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 #include "charclass_invlists.h"
92 #define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
93 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
100 # if defined(BUGGY_MSC6)
101 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
102 # pragma optimize("a",off)
103 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
104 # pragma optimize("w",on )
105 # endif /* BUGGY_MSC6 */
109 #define STATIC static
113 typedef struct RExC_state_t {
114 U32 flags; /* RXf_* are we folding, multilining? */
115 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
116 char *precomp; /* uncompiled string. */
117 REGEXP *rx_sv; /* The SV that is the regexp. */
118 regexp *rx; /* perl core regexp structure */
119 regexp_internal *rxi; /* internal data for regexp object pprivate field */
120 char *start; /* Start of input for compile */
121 char *end; /* End of input for compile */
122 char *parse; /* Input-scan pointer. */
123 I32 whilem_seen; /* number of WHILEM in this expr */
124 regnode *emit_start; /* Start of emitted-code area */
125 regnode *emit_bound; /* First regnode outside of the allocated space */
126 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
127 I32 naughty; /* How bad is this pattern? */
128 I32 sawback; /* Did we see \1, ...? */
130 I32 size; /* Code size. */
131 I32 npar; /* Capture buffer count, (OPEN). */
132 I32 cpar; /* Capture buffer count, (CLOSE). */
133 I32 nestroot; /* root parens we are in - used by accept */
136 regnode **open_parens; /* pointers to open parens */
137 regnode **close_parens; /* pointers to close parens */
138 regnode *opend; /* END node in program */
139 I32 utf8; /* whether the pattern is utf8 or not */
140 I32 orig_utf8; /* whether the pattern was originally in utf8 */
141 /* XXX use this for future optimisation of case
142 * where pattern must be upgraded to utf8. */
143 I32 uni_semantics; /* If a d charset modifier should use unicode
144 rules, even if the pattern is not in
146 HV *paren_names; /* Paren names */
148 regnode **recurse; /* Recurse regops */
149 I32 recurse_count; /* Number of recurse regops */
152 I32 override_recoding;
153 struct reg_code_block *code_blocks; /* positions of literal (?{})
155 int num_code_blocks; /* size of code_blocks[] */
156 int code_index; /* next code_blocks[] slot */
158 char *starttry; /* -Dr: where regtry was called. */
159 #define RExC_starttry (pRExC_state->starttry)
161 SV *runtime_code_qr; /* qr with the runtime code blocks */
163 const char *lastparse;
165 AV *paren_name_list; /* idx -> name */
166 #define RExC_lastparse (pRExC_state->lastparse)
167 #define RExC_lastnum (pRExC_state->lastnum)
168 #define RExC_paren_name_list (pRExC_state->paren_name_list)
172 #define RExC_flags (pRExC_state->flags)
173 #define RExC_pm_flags (pRExC_state->pm_flags)
174 #define RExC_precomp (pRExC_state->precomp)
175 #define RExC_rx_sv (pRExC_state->rx_sv)
176 #define RExC_rx (pRExC_state->rx)
177 #define RExC_rxi (pRExC_state->rxi)
178 #define RExC_start (pRExC_state->start)
179 #define RExC_end (pRExC_state->end)
180 #define RExC_parse (pRExC_state->parse)
181 #define RExC_whilem_seen (pRExC_state->whilem_seen)
182 #ifdef RE_TRACK_PATTERN_OFFSETS
183 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
185 #define RExC_emit (pRExC_state->emit)
186 #define RExC_emit_start (pRExC_state->emit_start)
187 #define RExC_emit_bound (pRExC_state->emit_bound)
188 #define RExC_naughty (pRExC_state->naughty)
189 #define RExC_sawback (pRExC_state->sawback)
190 #define RExC_seen (pRExC_state->seen)
191 #define RExC_size (pRExC_state->size)
192 #define RExC_npar (pRExC_state->npar)
193 #define RExC_nestroot (pRExC_state->nestroot)
194 #define RExC_extralen (pRExC_state->extralen)
195 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
196 #define RExC_utf8 (pRExC_state->utf8)
197 #define RExC_uni_semantics (pRExC_state->uni_semantics)
198 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
199 #define RExC_open_parens (pRExC_state->open_parens)
200 #define RExC_close_parens (pRExC_state->close_parens)
201 #define RExC_opend (pRExC_state->opend)
202 #define RExC_paren_names (pRExC_state->paren_names)
203 #define RExC_recurse (pRExC_state->recurse)
204 #define RExC_recurse_count (pRExC_state->recurse_count)
205 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
206 #define RExC_contains_locale (pRExC_state->contains_locale)
207 #define RExC_override_recoding (pRExC_state->override_recoding)
210 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
211 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
212 ((*s) == '{' && regcurly(s)))
215 #undef SPSTART /* dratted cpp namespace... */
218 * Flags to be passed up and down.
220 #define WORST 0 /* Worst case. */
221 #define HASWIDTH 0x01 /* Known to match non-null strings. */
223 /* Simple enough to be STAR/PLUS operand; in an EXACT node must be a single
224 * character, and if utf8, must be invariant. Note that this is not the same
225 * thing as REGNODE_SIMPLE */
227 #define SPSTART 0x04 /* Starts with * or +. */
228 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
229 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
231 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
233 /* whether trie related optimizations are enabled */
234 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
235 #define TRIE_STUDY_OPT
236 #define FULL_TRIE_STUDY
242 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
243 #define PBITVAL(paren) (1 << ((paren) & 7))
244 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
245 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
246 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
248 /* If not already in utf8, do a longjmp back to the beginning */
249 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
250 #define REQUIRE_UTF8 STMT_START { \
251 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
254 /* About scan_data_t.
256 During optimisation we recurse through the regexp program performing
257 various inplace (keyhole style) optimisations. In addition study_chunk
258 and scan_commit populate this data structure with information about
259 what strings MUST appear in the pattern. We look for the longest
260 string that must appear at a fixed location, and we look for the
261 longest string that may appear at a floating location. So for instance
266 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
267 strings (because they follow a .* construct). study_chunk will identify
268 both FOO and BAR as being the longest fixed and floating strings respectively.
270 The strings can be composites, for instance
274 will result in a composite fixed substring 'foo'.
276 For each string some basic information is maintained:
278 - offset or min_offset
279 This is the position the string must appear at, or not before.
280 It also implicitly (when combined with minlenp) tells us how many
281 characters must match before the string we are searching for.
282 Likewise when combined with minlenp and the length of the string it
283 tells us how many characters must appear after the string we have
287 Only used for floating strings. This is the rightmost point that
288 the string can appear at. If set to I32 max it indicates that the
289 string can occur infinitely far to the right.
292 A pointer to the minimum length of the pattern that the string
293 was found inside. This is important as in the case of positive
294 lookahead or positive lookbehind we can have multiple patterns
299 The minimum length of the pattern overall is 3, the minimum length
300 of the lookahead part is 3, but the minimum length of the part that
301 will actually match is 1. So 'FOO's minimum length is 3, but the
302 minimum length for the F is 1. This is important as the minimum length
303 is used to determine offsets in front of and behind the string being
304 looked for. Since strings can be composites this is the length of the
305 pattern at the time it was committed with a scan_commit. Note that
306 the length is calculated by study_chunk, so that the minimum lengths
307 are not known until the full pattern has been compiled, thus the
308 pointer to the value.
312 In the case of lookbehind the string being searched for can be
313 offset past the start point of the final matching string.
314 If this value was just blithely removed from the min_offset it would
315 invalidate some of the calculations for how many chars must match
316 before or after (as they are derived from min_offset and minlen and
317 the length of the string being searched for).
318 When the final pattern is compiled and the data is moved from the
319 scan_data_t structure into the regexp structure the information
320 about lookbehind is factored in, with the information that would
321 have been lost precalculated in the end_shift field for the
324 The fields pos_min and pos_delta are used to store the minimum offset
325 and the delta to the maximum offset at the current point in the pattern.
329 typedef struct scan_data_t {
330 /*I32 len_min; unused */
331 /*I32 len_delta; unused */
335 I32 last_end; /* min value, <0 unless valid. */
338 SV **longest; /* Either &l_fixed, or &l_float. */
339 SV *longest_fixed; /* longest fixed string found in pattern */
340 I32 offset_fixed; /* offset where it starts */
341 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
342 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
343 SV *longest_float; /* longest floating string found in pattern */
344 I32 offset_float_min; /* earliest point in string it can appear */
345 I32 offset_float_max; /* latest point in string it can appear */
346 I32 *minlen_float; /* pointer to the minlen relevant to the string */
347 I32 lookbehind_float; /* is the position of the string modified by LB */
351 struct regnode_charclass_class *start_class;
355 * Forward declarations for pregcomp()'s friends.
358 static const scan_data_t zero_scan_data =
359 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
361 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
362 #define SF_BEFORE_SEOL 0x0001
363 #define SF_BEFORE_MEOL 0x0002
364 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
365 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
368 # define SF_FIX_SHIFT_EOL (0+2)
369 # define SF_FL_SHIFT_EOL (0+4)
371 # define SF_FIX_SHIFT_EOL (+2)
372 # define SF_FL_SHIFT_EOL (+4)
375 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
376 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
378 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
379 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
380 #define SF_IS_INF 0x0040
381 #define SF_HAS_PAR 0x0080
382 #define SF_IN_PAR 0x0100
383 #define SF_HAS_EVAL 0x0200
384 #define SCF_DO_SUBSTR 0x0400
385 #define SCF_DO_STCLASS_AND 0x0800
386 #define SCF_DO_STCLASS_OR 0x1000
387 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
388 #define SCF_WHILEM_VISITED_POS 0x2000
390 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
391 #define SCF_SEEN_ACCEPT 0x8000
393 #define UTF cBOOL(RExC_utf8)
395 /* The enums for all these are ordered so things work out correctly */
396 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
397 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
398 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
399 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
400 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
401 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
402 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
404 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
406 #define OOB_NAMEDCLASS -1
408 /* There is no code point that is out-of-bounds, so this is problematic. But
409 * its only current use is to initialize a variable that is always set before
411 #define OOB_UNICODE 0xDEADBEEF
413 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
414 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
417 /* length of regex to show in messages that don't mark a position within */
418 #define RegexLengthToShowInErrorMessages 127
421 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
422 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
423 * op/pragma/warn/regcomp.
425 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
426 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
428 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
431 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
432 * arg. Show regex, up to a maximum length. If it's too long, chop and add
435 #define _FAIL(code) STMT_START { \
436 const char *ellipses = ""; \
437 IV len = RExC_end - RExC_precomp; \
440 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
441 if (len > RegexLengthToShowInErrorMessages) { \
442 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
443 len = RegexLengthToShowInErrorMessages - 10; \
449 #define FAIL(msg) _FAIL( \
450 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
451 msg, (int)len, RExC_precomp, ellipses))
453 #define FAIL2(msg,arg) _FAIL( \
454 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
455 arg, (int)len, RExC_precomp, ellipses))
458 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
460 #define Simple_vFAIL(m) STMT_START { \
461 const IV offset = RExC_parse - RExC_precomp; \
462 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
463 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
467 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
469 #define vFAIL(m) STMT_START { \
471 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
476 * Like Simple_vFAIL(), but accepts two arguments.
478 #define Simple_vFAIL2(m,a1) STMT_START { \
479 const IV offset = RExC_parse - RExC_precomp; \
480 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
481 (int)offset, RExC_precomp, RExC_precomp + offset); \
485 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
487 #define vFAIL2(m,a1) STMT_START { \
489 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
490 Simple_vFAIL2(m, a1); \
495 * Like Simple_vFAIL(), but accepts three arguments.
497 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
498 const IV offset = RExC_parse - RExC_precomp; \
499 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
500 (int)offset, RExC_precomp, RExC_precomp + offset); \
504 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
506 #define vFAIL3(m,a1,a2) STMT_START { \
508 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
509 Simple_vFAIL3(m, a1, a2); \
513 * Like Simple_vFAIL(), but accepts four arguments.
515 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
516 const IV offset = RExC_parse - RExC_precomp; \
517 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
518 (int)offset, RExC_precomp, RExC_precomp + offset); \
521 #define ckWARNreg(loc,m) STMT_START { \
522 const IV offset = loc - RExC_precomp; \
523 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
524 (int)offset, RExC_precomp, RExC_precomp + offset); \
527 #define ckWARNregdep(loc,m) STMT_START { \
528 const IV offset = loc - RExC_precomp; \
529 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
531 (int)offset, RExC_precomp, RExC_precomp + offset); \
534 #define ckWARN2regdep(loc,m, a1) STMT_START { \
535 const IV offset = loc - RExC_precomp; \
536 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
538 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
541 #define ckWARN2reg(loc, m, a1) STMT_START { \
542 const IV offset = loc - RExC_precomp; \
543 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
544 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
547 #define vWARN3(loc, m, a1, a2) STMT_START { \
548 const IV offset = loc - RExC_precomp; \
549 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
550 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
553 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
554 const IV offset = loc - RExC_precomp; \
555 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
556 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
559 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
560 const IV offset = loc - RExC_precomp; \
561 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
562 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
565 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
566 const IV offset = loc - RExC_precomp; \
567 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
568 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
571 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
572 const IV offset = loc - RExC_precomp; \
573 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
574 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
578 /* Allow for side effects in s */
579 #define REGC(c,s) STMT_START { \
580 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
583 /* Macros for recording node offsets. 20001227 mjd@plover.com
584 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
585 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
586 * Element 0 holds the number n.
587 * Position is 1 indexed.
589 #ifndef RE_TRACK_PATTERN_OFFSETS
590 #define Set_Node_Offset_To_R(node,byte)
591 #define Set_Node_Offset(node,byte)
592 #define Set_Cur_Node_Offset
593 #define Set_Node_Length_To_R(node,len)
594 #define Set_Node_Length(node,len)
595 #define Set_Node_Cur_Length(node)
596 #define Node_Offset(n)
597 #define Node_Length(n)
598 #define Set_Node_Offset_Length(node,offset,len)
599 #define ProgLen(ri) ri->u.proglen
600 #define SetProgLen(ri,x) ri->u.proglen = x
602 #define ProgLen(ri) ri->u.offsets[0]
603 #define SetProgLen(ri,x) ri->u.offsets[0] = x
604 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
606 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
607 __LINE__, (int)(node), (int)(byte))); \
609 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
611 RExC_offsets[2*(node)-1] = (byte); \
616 #define Set_Node_Offset(node,byte) \
617 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
618 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
620 #define Set_Node_Length_To_R(node,len) STMT_START { \
622 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
623 __LINE__, (int)(node), (int)(len))); \
625 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
627 RExC_offsets[2*(node)] = (len); \
632 #define Set_Node_Length(node,len) \
633 Set_Node_Length_To_R((node)-RExC_emit_start, len)
634 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
635 #define Set_Node_Cur_Length(node) \
636 Set_Node_Length(node, RExC_parse - parse_start)
638 /* Get offsets and lengths */
639 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
640 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
642 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
643 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
644 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
648 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
649 #define EXPERIMENTAL_INPLACESCAN
650 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
652 #define DEBUG_STUDYDATA(str,data,depth) \
653 DEBUG_OPTIMISE_MORE_r(if(data){ \
654 PerlIO_printf(Perl_debug_log, \
655 "%*s" str "Pos:%"IVdf"/%"IVdf \
656 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
657 (int)(depth)*2, "", \
658 (IV)((data)->pos_min), \
659 (IV)((data)->pos_delta), \
660 (UV)((data)->flags), \
661 (IV)((data)->whilem_c), \
662 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
663 is_inf ? "INF " : "" \
665 if ((data)->last_found) \
666 PerlIO_printf(Perl_debug_log, \
667 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
668 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
669 SvPVX_const((data)->last_found), \
670 (IV)((data)->last_end), \
671 (IV)((data)->last_start_min), \
672 (IV)((data)->last_start_max), \
673 ((data)->longest && \
674 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
675 SvPVX_const((data)->longest_fixed), \
676 (IV)((data)->offset_fixed), \
677 ((data)->longest && \
678 (data)->longest==&((data)->longest_float)) ? "*" : "", \
679 SvPVX_const((data)->longest_float), \
680 (IV)((data)->offset_float_min), \
681 (IV)((data)->offset_float_max) \
683 PerlIO_printf(Perl_debug_log,"\n"); \
686 static void clear_re(pTHX_ void *r);
688 /* Mark that we cannot extend a found fixed substring at this point.
689 Update the longest found anchored substring and the longest found
690 floating substrings if needed. */
693 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
695 const STRLEN l = CHR_SVLEN(data->last_found);
696 const STRLEN old_l = CHR_SVLEN(*data->longest);
697 GET_RE_DEBUG_FLAGS_DECL;
699 PERL_ARGS_ASSERT_SCAN_COMMIT;
701 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
702 SvSetMagicSV(*data->longest, data->last_found);
703 if (*data->longest == data->longest_fixed) {
704 data->offset_fixed = l ? data->last_start_min : data->pos_min;
705 if (data->flags & SF_BEFORE_EOL)
707 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
709 data->flags &= ~SF_FIX_BEFORE_EOL;
710 data->minlen_fixed=minlenp;
711 data->lookbehind_fixed=0;
713 else { /* *data->longest == data->longest_float */
714 data->offset_float_min = l ? data->last_start_min : data->pos_min;
715 data->offset_float_max = (l
716 ? data->last_start_max
717 : data->pos_min + data->pos_delta);
718 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
719 data->offset_float_max = I32_MAX;
720 if (data->flags & SF_BEFORE_EOL)
722 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
724 data->flags &= ~SF_FL_BEFORE_EOL;
725 data->minlen_float=minlenp;
726 data->lookbehind_float=0;
729 SvCUR_set(data->last_found, 0);
731 SV * const sv = data->last_found;
732 if (SvUTF8(sv) && SvMAGICAL(sv)) {
733 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
739 data->flags &= ~SF_BEFORE_EOL;
740 DEBUG_STUDYDATA("commit: ",data,0);
743 /* Can match anything (initialization) */
745 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
747 PERL_ARGS_ASSERT_CL_ANYTHING;
749 ANYOF_BITMAP_SETALL(cl);
750 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
751 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
753 /* If any portion of the regex is to operate under locale rules,
754 * initialization includes it. The reason this isn't done for all regexes
755 * is that the optimizer was written under the assumption that locale was
756 * all-or-nothing. Given the complexity and lack of documentation in the
757 * optimizer, and that there are inadequate test cases for locale, so many
758 * parts of it may not work properly, it is safest to avoid locale unless
760 if (RExC_contains_locale) {
761 ANYOF_CLASS_SETALL(cl); /* /l uses class */
762 cl->flags |= ANYOF_LOCALE;
765 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
769 /* Can match anything (initialization) */
771 S_cl_is_anything(const struct regnode_charclass_class *cl)
775 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
777 for (value = 0; value <= ANYOF_MAX; value += 2)
778 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
780 if (!(cl->flags & ANYOF_UNICODE_ALL))
782 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
787 /* Can match anything (initialization) */
789 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
791 PERL_ARGS_ASSERT_CL_INIT;
793 Zero(cl, 1, struct regnode_charclass_class);
795 cl_anything(pRExC_state, cl);
796 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
799 /* These two functions currently do the exact same thing */
800 #define cl_init_zero S_cl_init
802 /* 'AND' a given class with another one. Can create false positives. 'cl'
803 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
804 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
806 S_cl_and(struct regnode_charclass_class *cl,
807 const struct regnode_charclass_class *and_with)
809 PERL_ARGS_ASSERT_CL_AND;
811 assert(and_with->type == ANYOF);
813 /* I (khw) am not sure all these restrictions are necessary XXX */
814 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
815 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
816 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
817 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
818 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
821 if (and_with->flags & ANYOF_INVERT)
822 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
823 cl->bitmap[i] &= ~and_with->bitmap[i];
825 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
826 cl->bitmap[i] &= and_with->bitmap[i];
827 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
829 if (and_with->flags & ANYOF_INVERT) {
831 /* Here, the and'ed node is inverted. Get the AND of the flags that
832 * aren't affected by the inversion. Those that are affected are
833 * handled individually below */
834 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
835 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
836 cl->flags |= affected_flags;
838 /* We currently don't know how to deal with things that aren't in the
839 * bitmap, but we know that the intersection is no greater than what
840 * is already in cl, so let there be false positives that get sorted
841 * out after the synthetic start class succeeds, and the node is
842 * matched for real. */
844 /* The inversion of these two flags indicate that the resulting
845 * intersection doesn't have them */
846 if (and_with->flags & ANYOF_UNICODE_ALL) {
847 cl->flags &= ~ANYOF_UNICODE_ALL;
849 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
850 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
853 else { /* and'd node is not inverted */
854 U8 outside_bitmap_but_not_utf8; /* Temp variable */
856 if (! ANYOF_NONBITMAP(and_with)) {
858 /* Here 'and_with' doesn't match anything outside the bitmap
859 * (except possibly ANYOF_UNICODE_ALL), which means the
860 * intersection can't either, except for ANYOF_UNICODE_ALL, in
861 * which case we don't know what the intersection is, but it's no
862 * greater than what cl already has, so can just leave it alone,
863 * with possible false positives */
864 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
865 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
866 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
869 else if (! ANYOF_NONBITMAP(cl)) {
871 /* Here, 'and_with' does match something outside the bitmap, and cl
872 * doesn't have a list of things to match outside the bitmap. If
873 * cl can match all code points above 255, the intersection will
874 * be those above-255 code points that 'and_with' matches. If cl
875 * can't match all Unicode code points, it means that it can't
876 * match anything outside the bitmap (since the 'if' that got us
877 * into this block tested for that), so we leave the bitmap empty.
879 if (cl->flags & ANYOF_UNICODE_ALL) {
880 ARG_SET(cl, ARG(and_with));
882 /* and_with's ARG may match things that don't require UTF8.
883 * And now cl's will too, in spite of this being an 'and'. See
884 * the comments below about the kludge */
885 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
889 /* Here, both 'and_with' and cl match something outside the
890 * bitmap. Currently we do not do the intersection, so just match
891 * whatever cl had at the beginning. */
895 /* Take the intersection of the two sets of flags. However, the
896 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
897 * kludge around the fact that this flag is not treated like the others
898 * which are initialized in cl_anything(). The way the optimizer works
899 * is that the synthetic start class (SSC) is initialized to match
900 * anything, and then the first time a real node is encountered, its
901 * values are AND'd with the SSC's with the result being the values of
902 * the real node. However, there are paths through the optimizer where
903 * the AND never gets called, so those initialized bits are set
904 * inappropriately, which is not usually a big deal, as they just cause
905 * false positives in the SSC, which will just mean a probably
906 * imperceptible slow down in execution. However this bit has a
907 * higher false positive consequence in that it can cause utf8.pm,
908 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
909 * bigger slowdown and also causes significant extra memory to be used.
910 * In order to prevent this, the code now takes a different tack. The
911 * bit isn't set unless some part of the regular expression needs it,
912 * but once set it won't get cleared. This means that these extra
913 * modules won't get loaded unless there was some path through the
914 * pattern that would have required them anyway, and so any false
915 * positives that occur by not ANDing them out when they could be
916 * aren't as severe as they would be if we treated this bit like all
918 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
919 & ANYOF_NONBITMAP_NON_UTF8;
920 cl->flags &= and_with->flags;
921 cl->flags |= outside_bitmap_but_not_utf8;
925 /* 'OR' a given class with another one. Can create false positives. 'cl'
926 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
927 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
929 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
931 PERL_ARGS_ASSERT_CL_OR;
933 if (or_with->flags & ANYOF_INVERT) {
935 /* Here, the or'd node is to be inverted. This means we take the
936 * complement of everything not in the bitmap, but currently we don't
937 * know what that is, so give up and match anything */
938 if (ANYOF_NONBITMAP(or_with)) {
939 cl_anything(pRExC_state, cl);
942 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
943 * <= (B1 | !B2) | (CL1 | !CL2)
944 * which is wasteful if CL2 is small, but we ignore CL2:
945 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
946 * XXXX Can we handle case-fold? Unclear:
947 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
948 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
950 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
951 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
952 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
955 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
956 cl->bitmap[i] |= ~or_with->bitmap[i];
957 } /* XXXX: logic is complicated otherwise */
959 cl_anything(pRExC_state, cl);
962 /* And, we can just take the union of the flags that aren't affected
963 * by the inversion */
964 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
966 /* For the remaining flags:
967 ANYOF_UNICODE_ALL and inverted means to not match anything above
968 255, which means that the union with cl should just be
969 what cl has in it, so can ignore this flag
970 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
971 is 127-255 to match them, but then invert that, so the
972 union with cl should just be what cl has in it, so can
975 } else { /* 'or_with' is not inverted */
976 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
977 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
978 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
979 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
982 /* OR char bitmap and class bitmap separately */
983 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
984 cl->bitmap[i] |= or_with->bitmap[i];
985 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
986 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
987 cl->classflags[i] |= or_with->classflags[i];
988 cl->flags |= ANYOF_CLASS;
991 else { /* XXXX: logic is complicated, leave it along for a moment. */
992 cl_anything(pRExC_state, cl);
995 if (ANYOF_NONBITMAP(or_with)) {
997 /* Use the added node's outside-the-bit-map match if there isn't a
998 * conflict. If there is a conflict (both nodes match something
999 * outside the bitmap, but what they match outside is not the same
1000 * pointer, and hence not easily compared until XXX we extend
1001 * inversion lists this far), give up and allow the start class to
1002 * match everything outside the bitmap. If that stuff is all above
1003 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1004 if (! ANYOF_NONBITMAP(cl)) {
1005 ARG_SET(cl, ARG(or_with));
1007 else if (ARG(cl) != ARG(or_with)) {
1009 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1010 cl_anything(pRExC_state, cl);
1013 cl->flags |= ANYOF_UNICODE_ALL;
1018 /* Take the union */
1019 cl->flags |= or_with->flags;
1023 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1024 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1025 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1026 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1031 dump_trie(trie,widecharmap,revcharmap)
1032 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1033 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1035 These routines dump out a trie in a somewhat readable format.
1036 The _interim_ variants are used for debugging the interim
1037 tables that are used to generate the final compressed
1038 representation which is what dump_trie expects.
1040 Part of the reason for their existence is to provide a form
1041 of documentation as to how the different representations function.
1046 Dumps the final compressed table form of the trie to Perl_debug_log.
1047 Used for debugging make_trie().
1051 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1052 AV *revcharmap, U32 depth)
1055 SV *sv=sv_newmortal();
1056 int colwidth= widecharmap ? 6 : 4;
1058 GET_RE_DEBUG_FLAGS_DECL;
1060 PERL_ARGS_ASSERT_DUMP_TRIE;
1062 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1063 (int)depth * 2 + 2,"",
1064 "Match","Base","Ofs" );
1066 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1067 SV ** const tmp = av_fetch( revcharmap, state, 0);
1069 PerlIO_printf( Perl_debug_log, "%*s",
1071 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1072 PL_colors[0], PL_colors[1],
1073 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1074 PERL_PV_ESCAPE_FIRSTCHAR
1079 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1080 (int)depth * 2 + 2,"");
1082 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1083 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1084 PerlIO_printf( Perl_debug_log, "\n");
1086 for( state = 1 ; state < trie->statecount ; state++ ) {
1087 const U32 base = trie->states[ state ].trans.base;
1089 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1091 if ( trie->states[ state ].wordnum ) {
1092 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1094 PerlIO_printf( Perl_debug_log, "%6s", "" );
1097 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1102 while( ( base + ofs < trie->uniquecharcount ) ||
1103 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1104 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1107 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1109 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1110 if ( ( base + ofs >= trie->uniquecharcount ) &&
1111 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1112 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1114 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1116 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1118 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1122 PerlIO_printf( Perl_debug_log, "]");
1125 PerlIO_printf( Perl_debug_log, "\n" );
1127 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1128 for (word=1; word <= trie->wordcount; word++) {
1129 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1130 (int)word, (int)(trie->wordinfo[word].prev),
1131 (int)(trie->wordinfo[word].len));
1133 PerlIO_printf(Perl_debug_log, "\n" );
1136 Dumps a fully constructed but uncompressed trie in list form.
1137 List tries normally only are used for construction when the number of
1138 possible chars (trie->uniquecharcount) is very high.
1139 Used for debugging make_trie().
1142 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1143 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1147 SV *sv=sv_newmortal();
1148 int colwidth= widecharmap ? 6 : 4;
1149 GET_RE_DEBUG_FLAGS_DECL;
1151 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1153 /* print out the table precompression. */
1154 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1155 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1156 "------:-----+-----------------\n" );
1158 for( state=1 ; state < next_alloc ; state ++ ) {
1161 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1162 (int)depth * 2 + 2,"", (UV)state );
1163 if ( ! trie->states[ state ].wordnum ) {
1164 PerlIO_printf( Perl_debug_log, "%5s| ","");
1166 PerlIO_printf( Perl_debug_log, "W%4x| ",
1167 trie->states[ state ].wordnum
1170 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1171 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1173 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1175 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1176 PL_colors[0], PL_colors[1],
1177 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1178 PERL_PV_ESCAPE_FIRSTCHAR
1180 TRIE_LIST_ITEM(state,charid).forid,
1181 (UV)TRIE_LIST_ITEM(state,charid).newstate
1184 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1185 (int)((depth * 2) + 14), "");
1188 PerlIO_printf( Perl_debug_log, "\n");
1193 Dumps a fully constructed but uncompressed trie in table form.
1194 This is the normal DFA style state transition table, with a few
1195 twists to facilitate compression later.
1196 Used for debugging make_trie().
1199 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1200 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1205 SV *sv=sv_newmortal();
1206 int colwidth= widecharmap ? 6 : 4;
1207 GET_RE_DEBUG_FLAGS_DECL;
1209 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1212 print out the table precompression so that we can do a visual check
1213 that they are identical.
1216 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1218 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1219 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1221 PerlIO_printf( Perl_debug_log, "%*s",
1223 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1224 PL_colors[0], PL_colors[1],
1225 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1226 PERL_PV_ESCAPE_FIRSTCHAR
1232 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1234 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1235 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1238 PerlIO_printf( Perl_debug_log, "\n" );
1240 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1242 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1243 (int)depth * 2 + 2,"",
1244 (UV)TRIE_NODENUM( state ) );
1246 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1247 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1249 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1251 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1253 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1254 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1256 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1257 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1265 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1266 startbranch: the first branch in the whole branch sequence
1267 first : start branch of sequence of branch-exact nodes.
1268 May be the same as startbranch
1269 last : Thing following the last branch.
1270 May be the same as tail.
1271 tail : item following the branch sequence
1272 count : words in the sequence
1273 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1274 depth : indent depth
1276 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1278 A trie is an N'ary tree where the branches are determined by digital
1279 decomposition of the key. IE, at the root node you look up the 1st character and
1280 follow that branch repeat until you find the end of the branches. Nodes can be
1281 marked as "accepting" meaning they represent a complete word. Eg:
1285 would convert into the following structure. Numbers represent states, letters
1286 following numbers represent valid transitions on the letter from that state, if
1287 the number is in square brackets it represents an accepting state, otherwise it
1288 will be in parenthesis.
1290 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1294 (1) +-i->(6)-+-s->[7]
1296 +-s->(3)-+-h->(4)-+-e->[5]
1298 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1300 This shows that when matching against the string 'hers' we will begin at state 1
1301 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1302 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1303 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1304 single traverse. We store a mapping from accepting to state to which word was
1305 matched, and then when we have multiple possibilities we try to complete the
1306 rest of the regex in the order in which they occured in the alternation.
1308 The only prior NFA like behaviour that would be changed by the TRIE support is
1309 the silent ignoring of duplicate alternations which are of the form:
1311 / (DUPE|DUPE) X? (?{ ... }) Y /x
1313 Thus EVAL blocks following a trie may be called a different number of times with
1314 and without the optimisation. With the optimisations dupes will be silently
1315 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1316 the following demonstrates:
1318 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1320 which prints out 'word' three times, but
1322 'words'=~/(word|word|word)(?{ print $1 })S/
1324 which doesnt print it out at all. This is due to other optimisations kicking in.
1326 Example of what happens on a structural level:
1328 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1330 1: CURLYM[1] {1,32767}(18)
1341 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1342 and should turn into:
1344 1: CURLYM[1] {1,32767}(18)
1346 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1354 Cases where tail != last would be like /(?foo|bar)baz/:
1364 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1365 and would end up looking like:
1368 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1375 d = uvuni_to_utf8_flags(d, uv, 0);
1377 is the recommended Unicode-aware way of saying
1382 #define TRIE_STORE_REVCHAR(val) \
1385 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1386 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1387 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1388 SvCUR_set(zlopp, kapow - flrbbbbb); \
1391 av_push(revcharmap, zlopp); \
1393 char ooooff = (char)val; \
1394 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1398 #define TRIE_READ_CHAR STMT_START { \
1401 /* if it is UTF then it is either already folded, or does not need folding */ \
1402 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1404 else if (folder == PL_fold_latin1) { \
1405 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1406 if ( foldlen > 0 ) { \
1407 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1413 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1414 skiplen = UNISKIP(uvc); \
1415 foldlen -= skiplen; \
1416 scan = foldbuf + skiplen; \
1419 /* raw data, will be folded later if needed */ \
1427 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1428 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1429 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1430 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1432 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1433 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1434 TRIE_LIST_CUR( state )++; \
1437 #define TRIE_LIST_NEW(state) STMT_START { \
1438 Newxz( trie->states[ state ].trans.list, \
1439 4, reg_trie_trans_le ); \
1440 TRIE_LIST_CUR( state ) = 1; \
1441 TRIE_LIST_LEN( state ) = 4; \
1444 #define TRIE_HANDLE_WORD(state) STMT_START { \
1445 U16 dupe= trie->states[ state ].wordnum; \
1446 regnode * const noper_next = regnext( noper ); \
1449 /* store the word for dumping */ \
1451 if (OP(noper) != NOTHING) \
1452 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1454 tmp = newSVpvn_utf8( "", 0, UTF ); \
1455 av_push( trie_words, tmp ); \
1459 trie->wordinfo[curword].prev = 0; \
1460 trie->wordinfo[curword].len = wordlen; \
1461 trie->wordinfo[curword].accept = state; \
1463 if ( noper_next < tail ) { \
1465 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1466 trie->jump[curword] = (U16)(noper_next - convert); \
1468 jumper = noper_next; \
1470 nextbranch= regnext(cur); \
1474 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1475 /* chain, so that when the bits of chain are later */\
1476 /* linked together, the dups appear in the chain */\
1477 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1478 trie->wordinfo[dupe].prev = curword; \
1480 /* we haven't inserted this word yet. */ \
1481 trie->states[ state ].wordnum = curword; \
1486 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1487 ( ( base + charid >= ucharcount \
1488 && base + charid < ubound \
1489 && state == trie->trans[ base - ucharcount + charid ].check \
1490 && trie->trans[ base - ucharcount + charid ].next ) \
1491 ? trie->trans[ base - ucharcount + charid ].next \
1492 : ( state==1 ? special : 0 ) \
1496 #define MADE_JUMP_TRIE 2
1497 #define MADE_EXACT_TRIE 4
1500 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1503 /* first pass, loop through and scan words */
1504 reg_trie_data *trie;
1505 HV *widecharmap = NULL;
1506 AV *revcharmap = newAV();
1508 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1513 regnode *jumper = NULL;
1514 regnode *nextbranch = NULL;
1515 regnode *convert = NULL;
1516 U32 *prev_states; /* temp array mapping each state to previous one */
1517 /* we just use folder as a flag in utf8 */
1518 const U8 * folder = NULL;
1521 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1522 AV *trie_words = NULL;
1523 /* along with revcharmap, this only used during construction but both are
1524 * useful during debugging so we store them in the struct when debugging.
1527 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1528 STRLEN trie_charcount=0;
1530 SV *re_trie_maxbuff;
1531 GET_RE_DEBUG_FLAGS_DECL;
1533 PERL_ARGS_ASSERT_MAKE_TRIE;
1535 PERL_UNUSED_ARG(depth);
1542 case EXACTFU_TRICKYFOLD:
1543 case EXACTFU: folder = PL_fold_latin1; break;
1544 case EXACTF: folder = PL_fold; break;
1545 case EXACTFL: folder = PL_fold_locale; break;
1546 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1549 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1551 trie->startstate = 1;
1552 trie->wordcount = word_count;
1553 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1554 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1556 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1557 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1558 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1561 trie_words = newAV();
1564 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1565 if (!SvIOK(re_trie_maxbuff)) {
1566 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1568 DEBUG_TRIE_COMPILE_r({
1569 PerlIO_printf( Perl_debug_log,
1570 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1571 (int)depth * 2 + 2, "",
1572 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1573 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1577 /* Find the node we are going to overwrite */
1578 if ( first == startbranch && OP( last ) != BRANCH ) {
1579 /* whole branch chain */
1582 /* branch sub-chain */
1583 convert = NEXTOPER( first );
1586 /* -- First loop and Setup --
1588 We first traverse the branches and scan each word to determine if it
1589 contains widechars, and how many unique chars there are, this is
1590 important as we have to build a table with at least as many columns as we
1593 We use an array of integers to represent the character codes 0..255
1594 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1595 native representation of the character value as the key and IV's for the
1598 *TODO* If we keep track of how many times each character is used we can
1599 remap the columns so that the table compression later on is more
1600 efficient in terms of memory by ensuring the most common value is in the
1601 middle and the least common are on the outside. IMO this would be better
1602 than a most to least common mapping as theres a decent chance the most
1603 common letter will share a node with the least common, meaning the node
1604 will not be compressible. With a middle is most common approach the worst
1605 case is when we have the least common nodes twice.
1609 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1610 regnode *noper = NEXTOPER( cur );
1611 const U8 *uc = (U8*)STRING( noper );
1612 const U8 *e = uc + STR_LEN( noper );
1614 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1616 const U8 *scan = (U8*)NULL;
1617 U32 wordlen = 0; /* required init */
1619 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1621 if (OP(noper) == NOTHING) {
1622 regnode *noper_next= regnext(noper);
1623 if (noper_next != tail && OP(noper_next) == flags) {
1625 uc= (U8*)STRING(noper);
1626 e= uc + STR_LEN(noper);
1627 trie->minlen= STR_LEN(noper);
1634 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1635 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1636 regardless of encoding */
1637 if (OP( noper ) == EXACTFU_SS) {
1638 /* false positives are ok, so just set this */
1639 TRIE_BITMAP_SET(trie,0xDF);
1642 for ( ; uc < e ; uc += len ) {
1643 TRIE_CHARCOUNT(trie)++;
1648 U8 folded= folder[ (U8) uvc ];
1649 if ( !trie->charmap[ folded ] ) {
1650 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1651 TRIE_STORE_REVCHAR( folded );
1654 if ( !trie->charmap[ uvc ] ) {
1655 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1656 TRIE_STORE_REVCHAR( uvc );
1659 /* store the codepoint in the bitmap, and its folded
1661 TRIE_BITMAP_SET(trie, uvc);
1663 /* store the folded codepoint */
1664 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1667 /* store first byte of utf8 representation of
1668 variant codepoints */
1669 if (! UNI_IS_INVARIANT(uvc)) {
1670 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1673 set_bit = 0; /* We've done our bit :-) */
1678 widecharmap = newHV();
1680 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1683 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1685 if ( !SvTRUE( *svpp ) ) {
1686 sv_setiv( *svpp, ++trie->uniquecharcount );
1687 TRIE_STORE_REVCHAR(uvc);
1691 if( cur == first ) {
1692 trie->minlen = chars;
1693 trie->maxlen = chars;
1694 } else if (chars < trie->minlen) {
1695 trie->minlen = chars;
1696 } else if (chars > trie->maxlen) {
1697 trie->maxlen = chars;
1699 if (OP( noper ) == EXACTFU_SS) {
1700 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1701 if (trie->minlen > 1)
1704 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1705 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1706 * - We assume that any such sequence might match a 2 byte string */
1707 if (trie->minlen > 2 )
1711 } /* end first pass */
1712 DEBUG_TRIE_COMPILE_r(
1713 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1714 (int)depth * 2 + 2,"",
1715 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1716 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1717 (int)trie->minlen, (int)trie->maxlen )
1721 We now know what we are dealing with in terms of unique chars and
1722 string sizes so we can calculate how much memory a naive
1723 representation using a flat table will take. If it's over a reasonable
1724 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1725 conservative but potentially much slower representation using an array
1728 At the end we convert both representations into the same compressed
1729 form that will be used in regexec.c for matching with. The latter
1730 is a form that cannot be used to construct with but has memory
1731 properties similar to the list form and access properties similar
1732 to the table form making it both suitable for fast searches and
1733 small enough that its feasable to store for the duration of a program.
1735 See the comment in the code where the compressed table is produced
1736 inplace from the flat tabe representation for an explanation of how
1737 the compression works.
1742 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1745 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1747 Second Pass -- Array Of Lists Representation
1749 Each state will be represented by a list of charid:state records
1750 (reg_trie_trans_le) the first such element holds the CUR and LEN
1751 points of the allocated array. (See defines above).
1753 We build the initial structure using the lists, and then convert
1754 it into the compressed table form which allows faster lookups
1755 (but cant be modified once converted).
1758 STRLEN transcount = 1;
1760 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1761 "%*sCompiling trie using list compiler\n",
1762 (int)depth * 2 + 2, ""));
1764 trie->states = (reg_trie_state *)
1765 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1766 sizeof(reg_trie_state) );
1770 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1772 regnode *noper = NEXTOPER( cur );
1773 U8 *uc = (U8*)STRING( noper );
1774 const U8 *e = uc + STR_LEN( noper );
1775 U32 state = 1; /* required init */
1776 U16 charid = 0; /* sanity init */
1777 U8 *scan = (U8*)NULL; /* sanity init */
1778 STRLEN foldlen = 0; /* required init */
1779 U32 wordlen = 0; /* required init */
1780 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1783 if (OP(noper) == NOTHING) {
1784 regnode *noper_next= regnext(noper);
1785 if (noper_next != tail && OP(noper_next) == flags) {
1787 uc= (U8*)STRING(noper);
1788 e= uc + STR_LEN(noper);
1792 if (OP(noper) != NOTHING) {
1793 for ( ; uc < e ; uc += len ) {
1798 charid = trie->charmap[ uvc ];
1800 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1804 charid=(U16)SvIV( *svpp );
1807 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1814 if ( !trie->states[ state ].trans.list ) {
1815 TRIE_LIST_NEW( state );
1817 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1818 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1819 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1824 newstate = next_alloc++;
1825 prev_states[newstate] = state;
1826 TRIE_LIST_PUSH( state, charid, newstate );
1831 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1835 TRIE_HANDLE_WORD(state);
1837 } /* end second pass */
1839 /* next alloc is the NEXT state to be allocated */
1840 trie->statecount = next_alloc;
1841 trie->states = (reg_trie_state *)
1842 PerlMemShared_realloc( trie->states,
1844 * sizeof(reg_trie_state) );
1846 /* and now dump it out before we compress it */
1847 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1848 revcharmap, next_alloc,
1852 trie->trans = (reg_trie_trans *)
1853 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1860 for( state=1 ; state < next_alloc ; state ++ ) {
1864 DEBUG_TRIE_COMPILE_MORE_r(
1865 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1869 if (trie->states[state].trans.list) {
1870 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1874 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1875 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1876 if ( forid < minid ) {
1878 } else if ( forid > maxid ) {
1882 if ( transcount < tp + maxid - minid + 1) {
1884 trie->trans = (reg_trie_trans *)
1885 PerlMemShared_realloc( trie->trans,
1887 * sizeof(reg_trie_trans) );
1888 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1890 base = trie->uniquecharcount + tp - minid;
1891 if ( maxid == minid ) {
1893 for ( ; zp < tp ; zp++ ) {
1894 if ( ! trie->trans[ zp ].next ) {
1895 base = trie->uniquecharcount + zp - minid;
1896 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1897 trie->trans[ zp ].check = state;
1903 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1904 trie->trans[ tp ].check = state;
1909 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1910 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1911 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1912 trie->trans[ tid ].check = state;
1914 tp += ( maxid - minid + 1 );
1916 Safefree(trie->states[ state ].trans.list);
1919 DEBUG_TRIE_COMPILE_MORE_r(
1920 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1923 trie->states[ state ].trans.base=base;
1925 trie->lasttrans = tp + 1;
1929 Second Pass -- Flat Table Representation.
1931 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1932 We know that we will need Charcount+1 trans at most to store the data
1933 (one row per char at worst case) So we preallocate both structures
1934 assuming worst case.
1936 We then construct the trie using only the .next slots of the entry
1939 We use the .check field of the first entry of the node temporarily to
1940 make compression both faster and easier by keeping track of how many non
1941 zero fields are in the node.
1943 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1946 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1947 number representing the first entry of the node, and state as a
1948 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1949 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1950 are 2 entrys per node. eg:
1958 The table is internally in the right hand, idx form. However as we also
1959 have to deal with the states array which is indexed by nodenum we have to
1960 use TRIE_NODENUM() to convert.
1963 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1964 "%*sCompiling trie using table compiler\n",
1965 (int)depth * 2 + 2, ""));
1967 trie->trans = (reg_trie_trans *)
1968 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1969 * trie->uniquecharcount + 1,
1970 sizeof(reg_trie_trans) );
1971 trie->states = (reg_trie_state *)
1972 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1973 sizeof(reg_trie_state) );
1974 next_alloc = trie->uniquecharcount + 1;
1977 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1979 regnode *noper = NEXTOPER( cur );
1980 const U8 *uc = (U8*)STRING( noper );
1981 const U8 *e = uc + STR_LEN( noper );
1983 U32 state = 1; /* required init */
1985 U16 charid = 0; /* sanity init */
1986 U32 accept_state = 0; /* sanity init */
1987 U8 *scan = (U8*)NULL; /* sanity init */
1989 STRLEN foldlen = 0; /* required init */
1990 U32 wordlen = 0; /* required init */
1992 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1994 if (OP(noper) == NOTHING) {
1995 regnode *noper_next= regnext(noper);
1996 if (noper_next != tail && OP(noper_next) == flags) {
1998 uc= (U8*)STRING(noper);
1999 e= uc + STR_LEN(noper);
2003 if ( OP(noper) != NOTHING ) {
2004 for ( ; uc < e ; uc += len ) {
2009 charid = trie->charmap[ uvc ];
2011 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2012 charid = svpp ? (U16)SvIV(*svpp) : 0;
2016 if ( !trie->trans[ state + charid ].next ) {
2017 trie->trans[ state + charid ].next = next_alloc;
2018 trie->trans[ state ].check++;
2019 prev_states[TRIE_NODENUM(next_alloc)]
2020 = TRIE_NODENUM(state);
2021 next_alloc += trie->uniquecharcount;
2023 state = trie->trans[ state + charid ].next;
2025 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2027 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2030 accept_state = TRIE_NODENUM( state );
2031 TRIE_HANDLE_WORD(accept_state);
2033 } /* end second pass */
2035 /* and now dump it out before we compress it */
2036 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2038 next_alloc, depth+1));
2042 * Inplace compress the table.*
2044 For sparse data sets the table constructed by the trie algorithm will
2045 be mostly 0/FAIL transitions or to put it another way mostly empty.
2046 (Note that leaf nodes will not contain any transitions.)
2048 This algorithm compresses the tables by eliminating most such
2049 transitions, at the cost of a modest bit of extra work during lookup:
2051 - Each states[] entry contains a .base field which indicates the
2052 index in the state[] array wheres its transition data is stored.
2054 - If .base is 0 there are no valid transitions from that node.
2056 - If .base is nonzero then charid is added to it to find an entry in
2059 -If trans[states[state].base+charid].check!=state then the
2060 transition is taken to be a 0/Fail transition. Thus if there are fail
2061 transitions at the front of the node then the .base offset will point
2062 somewhere inside the previous nodes data (or maybe even into a node
2063 even earlier), but the .check field determines if the transition is
2067 The following process inplace converts the table to the compressed
2068 table: We first do not compress the root node 1,and mark all its
2069 .check pointers as 1 and set its .base pointer as 1 as well. This
2070 allows us to do a DFA construction from the compressed table later,
2071 and ensures that any .base pointers we calculate later are greater
2074 - We set 'pos' to indicate the first entry of the second node.
2076 - We then iterate over the columns of the node, finding the first and
2077 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2078 and set the .check pointers accordingly, and advance pos
2079 appropriately and repreat for the next node. Note that when we copy
2080 the next pointers we have to convert them from the original
2081 NODEIDX form to NODENUM form as the former is not valid post
2084 - If a node has no transitions used we mark its base as 0 and do not
2085 advance the pos pointer.
2087 - If a node only has one transition we use a second pointer into the
2088 structure to fill in allocated fail transitions from other states.
2089 This pointer is independent of the main pointer and scans forward
2090 looking for null transitions that are allocated to a state. When it
2091 finds one it writes the single transition into the "hole". If the
2092 pointer doesnt find one the single transition is appended as normal.
2094 - Once compressed we can Renew/realloc the structures to release the
2097 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2098 specifically Fig 3.47 and the associated pseudocode.
2102 const U32 laststate = TRIE_NODENUM( next_alloc );
2105 trie->statecount = laststate;
2107 for ( state = 1 ; state < laststate ; state++ ) {
2109 const U32 stateidx = TRIE_NODEIDX( state );
2110 const U32 o_used = trie->trans[ stateidx ].check;
2111 U32 used = trie->trans[ stateidx ].check;
2112 trie->trans[ stateidx ].check = 0;
2114 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2115 if ( flag || trie->trans[ stateidx + charid ].next ) {
2116 if ( trie->trans[ stateidx + charid ].next ) {
2118 for ( ; zp < pos ; zp++ ) {
2119 if ( ! trie->trans[ zp ].next ) {
2123 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2124 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2125 trie->trans[ zp ].check = state;
2126 if ( ++zp > pos ) pos = zp;
2133 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2135 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2136 trie->trans[ pos ].check = state;
2141 trie->lasttrans = pos + 1;
2142 trie->states = (reg_trie_state *)
2143 PerlMemShared_realloc( trie->states, laststate
2144 * sizeof(reg_trie_state) );
2145 DEBUG_TRIE_COMPILE_MORE_r(
2146 PerlIO_printf( Perl_debug_log,
2147 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2148 (int)depth * 2 + 2,"",
2149 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2152 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2155 } /* end table compress */
2157 DEBUG_TRIE_COMPILE_MORE_r(
2158 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2159 (int)depth * 2 + 2, "",
2160 (UV)trie->statecount,
2161 (UV)trie->lasttrans)
2163 /* resize the trans array to remove unused space */
2164 trie->trans = (reg_trie_trans *)
2165 PerlMemShared_realloc( trie->trans, trie->lasttrans
2166 * sizeof(reg_trie_trans) );
2168 { /* Modify the program and insert the new TRIE node */
2169 U8 nodetype =(U8)(flags & 0xFF);
2173 regnode *optimize = NULL;
2174 #ifdef RE_TRACK_PATTERN_OFFSETS
2177 U32 mjd_nodelen = 0;
2178 #endif /* RE_TRACK_PATTERN_OFFSETS */
2179 #endif /* DEBUGGING */
2181 This means we convert either the first branch or the first Exact,
2182 depending on whether the thing following (in 'last') is a branch
2183 or not and whther first is the startbranch (ie is it a sub part of
2184 the alternation or is it the whole thing.)
2185 Assuming its a sub part we convert the EXACT otherwise we convert
2186 the whole branch sequence, including the first.
2188 /* Find the node we are going to overwrite */
2189 if ( first != startbranch || OP( last ) == BRANCH ) {
2190 /* branch sub-chain */
2191 NEXT_OFF( first ) = (U16)(last - first);
2192 #ifdef RE_TRACK_PATTERN_OFFSETS
2194 mjd_offset= Node_Offset((convert));
2195 mjd_nodelen= Node_Length((convert));
2198 /* whole branch chain */
2200 #ifdef RE_TRACK_PATTERN_OFFSETS
2203 const regnode *nop = NEXTOPER( convert );
2204 mjd_offset= Node_Offset((nop));
2205 mjd_nodelen= Node_Length((nop));
2209 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2210 (int)depth * 2 + 2, "",
2211 (UV)mjd_offset, (UV)mjd_nodelen)
2214 /* But first we check to see if there is a common prefix we can
2215 split out as an EXACT and put in front of the TRIE node. */
2216 trie->startstate= 1;
2217 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2219 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2223 const U32 base = trie->states[ state ].trans.base;
2225 if ( trie->states[state].wordnum )
2228 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2229 if ( ( base + ofs >= trie->uniquecharcount ) &&
2230 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2231 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2233 if ( ++count > 1 ) {
2234 SV **tmp = av_fetch( revcharmap, ofs, 0);
2235 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2236 if ( state == 1 ) break;
2238 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2240 PerlIO_printf(Perl_debug_log,
2241 "%*sNew Start State=%"UVuf" Class: [",
2242 (int)depth * 2 + 2, "",
2245 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2246 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2248 TRIE_BITMAP_SET(trie,*ch);
2250 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2252 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2256 TRIE_BITMAP_SET(trie,*ch);
2258 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2259 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2265 SV **tmp = av_fetch( revcharmap, idx, 0);
2267 char *ch = SvPV( *tmp, len );
2269 SV *sv=sv_newmortal();
2270 PerlIO_printf( Perl_debug_log,
2271 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2272 (int)depth * 2 + 2, "",
2274 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2275 PL_colors[0], PL_colors[1],
2276 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2277 PERL_PV_ESCAPE_FIRSTCHAR
2282 OP( convert ) = nodetype;
2283 str=STRING(convert);
2286 STR_LEN(convert) += len;
2292 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2297 trie->prefixlen = (state-1);
2299 regnode *n = convert+NODE_SZ_STR(convert);
2300 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2301 trie->startstate = state;
2302 trie->minlen -= (state - 1);
2303 trie->maxlen -= (state - 1);
2305 /* At least the UNICOS C compiler choked on this
2306 * being argument to DEBUG_r(), so let's just have
2309 #ifdef PERL_EXT_RE_BUILD
2315 regnode *fix = convert;
2316 U32 word = trie->wordcount;
2318 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2319 while( ++fix < n ) {
2320 Set_Node_Offset_Length(fix, 0, 0);
2323 SV ** const tmp = av_fetch( trie_words, word, 0 );
2325 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2326 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2328 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2336 NEXT_OFF(convert) = (U16)(tail - convert);
2337 DEBUG_r(optimize= n);
2343 if ( trie->maxlen ) {
2344 NEXT_OFF( convert ) = (U16)(tail - convert);
2345 ARG_SET( convert, data_slot );
2346 /* Store the offset to the first unabsorbed branch in
2347 jump[0], which is otherwise unused by the jump logic.
2348 We use this when dumping a trie and during optimisation. */
2350 trie->jump[0] = (U16)(nextbranch - convert);
2352 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2353 * and there is a bitmap
2354 * and the first "jump target" node we found leaves enough room
2355 * then convert the TRIE node into a TRIEC node, with the bitmap
2356 * embedded inline in the opcode - this is hypothetically faster.
2358 if ( !trie->states[trie->startstate].wordnum
2360 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2362 OP( convert ) = TRIEC;
2363 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2364 PerlMemShared_free(trie->bitmap);
2367 OP( convert ) = TRIE;
2369 /* store the type in the flags */
2370 convert->flags = nodetype;
2374 + regarglen[ OP( convert ) ];
2376 /* XXX We really should free up the resource in trie now,
2377 as we won't use them - (which resources?) dmq */
2379 /* needed for dumping*/
2380 DEBUG_r(if (optimize) {
2381 regnode *opt = convert;
2383 while ( ++opt < optimize) {
2384 Set_Node_Offset_Length(opt,0,0);
2387 Try to clean up some of the debris left after the
2390 while( optimize < jumper ) {
2391 mjd_nodelen += Node_Length((optimize));
2392 OP( optimize ) = OPTIMIZED;
2393 Set_Node_Offset_Length(optimize,0,0);
2396 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2398 } /* end node insert */
2400 /* Finish populating the prev field of the wordinfo array. Walk back
2401 * from each accept state until we find another accept state, and if
2402 * so, point the first word's .prev field at the second word. If the
2403 * second already has a .prev field set, stop now. This will be the
2404 * case either if we've already processed that word's accept state,
2405 * or that state had multiple words, and the overspill words were
2406 * already linked up earlier.
2413 for (word=1; word <= trie->wordcount; word++) {
2415 if (trie->wordinfo[word].prev)
2417 state = trie->wordinfo[word].accept;
2419 state = prev_states[state];
2422 prev = trie->states[state].wordnum;
2426 trie->wordinfo[word].prev = prev;
2428 Safefree(prev_states);
2432 /* and now dump out the compressed format */
2433 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2435 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2437 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2438 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2440 SvREFCNT_dec(revcharmap);
2444 : trie->startstate>1
2450 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2452 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2454 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2455 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2458 We find the fail state for each state in the trie, this state is the longest proper
2459 suffix of the current state's 'word' that is also a proper prefix of another word in our
2460 trie. State 1 represents the word '' and is thus the default fail state. This allows
2461 the DFA not to have to restart after its tried and failed a word at a given point, it
2462 simply continues as though it had been matching the other word in the first place.
2464 'abcdgu'=~/abcdefg|cdgu/
2465 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2466 fail, which would bring us to the state representing 'd' in the second word where we would
2467 try 'g' and succeed, proceeding to match 'cdgu'.
2469 /* add a fail transition */
2470 const U32 trie_offset = ARG(source);
2471 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2473 const U32 ucharcount = trie->uniquecharcount;
2474 const U32 numstates = trie->statecount;
2475 const U32 ubound = trie->lasttrans + ucharcount;
2479 U32 base = trie->states[ 1 ].trans.base;
2482 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2483 GET_RE_DEBUG_FLAGS_DECL;
2485 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2487 PERL_UNUSED_ARG(depth);
2491 ARG_SET( stclass, data_slot );
2492 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2493 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2494 aho->trie=trie_offset;
2495 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2496 Copy( trie->states, aho->states, numstates, reg_trie_state );
2497 Newxz( q, numstates, U32);
2498 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2501 /* initialize fail[0..1] to be 1 so that we always have
2502 a valid final fail state */
2503 fail[ 0 ] = fail[ 1 ] = 1;
2505 for ( charid = 0; charid < ucharcount ; charid++ ) {
2506 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2508 q[ q_write ] = newstate;
2509 /* set to point at the root */
2510 fail[ q[ q_write++ ] ]=1;
2513 while ( q_read < q_write) {
2514 const U32 cur = q[ q_read++ % numstates ];
2515 base = trie->states[ cur ].trans.base;
2517 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2518 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2520 U32 fail_state = cur;
2523 fail_state = fail[ fail_state ];
2524 fail_base = aho->states[ fail_state ].trans.base;
2525 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2527 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2528 fail[ ch_state ] = fail_state;
2529 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2531 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2533 q[ q_write++ % numstates] = ch_state;
2537 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2538 when we fail in state 1, this allows us to use the
2539 charclass scan to find a valid start char. This is based on the principle
2540 that theres a good chance the string being searched contains lots of stuff
2541 that cant be a start char.
2543 fail[ 0 ] = fail[ 1 ] = 0;
2544 DEBUG_TRIE_COMPILE_r({
2545 PerlIO_printf(Perl_debug_log,
2546 "%*sStclass Failtable (%"UVuf" states): 0",
2547 (int)(depth * 2), "", (UV)numstates
2549 for( q_read=1; q_read<numstates; q_read++ ) {
2550 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2552 PerlIO_printf(Perl_debug_log, "\n");
2555 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2560 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2561 * These need to be revisited when a newer toolchain becomes available.
2563 #if defined(__sparc64__) && defined(__GNUC__)
2564 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2565 # undef SPARC64_GCC_WORKAROUND
2566 # define SPARC64_GCC_WORKAROUND 1
2570 #define DEBUG_PEEP(str,scan,depth) \
2571 DEBUG_OPTIMISE_r({if (scan){ \
2572 SV * const mysv=sv_newmortal(); \
2573 regnode *Next = regnext(scan); \
2574 regprop(RExC_rx, mysv, scan); \
2575 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2576 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2577 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2581 /* The below joins as many adjacent EXACTish nodes as possible into a single
2582 * one, and looks for problematic sequences of characters whose folds vs.
2583 * non-folds have sufficiently different lengths, that the optimizer would be
2584 * fooled into rejecting legitimate matches of them, and the trie construction
2585 * code needs to handle specially. The joining is only done if:
2586 * 1) there is room in the current conglomerated node to entirely contain the
2588 * 2) they are the exact same node type
2590 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2591 * these get optimized out
2593 * If there are problematic code sequences, *min_subtract is set to the delta
2594 * that the minimum size of the node can be less than its actual size. And,
2595 * the node type of the result is changed to reflect that it contains these
2598 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2599 * and contains LATIN SMALL LETTER SHARP S
2601 * This is as good a place as any to discuss the design of handling these
2602 * problematic sequences. It's been wrong in Perl for a very long time. There
2603 * are three code points currently in Unicode whose folded lengths differ so
2604 * much from the un-folded lengths that it causes problems for the optimizer
2605 * and trie construction. Why only these are problematic, and not others where
2606 * lengths also differ is something I (khw) do not understand. New versions of
2607 * Unicode might add more such code points. Hopefully the logic in
2608 * fold_grind.t that figures out what to test (in part by verifying that each
2609 * size-combination gets tested) will catch any that do come along, so they can
2610 * be added to the special handling below. The chances of new ones are
2611 * actually rather small, as most, if not all, of the world's scripts that have
2612 * casefolding have already been encoded by Unicode. Also, a number of
2613 * Unicode's decisions were made to allow compatibility with pre-existing
2614 * standards, and almost all of those have already been dealt with. These
2615 * would otherwise be the most likely candidates for generating further tricky
2616 * sequences. In other words, Unicode by itself is unlikely to add new ones
2617 * unless it is for compatibility with pre-existing standards, and there aren't
2618 * many of those left.
2620 * The previous designs for dealing with these involved assigning a special
2621 * node for them. This approach doesn't work, as evidenced by this example:
2622 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2623 * Both these fold to "sss", but if the pattern is parsed to create a node
2624 * that would match just the \xDF, it won't be able to handle the case where a
2625 * successful match would have to cross the node's boundary. The new approach
2626 * that hopefully generally solves the problem generates an EXACTFU_SS node
2629 * There are a number of components to the approach (a lot of work for just
2630 * three code points!):
2631 * 1) This routine examines each EXACTFish node that could contain the
2632 * problematic sequences. It returns in *min_subtract how much to
2633 * subtract from the the actual length of the string to get a real minimum
2634 * for one that could match it. This number is usually 0 except for the
2635 * problematic sequences. This delta is used by the caller to adjust the
2636 * min length of the match, and the delta between min and max, so that the
2637 * optimizer doesn't reject these possibilities based on size constraints.
2638 * 2) These sequences require special handling by the trie code, so this code
2639 * changes the joined node type to special ops: EXACTFU_TRICKYFOLD and
2641 * 3) This is sufficient for the two Greek sequences (described below), but
2642 * the one involving the Sharp s (\xDF) needs more. The node type
2643 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2644 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2645 * case where there is a possible fold length change. That means that a
2646 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2647 * itself with length changes, and so can be processed faster. regexec.c
2648 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2649 * is pre-folded by regcomp.c. This saves effort in regex matching.
2650 * However, the pre-folding isn't done for non-UTF8 patterns because the
2651 * fold of the MICRO SIGN requires UTF-8. Also what EXACTF and EXACTFL
2652 * nodes fold to isn't known until runtime. The fold possibilities for
2653 * the non-UTF8 patterns are quite simple, except for the sharp s. All
2654 * the ones that don't involve a UTF-8 target string are members of a
2655 * fold-pair, and arrays are set up for all of them so that the other
2656 * member of the pair can be found quickly. Code elsewhere in this file
2657 * makes sure that in EXACTFU nodes, the sharp s gets folded to 'ss', even
2658 * if the pattern isn't UTF-8. This avoids the issues described in the
2660 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2661 * 'ss' or not is not knowable at compile time. It will match iff the
2662 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2663 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2664 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2665 * described in item 3). An assumption that the optimizer part of
2666 * regexec.c (probably unwittingly) makes is that a character in the
2667 * pattern corresponds to at most a single character in the target string.
2668 * (And I do mean character, and not byte here, unlike other parts of the
2669 * documentation that have never been updated to account for multibyte
2670 * Unicode.) This assumption is wrong only in this case, as all other
2671 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2672 * virtue of having this file pre-fold UTF-8 patterns. I'm
2673 * reluctant to try to change this assumption, so instead the code punts.
2674 * This routine examines EXACTF nodes for the sharp s, and returns a
2675 * boolean indicating whether or not the node is an EXACTF node that
2676 * contains a sharp s. When it is true, the caller sets a flag that later
2677 * causes the optimizer in this file to not set values for the floating
2678 * and fixed string lengths, and thus avoids the optimizer code in
2679 * regexec.c that makes the invalid assumption. Thus, there is no
2680 * optimization based on string lengths for EXACTF nodes that contain the
2681 * sharp s. This only happens for /id rules (which means the pattern
2685 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2686 if (PL_regkind[OP(scan)] == EXACT) \
2687 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2690 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) {
2691 /* Merge several consecutive EXACTish nodes into one. */
2692 regnode *n = regnext(scan);
2694 regnode *next = scan + NODE_SZ_STR(scan);
2698 regnode *stop = scan;
2699 GET_RE_DEBUG_FLAGS_DECL;
2701 PERL_UNUSED_ARG(depth);
2704 PERL_ARGS_ASSERT_JOIN_EXACT;
2705 #ifndef EXPERIMENTAL_INPLACESCAN
2706 PERL_UNUSED_ARG(flags);
2707 PERL_UNUSED_ARG(val);
2709 DEBUG_PEEP("join",scan,depth);
2711 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2712 * EXACT ones that are mergeable to the current one. */
2714 && (PL_regkind[OP(n)] == NOTHING
2715 || (stringok && OP(n) == OP(scan)))
2717 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2720 if (OP(n) == TAIL || n > next)
2722 if (PL_regkind[OP(n)] == NOTHING) {
2723 DEBUG_PEEP("skip:",n,depth);
2724 NEXT_OFF(scan) += NEXT_OFF(n);
2725 next = n + NODE_STEP_REGNODE;
2732 else if (stringok) {
2733 const unsigned int oldl = STR_LEN(scan);
2734 regnode * const nnext = regnext(n);
2736 /* XXX I (khw) kind of doubt that this works on platforms where
2737 * U8_MAX is above 255 because of lots of other assumptions */
2738 if (oldl + STR_LEN(n) > U8_MAX)
2741 DEBUG_PEEP("merg",n,depth);
2744 NEXT_OFF(scan) += NEXT_OFF(n);
2745 STR_LEN(scan) += STR_LEN(n);
2746 next = n + NODE_SZ_STR(n);
2747 /* Now we can overwrite *n : */
2748 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2756 #ifdef EXPERIMENTAL_INPLACESCAN
2757 if (flags && !NEXT_OFF(n)) {
2758 DEBUG_PEEP("atch", val, depth);
2759 if (reg_off_by_arg[OP(n)]) {
2760 ARG_SET(n, val - n);
2763 NEXT_OFF(n) = val - n;
2771 *has_exactf_sharp_s = FALSE;
2773 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2774 * can now analyze for sequences of problematic code points. (Prior to
2775 * this final joining, sequences could have been split over boundaries, and
2776 * hence missed). The sequences only happen in folding, hence for any
2777 * non-EXACT EXACTish node */
2778 if (OP(scan) != EXACT) {
2780 U8 * s0 = (U8*) STRING(scan);
2781 U8 * const s_end = s0 + STR_LEN(scan);
2783 /* The below is perhaps overboard, but this allows us to save a test
2784 * each time through the loop at the expense of a mask. This is
2785 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2786 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2787 * This uses an exclusive 'or' to find that bit and then inverts it to
2788 * form a mask, with just a single 0, in the bit position where 'S' and
2790 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2791 const U8 s_masked = 's' & S_or_s_mask;
2793 /* One pass is made over the node's string looking for all the
2794 * possibilities. to avoid some tests in the loop, there are two main
2795 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2799 /* There are two problematic Greek code points in Unicode
2802 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2803 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2809 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2810 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2812 * This means that in case-insensitive matching (or "loose
2813 * matching", as Unicode calls it), an EXACTF of length six (the
2814 * UTF-8 encoded byte length of the above casefolded versions) can
2815 * match a target string of length two (the byte length of UTF-8
2816 * encoded U+0390 or U+03B0). This would rather mess up the
2817 * minimum length computation. (there are other code points that
2818 * also fold to these two sequences, but the delta is smaller)
2820 * If these sequences are found, the minimum length is decreased by
2821 * four (six minus two).
2823 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2824 * LETTER SHARP S. We decrease the min length by 1 for each
2825 * occurrence of 'ss' found */
2827 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2828 # define U390_first_byte 0xb4
2829 const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
2830 # define U3B0_first_byte 0xb5
2831 const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
2833 # define U390_first_byte 0xce
2834 const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
2835 # define U3B0_first_byte 0xcf
2836 const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
2838 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2839 yields a net of 0 */
2840 /* Examine the string for one of the problematic sequences */
2842 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2843 * sequence we are looking for is 2 */
2847 /* Look for the first byte in each problematic sequence */
2849 /* We don't have to worry about other things that fold to
2850 * 's' (such as the long s, U+017F), as all above-latin1
2851 * code points have been pre-folded */
2855 /* Current character is an 's' or 'S'. If next one is
2856 * as well, we have the dreaded sequence */
2857 if (((*(s+1) & S_or_s_mask) == s_masked)
2858 /* These two node types don't have special handling
2860 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2863 OP(scan) = EXACTFU_SS;
2864 s++; /* No need to look at this character again */
2868 case U390_first_byte:
2869 if (s_end - s >= len
2871 /* The 1's are because are skipping comparing the
2873 && memEQ(s + 1, U390_tail, len - 1))
2875 goto greek_sequence;
2879 case U3B0_first_byte:
2880 if (! (s_end - s >= len
2881 && memEQ(s + 1, U3B0_tail, len - 1)))
2888 /* This requires special handling by trie's, so change
2889 * the node type to indicate this. If EXACTFA and
2890 * EXACTFL were ever to be handled by trie's, this
2891 * would have to be changed. If this node has already
2892 * been changed to EXACTFU_SS in this loop, leave it as
2893 * is. (I (khw) think it doesn't matter in regexec.c
2894 * for UTF patterns, but no need to change it */
2895 if (OP(scan) == EXACTFU) {
2896 OP(scan) = EXACTFU_TRICKYFOLD;
2898 s += 6; /* We already know what this sequence is. Skip
2904 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2906 /* Here, the pattern is not UTF-8. We need to look only for the
2907 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2908 * in the final position. Otherwise we can stop looking 1 byte
2909 * earlier because have to find both the first and second 's' */
2910 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2912 for (s = s0; s < upper; s++) {
2917 && ((*(s+1) & S_or_s_mask) == s_masked))
2921 /* EXACTF nodes need to know that the minimum
2922 * length changed so that a sharp s in the string
2923 * can match this ss in the pattern, but they
2924 * remain EXACTF nodes, as they won't match this
2925 * unless the target string is is UTF-8, which we
2926 * don't know until runtime */
2927 if (OP(scan) != EXACTF) {
2928 OP(scan) = EXACTFU_SS;
2933 case LATIN_SMALL_LETTER_SHARP_S:
2934 if (OP(scan) == EXACTF) {
2935 *has_exactf_sharp_s = TRUE;
2944 /* Allow dumping but overwriting the collection of skipped
2945 * ops and/or strings with fake optimized ops */
2946 n = scan + NODE_SZ_STR(scan);
2954 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2958 /* REx optimizer. Converts nodes into quicker variants "in place".
2959 Finds fixed substrings. */
2961 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2962 to the position after last scanned or to NULL. */
2964 #define INIT_AND_WITHP \
2965 assert(!and_withp); \
2966 Newx(and_withp,1,struct regnode_charclass_class); \
2967 SAVEFREEPV(and_withp)
2969 /* this is a chain of data about sub patterns we are processing that
2970 need to be handled separately/specially in study_chunk. Its so
2971 we can simulate recursion without losing state. */
2973 typedef struct scan_frame {
2974 regnode *last; /* last node to process in this frame */
2975 regnode *next; /* next node to process when last is reached */
2976 struct scan_frame *prev; /*previous frame*/
2977 I32 stop; /* what stopparen do we use */
2981 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2983 #define CASE_SYNST_FNC(nAmE) \
2985 if (flags & SCF_DO_STCLASS_AND) { \
2986 for (value = 0; value < 256; value++) \
2987 if (!is_ ## nAmE ## _cp(value)) \
2988 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2991 for (value = 0; value < 256; value++) \
2992 if (is_ ## nAmE ## _cp(value)) \
2993 ANYOF_BITMAP_SET(data->start_class, value); \
2997 if (flags & SCF_DO_STCLASS_AND) { \
2998 for (value = 0; value < 256; value++) \
2999 if (is_ ## nAmE ## _cp(value)) \
3000 ANYOF_BITMAP_CLEAR(data->start_class, value); \
3003 for (value = 0; value < 256; value++) \
3004 if (!is_ ## nAmE ## _cp(value)) \
3005 ANYOF_BITMAP_SET(data->start_class, value); \
3012 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3013 I32 *minlenp, I32 *deltap,
3018 struct regnode_charclass_class *and_withp,
3019 U32 flags, U32 depth)
3020 /* scanp: Start here (read-write). */
3021 /* deltap: Write maxlen-minlen here. */
3022 /* last: Stop before this one. */
3023 /* data: string data about the pattern */
3024 /* stopparen: treat close N as END */
3025 /* recursed: which subroutines have we recursed into */
3026 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3029 I32 min = 0, pars = 0, code;
3030 regnode *scan = *scanp, *next;
3032 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3033 int is_inf_internal = 0; /* The studied chunk is infinite */
3034 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3035 scan_data_t data_fake;
3036 SV *re_trie_maxbuff = NULL;
3037 regnode *first_non_open = scan;
3038 I32 stopmin = I32_MAX;
3039 scan_frame *frame = NULL;
3040 GET_RE_DEBUG_FLAGS_DECL;
3042 PERL_ARGS_ASSERT_STUDY_CHUNK;
3045 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3049 while (first_non_open && OP(first_non_open) == OPEN)
3050 first_non_open=regnext(first_non_open);
3055 while ( scan && OP(scan) != END && scan < last ){
3056 UV min_subtract = 0; /* How much to subtract from the minimum node
3057 length to get a real minimum (because the
3058 folded version may be shorter) */
3059 bool has_exactf_sharp_s = FALSE;
3060 /* Peephole optimizer: */
3061 DEBUG_STUDYDATA("Peep:", data,depth);
3062 DEBUG_PEEP("Peep",scan,depth);
3064 /* Its not clear to khw or hv why this is done here, and not in the
3065 * clauses that deal with EXACT nodes. khw's guess is that it's
3066 * because of a previous design */
3067 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3069 /* Follow the next-chain of the current node and optimize
3070 away all the NOTHINGs from it. */
3071 if (OP(scan) != CURLYX) {
3072 const int max = (reg_off_by_arg[OP(scan)]
3074 /* I32 may be smaller than U16 on CRAYs! */
3075 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3076 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3080 /* Skip NOTHING and LONGJMP. */
3081 while ((n = regnext(n))
3082 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3083 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3084 && off + noff < max)
3086 if (reg_off_by_arg[OP(scan)])
3089 NEXT_OFF(scan) = off;
3094 /* The principal pseudo-switch. Cannot be a switch, since we
3095 look into several different things. */
3096 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3097 || OP(scan) == IFTHEN) {
3098 next = regnext(scan);
3100 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3102 if (OP(next) == code || code == IFTHEN) {
3103 /* NOTE - There is similar code to this block below for handling
3104 TRIE nodes on a re-study. If you change stuff here check there
3106 I32 max1 = 0, min1 = I32_MAX, num = 0;
3107 struct regnode_charclass_class accum;
3108 regnode * const startbranch=scan;
3110 if (flags & SCF_DO_SUBSTR)
3111 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3112 if (flags & SCF_DO_STCLASS)
3113 cl_init_zero(pRExC_state, &accum);
3115 while (OP(scan) == code) {
3116 I32 deltanext, minnext, f = 0, fake;
3117 struct regnode_charclass_class this_class;
3120 data_fake.flags = 0;
3122 data_fake.whilem_c = data->whilem_c;
3123 data_fake.last_closep = data->last_closep;
3126 data_fake.last_closep = &fake;
3128 data_fake.pos_delta = delta;
3129 next = regnext(scan);
3130 scan = NEXTOPER(scan);
3132 scan = NEXTOPER(scan);
3133 if (flags & SCF_DO_STCLASS) {
3134 cl_init(pRExC_state, &this_class);
3135 data_fake.start_class = &this_class;
3136 f = SCF_DO_STCLASS_AND;
3138 if (flags & SCF_WHILEM_VISITED_POS)
3139 f |= SCF_WHILEM_VISITED_POS;
3141 /* we suppose the run is continuous, last=next...*/
3142 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3144 stopparen, recursed, NULL, f,depth+1);
3147 if (max1 < minnext + deltanext)
3148 max1 = minnext + deltanext;
3149 if (deltanext == I32_MAX)
3150 is_inf = is_inf_internal = 1;
3152 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3154 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3155 if ( stopmin > minnext)
3156 stopmin = min + min1;
3157 flags &= ~SCF_DO_SUBSTR;
3159 data->flags |= SCF_SEEN_ACCEPT;
3162 if (data_fake.flags & SF_HAS_EVAL)
3163 data->flags |= SF_HAS_EVAL;
3164 data->whilem_c = data_fake.whilem_c;
3166 if (flags & SCF_DO_STCLASS)
3167 cl_or(pRExC_state, &accum, &this_class);
3169 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3171 if (flags & SCF_DO_SUBSTR) {
3172 data->pos_min += min1;
3173 data->pos_delta += max1 - min1;
3174 if (max1 != min1 || is_inf)
3175 data->longest = &(data->longest_float);
3178 delta += max1 - min1;
3179 if (flags & SCF_DO_STCLASS_OR) {
3180 cl_or(pRExC_state, data->start_class, &accum);
3182 cl_and(data->start_class, and_withp);
3183 flags &= ~SCF_DO_STCLASS;
3186 else if (flags & SCF_DO_STCLASS_AND) {
3188 cl_and(data->start_class, &accum);
3189 flags &= ~SCF_DO_STCLASS;
3192 /* Switch to OR mode: cache the old value of
3193 * data->start_class */
3195 StructCopy(data->start_class, and_withp,
3196 struct regnode_charclass_class);
3197 flags &= ~SCF_DO_STCLASS_AND;
3198 StructCopy(&accum, data->start_class,
3199 struct regnode_charclass_class);
3200 flags |= SCF_DO_STCLASS_OR;
3201 data->start_class->flags |= ANYOF_EOS;
3205 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3208 Assuming this was/is a branch we are dealing with: 'scan' now
3209 points at the item that follows the branch sequence, whatever
3210 it is. We now start at the beginning of the sequence and look
3217 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3219 If we can find such a subsequence we need to turn the first
3220 element into a trie and then add the subsequent branch exact
3221 strings to the trie.
3225 1. patterns where the whole set of branches can be converted.
3227 2. patterns where only a subset can be converted.
3229 In case 1 we can replace the whole set with a single regop
3230 for the trie. In case 2 we need to keep the start and end
3233 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3234 becomes BRANCH TRIE; BRANCH X;
3236 There is an additional case, that being where there is a
3237 common prefix, which gets split out into an EXACT like node
3238 preceding the TRIE node.
3240 If x(1..n)==tail then we can do a simple trie, if not we make
3241 a "jump" trie, such that when we match the appropriate word
3242 we "jump" to the appropriate tail node. Essentially we turn
3243 a nested if into a case structure of sorts.
3248 if (!re_trie_maxbuff) {
3249 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3250 if (!SvIOK(re_trie_maxbuff))
3251 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3253 if ( SvIV(re_trie_maxbuff)>=0 ) {
3255 regnode *first = (regnode *)NULL;
3256 regnode *last = (regnode *)NULL;
3257 regnode *tail = scan;
3262 SV * const mysv = sv_newmortal(); /* for dumping */
3264 /* var tail is used because there may be a TAIL
3265 regop in the way. Ie, the exacts will point to the
3266 thing following the TAIL, but the last branch will
3267 point at the TAIL. So we advance tail. If we
3268 have nested (?:) we may have to move through several
3272 while ( OP( tail ) == TAIL ) {
3273 /* this is the TAIL generated by (?:) */
3274 tail = regnext( tail );
3278 DEBUG_TRIE_COMPILE_r({
3279 regprop(RExC_rx, mysv, tail );
3280 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3281 (int)depth * 2 + 2, "",
3282 "Looking for TRIE'able sequences. Tail node is: ",
3283 SvPV_nolen_const( mysv )
3289 Step through the branches
3290 cur represents each branch,
3291 noper is the first thing to be matched as part of that branch
3292 noper_next is the regnext() of that node.
3294 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3295 via a "jump trie" but we also support building with NOJUMPTRIE,
3296 which restricts the trie logic to structures like /FOO|BAR/.
3298 If noper is a trieable nodetype then the branch is a possible optimization
3299 target. If we are building under NOJUMPTRIE then we require that noper_next
3300 is the same as scan (our current position in the regex program).
3302 Once we have two or more consecutive such branches we can create a
3303 trie of the EXACT's contents and stitch it in place into the program.
3305 If the sequence represents all of the branches in the alternation we
3306 replace the entire thing with a single TRIE node.
3308 Otherwise when it is a subsequence we need to stitch it in place and
3309 replace only the relevant branches. This means the first branch has
3310 to remain as it is used by the alternation logic, and its next pointer,
3311 and needs to be repointed at the item on the branch chain following
3312 the last branch we have optimized away.
3314 This could be either a BRANCH, in which case the subsequence is internal,
3315 or it could be the item following the branch sequence in which case the
3316 subsequence is at the end (which does not necessarily mean the first node
3317 is the start of the alternation).
3319 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3322 ----------------+-----------
3326 EXACTFU_SS | EXACTFU
3327 EXACTFU_TRICKYFOLD | EXACTFU
3332 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3333 ( EXACT == (X) ) ? EXACT : \
3334 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3337 /* dont use tail as the end marker for this traverse */
3338 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3339 regnode * const noper = NEXTOPER( cur );
3340 U8 noper_type = OP( noper );
3341 U8 noper_trietype = TRIE_TYPE( noper_type );
3342 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3343 regnode * const noper_next = regnext( noper );
3344 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3345 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3348 DEBUG_TRIE_COMPILE_r({
3349 regprop(RExC_rx, mysv, cur);
3350 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3351 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3353 regprop(RExC_rx, mysv, noper);
3354 PerlIO_printf( Perl_debug_log, " -> %s",
3355 SvPV_nolen_const(mysv));
3358 regprop(RExC_rx, mysv, noper_next );
3359 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3360 SvPV_nolen_const(mysv));
3362 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3363 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3364 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3368 /* Is noper a trieable nodetype that can be merged with the
3369 * current trie (if there is one)? */
3373 ( noper_trietype == NOTHING)
3374 || ( trietype == NOTHING )
3375 || ( trietype == noper_trietype )
3378 && noper_next == tail
3382 /* Handle mergable triable node
3383 * Either we are the first node in a new trieable sequence,
3384 * in which case we do some bookkeeping, otherwise we update
3385 * the end pointer. */
3388 if ( noper_trietype == NOTHING ) {
3389 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3390 regnode * const noper_next = regnext( noper );
3391 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3392 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3395 if ( noper_next_trietype ) {
3396 trietype = noper_next_trietype;
3397 } else if (noper_next_type) {
3398 /* a NOTHING regop is 1 regop wide. We need at least two
3399 * for a trie so we can't merge this in */
3403 trietype = noper_trietype;
3406 if ( trietype == NOTHING )
3407 trietype = noper_trietype;
3412 } /* end handle mergable triable node */
3414 /* handle unmergable node -
3415 * noper may either be a triable node which can not be tried
3416 * together with the current trie, or a non triable node */
3418 /* If last is set and trietype is not NOTHING then we have found
3419 * at least two triable branch sequences in a row of a similar
3420 * trietype so we can turn them into a trie. If/when we
3421 * allow NOTHING to start a trie sequence this condition will be
3422 * required, and it isn't expensive so we leave it in for now. */
3423 if ( trietype != NOTHING )
3424 make_trie( pRExC_state,
3425 startbranch, first, cur, tail, count,
3426 trietype, depth+1 );
3427 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3431 && noper_next == tail
3434 /* noper is triable, so we can start a new trie sequence */
3437 trietype = noper_trietype;
3439 /* if we already saw a first but the current node is not triable then we have
3440 * to reset the first information. */
3445 } /* end handle unmergable node */
3446 } /* loop over branches */
3447 DEBUG_TRIE_COMPILE_r({
3448 regprop(RExC_rx, mysv, cur);
3449 PerlIO_printf( Perl_debug_log,
3450 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3451 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3455 if ( trietype != NOTHING ) {
3456 /* the last branch of the sequence was part of a trie,
3457 * so we have to construct it here outside of the loop
3459 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3460 #ifdef TRIE_STUDY_OPT
3461 if ( ((made == MADE_EXACT_TRIE &&
3462 startbranch == first)
3463 || ( first_non_open == first )) &&
3465 flags |= SCF_TRIE_RESTUDY;
3466 if ( startbranch == first
3469 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3474 /* at this point we know whatever we have is a NOTHING sequence/branch
3475 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3477 if ( startbranch == first ) {
3479 /* the entire thing is a NOTHING sequence, something like this:
3480 * (?:|) So we can turn it into a plain NOTHING op. */
3481 DEBUG_TRIE_COMPILE_r({
3482 regprop(RExC_rx, mysv, cur);
3483 PerlIO_printf( Perl_debug_log,
3484 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3485 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3488 OP(startbranch)= NOTHING;
3489 NEXT_OFF(startbranch)= tail - startbranch;
3490 for ( opt= startbranch + 1; opt < tail ; opt++ )
3494 } /* end if ( last) */
3495 } /* TRIE_MAXBUF is non zero */
3500 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3501 scan = NEXTOPER(NEXTOPER(scan));
3502 } else /* single branch is optimized. */
3503 scan = NEXTOPER(scan);
3505 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3506 scan_frame *newframe = NULL;
3511 if (OP(scan) != SUSPEND) {
3512 /* set the pointer */
3513 if (OP(scan) == GOSUB) {
3515 RExC_recurse[ARG2L(scan)] = scan;
3516 start = RExC_open_parens[paren-1];
3517 end = RExC_close_parens[paren-1];
3520 start = RExC_rxi->program + 1;
3524 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3525 SAVEFREEPV(recursed);
3527 if (!PAREN_TEST(recursed,paren+1)) {
3528 PAREN_SET(recursed,paren+1);
3529 Newx(newframe,1,scan_frame);
3531 if (flags & SCF_DO_SUBSTR) {
3532 SCAN_COMMIT(pRExC_state,data,minlenp);
3533 data->longest = &(data->longest_float);
3535 is_inf = is_inf_internal = 1;
3536 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3537 cl_anything(pRExC_state, data->start_class);
3538 flags &= ~SCF_DO_STCLASS;
3541 Newx(newframe,1,scan_frame);
3544 end = regnext(scan);
3549 SAVEFREEPV(newframe);
3550 newframe->next = regnext(scan);
3551 newframe->last = last;
3552 newframe->stop = stopparen;
3553 newframe->prev = frame;
3563 else if (OP(scan) == EXACT) {
3564 I32 l = STR_LEN(scan);
3567 const U8 * const s = (U8*)STRING(scan);
3568 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3569 l = utf8_length(s, s + l);
3571 uc = *((U8*)STRING(scan));
3574 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3575 /* The code below prefers earlier match for fixed
3576 offset, later match for variable offset. */
3577 if (data->last_end == -1) { /* Update the start info. */
3578 data->last_start_min = data->pos_min;
3579 data->last_start_max = is_inf
3580 ? I32_MAX : data->pos_min + data->pos_delta;
3582 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3584 SvUTF8_on(data->last_found);
3586 SV * const sv = data->last_found;
3587 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3588 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3589 if (mg && mg->mg_len >= 0)
3590 mg->mg_len += utf8_length((U8*)STRING(scan),
3591 (U8*)STRING(scan)+STR_LEN(scan));
3593 data->last_end = data->pos_min + l;
3594 data->pos_min += l; /* As in the first entry. */
3595 data->flags &= ~SF_BEFORE_EOL;
3597 if (flags & SCF_DO_STCLASS_AND) {
3598 /* Check whether it is compatible with what we know already! */
3602 /* If compatible, we or it in below. It is compatible if is
3603 * in the bitmp and either 1) its bit or its fold is set, or 2)
3604 * it's for a locale. Even if there isn't unicode semantics
3605 * here, at runtime there may be because of matching against a
3606 * utf8 string, so accept a possible false positive for
3607 * latin1-range folds */
3609 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3610 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3611 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
3612 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3617 ANYOF_CLASS_ZERO(data->start_class);
3618 ANYOF_BITMAP_ZERO(data->start_class);
3620 ANYOF_BITMAP_SET(data->start_class, uc);
3621 else if (uc >= 0x100) {
3624 /* Some Unicode code points fold to the Latin1 range; as
3625 * XXX temporary code, instead of figuring out if this is
3626 * one, just assume it is and set all the start class bits
3627 * that could be some such above 255 code point's fold
3628 * which will generate fals positives. As the code
3629 * elsewhere that does compute the fold settles down, it
3630 * can be extracted out and re-used here */
3631 for (i = 0; i < 256; i++){
3632 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3633 ANYOF_BITMAP_SET(data->start_class, i);
3637 data->start_class->flags &= ~ANYOF_EOS;
3639 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3641 else if (flags & SCF_DO_STCLASS_OR) {
3642 /* false positive possible if the class is case-folded */
3644 ANYOF_BITMAP_SET(data->start_class, uc);
3646 data->start_class->flags |= ANYOF_UNICODE_ALL;
3647 data->start_class->flags &= ~ANYOF_EOS;
3648 cl_and(data->start_class, and_withp);
3650 flags &= ~SCF_DO_STCLASS;
3652 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3653 I32 l = STR_LEN(scan);
3654 UV uc = *((U8*)STRING(scan));
3656 /* Search for fixed substrings supports EXACT only. */
3657 if (flags & SCF_DO_SUBSTR) {
3659 SCAN_COMMIT(pRExC_state, data, minlenp);
3662 const U8 * const s = (U8 *)STRING(scan);
3663 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3664 l = utf8_length(s, s + l);
3666 else if (has_exactf_sharp_s) {
3667 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3669 min += l - min_subtract;
3673 delta += min_subtract;
3674 if (flags & SCF_DO_SUBSTR) {
3675 data->pos_min += l - min_subtract;
3676 if (data->pos_min < 0) {
3679 data->pos_delta += min_subtract;
3681 data->longest = &(data->longest_float);
3684 if (flags & SCF_DO_STCLASS_AND) {
3685 /* Check whether it is compatible with what we know already! */
3688 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3689 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3690 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3694 ANYOF_CLASS_ZERO(data->start_class);
3695 ANYOF_BITMAP_ZERO(data->start_class);
3697 ANYOF_BITMAP_SET(data->start_class, uc);
3698 data->start_class->flags &= ~ANYOF_EOS;
3699 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
3700 if (OP(scan) == EXACTFL) {
3701 /* XXX This set is probably no longer necessary, and
3702 * probably wrong as LOCALE now is on in the initial
3704 data->start_class->flags |= ANYOF_LOCALE;
3708 /* Also set the other member of the fold pair. In case
3709 * that unicode semantics is called for at runtime, use
3710 * the full latin1 fold. (Can't do this for locale,
3711 * because not known until runtime) */
3712 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3714 /* All other (EXACTFL handled above) folds except under
3715 * /iaa that include s, S, and sharp_s also may include
3717 if (OP(scan) != EXACTFA) {
3718 if (uc == 's' || uc == 'S') {
3719 ANYOF_BITMAP_SET(data->start_class,
3720 LATIN_SMALL_LETTER_SHARP_S);
3722 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3723 ANYOF_BITMAP_SET(data->start_class, 's');
3724 ANYOF_BITMAP_SET(data->start_class, 'S');
3729 else if (uc >= 0x100) {
3731 for (i = 0; i < 256; i++){
3732 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3733 ANYOF_BITMAP_SET(data->start_class, i);
3738 else if (flags & SCF_DO_STCLASS_OR) {
3739 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
3740 /* false positive possible if the class is case-folded.
3741 Assume that the locale settings are the same... */
3743 ANYOF_BITMAP_SET(data->start_class, uc);
3744 if (OP(scan) != EXACTFL) {
3746 /* And set the other member of the fold pair, but
3747 * can't do that in locale because not known until
3749 ANYOF_BITMAP_SET(data->start_class,
3750 PL_fold_latin1[uc]);
3752 /* All folds except under /iaa that include s, S,
3753 * and sharp_s also may include the others */
3754 if (OP(scan) != EXACTFA) {
3755 if (uc == 's' || uc == 'S') {
3756 ANYOF_BITMAP_SET(data->start_class,
3757 LATIN_SMALL_LETTER_SHARP_S);
3759 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3760 ANYOF_BITMAP_SET(data->start_class, 's');
3761 ANYOF_BITMAP_SET(data->start_class, 'S');
3766 data->start_class->flags &= ~ANYOF_EOS;
3768 cl_and(data->start_class, and_withp);
3770 flags &= ~SCF_DO_STCLASS;
3772 else if (REGNODE_VARIES(OP(scan))) {
3773 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3774 I32 f = flags, pos_before = 0;
3775 regnode * const oscan = scan;
3776 struct regnode_charclass_class this_class;
3777 struct regnode_charclass_class *oclass = NULL;
3778 I32 next_is_eval = 0;
3780 switch (PL_regkind[OP(scan)]) {
3781 case WHILEM: /* End of (?:...)* . */
3782 scan = NEXTOPER(scan);
3785 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3786 next = NEXTOPER(scan);
3787 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3789 maxcount = REG_INFTY;
3790 next = regnext(scan);
3791 scan = NEXTOPER(scan);
3795 if (flags & SCF_DO_SUBSTR)
3800 if (flags & SCF_DO_STCLASS) {
3802 maxcount = REG_INFTY;
3803 next = regnext(scan);
3804 scan = NEXTOPER(scan);
3807 is_inf = is_inf_internal = 1;
3808 scan = regnext(scan);
3809 if (flags & SCF_DO_SUBSTR) {
3810 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3811 data->longest = &(data->longest_float);
3813 goto optimize_curly_tail;
3815 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3816 && (scan->flags == stopparen))
3821 mincount = ARG1(scan);
3822 maxcount = ARG2(scan);
3824 next = regnext(scan);
3825 if (OP(scan) == CURLYX) {
3826 I32 lp = (data ? *(data->last_closep) : 0);
3827 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3829 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3830 next_is_eval = (OP(scan) == EVAL);
https://perl5.git.perl.org / perl5.git / blob