906 * false positives in the SSC, which will just mean a probably
907 * imperceptible slow down in execution. However this bit has a
908 * higher false positive consequence in that it can cause utf8.pm,
909 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
910 * bigger slowdown and also causes significant extra memory to be used.
911 * In order to prevent this, the code now takes a different tack. The
912 * bit isn't set unless some part of the regular expression needs it,
913 * but once set it won't get cleared. This means that these extra
914 * modules won't get loaded unless there was some path through the
915 * pattern that would have required them anyway, and so any false
916 * positives that occur by not ANDing them out when they could be
917 * aren't as severe as they would be if we treated this bit like all
919 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
920 & ANYOF_NONBITMAP_NON_UTF8;
921 cl->flags &= and_with->flags;
922 cl->flags |= outside_bitmap_but_not_utf8;
926 /* 'OR' a given class with another one. Can create false positives. 'cl'
927 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
928 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
930 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
932 PERL_ARGS_ASSERT_CL_OR;
934 if (or_with->flags & ANYOF_INVERT) {
936 /* Here, the or'd node is to be inverted. This means we take the
937 * complement of everything not in the bitmap, but currently we don't
938 * know what that is, so give up and match anything */
939 if (ANYOF_NONBITMAP(or_with)) {
940 cl_anything(pRExC_state, cl);
943 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
944 * <= (B1 | !B2) | (CL1 | !CL2)
945 * which is wasteful if CL2 is small, but we ignore CL2:
946 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
947 * XXXX Can we handle case-fold? Unclear:
948 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
949 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
951 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
952 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
953 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
956 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
957 cl->bitmap[i] |= ~or_with->bitmap[i];
958 } /* XXXX: logic is complicated otherwise */
960 cl_anything(pRExC_state, cl);
963 /* And, we can just take the union of the flags that aren't affected
964 * by the inversion */
965 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
967 /* For the remaining flags:
968 ANYOF_UNICODE_ALL and inverted means to not match anything above
969 255, which means that the union with cl should just be
970 what cl has in it, so can ignore this flag
971 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
972 is 127-255 to match them, but then invert that, so the
973 union with cl should just be what cl has in it, so can
976 } else { /* 'or_with' is not inverted */
977 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
978 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
979 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
980 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
983 /* OR char bitmap and class bitmap separately */
984 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
985 cl->bitmap[i] |= or_with->bitmap[i];
986 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
987 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
988 cl->classflags[i] |= or_with->classflags[i];
989 cl->flags |= ANYOF_CLASS;
992 else { /* XXXX: logic is complicated, leave it along for a moment. */
993 cl_anything(pRExC_state, cl);
996 if (ANYOF_NONBITMAP(or_with)) {
998 /* Use the added node's outside-the-bit-map match if there isn't a
999 * conflict. If there is a conflict (both nodes match something
1000 * outside the bitmap, but what they match outside is not the same
1001 * pointer, and hence not easily compared until XXX we extend
1002 * inversion lists this far), give up and allow the start class to
1003 * match everything outside the bitmap. If that stuff is all above
1004 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1005 if (! ANYOF_NONBITMAP(cl)) {
1006 ARG_SET(cl, ARG(or_with));
1008 else if (ARG(cl) != ARG(or_with)) {
1010 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1011 cl_anything(pRExC_state, cl);
1014 cl->flags |= ANYOF_UNICODE_ALL;
1019 /* Take the union */
1020 cl->flags |= or_with->flags;
1024 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1025 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1026 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1027 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1032 dump_trie(trie,widecharmap,revcharmap)
1033 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1034 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1036 These routines dump out a trie in a somewhat readable format.
1037 The _interim_ variants are used for debugging the interim
1038 tables that are used to generate the final compressed
1039 representation which is what dump_trie expects.
1041 Part of the reason for their existence is to provide a form
1042 of documentation as to how the different representations function.
1047 Dumps the final compressed table form of the trie to Perl_debug_log.
1048 Used for debugging make_trie().
1052 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1053 AV *revcharmap, U32 depth)
1056 SV *sv=sv_newmortal();
1057 int colwidth= widecharmap ? 6 : 4;
1059 GET_RE_DEBUG_FLAGS_DECL;
1061 PERL_ARGS_ASSERT_DUMP_TRIE;
1063 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1064 (int)depth * 2 + 2,"",
1065 "Match","Base","Ofs" );
1067 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1068 SV ** const tmp = av_fetch( revcharmap, state, 0);
1070 PerlIO_printf( Perl_debug_log, "%*s",
1072 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1073 PL_colors[0], PL_colors[1],
1074 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1075 PERL_PV_ESCAPE_FIRSTCHAR
1080 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1081 (int)depth * 2 + 2,"");
1083 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1084 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1085 PerlIO_printf( Perl_debug_log, "\n");
1087 for( state = 1 ; state < trie->statecount ; state++ ) {
1088 const U32 base = trie->states[ state ].trans.base;
1090 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1092 if ( trie->states[ state ].wordnum ) {
1093 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1095 PerlIO_printf( Perl_debug_log, "%6s", "" );
1098 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1103 while( ( base + ofs < trie->uniquecharcount ) ||
1104 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1105 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1108 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1110 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1111 if ( ( base + ofs >= trie->uniquecharcount ) &&
1112 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1113 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1115 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1117 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1119 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1123 PerlIO_printf( Perl_debug_log, "]");
1126 PerlIO_printf( Perl_debug_log, "\n" );
1128 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1129 for (word=1; word <= trie->wordcount; word++) {
1130 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1131 (int)word, (int)(trie->wordinfo[word].prev),
1132 (int)(trie->wordinfo[word].len));
1134 PerlIO_printf(Perl_debug_log, "\n" );
1137 Dumps a fully constructed but uncompressed trie in list form.
1138 List tries normally only are used for construction when the number of
1139 possible chars (trie->uniquecharcount) is very high.
1140 Used for debugging make_trie().
1143 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1144 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1148 SV *sv=sv_newmortal();
1149 int colwidth= widecharmap ? 6 : 4;
1150 GET_RE_DEBUG_FLAGS_DECL;
1152 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1154 /* print out the table precompression. */
1155 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1156 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1157 "------:-----+-----------------\n" );
1159 for( state=1 ; state < next_alloc ; state ++ ) {
1162 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1163 (int)depth * 2 + 2,"", (UV)state );
1164 if ( ! trie->states[ state ].wordnum ) {
1165 PerlIO_printf( Perl_debug_log, "%5s| ","");
1167 PerlIO_printf( Perl_debug_log, "W%4x| ",
1168 trie->states[ state ].wordnum
1171 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1172 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1174 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1176 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1177 PL_colors[0], PL_colors[1],
1178 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1179 PERL_PV_ESCAPE_FIRSTCHAR
1181 TRIE_LIST_ITEM(state,charid).forid,
1182 (UV)TRIE_LIST_ITEM(state,charid).newstate
1185 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1186 (int)((depth * 2) + 14), "");
1189 PerlIO_printf( Perl_debug_log, "\n");
1194 Dumps a fully constructed but uncompressed trie in table form.
1195 This is the normal DFA style state transition table, with a few
1196 twists to facilitate compression later.
1197 Used for debugging make_trie().
1200 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1201 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1206 SV *sv=sv_newmortal();
1207 int colwidth= widecharmap ? 6 : 4;
1208 GET_RE_DEBUG_FLAGS_DECL;
1210 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1213 print out the table precompression so that we can do a visual check
1214 that they are identical.
1217 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1219 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1220 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1222 PerlIO_printf( Perl_debug_log, "%*s",
1224 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1225 PL_colors[0], PL_colors[1],
1226 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1227 PERL_PV_ESCAPE_FIRSTCHAR
1233 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1235 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1236 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1239 PerlIO_printf( Perl_debug_log, "\n" );
1241 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1243 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1244 (int)depth * 2 + 2,"",
1245 (UV)TRIE_NODENUM( state ) );
1247 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1248 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1250 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1252 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1254 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1255 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1257 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1258 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1266 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1267 startbranch: the first branch in the whole branch sequence
1268 first : start branch of sequence of branch-exact nodes.
1269 May be the same as startbranch
1270 last : Thing following the last branch.
1271 May be the same as tail.
1272 tail : item following the branch sequence
1273 count : words in the sequence
1274 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1275 depth : indent depth
1277 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1279 A trie is an N'ary tree where the branches are determined by digital
1280 decomposition of the key. IE, at the root node you look up the 1st character and
1281 follow that branch repeat until you find the end of the branches. Nodes can be
1282 marked as "accepting" meaning they represent a complete word. Eg:
1286 would convert into the following structure. Numbers represent states, letters
1287 following numbers represent valid transitions on the letter from that state, if
1288 the number is in square brackets it represents an accepting state, otherwise it
1289 will be in parenthesis.
1291 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1295 (1) +-i->(6)-+-s->[7]
1297 +-s->(3)-+-h->(4)-+-e->[5]
1299 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1301 This shows that when matching against the string 'hers' we will begin at state 1
1302 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1303 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1304 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1305 single traverse. We store a mapping from accepting to state to which word was
1306 matched, and then when we have multiple possibilities we try to complete the
1307 rest of the regex in the order in which they occured in the alternation.
1309 The only prior NFA like behaviour that would be changed by the TRIE support is
1310 the silent ignoring of duplicate alternations which are of the form:
1312 / (DUPE|DUPE) X? (?{ ... }) Y /x
1314 Thus EVAL blocks following a trie may be called a different number of times with
1315 and without the optimisation. With the optimisations dupes will be silently
1316 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1317 the following demonstrates:
1319 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1321 which prints out 'word' three times, but
1323 'words'=~/(word|word|word)(?{ print $1 })S/
1325 which doesnt print it out at all. This is due to other optimisations kicking in.
1327 Example of what happens on a structural level:
1329 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1331 1: CURLYM[1] {1,32767}(18)
1342 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1343 and should turn into:
1345 1: CURLYM[1] {1,32767}(18)
1347 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1355 Cases where tail != last would be like /(?foo|bar)baz/:
1365 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1366 and would end up looking like:
1369 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1376 d = uvuni_to_utf8_flags(d, uv, 0);
1378 is the recommended Unicode-aware way of saying
1383 #define TRIE_STORE_REVCHAR(val) \
1386 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1387 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1388 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1389 SvCUR_set(zlopp, kapow - flrbbbbb); \
1392 av_push(revcharmap, zlopp); \
1394 char ooooff = (char)val; \
1395 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1399 #define TRIE_READ_CHAR STMT_START { \
1402 /* if it is UTF then it is either already folded, or does not need folding */ \
1403 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1405 else if (folder == PL_fold_latin1) { \
1406 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1407 if ( foldlen > 0 ) { \
1408 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1414 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1415 skiplen = UNISKIP(uvc); \
1416 foldlen -= skiplen; \
1417 scan = foldbuf + skiplen; \
1420 /* raw data, will be folded later if needed */ \
1428 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1429 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1430 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1431 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1433 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1434 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1435 TRIE_LIST_CUR( state )++; \
1438 #define TRIE_LIST_NEW(state) STMT_START { \
1439 Newxz( trie->states[ state ].trans.list, \
1440 4, reg_trie_trans_le ); \
1441 TRIE_LIST_CUR( state ) = 1; \
1442 TRIE_LIST_LEN( state ) = 4; \
1445 #define TRIE_HANDLE_WORD(state) STMT_START { \
1446 U16 dupe= trie->states[ state ].wordnum; \
1447 regnode * const noper_next = regnext( noper ); \
1450 /* store the word for dumping */ \
1452 if (OP(noper) != NOTHING) \
1453 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1455 tmp = newSVpvn_utf8( "", 0, UTF ); \
1456 av_push( trie_words, tmp ); \
1460 trie->wordinfo[curword].prev = 0; \
1461 trie->wordinfo[curword].len = wordlen; \
1462 trie->wordinfo[curword].accept = state; \
1464 if ( noper_next < tail ) { \
1466 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1467 trie->jump[curword] = (U16)(noper_next - convert); \
1469 jumper = noper_next; \
1471 nextbranch= regnext(cur); \
1475 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1476 /* chain, so that when the bits of chain are later */\
1477 /* linked together, the dups appear in the chain */\
1478 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1479 trie->wordinfo[dupe].prev = curword; \
1481 /* we haven't inserted this word yet. */ \
1482 trie->states[ state ].wordnum = curword; \
1487 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1488 ( ( base + charid >= ucharcount \
1489 && base + charid < ubound \
1490 && state == trie->trans[ base - ucharcount + charid ].check \
1491 && trie->trans[ base - ucharcount + charid ].next ) \
1492 ? trie->trans[ base - ucharcount + charid ].next \
1493 : ( state==1 ? special : 0 ) \
1497 #define MADE_JUMP_TRIE 2
1498 #define MADE_EXACT_TRIE 4
1501 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1504 /* first pass, loop through and scan words */
1505 reg_trie_data *trie;
1506 HV *widecharmap = NULL;
1507 AV *revcharmap = newAV();
1509 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1514 regnode *jumper = NULL;
1515 regnode *nextbranch = NULL;
1516 regnode *convert = NULL;
1517 U32 *prev_states; /* temp array mapping each state to previous one */
1518 /* we just use folder as a flag in utf8 */
1519 const U8 * folder = NULL;
1522 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1523 AV *trie_words = NULL;
1524 /* along with revcharmap, this only used during construction but both are
1525 * useful during debugging so we store them in the struct when debugging.
1528 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1529 STRLEN trie_charcount=0;
1531 SV *re_trie_maxbuff;
1532 GET_RE_DEBUG_FLAGS_DECL;
1534 PERL_ARGS_ASSERT_MAKE_TRIE;
1536 PERL_UNUSED_ARG(depth);
1543 case EXACTFU_TRICKYFOLD:
1544 case EXACTFU: folder = PL_fold_latin1; break;
1545 case EXACTF: folder = PL_fold; break;
1546 case EXACTFL: folder = PL_fold_locale; break;
1547 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1550 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1552 trie->startstate = 1;
1553 trie->wordcount = word_count;
1554 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1555 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1557 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1558 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1559 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1562 trie_words = newAV();
1565 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1566 if (!SvIOK(re_trie_maxbuff)) {
1567 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1569 DEBUG_TRIE_COMPILE_r({
1570 PerlIO_printf( Perl_debug_log,
1571 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1572 (int)depth * 2 + 2, "",
1573 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1574 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1578 /* Find the node we are going to overwrite */
1579 if ( first == startbranch && OP( last ) != BRANCH ) {
1580 /* whole branch chain */
1583 /* branch sub-chain */
1584 convert = NEXTOPER( first );
1587 /* -- First loop and Setup --
1589 We first traverse the branches and scan each word to determine if it
1590 contains widechars, and how many unique chars there are, this is
1591 important as we have to build a table with at least as many columns as we
1594 We use an array of integers to represent the character codes 0..255
1595 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1596 native representation of the character value as the key and IV's for the
1599 *TODO* If we keep track of how many times each character is used we can
1600 remap the columns so that the table compression later on is more
1601 efficient in terms of memory by ensuring the most common value is in the
1602 middle and the least common are on the outside. IMO this would be better
1603 than a most to least common mapping as theres a decent chance the most
1604 common letter will share a node with the least common, meaning the node
1605 will not be compressible. With a middle is most common approach the worst
1606 case is when we have the least common nodes twice.
1610 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1611 regnode *noper = NEXTOPER( cur );
1612 const U8 *uc = (U8*)STRING( noper );
1613 const U8 *e = uc + STR_LEN( noper );
1615 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1617 const U8 *scan = (U8*)NULL;
1618 U32 wordlen = 0; /* required init */
1620 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1622 if (OP(noper) == NOTHING) {
1623 regnode *noper_next= regnext(noper);
1624 if (noper_next != tail && OP(noper_next) == flags) {
1626 uc= (U8*)STRING(noper);
1627 e= uc + STR_LEN(noper);
1628 trie->minlen= STR_LEN(noper);
1635 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1636 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1637 regardless of encoding */
1638 if (OP( noper ) == EXACTFU_SS) {
1639 /* false positives are ok, so just set this */
1640 TRIE_BITMAP_SET(trie,0xDF);
1643 for ( ; uc < e ; uc += len ) {
1644 TRIE_CHARCOUNT(trie)++;
1649 U8 folded= folder[ (U8) uvc ];
1650 if ( !trie->charmap[ folded ] ) {
1651 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1652 TRIE_STORE_REVCHAR( folded );
1655 if ( !trie->charmap[ uvc ] ) {
1656 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1657 TRIE_STORE_REVCHAR( uvc );
1660 /* store the codepoint in the bitmap, and its folded
1662 TRIE_BITMAP_SET(trie, uvc);
1664 /* store the folded codepoint */
1665 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1668 /* store first byte of utf8 representation of
1669 variant codepoints */
1670 if (! UNI_IS_INVARIANT(uvc)) {
1671 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1674 set_bit = 0; /* We've done our bit :-) */
1679 widecharmap = newHV();
1681 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1684 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1686 if ( !SvTRUE( *svpp ) ) {
1687 sv_setiv( *svpp, ++trie->uniquecharcount );
1688 TRIE_STORE_REVCHAR(uvc);
1692 if( cur == first ) {
1693 trie->minlen = chars;
1694 trie->maxlen = chars;
1695 } else if (chars < trie->minlen) {
1696 trie->minlen = chars;
1697 } else if (chars > trie->maxlen) {
1698 trie->maxlen = chars;
1700 if (OP( noper ) == EXACTFU_SS) {
1701 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1702 if (trie->minlen > 1)
1705 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1706 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1707 * - We assume that any such sequence might match a 2 byte string */
1708 if (trie->minlen > 2 )
1712 } /* end first pass */
1713 DEBUG_TRIE_COMPILE_r(
1714 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1715 (int)depth * 2 + 2,"",
1716 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1717 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1718 (int)trie->minlen, (int)trie->maxlen )
1722 We now know what we are dealing with in terms of unique chars and
1723 string sizes so we can calculate how much memory a naive
1724 representation using a flat table will take. If it's over a reasonable
1725 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1726 conservative but potentially much slower representation using an array
1729 At the end we convert both representations into the same compressed
1730 form that will be used in regexec.c for matching with. The latter
1731 is a form that cannot be used to construct with but has memory
1732 properties similar to the list form and access properties similar
1733 to the table form making it both suitable for fast searches and
1734 small enough that its feasable to store for the duration of a program.
1736 See the comment in the code where the compressed table is produced
1737 inplace from the flat tabe representation for an explanation of how
1738 the compression works.
1743 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1746 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1748 Second Pass -- Array Of Lists Representation
1750 Each state will be represented by a list of charid:state records
1751 (reg_trie_trans_le) the first such element holds the CUR and LEN
1752 points of the allocated array. (See defines above).
1754 We build the initial structure using the lists, and then convert
1755 it into the compressed table form which allows faster lookups
1756 (but cant be modified once converted).
1759 STRLEN transcount = 1;
1761 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1762 "%*sCompiling trie using list compiler\n",
1763 (int)depth * 2 + 2, ""));
1765 trie->states = (reg_trie_state *)
1766 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1767 sizeof(reg_trie_state) );
1771 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1773 regnode *noper = NEXTOPER( cur );
1774 U8 *uc = (U8*)STRING( noper );
1775 const U8 *e = uc + STR_LEN( noper );
1776 U32 state = 1; /* required init */
1777 U16 charid = 0; /* sanity init */
1778 U8 *scan = (U8*)NULL; /* sanity init */
1779 STRLEN foldlen = 0; /* required init */
1780 U32 wordlen = 0; /* required init */
1781 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1784 if (OP(noper) == NOTHING) {
1785 regnode *noper_next= regnext(noper);
1786 if (noper_next != tail && OP(noper_next) == flags) {
1788 uc= (U8*)STRING(noper);
1789 e= uc + STR_LEN(noper);
1793 if (OP(noper) != NOTHING) {
1794 for ( ; uc < e ; uc += len ) {
1799 charid = trie->charmap[ uvc ];
1801 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1805 charid=(U16)SvIV( *svpp );
1808 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1815 if ( !trie->states[ state ].trans.list ) {
1816 TRIE_LIST_NEW( state );
1818 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1819 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1820 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1825 newstate = next_alloc++;
1826 prev_states[newstate] = state;
1827 TRIE_LIST_PUSH( state, charid, newstate );
1832 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1836 TRIE_HANDLE_WORD(state);
1838 } /* end second pass */
1840 /* next alloc is the NEXT state to be allocated */
1841 trie->statecount = next_alloc;
1842 trie->states = (reg_trie_state *)
1843 PerlMemShared_realloc( trie->states,
1845 * sizeof(reg_trie_state) );
1847 /* and now dump it out before we compress it */
1848 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1849 revcharmap, next_alloc,
1853 trie->trans = (reg_trie_trans *)
1854 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1861 for( state=1 ; state < next_alloc ; state ++ ) {
1865 DEBUG_TRIE_COMPILE_MORE_r(
1866 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1870 if (trie->states[state].trans.list) {
1871 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1875 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1876 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1877 if ( forid < minid ) {
1879 } else if ( forid > maxid ) {
1883 if ( transcount < tp + maxid - minid + 1) {
1885 trie->trans = (reg_trie_trans *)
1886 PerlMemShared_realloc( trie->trans,
1888 * sizeof(reg_trie_trans) );
1889 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1891 base = trie->uniquecharcount + tp - minid;
1892 if ( maxid == minid ) {
1894 for ( ; zp < tp ; zp++ ) {
1895 if ( ! trie->trans[ zp ].next ) {
1896 base = trie->uniquecharcount + zp - minid;
1897 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1898 trie->trans[ zp ].check = state;
1904 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1905 trie->trans[ tp ].check = state;
1910 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1911 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1912 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1913 trie->trans[ tid ].check = state;
1915 tp += ( maxid - minid + 1 );
1917 Safefree(trie->states[ state ].trans.list);
1920 DEBUG_TRIE_COMPILE_MORE_r(
1921 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1924 trie->states[ state ].trans.base=base;
1926 trie->lasttrans = tp + 1;
1930 Second Pass -- Flat Table Representation.
1932 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1933 We know that we will need Charcount+1 trans at most to store the data
1934 (one row per char at worst case) So we preallocate both structures
1935 assuming worst case.
1937 We then construct the trie using only the .next slots of the entry
1940 We use the .check field of the first entry of the node temporarily to
1941 make compression both faster and easier by keeping track of how many non
1942 zero fields are in the node.
1944 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1947 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1948 number representing the first entry of the node, and state as a
1949 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1950 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1951 are 2 entrys per node. eg:
1959 The table is internally in the right hand, idx form. However as we also
1960 have to deal with the states array which is indexed by nodenum we have to
1961 use TRIE_NODENUM() to convert.
1964 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1965 "%*sCompiling trie using table compiler\n",
1966 (int)depth * 2 + 2, ""));
1968 trie->trans = (reg_trie_trans *)
1969 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1970 * trie->uniquecharcount + 1,
1971 sizeof(reg_trie_trans) );
1972 trie->states = (reg_trie_state *)
1973 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1974 sizeof(reg_trie_state) );
1975 next_alloc = trie->uniquecharcount + 1;
1978 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1980 regnode *noper = NEXTOPER( cur );
1981 const U8 *uc = (U8*)STRING( noper );
1982 const U8 *e = uc + STR_LEN( noper );
1984 U32 state = 1; /* required init */
1986 U16 charid = 0; /* sanity init */
1987 U32 accept_state = 0; /* sanity init */
1988 U8 *scan = (U8*)NULL; /* sanity init */
1990 STRLEN foldlen = 0; /* required init */
1991 U32 wordlen = 0; /* required init */
1993 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1995 if (OP(noper) == NOTHING) {
1996 regnode *noper_next= regnext(noper);
1997 if (noper_next != tail && OP(noper_next) == flags) {
1999 uc= (U8*)STRING(noper);
2000 e= uc + STR_LEN(noper);
2004 if ( OP(noper) != NOTHING ) {
2005 for ( ; uc < e ; uc += len ) {
2010 charid = trie->charmap[ uvc ];
2012 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2013 charid = svpp ? (U16)SvIV(*svpp) : 0;
2017 if ( !trie->trans[ state + charid ].next ) {
2018 trie->trans[ state + charid ].next = next_alloc;
2019 trie->trans[ state ].check++;
2020 prev_states[TRIE_NODENUM(next_alloc)]
2021 = TRIE_NODENUM(state);
2022 next_alloc += trie->uniquecharcount;
2024 state = trie->trans[ state + charid ].next;
2026 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2028 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2031 accept_state = TRIE_NODENUM( state );
2032 TRIE_HANDLE_WORD(accept_state);
2034 } /* end second pass */
2036 /* and now dump it out before we compress it */
2037 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2039 next_alloc, depth+1));
2043 * Inplace compress the table.*
2045 For sparse data sets the table constructed by the trie algorithm will
2046 be mostly 0/FAIL transitions or to put it another way mostly empty.
2047 (Note that leaf nodes will not contain any transitions.)
2049 This algorithm compresses the tables by eliminating most such
2050 transitions, at the cost of a modest bit of extra work during lookup:
2052 - Each states[] entry contains a .base field which indicates the
2053 index in the state[] array wheres its transition data is stored.
2055 - If .base is 0 there are no valid transitions from that node.
2057 - If .base is nonzero then charid is added to it to find an entry in
2060 -If trans[states[state].base+charid].check!=state then the
2061 transition is taken to be a 0/Fail transition. Thus if there are fail
2062 transitions at the front of the node then the .base offset will point
2063 somewhere inside the previous nodes data (or maybe even into a node
2064 even earlier), but the .check field determines if the transition is
2068 The following process inplace converts the table to the compressed
2069 table: We first do not compress the root node 1,and mark all its
2070 .check pointers as 1 and set its .base pointer as 1 as well. This
2071 allows us to do a DFA construction from the compressed table later,
2072 and ensures that any .base pointers we calculate later are greater
2075 - We set 'pos' to indicate the first entry of the second node.
2077 - We then iterate over the columns of the node, finding the first and
2078 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2079 and set the .check pointers accordingly, and advance pos
2080 appropriately and repreat for the next node. Note that when we copy
2081 the next pointers we have to convert them from the original
2082 NODEIDX form to NODENUM form as the former is not valid post
2085 - If a node has no transitions used we mark its base as 0 and do not
2086 advance the pos pointer.
2088 - If a node only has one transition we use a second pointer into the
2089 structure to fill in allocated fail transitions from other states.
2090 This pointer is independent of the main pointer and scans forward
2091 looking for null transitions that are allocated to a state. When it
2092 finds one it writes the single transition into the "hole". If the
2093 pointer doesnt find one the single transition is appended as normal.
2095 - Once compressed we can Renew/realloc the structures to release the
2098 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2099 specifically Fig 3.47 and the associated pseudocode.
2103 const U32 laststate = TRIE_NODENUM( next_alloc );
2106 trie->statecount = laststate;
2108 for ( state = 1 ; state < laststate ; state++ ) {
2110 const U32 stateidx = TRIE_NODEIDX( state );
2111 const U32 o_used = trie->trans[ stateidx ].check;
2112 U32 used = trie->trans[ stateidx ].check;
2113 trie->trans[ stateidx ].check = 0;
2115 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2116 if ( flag || trie->trans[ stateidx + charid ].next ) {
2117 if ( trie->trans[ stateidx + charid ].next ) {
2119 for ( ; zp < pos ; zp++ ) {
2120 if ( ! trie->trans[ zp ].next ) {
2124 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2125 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2126 trie->trans[ zp ].check = state;
2127 if ( ++zp > pos ) pos = zp;
2134 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2136 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2137 trie->trans[ pos ].check = state;
2142 trie->lasttrans = pos + 1;
2143 trie->states = (reg_trie_state *)
2144 PerlMemShared_realloc( trie->states, laststate
2145 * sizeof(reg_trie_state) );
2146 DEBUG_TRIE_COMPILE_MORE_r(
2147 PerlIO_printf( Perl_debug_log,
2148 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2149 (int)depth * 2 + 2,"",
2150 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2153 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2156 } /* end table compress */
2158 DEBUG_TRIE_COMPILE_MORE_r(
2159 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2160 (int)depth * 2 + 2, "",
2161 (UV)trie->statecount,
2162 (UV)trie->lasttrans)
2164 /* resize the trans array to remove unused space */
2165 trie->trans = (reg_trie_trans *)
2166 PerlMemShared_realloc( trie->trans, trie->lasttrans
2167 * sizeof(reg_trie_trans) );
2169 { /* Modify the program and insert the new TRIE node */
2170 U8 nodetype =(U8)(flags & 0xFF);
2174 regnode *optimize = NULL;
2175 #ifdef RE_TRACK_PATTERN_OFFSETS
2178 U32 mjd_nodelen = 0;
2179 #endif /* RE_TRACK_PATTERN_OFFSETS */
2180 #endif /* DEBUGGING */
2182 This means we convert either the first branch or the first Exact,
2183 depending on whether the thing following (in 'last') is a branch
2184 or not and whther first is the startbranch (ie is it a sub part of
2185 the alternation or is it the whole thing.)
2186 Assuming its a sub part we convert the EXACT otherwise we convert
2187 the whole branch sequence, including the first.
2189 /* Find the node we are going to overwrite */
2190 if ( first != startbranch || OP( last ) == BRANCH ) {
2191 /* branch sub-chain */
2192 NEXT_OFF( first ) = (U16)(last - first);
2193 #ifdef RE_TRACK_PATTERN_OFFSETS
2195 mjd_offset= Node_Offset((convert));
2196 mjd_nodelen= Node_Length((convert));
2199 /* whole branch chain */
2201 #ifdef RE_TRACK_PATTERN_OFFSETS
2204 const regnode *nop = NEXTOPER( convert );
2205 mjd_offset= Node_Offset((nop));
2206 mjd_nodelen= Node_Length((nop));
2210 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2211 (int)depth * 2 + 2, "",
2212 (UV)mjd_offset, (UV)mjd_nodelen)
2215 /* But first we check to see if there is a common prefix we can
2216 split out as an EXACT and put in front of the TRIE node. */
2217 trie->startstate= 1;
2218 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2220 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2224 const U32 base = trie->states[ state ].trans.base;
2226 if ( trie->states[state].wordnum )
2229 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2230 if ( ( base + ofs >= trie->uniquecharcount ) &&
2231 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2232 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2234 if ( ++count > 1 ) {
2235 SV **tmp = av_fetch( revcharmap, ofs, 0);
2236 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2237 if ( state == 1 ) break;
2239 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2241 PerlIO_printf(Perl_debug_log,
2242 "%*sNew Start State=%"UVuf" Class: [",
2243 (int)depth * 2 + 2, "",
2246 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2247 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2249 TRIE_BITMAP_SET(trie,*ch);
2251 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2253 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2257 TRIE_BITMAP_SET(trie,*ch);
2259 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2260 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2266 SV **tmp = av_fetch( revcharmap, idx, 0);
2268 char *ch = SvPV( *tmp, len );
2270 SV *sv=sv_newmortal();
2271 PerlIO_printf( Perl_debug_log,
2272 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2273 (int)depth * 2 + 2, "",
2275 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2276 PL_colors[0], PL_colors[1],
2277 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2278 PERL_PV_ESCAPE_FIRSTCHAR
2283 OP( convert ) = nodetype;
2284 str=STRING(convert);
2287 STR_LEN(convert) += len;
2293 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2298 trie->prefixlen = (state-1);
2300 regnode *n = convert+NODE_SZ_STR(convert);
2301 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2302 trie->startstate = state;
2303 trie->minlen -= (state - 1);
2304 trie->maxlen -= (state - 1);
2306 /* At least the UNICOS C compiler choked on this
2307 * being argument to DEBUG_r(), so let's just have
2310 #ifdef PERL_EXT_RE_BUILD
2316 regnode *fix = convert;
2317 U32 word = trie->wordcount;
2319 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2320 while( ++fix < n ) {
2321 Set_Node_Offset_Length(fix, 0, 0);
2324 SV ** const tmp = av_fetch( trie_words, word, 0 );
2326 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2327 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2329 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2337 NEXT_OFF(convert) = (U16)(tail - convert);
2338 DEBUG_r(optimize= n);
2344 if ( trie->maxlen ) {
2345 NEXT_OFF( convert ) = (U16)(tail - convert);
2346 ARG_SET( convert, data_slot );
2347 /* Store the offset to the first unabsorbed branch in
2348 jump[0], which is otherwise unused by the jump logic.
2349 We use this when dumping a trie and during optimisation. */
2351 trie->jump[0] = (U16)(nextbranch - convert);
2353 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2354 * and there is a bitmap
2355 * and the first "jump target" node we found leaves enough room
2356 * then convert the TRIE node into a TRIEC node, with the bitmap
2357 * embedded inline in the opcode - this is hypothetically faster.
2359 if ( !trie->states[trie->startstate].wordnum
2361 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2363 OP( convert ) = TRIEC;
2364 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2365 PerlMemShared_free(trie->bitmap);
2368 OP( convert ) = TRIE;
2370 /* store the type in the flags */
2371 convert->flags = nodetype;
2375 + regarglen[ OP( convert ) ];
2377 /* XXX We really should free up the resource in trie now,
2378 as we won't use them - (which resources?) dmq */
2380 /* needed for dumping*/
2381 DEBUG_r(if (optimize) {
2382 regnode *opt = convert;
2384 while ( ++opt < optimize) {
2385 Set_Node_Offset_Length(opt,0,0);
2388 Try to clean up some of the debris left after the
2391 while( optimize < jumper ) {
2392 mjd_nodelen += Node_Length((optimize));
2393 OP( optimize ) = OPTIMIZED;
2394 Set_Node_Offset_Length(optimize,0,0);
2397 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2399 } /* end node insert */
2401 /* Finish populating the prev field of the wordinfo array. Walk back
2402 * from each accept state until we find another accept state, and if
2403 * so, point the first word's .prev field at the second word. If the
2404 * second already has a .prev field set, stop now. This will be the
2405 * case either if we've already processed that word's accept state,
2406 * or that state had multiple words, and the overspill words were
2407 * already linked up earlier.
2414 for (word=1; word <= trie->wordcount; word++) {
2416 if (trie->wordinfo[word].prev)
2418 state = trie->wordinfo[word].accept;
2420 state = prev_states[state];
2423 prev = trie->states[state].wordnum;
2427 trie->wordinfo[word].prev = prev;
2429 Safefree(prev_states);
2433 /* and now dump out the compressed format */
2434 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2436 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2438 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2439 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2441 SvREFCNT_dec(revcharmap);
2445 : trie->startstate>1
2451 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2453 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2455 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2456 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2459 We find the fail state for each state in the trie, this state is the longest proper
2460 suffix of the current state's 'word' that is also a proper prefix of another word in our
2461 trie. State 1 represents the word '' and is thus the default fail state. This allows
2462 the DFA not to have to restart after its tried and failed a word at a given point, it
2463 simply continues as though it had been matching the other word in the first place.
2465 'abcdgu'=~/abcdefg|cdgu/
2466 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2467 fail, which would bring us to the state representing 'd' in the second word where we would
2468 try 'g' and succeed, proceeding to match 'cdgu'.
2470 /* add a fail transition */
2471 const U32 trie_offset = ARG(source);
2472 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2474 const U32 ucharcount = trie->uniquecharcount;
2475 const U32 numstates = trie->statecount;
2476 const U32 ubound = trie->lasttrans + ucharcount;
2480 U32 base = trie->states[ 1 ].trans.base;
2483 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2484 GET_RE_DEBUG_FLAGS_DECL;
2486 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2488 PERL_UNUSED_ARG(depth);
2492 ARG_SET( stclass, data_slot );
2493 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2494 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2495 aho->trie=trie_offset;
2496 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2497 Copy( trie->states, aho->states, numstates, reg_trie_state );
2498 Newxz( q, numstates, U32);
2499 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2502 /* initialize fail[0..1] to be 1 so that we always have
2503 a valid final fail state */
2504 fail[ 0 ] = fail[ 1 ] = 1;
2506 for ( charid = 0; charid < ucharcount ; charid++ ) {
2507 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2509 q[ q_write ] = newstate;
2510 /* set to point at the root */
2511 fail[ q[ q_write++ ] ]=1;
2514 while ( q_read < q_write) {
2515 const U32 cur = q[ q_read++ % numstates ];
2516 base = trie->states[ cur ].trans.base;
2518 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2519 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2521 U32 fail_state = cur;
2524 fail_state = fail[ fail_state ];
2525 fail_base = aho->states[ fail_state ].trans.base;
2526 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2528 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2529 fail[ ch_state ] = fail_state;
2530 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2532 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2534 q[ q_write++ % numstates] = ch_state;
2538 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2539 when we fail in state 1, this allows us to use the
2540 charclass scan to find a valid start char. This is based on the principle
2541 that theres a good chance the string being searched contains lots of stuff
2542 that cant be a start char.
2544 fail[ 0 ] = fail[ 1 ] = 0;
2545 DEBUG_TRIE_COMPILE_r({
2546 PerlIO_printf(Perl_debug_log,
2547 "%*sStclass Failtable (%"UVuf" states): 0",
2548 (int)(depth * 2), "", (UV)numstates
2550 for( q_read=1; q_read<numstates; q_read++ ) {
2551 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2553 PerlIO_printf(Perl_debug_log, "\n");
2556 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2561 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2562 * These need to be revisited when a newer toolchain becomes available.
2564 #if defined(__sparc64__) && defined(__GNUC__)
2565 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2566 # undef SPARC64_GCC_WORKAROUND
2567 # define SPARC64_GCC_WORKAROUND 1
2571 #define DEBUG_PEEP(str,scan,depth) \
2572 DEBUG_OPTIMISE_r({if (scan){ \
2573 SV * const mysv=sv_newmortal(); \
2574 regnode *Next = regnext(scan); \
2575 regprop(RExC_rx, mysv, scan); \
2576 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2577 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2578 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2582 /* The below joins as many adjacent EXACTish nodes as possible into a single
2583 * one, and looks for problematic sequences of characters whose folds vs.
2584 * non-folds have sufficiently different lengths, that the optimizer would be
2585 * fooled into rejecting legitimate matches of them, and the trie construction
2586 * code needs to handle specially. The joining is only done if:
2587 * 1) there is room in the current conglomerated node to entirely contain the
2589 * 2) they are the exact same node type
2591 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2592 * these get optimized out
2594 * If there are problematic code sequences, *min_subtract is set to the delta
2595 * that the minimum size of the node can be less than its actual size. And,
2596 * the node type of the result is changed to reflect that it contains these
2599 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2600 * and contains LATIN SMALL LETTER SHARP S
2602 * This is as good a place as any to discuss the design of handling these
2603 * problematic sequences. It's been wrong in Perl for a very long time. There
2604 * are three code points currently in Unicode whose folded lengths differ so
2605 * much from the un-folded lengths that it causes problems for the optimizer
2606 * and trie construction. Why only these are problematic, and not others where
2607 * lengths also differ is something I (khw) do not understand. New versions of
2608 * Unicode might add more such code points. Hopefully the logic in
2609 * fold_grind.t that figures out what to test (in part by verifying that each
2610 * size-combination gets tested) will catch any that do come along, so they can
2611 * be added to the special handling below. The chances of new ones are
2612 * actually rather small, as most, if not all, of the world's scripts that have
2613 * casefolding have already been encoded by Unicode. Also, a number of
2614 * Unicode's decisions were made to allow compatibility with pre-existing
2615 * standards, and almost all of those have already been dealt with. These
2616 * would otherwise be the most likely candidates for generating further tricky
2617 * sequences. In other words, Unicode by itself is unlikely to add new ones
2618 * unless it is for compatibility with pre-existing standards, and there aren't
2619 * many of those left.
2621 * The previous designs for dealing with these involved assigning a special
2622 * node for them. This approach doesn't work, as evidenced by this example:
2623 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2624 * Both these fold to "sss", but if the pattern is parsed to create a node
2625 * that would match just the \xDF, it won't be able to handle the case where a
2626 * successful match would have to cross the node's boundary. The new approach
2627 * that hopefully generally solves the problem generates an EXACTFU_SS node
2630 * There are a number of components to the approach (a lot of work for just
2631 * three code points!):
2632 * 1) This routine examines each EXACTFish node that could contain the
2633 * problematic sequences. It returns in *min_subtract how much to
2634 * subtract from the the actual length of the string to get a real minimum
2635 * for one that could match it. This number is usually 0 except for the
2636 * problematic sequences. This delta is used by the caller to adjust the
2637 * min length of the match, and the delta between min and max, so that the
2638 * optimizer doesn't reject these possibilities based on size constraints.
2639 * 2) These sequences require special handling by the trie code, so this code
2640 * changes the joined node type to special ops: EXACTFU_TRICKYFOLD and
2642 * 3) This is sufficient for the two Greek sequences (described below), but
2643 * the one involving the Sharp s (\xDF) needs more. The node type
2644 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2645 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2646 * case where there is a possible fold length change. That means that a
2647 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2648 * itself with length changes, and so can be processed faster. regexec.c
2649 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2650 * is pre-folded by regcomp.c. This saves effort in regex matching.
2651 * However, the pre-folding isn't done for non-UTF8 patterns because the
2652 * fold of the MICRO SIGN requires UTF-8. Also what EXACTF and EXACTFL
2653 * nodes fold to isn't known until runtime. The fold possibilities for
2654 * the non-UTF8 patterns are quite simple, except for the sharp s. All
2655 * the ones that don't involve a UTF-8 target string are members of a
2656 * fold-pair, and arrays are set up for all of them so that the other
2657 * member of the pair can be found quickly. Code elsewhere in this file
2658 * makes sure that in EXACTFU nodes, the sharp s gets folded to 'ss', even
2659 * if the pattern isn't UTF-8. This avoids the issues described in the
2661 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2662 * 'ss' or not is not knowable at compile time. It will match iff the
2663 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2664 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2665 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2666 * described in item 3). An assumption that the optimizer part of
2667 * regexec.c (probably unwittingly) makes is that a character in the
2668 * pattern corresponds to at most a single character in the target string.
2669 * (And I do mean character, and not byte here, unlike other parts of the
2670 * documentation that have never been updated to account for multibyte
2671 * Unicode.) This assumption is wrong only in this case, as all other
2672 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2673 * virtue of having this file pre-fold UTF-8 patterns. I'm
2674 * reluctant to try to change this assumption, so instead the code punts.
2675 * This routine examines EXACTF nodes for the sharp s, and returns a
2676 * boolean indicating whether or not the node is an EXACTF node that
2677 * contains a sharp s. When it is true, the caller sets a flag that later
2678 * causes the optimizer in this file to not set values for the floating
2679 * and fixed string lengths, and thus avoids the optimizer code in
2680 * regexec.c that makes the invalid assumption. Thus, there is no
2681 * optimization based on string lengths for EXACTF nodes that contain the
2682 * sharp s. This only happens for /id rules (which means the pattern
2686 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2687 if (PL_regkind[OP(scan)] == EXACT) \
2688 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2691 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) {
2692 /* Merge several consecutive EXACTish nodes into one. */
2693 regnode *n = regnext(scan);
2695 regnode *next = scan + NODE_SZ_STR(scan);
2699 regnode *stop = scan;
2700 GET_RE_DEBUG_FLAGS_DECL;
2702 PERL_UNUSED_ARG(depth);
2705 PERL_ARGS_ASSERT_JOIN_EXACT;
2706 #ifndef EXPERIMENTAL_INPLACESCAN
2707 PERL_UNUSED_ARG(flags);
2708 PERL_UNUSED_ARG(val);
2710 DEBUG_PEEP("join",scan,depth);
2712 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2713 * EXACT ones that are mergeable to the current one. */
2715 && (PL_regkind[OP(n)] == NOTHING
2716 || (stringok && OP(n) == OP(scan)))
2718 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2721 if (OP(n) == TAIL || n > next)
2723 if (PL_regkind[OP(n)] == NOTHING) {
2724 DEBUG_PEEP("skip:",n,depth);
2725 NEXT_OFF(scan) += NEXT_OFF(n);
2726 next = n + NODE_STEP_REGNODE;
2733 else if (stringok) {
2734 const unsigned int oldl = STR_LEN(scan);
2735 regnode * const nnext = regnext(n);
2737 /* XXX I (khw) kind of doubt that this works on platforms where
2738 * U8_MAX is above 255 because of lots of other assumptions */
2739 if (oldl + STR_LEN(n) > U8_MAX)
2742 DEBUG_PEEP("merg",n,depth);
2745 NEXT_OFF(scan) += NEXT_OFF(n);
2746 STR_LEN(scan) += STR_LEN(n);
2747 next = n + NODE_SZ_STR(n);
2748 /* Now we can overwrite *n : */
2749 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2757 #ifdef EXPERIMENTAL_INPLACESCAN
2758 if (flags && !NEXT_OFF(n)) {
2759 DEBUG_PEEP("atch", val, depth);
2760 if (reg_off_by_arg[OP(n)]) {
2761 ARG_SET(n, val - n);
2764 NEXT_OFF(n) = val - n;
2772 *has_exactf_sharp_s = FALSE;
2774 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2775 * can now analyze for sequences of problematic code points. (Prior to
2776 * this final joining, sequences could have been split over boundaries, and
2777 * hence missed). The sequences only happen in folding, hence for any
2778 * non-EXACT EXACTish node */
2779 if (OP(scan) != EXACT) {
2781 U8 * s0 = (U8*) STRING(scan);
2782 U8 * const s_end = s0 + STR_LEN(scan);
2784 /* The below is perhaps overboard, but this allows us to save a test
2785 * each time through the loop at the expense of a mask. This is
2786 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2787 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2788 * This uses an exclusive 'or' to find that bit and then inverts it to
2789 * form a mask, with just a single 0, in the bit position where 'S' and
2791 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2792 const U8 s_masked = 's' & S_or_s_mask;
2794 /* One pass is made over the node's string looking for all the
2795 * possibilities. to avoid some tests in the loop, there are two main
2796 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2800 /* There are two problematic Greek code points in Unicode
2803 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2804 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2810 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2811 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2813 * This means that in case-insensitive matching (or "loose
2814 * matching", as Unicode calls it), an EXACTF of length six (the
2815 * UTF-8 encoded byte length of the above casefolded versions) can
2816 * match a target string of length two (the byte length of UTF-8
2817 * encoded U+0390 or U+03B0). This would rather mess up the
2818 * minimum length computation. (there are other code points that
2819 * also fold to these two sequences, but the delta is smaller)
2821 * If these sequences are found, the minimum length is decreased by
2822 * four (six minus two).
2824 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2825 * LETTER SHARP S. We decrease the min length by 1 for each
2826 * occurrence of 'ss' found */
2828 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2829 # define U390_first_byte 0xb4
2830 const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
2831 # define U3B0_first_byte 0xb5
2832 const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
2834 # define U390_first_byte 0xce
2835 const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
2836 # define U3B0_first_byte 0xcf
2837 const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
2839 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2840 yields a net of 0 */
2841 /* Examine the string for one of the problematic sequences */
2843 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2844 * sequence we are looking for is 2 */
2848 /* Look for the first byte in each problematic sequence */
2850 /* We don't have to worry about other things that fold to
2851 * 's' (such as the long s, U+017F), as all above-latin1
2852 * code points have been pre-folded */
2856 /* Current character is an 's' or 'S'. If next one is
2857 * as well, we have the dreaded sequence */
2858 if (((*(s+1) & S_or_s_mask) == s_masked)
2859 /* These two node types don't have special handling
2861 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2864 OP(scan) = EXACTFU_SS;
2865 s++; /* No need to look at this character again */
2869 case U390_first_byte:
2870 if (s_end - s >= len
2872 /* The 1's are because are skipping comparing the
2874 && memEQ(s + 1, U390_tail, len - 1))
2876 goto greek_sequence;
2880 case U3B0_first_byte:
2881 if (! (s_end - s >= len
2882 && memEQ(s + 1, U3B0_tail, len - 1)))
2889 /* This requires special handling by trie's, so change
2890 * the node type to indicate this. If EXACTFA and
2891 * EXACTFL were ever to be handled by trie's, this
2892 * would have to be changed. If this node has already
2893 * been changed to EXACTFU_SS in this loop, leave it as
2894 * is. (I (khw) think it doesn't matter in regexec.c
2895 * for UTF patterns, but no need to change it */
2896 if (OP(scan) == EXACTFU) {
2897 OP(scan) = EXACTFU_TRICKYFOLD;
2899 s += 6; /* We already know what this sequence is. Skip
2905 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2907 /* Here, the pattern is not UTF-8. We need to look only for the
2908 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2909 * in the final position. Otherwise we can stop looking 1 byte
2910 * earlier because have to find both the first and second 's' */
2911 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2913 for (s = s0; s < upper; s++) {
2918 && ((*(s+1) & S_or_s_mask) == s_masked))
2922 /* EXACTF nodes need to know that the minimum
2923 * length changed so that a sharp s in the string
2924 * can match this ss in the pattern, but they
2925 * remain EXACTF nodes, as they won't match this
2926 * unless the target string is is UTF-8, which we
2927 * don't know until runtime */
2928 if (OP(scan) != EXACTF) {
2929 OP(scan) = EXACTFU_SS;
2934 case LATIN_SMALL_LETTER_SHARP_S:
2935 if (OP(scan) == EXACTF) {
2936 *has_exactf_sharp_s = TRUE;
2945 /* Allow dumping but overwriting the collection of skipped
2946 * ops and/or strings with fake optimized ops */
2947 n = scan + NODE_SZ_STR(scan);
2955 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2959 /* REx optimizer. Converts nodes into quicker variants "in place".
2960 Finds fixed substrings. */
2962 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2963 to the position after last scanned or to NULL. */
2965 #define INIT_AND_WITHP \
2966 assert(!and_withp); \
2967 Newx(and_withp,1,struct regnode_charclass_class); \
2968 SAVEFREEPV(and_withp)
2970 /* this is a chain of data about sub patterns we are processing that
2971 need to be handled separately/specially in study_chunk. Its so
2972 we can simulate recursion without losing state. */
2974 typedef struct scan_frame {
2975 regnode *last; /* last node to process in this frame */
2976 regnode *next; /* next node to process when last is reached */
2977 struct scan_frame *prev; /*previous frame*/
2978 I32 stop; /* what stopparen do we use */
2982 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2984 #define CASE_SYNST_FNC(nAmE) \
2986 if (flags & SCF_DO_STCLASS_AND) { \
2987 for (value = 0; value < 256; value++) \
2988 if (!is_ ## nAmE ## _cp(value)) \
2989 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2992 for (value = 0; value < 256; value++) \
2993 if (is_ ## nAmE ## _cp(value)) \
2994 ANYOF_BITMAP_SET(data->start_class, value); \
2998 if (flags & SCF_DO_STCLASS_AND) { \
2999 for (value = 0; value < 256; value++) \
3000 if (is_ ## nAmE ## _cp(value)) \
3001 ANYOF_BITMAP_CLEAR(data->start_class, value); \
3004 for (value = 0; value < 256; value++) \
3005 if (!is_ ## nAmE ## _cp(value)) \
3006 ANYOF_BITMAP_SET(data->start_class, value); \
3013 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3014 I32 *minlenp, I32 *deltap,
3019 struct regnode_charclass_class *and_withp,
3020 U32 flags, U32 depth)
3021 /* scanp: Start here (read-write). */
3022 /* deltap: Write maxlen-minlen here. */
3023 /* last: Stop before this one. */
3024 /* data: string data about the pattern */
3025 /* stopparen: treat close N as END */
3026 /* recursed: which subroutines have we recursed into */
3027 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3030 I32 min = 0, pars = 0, code;
3031 regnode *scan = *scanp, *next;
3033 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3034 int is_inf_internal = 0; /* The studied chunk is infinite */
3035 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3036 scan_data_t data_fake;
3037 SV *re_trie_maxbuff = NULL;
3038 regnode *first_non_open = scan;
3039 I32 stopmin = I32_MAX;
3040 scan_frame *frame = NULL;
3041 GET_RE_DEBUG_FLAGS_DECL;
3043 PERL_ARGS_ASSERT_STUDY_CHUNK;
3046 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3050 while (first_non_open && OP(first_non_open) == OPEN)
3051 first_non_open=regnext(first_non_open);
3056 while ( scan && OP(scan) != END && scan < last ){
3057 UV min_subtract = 0; /* How much to subtract from the minimum node
3058 length to get a real minimum (because the
3059 folded version may be shorter) */
3060 bool has_exactf_sharp_s = FALSE;
3061 /* Peephole optimizer: */
3062 DEBUG_STUDYDATA("Peep:", data,depth);
3063 DEBUG_PEEP("Peep",scan,depth);
3065 /* Its not clear to khw or hv why this is done here, and not in the
3066 * clauses that deal with EXACT nodes. khw's guess is that it's
3067 * because of a previous design */
3068 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3070 /* Follow the next-chain of the current node and optimize
3071 away all the NOTHINGs from it. */
3072 if (OP(scan) != CURLYX) {
3073 const int max = (reg_off_by_arg[OP(scan)]
3075 /* I32 may be smaller than U16 on CRAYs! */
3076 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3077 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3081 /* Skip NOTHING and LONGJMP. */
3082 while ((n = regnext(n))
3083 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3084 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3085 && off + noff < max)
3087 if (reg_off_by_arg[OP(scan)])
3090 NEXT_OFF(scan) = off;
3095 /* The principal pseudo-switch. Cannot be a switch, since we
3096 look into several different things. */
3097 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3098 || OP(scan) == IFTHEN) {
3099 next = regnext(scan);
3101 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3103 if (OP(next) == code || code == IFTHEN) {
3104 /* NOTE - There is similar code to this block below for handling
3105 TRIE nodes on a re-study. If you change stuff here check there
3107 I32 max1 = 0, min1 = I32_MAX, num = 0;
3108 struct regnode_charclass_class accum;
3109 regnode * const startbranch=scan;
3111 if (flags & SCF_DO_SUBSTR)
3112 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3113 if (flags & SCF_DO_STCLASS)
3114 cl_init_zero(pRExC_state, &accum);
3116 while (OP(scan) == code) {
3117 I32 deltanext, minnext, f = 0, fake;
3118 struct regnode_charclass_class this_class;
3121 data_fake.flags = 0;
3123 data_fake.whilem_c = data->whilem_c;
3124 data_fake.last_closep = data->last_closep;
3127 data_fake.last_closep = &fake;
3129 data_fake.pos_delta = delta;
3130 next = regnext(scan);
3131 scan = NEXTOPER(scan);
3133 scan = NEXTOPER(scan);
3134 if (flags & SCF_DO_STCLASS) {
3135 cl_init(pRExC_state, &this_class);
3136 data_fake.start_class = &this_class;
3137 f = SCF_DO_STCLASS_AND;
3139 if (flags & SCF_WHILEM_VISITED_POS)
3140 f |= SCF_WHILEM_VISITED_POS;
3142 /* we suppose the run is continuous, last=next...*/
3143 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3145 stopparen, recursed, NULL, f,depth+1);
3148 if (max1 < minnext + deltanext)
3149 max1 = minnext + deltanext;
3150 if (deltanext == I32_MAX)
3151 is_inf = is_inf_internal = 1;
3153 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3155 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3156 if ( stopmin > minnext)
3157 stopmin = min + min1;
3158 flags &= ~SCF_DO_SUBSTR;
3160 data->flags |= SCF_SEEN_ACCEPT;
3163 if (data_fake.flags & SF_HAS_EVAL)
3164 data->flags |= SF_HAS_EVAL;
3165 data->whilem_c = data_fake.whilem_c;
3167 if (flags & SCF_DO_STCLASS)
3168 cl_or(pRExC_state, &accum, &this_class);
3170 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3172 if (flags & SCF_DO_SUBSTR) {
3173 data->pos_min += min1;
3174 data->pos_delta += max1 - min1;
3175 if (max1 != min1 || is_inf)
3176 data->longest = &(data->longest_float);
3179 delta += max1 - min1;
3180 if (flags & SCF_DO_STCLASS_OR) {
3181 cl_or(pRExC_state, data->start_class, &accum);
3183 cl_and(data->start_class, and_withp);
3184 flags &= ~SCF_DO_STCLASS;
3187 else if (flags & SCF_DO_STCLASS_AND) {
3189 cl_and(data->start_class, &accum);
3190 flags &= ~SCF_DO_STCLASS;
3193 /* Switch to OR mode: cache the old value of
3194 * data->start_class */
3196 StructCopy(data->start_class, and_withp,
3197 struct regnode_charclass_class);
3198 flags &= ~SCF_DO_STCLASS_AND;
3199 StructCopy(&accum, data->start_class,
3200 struct regnode_charclass_class);
3201 flags |= SCF_DO_STCLASS_OR;
3202 data->start_class->flags |= ANYOF_EOS;
3206 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3209 Assuming this was/is a branch we are dealing with: 'scan' now
3210 points at the item that follows the branch sequence, whatever
3211 it is. We now start at the beginning of the sequence and look
3218 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3220 If we can find such a subsequence we need to turn the first
3221 element into a trie and then add the subsequent branch exact
3222 strings to the trie.
3226 1. patterns where the whole set of branches can be converted.
3228 2. patterns where only a subset can be converted.
3230 In case 1 we can replace the whole set with a single regop
3231 for the trie. In case 2 we need to keep the start and end
3234 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3235 becomes BRANCH TRIE; BRANCH X;
3237 There is an additional case, that being where there is a
3238 common prefix, which gets split out into an EXACT like node
3239 preceding the TRIE node.
3241 If x(1..n)==tail then we can do a simple trie, if not we make
3242 a "jump" trie, such that when we match the appropriate word
3243 we "jump" to the appropriate tail node. Essentially we turn
3244 a nested if into a case structure of sorts.
3249 if (!re_trie_maxbuff) {
3250 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3251 if (!SvIOK(re_trie_maxbuff))
3252 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3254 if ( SvIV(re_trie_maxbuff)>=0 ) {
3256 regnode *first = (regnode *)NULL;
3257 regnode *last = (regnode *)NULL;
3258 regnode *tail = scan;
3263 SV * const mysv = sv_newmortal(); /* for dumping */
3265 /* var tail is used because there may be a TAIL
3266 regop in the way. Ie, the exacts will point to the
3267 thing following the TAIL, but the last branch will
3268 point at the TAIL. So we advance tail. If we
3269 have nested (?:) we may have to move through several
3273 while ( OP( tail ) == TAIL ) {
3274 /* this is the TAIL generated by (?:) */
3275 tail = regnext( tail );
3279 DEBUG_TRIE_COMPILE_r({
3280 regprop(RExC_rx, mysv, tail );
3281 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3282 (int)depth * 2 + 2, "",
3283 "Looking for TRIE'able sequences. Tail node is: ",
3284 SvPV_nolen_const( mysv )
3290 Step through the branches
3291 cur represents each branch,
3292 noper is the first thing to be matched as part of that branch
3293 noper_next is the regnext() of that node.
3295 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3296 via a "jump trie" but we also support building with NOJUMPTRIE,
3297 which restricts the trie logic to structures like /FOO|BAR/.
3299 If noper is a trieable nodetype then the branch is a possible optimization
3300 target. If we are building under NOJUMPTRIE then we require that noper_next
3301 is the same as scan (our current position in the regex program).
3303 Once we have two or more consecutive such branches we can create a
3304 trie of the EXACT's contents and stitch it in place into the program.
3306 If the sequence represents all of the branches in the alternation we
3307 replace the entire thing with a single TRIE node.
3309 Otherwise when it is a subsequence we need to stitch it in place and
3310 replace only the relevant branches. This means the first branch has
3311 to remain as it is used by the alternation logic, and its next pointer,
3312 and needs to be repointed at the item on the branch chain following
3313 the last branch we have optimized away.
3315 This could be either a BRANCH, in which case the subsequence is internal,
3316 or it could be the item following the branch sequence in which case the
3317 subsequence is at the end (which does not necessarily mean the first node
3318 is the start of the alternation).
3320 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3323 ----------------+-----------
3327 EXACTFU_SS | EXACTFU
3328 EXACTFU_TRICKYFOLD | EXACTFU
3333 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3334 ( EXACT == (X) ) ? EXACT : \
3335 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3338 /* dont use tail as the end marker for this traverse */
3339 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3340 regnode * const noper = NEXTOPER( cur );
3341 U8 noper_type = OP( noper );
3342 U8 noper_trietype = TRIE_TYPE( noper_type );
3343 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3344 regnode * const noper_next = regnext( noper );
3345 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3346 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3349 DEBUG_TRIE_COMPILE_r({
3350 regprop(RExC_rx, mysv, cur);
3351 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3352 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3354 regprop(RExC_rx, mysv, noper);
3355 PerlIO_printf( Perl_debug_log, " -> %s",
3356 SvPV_nolen_const(mysv));
3359 regprop(RExC_rx, mysv, noper_next );
3360 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3361 SvPV_nolen_const(mysv));
3363 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3364 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3365 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3369 /* Is noper a trieable nodetype that can be merged with the
3370 * current trie (if there is one)? */
3374 ( noper_trietype == NOTHING)
3375 || ( trietype == NOTHING )
3376 || ( trietype == noper_trietype )
3379 && noper_next == tail
3383 /* Handle mergable triable node
3384 * Either we are the first node in a new trieable sequence,
3385 * in which case we do some bookkeeping, otherwise we update
3386 * the end pointer. */
3389 if ( noper_trietype == NOTHING ) {
3390 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3391 regnode * const noper_next = regnext( noper );
3392 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3393 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3396 if ( noper_next_trietype ) {
3397 trietype = noper_next_trietype;
3398 } else if (noper_next_type) {
3399 /* a NOTHING regop is 1 regop wide. We need at least two
3400 * for a trie so we can't merge this in */
3404 trietype = noper_trietype;
3407 if ( trietype == NOTHING )
3408 trietype = noper_trietype;
3413 } /* end handle mergable triable node */
3415 /* handle unmergable node -
3416 * noper may either be a triable node which can not be tried
3417 * together with the current trie, or a non triable node */
3419 /* If last is set and trietype is not NOTHING then we have found
3420 * at least two triable branch sequences in a row of a similar
3421 * trietype so we can turn them into a trie. If/when we
3422 * allow NOTHING to start a trie sequence this condition will be
3423 * required, and it isn't expensive so we leave it in for now. */
3424 if ( trietype != NOTHING )
3425 make_trie( pRExC_state,
3426 startbranch, first, cur, tail, count,
3427 trietype, depth+1 );
3428 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3432 && noper_next == tail
3435 /* noper is triable, so we can start a new trie sequence */
3438 trietype = noper_trietype;
3440 /* if we already saw a first but the current node is not triable then we have
3441 * to reset the first information. */
3446 } /* end handle unmergable node */
3447 } /* loop over branches */
3448 DEBUG_TRIE_COMPILE_r({
3449 regprop(RExC_rx, mysv, cur);
3450 PerlIO_printf( Perl_debug_log,
3451 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3452 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3456 if ( trietype != NOTHING ) {
3457 /* the last branch of the sequence was part of a trie,
3458 * so we have to construct it here outside of the loop
3460 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3461 #ifdef TRIE_STUDY_OPT
3462 if ( ((made == MADE_EXACT_TRIE &&
3463 startbranch == first)
3464 || ( first_non_open == first )) &&
3466 flags |= SCF_TRIE_RESTUDY;
3467 if ( startbranch == first
3470 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3475 /* at this point we know whatever we have is a NOTHING sequence/branch
3476 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3478 if ( startbranch == first ) {
3480 /* the entire thing is a NOTHING sequence, something like this:
3481 * (?:|) So we can turn it into a plain NOTHING op. */
3482 DEBUG_TRIE_COMPILE_r({
3483 regprop(RExC_rx, mysv, cur);
3484 PerlIO_printf( Perl_debug_log,
3485 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3486 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3489 OP(startbranch)= NOTHING;
3490 NEXT_OFF(startbranch)= tail - startbranch;
3491 for ( opt= startbranch + 1; opt < tail ; opt++ )
3495 } /* end if ( last) */
3496 } /* TRIE_MAXBUF is non zero */
3501 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3502 scan = NEXTOPER(NEXTOPER(scan));
3503 } else /* single branch is optimized. */
3504 scan = NEXTOPER(scan);
3506 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3507 scan_frame *newframe = NULL;
3512 if (OP(scan) != SUSPEND) {
3513 /* set the pointer */
3514 if (OP(scan) == GOSUB) {
3516 RExC_recurse[ARG2L(scan)] = scan;
3517 start = RExC_open_parens[paren-1];
3518 end = RExC_close_parens[paren-1];
3521 start = RExC_rxi->program + 1;
3525 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3526 SAVEFREEPV(recursed);
3528 if (!PAREN_TEST(recursed,paren+1)) {
3529 PAREN_SET(recursed,paren+1);
3530 Newx(newframe,1,scan_frame);
3532 if (flags & SCF_DO_SUBSTR) {
3533 SCAN_COMMIT(pRExC_state,data,minlenp);
3534 data->longest = &(data->longest_float);
3536 is_inf = is_inf_internal = 1;
3537 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3538 cl_anything(pRExC_state, data->start_class);
3539 flags &= ~SCF_DO_STCLASS;
3542 Newx(newframe,1,scan_frame);
3545 end = regnext(scan);
3550 SAVEFREEPV(newframe);
3551 newframe->next = regnext(scan);
3552 newframe->last = last;
3553 newframe->stop = stopparen;
3554 newframe->prev = frame;
3564 else if (OP(scan) == EXACT) {
3565 I32 l = STR_LEN(scan);
3568 const U8 * const s = (U8*)STRING(scan);
3569 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3570 l = utf8_length(s, s + l);
3572 uc = *((U8*)STRING(scan));
3575 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3576 /* The code below prefers earlier match for fixed
3577 offset, later match for variable offset. */
3578 if (data->last_end == -1) { /* Update the start info. */
3579 data->last_start_min = data->pos_min;
3580 data->last_start_max = is_inf
3581 ? I32_MAX : data->pos_min + data->pos_delta;
3583 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3585 SvUTF8_on(data->last_found);
3587 SV * const sv = data->last_found;
3588 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3589 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3590 if (mg && mg->mg_len >= 0)
3591 mg->mg_len += utf8_length((U8*)STRING(scan),
3592 (U8*)STRING(scan)+STR_LEN(scan));
3594 data->last_end = data->pos_min + l;
3595 data->pos_min += l; /* As in the first entry. */
3596 data->flags &= ~SF_BEFORE_EOL;
3598 if (flags & SCF_DO_STCLASS_AND) {
3599 /* Check whether it is compatible with what we know already! */
3603 /* If compatible, we or it in below. It is compatible if is
3604 * in the bitmp and either 1) its bit or its fold is set, or 2)
3605 * it's for a locale. Even if there isn't unicode semantics
3606 * here, at runtime there may be because of matching against a
3607 * utf8 string, so accept a possible false positive for
3608 * latin1-range folds */
3610 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3611 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3612 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
3613 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3618 ANYOF_CLASS_ZERO(data->start_class);
3619 ANYOF_BITMAP_ZERO(data->start_class);
3621 ANYOF_BITMAP_SET(data->start_class, uc);
3622 else if (uc >= 0x100) {
3625 /* Some Unicode code points fold to the Latin1 range; as
3626 * XXX temporary code, instead of figuring out if this is
3627 * one, just assume it is and set all the start class bits
3628 * that could be some such above 255 code point's fold
3629 * which will generate fals positives. As the code
3630 * elsewhere that does compute the fold settles down, it
3631 * can be extracted out and re-used here */
3632 for (i = 0; i < 256; i++){
3633 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3634 ANYOF_BITMAP_SET(data->start_class, i);
3638 data->start_class->flags &= ~ANYOF_EOS;
3640 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3642 else if (flags & SCF_DO_STCLASS_OR) {
3643 /* false positive possible if the class is case-folded */
3645 ANYOF_BITMAP_SET(data->start_class, uc);
3647 data->start_class->flags |= ANYOF_UNICODE_ALL;
3648 data->start_class->flags &= ~ANYOF_EOS;
3649 cl_and(data->start_class, and_withp);
3651 flags &= ~SCF_DO_STCLASS;
3653 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3654 I32 l = STR_LEN(scan);
3655 UV uc = *((U8*)STRING(scan));
3657 /* Search for fixed substrings supports EXACT only. */
3658 if (flags & SCF_DO_SUBSTR) {
3660 SCAN_COMMIT(pRExC_state, data, minlenp);
3663 const U8 * const s = (U8 *)STRING(scan);
3664 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3665 l = utf8_length(s, s + l);
3667 else if (has_exactf_sharp_s) {
3668 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3670 min += l - min_subtract;
3674 delta += min_subtract;
3675 if (flags & SCF_DO_SUBSTR) {
3676 data->pos_min += l - min_subtract;
3677 if (data->pos_min < 0) {
3680 data->pos_delta += min_subtract;
3682 data->longest = &(data->longest_float);
3685 if (flags & SCF_DO_STCLASS_AND) {
3686 /* Check whether it is compatible with what we know already! */
3689 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3690 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3691 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3695 ANYOF_CLASS_ZERO(data->start_class);
3696 ANYOF_BITMAP_ZERO(data->start_class);
3698 ANYOF_BITMAP_SET(data->start_class, uc);
3699 data->start_class->flags &= ~ANYOF_EOS;
3700 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
3701 if (OP(scan) == EXACTFL) {
3702 /* XXX This set is probably no longer necessary, and
3703 * probably wrong as LOCALE now is on in the initial
3705 data->start_class->flags |= ANYOF_LOCALE;
3709 /* Also set the other member of the fold pair. In case
3710 * that unicode semantics is called for at runtime, use
3711 * the full latin1 fold. (Can't do this for locale,
3712 * because not known until runtime) */
3713 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3715 /* All other (EXACTFL handled above) folds except under
3716 * /iaa that include s, S, and sharp_s also may include
3718 if (OP(scan) != EXACTFA) {
3719 if (uc == 's' || uc == 'S') {
3720 ANYOF_BITMAP_SET(data->start_class,
3721 LATIN_SMALL_LETTER_SHARP_S);
3723 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3724 ANYOF_BITMAP_SET(data->start_class, 's');
3725 ANYOF_BITMAP_SET(data->start_class, 'S');
3730 else if (uc >= 0x100) {
3732 for (i = 0; i < 256; i++){
3733 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3734 ANYOF_BITMAP_SET(data->start_class, i);
3739 else if (flags & SCF_DO_STCLASS_OR) {
3740 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
3741 /* false positive possible if the class is case-folded.
3742 Assume that the locale settings are the same... */
3744 ANYOF_BITMAP_SET(data->start_class, uc);
3745 if (OP(scan) != EXACTFL) {
3747 /* And set the other member of the fold pair, but
3748 * can't do that in locale because not known until
3750 ANYOF_BITMAP_SET(data->start_class,
3751 PL_fold_latin1[uc]);
3753 /* All folds except under /iaa that include s, S,
3754 * and sharp_s also may include the others */
3755 if (OP(scan) != EXACTFA) {
3756 if (uc == 's' || uc == 'S') {
3757 ANYOF_BITMAP_SET(data->start_class,
3758 LATIN_SMALL_LETTER_SHARP_S);
3760 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3761 ANYOF_BITMAP_SET(data->start_class, 's');
3762 ANYOF_BITMAP_SET(data->start_class, 'S');
3767 data->start_class->flags &= ~ANYOF_EOS;
3769 cl_and(data->start_class, and_withp);
3771 flags &= ~SCF_DO_STCLASS;
3773 else if (REGNODE_VARIES(OP(scan))) {
3774 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3775 I32 f = flags, pos_before = 0;
3776 regnode * const oscan = scan;
3777 struct regnode_charclass_class this_class;
3778 struct regnode_charclass_class *oclass = NULL;
3779 I32 next_is_eval = 0;
3781 switch (PL_regkind[OP(scan)]) {
3782 case WHILEM: /* End of (?:...)* . */
3783 scan = NEXTOPER(scan);
3786 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3787 next = NEXTOPER(scan);
3788 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3790 maxcount = REG_INFTY;
3791 next = regnext(scan);
3792 scan = NEXTOPER(scan);
3796 if (flags & SCF_DO_SUBSTR)
3801 if (flags & SCF_DO_STCLASS) {
3803 maxcount = REG_INFTY;
3804 next = regnext(scan);
3805 scan = NEXTOPER(scan);
3808 is_inf = is_inf_internal = 1;
3809 scan = regnext(scan);
3810 if (flags & SCF_DO_SUBSTR) {
3811 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3812 data->longest = &(data->longest_float);
3814 goto optimize_curly_tail;
3816 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3817 && (scan->flags == stopparen))
3822 mincount = ARG1(scan);
3823 maxcount = ARG2(scan);
3825 next = regnext(scan);
3826 if (OP(scan) == CURLYX) {
3827 I32 lp = (data ? *(data->last_closep) : 0);
3828 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3830 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3831 next_is_eval = (OP(scan) == EVAL);
https://perl5.git.perl.org / perl5.git / blob