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"
91 #include "inline_invlist.c"
92 #include "unicode_constants.h"
100 #define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
101 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
102 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
109 # if defined(BUGGY_MSC6)
110 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
111 # pragma optimize("a",off)
112 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
113 # pragma optimize("w",on )
114 # endif /* BUGGY_MSC6 */
118 #define STATIC static
122 typedef struct RExC_state_t {
123 U32 flags; /* RXf_* are we folding, multilining? */
124 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
125 char *precomp; /* uncompiled string. */
126 REGEXP *rx_sv; /* The SV that is the regexp. */
127 regexp *rx; /* perl core regexp structure */
128 regexp_internal *rxi; /* internal data for regexp object pprivate field */
129 char *start; /* Start of input for compile */
130 char *end; /* End of input for compile */
131 char *parse; /* Input-scan pointer. */
132 I32 whilem_seen; /* number of WHILEM in this expr */
133 regnode *emit_start; /* Start of emitted-code area */
134 regnode *emit_bound; /* First regnode outside of the allocated space */
135 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
136 I32 naughty; /* How bad is this pattern? */
137 I32 sawback; /* Did we see \1, ...? */
139 I32 size; /* Code size. */
140 I32 npar; /* Capture buffer count, (OPEN). */
141 I32 cpar; /* Capture buffer count, (CLOSE). */
142 I32 nestroot; /* root parens we are in - used by accept */
145 regnode **open_parens; /* pointers to open parens */
146 regnode **close_parens; /* pointers to close parens */
147 regnode *opend; /* END node in program */
148 I32 utf8; /* whether the pattern is utf8 or not */
149 I32 orig_utf8; /* whether the pattern was originally in utf8 */
150 /* XXX use this for future optimisation of case
151 * where pattern must be upgraded to utf8. */
152 I32 uni_semantics; /* If a d charset modifier should use unicode
153 rules, even if the pattern is not in
155 HV *paren_names; /* Paren names */
157 regnode **recurse; /* Recurse regops */
158 I32 recurse_count; /* Number of recurse regops */
161 I32 override_recoding;
162 I32 in_multi_char_class;
163 struct reg_code_block *code_blocks; /* positions of literal (?{})
165 int num_code_blocks; /* size of code_blocks[] */
166 int code_index; /* next code_blocks[] slot */
168 char *starttry; /* -Dr: where regtry was called. */
169 #define RExC_starttry (pRExC_state->starttry)
171 SV *runtime_code_qr; /* qr with the runtime code blocks */
173 const char *lastparse;
175 AV *paren_name_list; /* idx -> name */
176 #define RExC_lastparse (pRExC_state->lastparse)
177 #define RExC_lastnum (pRExC_state->lastnum)
178 #define RExC_paren_name_list (pRExC_state->paren_name_list)
182 #define RExC_flags (pRExC_state->flags)
183 #define RExC_pm_flags (pRExC_state->pm_flags)
184 #define RExC_precomp (pRExC_state->precomp)
185 #define RExC_rx_sv (pRExC_state->rx_sv)
186 #define RExC_rx (pRExC_state->rx)
187 #define RExC_rxi (pRExC_state->rxi)
188 #define RExC_start (pRExC_state->start)
189 #define RExC_end (pRExC_state->end)
190 #define RExC_parse (pRExC_state->parse)
191 #define RExC_whilem_seen (pRExC_state->whilem_seen)
192 #ifdef RE_TRACK_PATTERN_OFFSETS
193 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
195 #define RExC_emit (pRExC_state->emit)
196 #define RExC_emit_start (pRExC_state->emit_start)
197 #define RExC_emit_bound (pRExC_state->emit_bound)
198 #define RExC_naughty (pRExC_state->naughty)
199 #define RExC_sawback (pRExC_state->sawback)
200 #define RExC_seen (pRExC_state->seen)
201 #define RExC_size (pRExC_state->size)
202 #define RExC_npar (pRExC_state->npar)
203 #define RExC_nestroot (pRExC_state->nestroot)
204 #define RExC_extralen (pRExC_state->extralen)
205 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
206 #define RExC_utf8 (pRExC_state->utf8)
207 #define RExC_uni_semantics (pRExC_state->uni_semantics)
208 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
209 #define RExC_open_parens (pRExC_state->open_parens)
210 #define RExC_close_parens (pRExC_state->close_parens)
211 #define RExC_opend (pRExC_state->opend)
212 #define RExC_paren_names (pRExC_state->paren_names)
213 #define RExC_recurse (pRExC_state->recurse)
214 #define RExC_recurse_count (pRExC_state->recurse_count)
215 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
216 #define RExC_contains_locale (pRExC_state->contains_locale)
217 #define RExC_override_recoding (pRExC_state->override_recoding)
218 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
221 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
222 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
223 ((*s) == '{' && regcurly(s)))
226 #undef SPSTART /* dratted cpp namespace... */
229 * Flags to be passed up and down.
231 #define WORST 0 /* Worst case. */
232 #define HASWIDTH 0x01 /* Known to match non-null strings. */
234 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
235 * character. (There needs to be a case: in the switch statement in regexec.c
236 * for any node marked SIMPLE.) Note that this is not the same thing as
239 #define SPSTART 0x04 /* Starts with * or + */
240 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
241 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
243 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
245 /* whether trie related optimizations are enabled */
246 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
247 #define TRIE_STUDY_OPT
248 #define FULL_TRIE_STUDY
254 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
255 #define PBITVAL(paren) (1 << ((paren) & 7))
256 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
257 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
258 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
260 /* If not already in utf8, do a longjmp back to the beginning */
261 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
262 #define REQUIRE_UTF8 STMT_START { \
263 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
266 /* About scan_data_t.
268 During optimisation we recurse through the regexp program performing
269 various inplace (keyhole style) optimisations. In addition study_chunk
270 and scan_commit populate this data structure with information about
271 what strings MUST appear in the pattern. We look for the longest
272 string that must appear at a fixed location, and we look for the
273 longest string that may appear at a floating location. So for instance
278 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
279 strings (because they follow a .* construct). study_chunk will identify
280 both FOO and BAR as being the longest fixed and floating strings respectively.
282 The strings can be composites, for instance
286 will result in a composite fixed substring 'foo'.
288 For each string some basic information is maintained:
290 - offset or min_offset
291 This is the position the string must appear at, or not before.
292 It also implicitly (when combined with minlenp) tells us how many
293 characters must match before the string we are searching for.
294 Likewise when combined with minlenp and the length of the string it
295 tells us how many characters must appear after the string we have
299 Only used for floating strings. This is the rightmost point that
300 the string can appear at. If set to I32 max it indicates that the
301 string can occur infinitely far to the right.
304 A pointer to the minimum number of characters of the pattern that the
305 string was found inside. This is important as in the case of positive
306 lookahead or positive lookbehind we can have multiple patterns
311 The minimum length of the pattern overall is 3, the minimum length
312 of the lookahead part is 3, but the minimum length of the part that
313 will actually match is 1. So 'FOO's minimum length is 3, but the
314 minimum length for the F is 1. This is important as the minimum length
315 is used to determine offsets in front of and behind the string being
316 looked for. Since strings can be composites this is the length of the
317 pattern at the time it was committed with a scan_commit. Note that
318 the length is calculated by study_chunk, so that the minimum lengths
319 are not known until the full pattern has been compiled, thus the
320 pointer to the value.
324 In the case of lookbehind the string being searched for can be
325 offset past the start point of the final matching string.
326 If this value was just blithely removed from the min_offset it would
327 invalidate some of the calculations for how many chars must match
328 before or after (as they are derived from min_offset and minlen and
329 the length of the string being searched for).
330 When the final pattern is compiled and the data is moved from the
331 scan_data_t structure into the regexp structure the information
332 about lookbehind is factored in, with the information that would
333 have been lost precalculated in the end_shift field for the
336 The fields pos_min and pos_delta are used to store the minimum offset
337 and the delta to the maximum offset at the current point in the pattern.
341 typedef struct scan_data_t {
342 /*I32 len_min; unused */
343 /*I32 len_delta; unused */
347 I32 last_end; /* min value, <0 unless valid. */
350 SV **longest; /* Either &l_fixed, or &l_float. */
351 SV *longest_fixed; /* longest fixed string found in pattern */
352 I32 offset_fixed; /* offset where it starts */
353 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
354 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
355 SV *longest_float; /* longest floating string found in pattern */
356 I32 offset_float_min; /* earliest point in string it can appear */
357 I32 offset_float_max; /* latest point in string it can appear */
358 I32 *minlen_float; /* pointer to the minlen relevant to the string */
359 I32 lookbehind_float; /* is the position of the string modified by LB */
363 struct regnode_charclass_class *start_class;
367 * Forward declarations for pregcomp()'s friends.
370 static const scan_data_t zero_scan_data =
371 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
373 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
374 #define SF_BEFORE_SEOL 0x0001
375 #define SF_BEFORE_MEOL 0x0002
376 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
377 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
380 # define SF_FIX_SHIFT_EOL (0+2)
381 # define SF_FL_SHIFT_EOL (0+4)
383 # define SF_FIX_SHIFT_EOL (+2)
384 # define SF_FL_SHIFT_EOL (+4)
387 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
388 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
390 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
391 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
392 #define SF_IS_INF 0x0040
393 #define SF_HAS_PAR 0x0080
394 #define SF_IN_PAR 0x0100
395 #define SF_HAS_EVAL 0x0200
396 #define SCF_DO_SUBSTR 0x0400
397 #define SCF_DO_STCLASS_AND 0x0800
398 #define SCF_DO_STCLASS_OR 0x1000
399 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
400 #define SCF_WHILEM_VISITED_POS 0x2000
402 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
403 #define SCF_SEEN_ACCEPT 0x8000
405 #define UTF cBOOL(RExC_utf8)
407 /* The enums for all these are ordered so things work out correctly */
408 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
409 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
410 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
411 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
412 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
413 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
414 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
416 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
418 #define OOB_NAMEDCLASS -1
420 /* There is no code point that is out-of-bounds, so this is problematic. But
421 * its only current use is to initialize a variable that is always set before
423 #define OOB_UNICODE 0xDEADBEEF
425 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
426 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
429 /* length of regex to show in messages that don't mark a position within */
430 #define RegexLengthToShowInErrorMessages 127
433 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
434 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
435 * op/pragma/warn/regcomp.
437 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
438 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
440 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
443 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
444 * arg. Show regex, up to a maximum length. If it's too long, chop and add
447 #define _FAIL(code) STMT_START { \
448 const char *ellipses = ""; \
449 IV len = RExC_end - RExC_precomp; \
452 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
453 if (len > RegexLengthToShowInErrorMessages) { \
454 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
455 len = RegexLengthToShowInErrorMessages - 10; \
461 #define FAIL(msg) _FAIL( \
462 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
463 msg, (int)len, RExC_precomp, ellipses))
465 #define FAIL2(msg,arg) _FAIL( \
466 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
467 arg, (int)len, RExC_precomp, ellipses))
470 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
472 #define Simple_vFAIL(m) STMT_START { \
473 const IV offset = RExC_parse - RExC_precomp; \
474 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
475 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
479 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
481 #define vFAIL(m) STMT_START { \
483 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
488 * Like Simple_vFAIL(), but accepts two arguments.
490 #define Simple_vFAIL2(m,a1) STMT_START { \
491 const IV offset = RExC_parse - RExC_precomp; \
492 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
493 (int)offset, RExC_precomp, RExC_precomp + offset); \
497 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
499 #define vFAIL2(m,a1) STMT_START { \
501 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
502 Simple_vFAIL2(m, a1); \
507 * Like Simple_vFAIL(), but accepts three arguments.
509 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
510 const IV offset = RExC_parse - RExC_precomp; \
511 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
512 (int)offset, RExC_precomp, RExC_precomp + offset); \
516 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
518 #define vFAIL3(m,a1,a2) STMT_START { \
520 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
521 Simple_vFAIL3(m, a1, a2); \
525 * Like Simple_vFAIL(), but accepts four arguments.
527 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
528 const IV offset = RExC_parse - RExC_precomp; \
529 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
530 (int)offset, RExC_precomp, RExC_precomp + offset); \
533 #define ckWARNreg(loc,m) STMT_START { \
534 const IV offset = loc - RExC_precomp; \
535 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
536 (int)offset, RExC_precomp, RExC_precomp + offset); \
539 #define ckWARNregdep(loc,m) STMT_START { \
540 const IV offset = loc - RExC_precomp; \
541 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
543 (int)offset, RExC_precomp, RExC_precomp + offset); \
546 #define ckWARN2regdep(loc,m, a1) STMT_START { \
547 const IV offset = loc - RExC_precomp; \
548 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
550 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
553 #define ckWARN2reg(loc, m, a1) STMT_START { \
554 const IV offset = loc - RExC_precomp; \
555 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
556 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
559 #define vWARN3(loc, m, a1, a2) STMT_START { \
560 const IV offset = loc - RExC_precomp; \
561 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
562 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
565 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
566 const IV offset = loc - RExC_precomp; \
567 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
568 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
571 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
572 const IV offset = loc - RExC_precomp; \
573 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
574 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
577 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
578 const IV offset = loc - RExC_precomp; \
579 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
580 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
583 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
584 const IV offset = loc - RExC_precomp; \
585 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
586 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
590 /* Allow for side effects in s */
591 #define REGC(c,s) STMT_START { \
592 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
595 /* Macros for recording node offsets. 20001227 mjd@plover.com
596 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
597 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
598 * Element 0 holds the number n.
599 * Position is 1 indexed.
601 #ifndef RE_TRACK_PATTERN_OFFSETS
602 #define Set_Node_Offset_To_R(node,byte)
603 #define Set_Node_Offset(node,byte)
604 #define Set_Cur_Node_Offset
605 #define Set_Node_Length_To_R(node,len)
606 #define Set_Node_Length(node,len)
607 #define Set_Node_Cur_Length(node)
608 #define Node_Offset(n)
609 #define Node_Length(n)
610 #define Set_Node_Offset_Length(node,offset,len)
611 #define ProgLen(ri) ri->u.proglen
612 #define SetProgLen(ri,x) ri->u.proglen = x
614 #define ProgLen(ri) ri->u.offsets[0]
615 #define SetProgLen(ri,x) ri->u.offsets[0] = x
616 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
618 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
619 __LINE__, (int)(node), (int)(byte))); \
621 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
623 RExC_offsets[2*(node)-1] = (byte); \
628 #define Set_Node_Offset(node,byte) \
629 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
630 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
632 #define Set_Node_Length_To_R(node,len) STMT_START { \
634 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
635 __LINE__, (int)(node), (int)(len))); \
637 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
639 RExC_offsets[2*(node)] = (len); \
644 #define Set_Node_Length(node,len) \
645 Set_Node_Length_To_R((node)-RExC_emit_start, len)
646 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
647 #define Set_Node_Cur_Length(node) \
648 Set_Node_Length(node, RExC_parse - parse_start)
650 /* Get offsets and lengths */
651 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
652 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
654 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
655 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
656 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
660 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
661 #define EXPERIMENTAL_INPLACESCAN
662 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
664 #define DEBUG_STUDYDATA(str,data,depth) \
665 DEBUG_OPTIMISE_MORE_r(if(data){ \
666 PerlIO_printf(Perl_debug_log, \
667 "%*s" str "Pos:%"IVdf"/%"IVdf \
668 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
669 (int)(depth)*2, "", \
670 (IV)((data)->pos_min), \
671 (IV)((data)->pos_delta), \
672 (UV)((data)->flags), \
673 (IV)((data)->whilem_c), \
674 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
675 is_inf ? "INF " : "" \
677 if ((data)->last_found) \
678 PerlIO_printf(Perl_debug_log, \
679 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
680 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
681 SvPVX_const((data)->last_found), \
682 (IV)((data)->last_end), \
683 (IV)((data)->last_start_min), \
684 (IV)((data)->last_start_max), \
685 ((data)->longest && \
686 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
687 SvPVX_const((data)->longest_fixed), \
688 (IV)((data)->offset_fixed), \
689 ((data)->longest && \
690 (data)->longest==&((data)->longest_float)) ? "*" : "", \
691 SvPVX_const((data)->longest_float), \
692 (IV)((data)->offset_float_min), \
693 (IV)((data)->offset_float_max) \
695 PerlIO_printf(Perl_debug_log,"\n"); \
698 static void clear_re(pTHX_ void *r);
700 /* Mark that we cannot extend a found fixed substring at this point.
701 Update the longest found anchored substring and the longest found
702 floating substrings if needed. */
705 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
707 const STRLEN l = CHR_SVLEN(data->last_found);
708 const STRLEN old_l = CHR_SVLEN(*data->longest);
709 GET_RE_DEBUG_FLAGS_DECL;
711 PERL_ARGS_ASSERT_SCAN_COMMIT;
713 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
714 SvSetMagicSV(*data->longest, data->last_found);
715 if (*data->longest == data->longest_fixed) {
716 data->offset_fixed = l ? data->last_start_min : data->pos_min;
717 if (data->flags & SF_BEFORE_EOL)
719 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
721 data->flags &= ~SF_FIX_BEFORE_EOL;
722 data->minlen_fixed=minlenp;
723 data->lookbehind_fixed=0;
725 else { /* *data->longest == data->longest_float */
726 data->offset_float_min = l ? data->last_start_min : data->pos_min;
727 data->offset_float_max = (l
728 ? data->last_start_max
729 : data->pos_min + data->pos_delta);
730 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
731 data->offset_float_max = I32_MAX;
732 if (data->flags & SF_BEFORE_EOL)
734 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
736 data->flags &= ~SF_FL_BEFORE_EOL;
737 data->minlen_float=minlenp;
738 data->lookbehind_float=0;
741 SvCUR_set(data->last_found, 0);
743 SV * const sv = data->last_found;
744 if (SvUTF8(sv) && SvMAGICAL(sv)) {
745 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
751 data->flags &= ~SF_BEFORE_EOL;
752 DEBUG_STUDYDATA("commit: ",data,0);
755 /* Can match anything (initialization) */
757 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
759 PERL_ARGS_ASSERT_CL_ANYTHING;
761 ANYOF_BITMAP_SETALL(cl);
762 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
763 |ANYOF_NON_UTF8_LATIN1_ALL;
765 /* If any portion of the regex is to operate under locale rules,
766 * initialization includes it. The reason this isn't done for all regexes
767 * is that the optimizer was written under the assumption that locale was
768 * all-or-nothing. Given the complexity and lack of documentation in the
769 * optimizer, and that there are inadequate test cases for locale, so many
770 * parts of it may not work properly, it is safest to avoid locale unless
772 if (RExC_contains_locale) {
773 ANYOF_CLASS_SETALL(cl); /* /l uses class */
774 cl->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
777 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
781 /* Can match anything (initialization) */
783 S_cl_is_anything(const struct regnode_charclass_class *cl)
787 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
789 for (value = 0; value <= ANYOF_MAX; value += 2)
790 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
792 if (!(cl->flags & ANYOF_UNICODE_ALL))
794 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
799 /* Can match anything (initialization) */
801 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
803 PERL_ARGS_ASSERT_CL_INIT;
805 Zero(cl, 1, struct regnode_charclass_class);
807 cl_anything(pRExC_state, cl);
808 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
811 /* These two functions currently do the exact same thing */
812 #define cl_init_zero S_cl_init
814 /* 'AND' a given class with another one. Can create false positives. 'cl'
815 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
816 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
818 S_cl_and(struct regnode_charclass_class *cl,
819 const struct regnode_charclass_class *and_with)
821 PERL_ARGS_ASSERT_CL_AND;
823 assert(and_with->type == ANYOF);
825 /* I (khw) am not sure all these restrictions are necessary XXX */
826 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
827 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
828 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
829 && !(and_with->flags & ANYOF_LOC_FOLD)
830 && !(cl->flags & ANYOF_LOC_FOLD)) {
833 if (and_with->flags & ANYOF_INVERT)
834 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
835 cl->bitmap[i] &= ~and_with->bitmap[i];
837 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
838 cl->bitmap[i] &= and_with->bitmap[i];
839 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
841 if (and_with->flags & ANYOF_INVERT) {
843 /* Here, the and'ed node is inverted. Get the AND of the flags that
844 * aren't affected by the inversion. Those that are affected are
845 * handled individually below */
846 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
847 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
848 cl->flags |= affected_flags;
850 /* We currently don't know how to deal with things that aren't in the
851 * bitmap, but we know that the intersection is no greater than what
852 * is already in cl, so let there be false positives that get sorted
853 * out after the synthetic start class succeeds, and the node is
854 * matched for real. */
856 /* The inversion of these two flags indicate that the resulting
857 * intersection doesn't have them */
858 if (and_with->flags & ANYOF_UNICODE_ALL) {
859 cl->flags &= ~ANYOF_UNICODE_ALL;
861 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
862 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
865 else { /* and'd node is not inverted */
866 U8 outside_bitmap_but_not_utf8; /* Temp variable */
868 if (! ANYOF_NONBITMAP(and_with)) {
870 /* Here 'and_with' doesn't match anything outside the bitmap
871 * (except possibly ANYOF_UNICODE_ALL), which means the
872 * intersection can't either, except for ANYOF_UNICODE_ALL, in
873 * which case we don't know what the intersection is, but it's no
874 * greater than what cl already has, so can just leave it alone,
875 * with possible false positives */
876 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
877 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
878 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
881 else if (! ANYOF_NONBITMAP(cl)) {
883 /* Here, 'and_with' does match something outside the bitmap, and cl
884 * doesn't have a list of things to match outside the bitmap. If
885 * cl can match all code points above 255, the intersection will
886 * be those above-255 code points that 'and_with' matches. If cl
887 * can't match all Unicode code points, it means that it can't
888 * match anything outside the bitmap (since the 'if' that got us
889 * into this block tested for that), so we leave the bitmap empty.
891 if (cl->flags & ANYOF_UNICODE_ALL) {
892 ARG_SET(cl, ARG(and_with));
894 /* and_with's ARG may match things that don't require UTF8.
895 * And now cl's will too, in spite of this being an 'and'. See
896 * the comments below about the kludge */
897 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
901 /* Here, both 'and_with' and cl match something outside the
902 * bitmap. Currently we do not do the intersection, so just match
903 * whatever cl had at the beginning. */
907 /* Take the intersection of the two sets of flags. However, the
908 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
909 * kludge around the fact that this flag is not treated like the others
910 * which are initialized in cl_anything(). The way the optimizer works
911 * is that the synthetic start class (SSC) is initialized to match
912 * anything, and then the first time a real node is encountered, its
913 * values are AND'd with the SSC's with the result being the values of
914 * the real node. However, there are paths through the optimizer where
915 * the AND never gets called, so those initialized bits are set
916 * inappropriately, which is not usually a big deal, as they just cause
917 * false positives in the SSC, which will just mean a probably
918 * imperceptible slow down in execution. However this bit has a
919 * higher false positive consequence in that it can cause utf8.pm,
920 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
921 * bigger slowdown and also causes significant extra memory to be used.
922 * In order to prevent this, the code now takes a different tack. The
923 * bit isn't set unless some part of the regular expression needs it,
924 * but once set it won't get cleared. This means that these extra
925 * modules won't get loaded unless there was some path through the
926 * pattern that would have required them anyway, and so any false
927 * positives that occur by not ANDing them out when they could be
928 * aren't as severe as they would be if we treated this bit like all
930 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
931 & ANYOF_NONBITMAP_NON_UTF8;
932 cl->flags &= and_with->flags;
933 cl->flags |= outside_bitmap_but_not_utf8;
937 /* 'OR' a given class with another one. Can create false positives. 'cl'
938 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
939 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
941 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
943 PERL_ARGS_ASSERT_CL_OR;
945 if (or_with->flags & ANYOF_INVERT) {
947 /* Here, the or'd node is to be inverted. This means we take the
948 * complement of everything not in the bitmap, but currently we don't
949 * know what that is, so give up and match anything */
950 if (ANYOF_NONBITMAP(or_with)) {
951 cl_anything(pRExC_state, cl);
954 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
955 * <= (B1 | !B2) | (CL1 | !CL2)
956 * which is wasteful if CL2 is small, but we ignore CL2:
957 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
958 * XXXX Can we handle case-fold? Unclear:
959 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
960 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
962 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
963 && !(or_with->flags & ANYOF_LOC_FOLD)
964 && !(cl->flags & ANYOF_LOC_FOLD) ) {
967 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
968 cl->bitmap[i] |= ~or_with->bitmap[i];
969 } /* XXXX: logic is complicated otherwise */
971 cl_anything(pRExC_state, cl);
974 /* And, we can just take the union of the flags that aren't affected
975 * by the inversion */
976 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
978 /* For the remaining flags:
979 ANYOF_UNICODE_ALL and inverted means to not match anything above
980 255, which means that the union with cl should just be
981 what cl has in it, so can ignore this flag
982 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
983 is 127-255 to match them, but then invert that, so the
984 union with cl should just be what cl has in it, so can
987 } else { /* 'or_with' is not inverted */
988 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
989 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
990 && (!(or_with->flags & ANYOF_LOC_FOLD)
991 || (cl->flags & ANYOF_LOC_FOLD)) ) {
994 /* OR char bitmap and class bitmap separately */
995 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
996 cl->bitmap[i] |= or_with->bitmap[i];
997 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
998 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
999 cl->classflags[i] |= or_with->classflags[i];
1000 cl->flags |= ANYOF_CLASS;
1003 else { /* XXXX: logic is complicated, leave it along for a moment. */
1004 cl_anything(pRExC_state, cl);
1007 if (ANYOF_NONBITMAP(or_with)) {
1009 /* Use the added node's outside-the-bit-map match if there isn't a
1010 * conflict. If there is a conflict (both nodes match something
1011 * outside the bitmap, but what they match outside is not the same
1012 * pointer, and hence not easily compared until XXX we extend
1013 * inversion lists this far), give up and allow the start class to
1014 * match everything outside the bitmap. If that stuff is all above
1015 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1016 if (! ANYOF_NONBITMAP(cl)) {
1017 ARG_SET(cl, ARG(or_with));
1019 else if (ARG(cl) != ARG(or_with)) {
1021 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1022 cl_anything(pRExC_state, cl);
1025 cl->flags |= ANYOF_UNICODE_ALL;
1030 /* Take the union */
1031 cl->flags |= or_with->flags;
1035 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1036 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1037 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1038 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1043 dump_trie(trie,widecharmap,revcharmap)
1044 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1045 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1047 These routines dump out a trie in a somewhat readable format.
1048 The _interim_ variants are used for debugging the interim
1049 tables that are used to generate the final compressed
1050 representation which is what dump_trie expects.
1052 Part of the reason for their existence is to provide a form
1053 of documentation as to how the different representations function.
1058 Dumps the final compressed table form of the trie to Perl_debug_log.
1059 Used for debugging make_trie().
1063 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1064 AV *revcharmap, U32 depth)
1067 SV *sv=sv_newmortal();
1068 int colwidth= widecharmap ? 6 : 4;
1070 GET_RE_DEBUG_FLAGS_DECL;
1072 PERL_ARGS_ASSERT_DUMP_TRIE;
1074 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1075 (int)depth * 2 + 2,"",
1076 "Match","Base","Ofs" );
1078 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1079 SV ** const tmp = av_fetch( revcharmap, state, 0);
1081 PerlIO_printf( Perl_debug_log, "%*s",
1083 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1084 PL_colors[0], PL_colors[1],
1085 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1086 PERL_PV_ESCAPE_FIRSTCHAR
1091 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1092 (int)depth * 2 + 2,"");
1094 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1095 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1096 PerlIO_printf( Perl_debug_log, "\n");
1098 for( state = 1 ; state < trie->statecount ; state++ ) {
1099 const U32 base = trie->states[ state ].trans.base;
1101 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1103 if ( trie->states[ state ].wordnum ) {
1104 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1106 PerlIO_printf( Perl_debug_log, "%6s", "" );
1109 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1114 while( ( base + ofs < trie->uniquecharcount ) ||
1115 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1116 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1119 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1121 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1122 if ( ( base + ofs >= trie->uniquecharcount ) &&
1123 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1124 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1126 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1128 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1130 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1134 PerlIO_printf( Perl_debug_log, "]");
1137 PerlIO_printf( Perl_debug_log, "\n" );
1139 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1140 for (word=1; word <= trie->wordcount; word++) {
1141 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1142 (int)word, (int)(trie->wordinfo[word].prev),
1143 (int)(trie->wordinfo[word].len));
1145 PerlIO_printf(Perl_debug_log, "\n" );
1148 Dumps a fully constructed but uncompressed trie in list form.
1149 List tries normally only are used for construction when the number of
1150 possible chars (trie->uniquecharcount) is very high.
1151 Used for debugging make_trie().
1154 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1155 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1159 SV *sv=sv_newmortal();
1160 int colwidth= widecharmap ? 6 : 4;
1161 GET_RE_DEBUG_FLAGS_DECL;
1163 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1165 /* print out the table precompression. */
1166 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1167 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1168 "------:-----+-----------------\n" );
1170 for( state=1 ; state < next_alloc ; state ++ ) {
1173 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1174 (int)depth * 2 + 2,"", (UV)state );
1175 if ( ! trie->states[ state ].wordnum ) {
1176 PerlIO_printf( Perl_debug_log, "%5s| ","");
1178 PerlIO_printf( Perl_debug_log, "W%4x| ",
1179 trie->states[ state ].wordnum
1182 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1183 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1185 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1187 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1188 PL_colors[0], PL_colors[1],
1189 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1190 PERL_PV_ESCAPE_FIRSTCHAR
1192 TRIE_LIST_ITEM(state,charid).forid,
1193 (UV)TRIE_LIST_ITEM(state,charid).newstate
1196 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1197 (int)((depth * 2) + 14), "");
1200 PerlIO_printf( Perl_debug_log, "\n");
1205 Dumps a fully constructed but uncompressed trie in table form.
1206 This is the normal DFA style state transition table, with a few
1207 twists to facilitate compression later.
1208 Used for debugging make_trie().
1211 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1212 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1217 SV *sv=sv_newmortal();
1218 int colwidth= widecharmap ? 6 : 4;
1219 GET_RE_DEBUG_FLAGS_DECL;
1221 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1224 print out the table precompression so that we can do a visual check
1225 that they are identical.
1228 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1230 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1231 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1233 PerlIO_printf( Perl_debug_log, "%*s",
1235 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1236 PL_colors[0], PL_colors[1],
1237 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1238 PERL_PV_ESCAPE_FIRSTCHAR
1244 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1246 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1247 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1250 PerlIO_printf( Perl_debug_log, "\n" );
1252 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1254 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1255 (int)depth * 2 + 2,"",
1256 (UV)TRIE_NODENUM( state ) );
1258 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1259 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1261 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1263 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1265 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1266 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1268 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1269 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1277 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1278 startbranch: the first branch in the whole branch sequence
1279 first : start branch of sequence of branch-exact nodes.
1280 May be the same as startbranch
1281 last : Thing following the last branch.
1282 May be the same as tail.
1283 tail : item following the branch sequence
1284 count : words in the sequence
1285 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1286 depth : indent depth
1288 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1290 A trie is an N'ary tree where the branches are determined by digital
1291 decomposition of the key. IE, at the root node you look up the 1st character and
1292 follow that branch repeat until you find the end of the branches. Nodes can be
1293 marked as "accepting" meaning they represent a complete word. Eg:
1297 would convert into the following structure. Numbers represent states, letters
1298 following numbers represent valid transitions on the letter from that state, if
1299 the number is in square brackets it represents an accepting state, otherwise it
1300 will be in parenthesis.
1302 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1306 (1) +-i->(6)-+-s->[7]
1308 +-s->(3)-+-h->(4)-+-e->[5]
1310 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1312 This shows that when matching against the string 'hers' we will begin at state 1
1313 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1314 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1315 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1316 single traverse. We store a mapping from accepting to state to which word was
1317 matched, and then when we have multiple possibilities we try to complete the
1318 rest of the regex in the order in which they occured in the alternation.
1320 The only prior NFA like behaviour that would be changed by the TRIE support is
1321 the silent ignoring of duplicate alternations which are of the form:
1323 / (DUPE|DUPE) X? (?{ ... }) Y /x
1325 Thus EVAL blocks following a trie may be called a different number of times with
1326 and without the optimisation. With the optimisations dupes will be silently
1327 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1328 the following demonstrates:
1330 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1332 which prints out 'word' three times, but
1334 'words'=~/(word|word|word)(?{ print $1 })S/
1336 which doesnt print it out at all. This is due to other optimisations kicking in.
1338 Example of what happens on a structural level:
1340 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1342 1: CURLYM[1] {1,32767}(18)
1353 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1354 and should turn into:
1356 1: CURLYM[1] {1,32767}(18)
1358 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1366 Cases where tail != last would be like /(?foo|bar)baz/:
1376 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1377 and would end up looking like:
1380 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1387 d = uvuni_to_utf8_flags(d, uv, 0);
1389 is the recommended Unicode-aware way of saying
1394 #define TRIE_STORE_REVCHAR(val) \
1397 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1398 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1399 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1400 SvCUR_set(zlopp, kapow - flrbbbbb); \
1403 av_push(revcharmap, zlopp); \
1405 char ooooff = (char)val; \
1406 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1410 #define TRIE_READ_CHAR STMT_START { \
1413 /* if it is UTF then it is either already folded, or does not need folding */ \
1414 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1416 else if (folder == PL_fold_latin1) { \
1417 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1418 if ( foldlen > 0 ) { \
1419 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1425 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1426 skiplen = UNISKIP(uvc); \
1427 foldlen -= skiplen; \
1428 scan = foldbuf + skiplen; \
1431 /* raw data, will be folded later if needed */ \
1439 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1440 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1441 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1442 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1444 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1445 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1446 TRIE_LIST_CUR( state )++; \
1449 #define TRIE_LIST_NEW(state) STMT_START { \
1450 Newxz( trie->states[ state ].trans.list, \
1451 4, reg_trie_trans_le ); \
1452 TRIE_LIST_CUR( state ) = 1; \
1453 TRIE_LIST_LEN( state ) = 4; \
1456 #define TRIE_HANDLE_WORD(state) STMT_START { \
1457 U16 dupe= trie->states[ state ].wordnum; \
1458 regnode * const noper_next = regnext( noper ); \
1461 /* store the word for dumping */ \
1463 if (OP(noper) != NOTHING) \
1464 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1466 tmp = newSVpvn_utf8( "", 0, UTF ); \
1467 av_push( trie_words, tmp ); \
1471 trie->wordinfo[curword].prev = 0; \
1472 trie->wordinfo[curword].len = wordlen; \
1473 trie->wordinfo[curword].accept = state; \
1475 if ( noper_next < tail ) { \
1477 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1478 trie->jump[curword] = (U16)(noper_next - convert); \
1480 jumper = noper_next; \
1482 nextbranch= regnext(cur); \
1486 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1487 /* chain, so that when the bits of chain are later */\
1488 /* linked together, the dups appear in the chain */\
1489 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1490 trie->wordinfo[dupe].prev = curword; \
1492 /* we haven't inserted this word yet. */ \
1493 trie->states[ state ].wordnum = curword; \
1498 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1499 ( ( base + charid >= ucharcount \
1500 && base + charid < ubound \
1501 && state == trie->trans[ base - ucharcount + charid ].check \
1502 && trie->trans[ base - ucharcount + charid ].next ) \
1503 ? trie->trans[ base - ucharcount + charid ].next \
1504 : ( state==1 ? special : 0 ) \
1508 #define MADE_JUMP_TRIE 2
1509 #define MADE_EXACT_TRIE 4
1512 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1515 /* first pass, loop through and scan words */
1516 reg_trie_data *trie;
1517 HV *widecharmap = NULL;
1518 AV *revcharmap = newAV();
1520 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1525 regnode *jumper = NULL;
1526 regnode *nextbranch = NULL;
1527 regnode *convert = NULL;
1528 U32 *prev_states; /* temp array mapping each state to previous one */
1529 /* we just use folder as a flag in utf8 */
1530 const U8 * folder = NULL;
1533 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1534 AV *trie_words = NULL;
1535 /* along with revcharmap, this only used during construction but both are
1536 * useful during debugging so we store them in the struct when debugging.
1539 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1540 STRLEN trie_charcount=0;
1542 SV *re_trie_maxbuff;
1543 GET_RE_DEBUG_FLAGS_DECL;
1545 PERL_ARGS_ASSERT_MAKE_TRIE;
1547 PERL_UNUSED_ARG(depth);
1554 case EXACTFU_TRICKYFOLD:
1555 case EXACTFU: folder = PL_fold_latin1; break;
1556 case EXACTF: folder = PL_fold; break;
1557 case EXACTFL: folder = PL_fold_locale; break;
1558 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1561 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1563 trie->startstate = 1;
1564 trie->wordcount = word_count;
1565 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1566 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1568 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1569 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1570 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1573 trie_words = newAV();
1576 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1577 if (!SvIOK(re_trie_maxbuff)) {
1578 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1580 DEBUG_TRIE_COMPILE_r({
1581 PerlIO_printf( Perl_debug_log,
1582 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1583 (int)depth * 2 + 2, "",
1584 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1585 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1589 /* Find the node we are going to overwrite */
1590 if ( first == startbranch && OP( last ) != BRANCH ) {
1591 /* whole branch chain */
1594 /* branch sub-chain */
1595 convert = NEXTOPER( first );
1598 /* -- First loop and Setup --
1600 We first traverse the branches and scan each word to determine if it
1601 contains widechars, and how many unique chars there are, this is
1602 important as we have to build a table with at least as many columns as we
1605 We use an array of integers to represent the character codes 0..255
1606 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1607 native representation of the character value as the key and IV's for the
1610 *TODO* If we keep track of how many times each character is used we can
1611 remap the columns so that the table compression later on is more
1612 efficient in terms of memory by ensuring the most common value is in the
1613 middle and the least common are on the outside. IMO this would be better
1614 than a most to least common mapping as theres a decent chance the most
1615 common letter will share a node with the least common, meaning the node
1616 will not be compressible. With a middle is most common approach the worst
1617 case is when we have the least common nodes twice.
1621 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1622 regnode *noper = NEXTOPER( cur );
1623 const U8 *uc = (U8*)STRING( noper );
1624 const U8 *e = uc + STR_LEN( noper );
1626 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1628 const U8 *scan = (U8*)NULL;
1629 U32 wordlen = 0; /* required init */
1631 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1633 if (OP(noper) == NOTHING) {
1634 regnode *noper_next= regnext(noper);
1635 if (noper_next != tail && OP(noper_next) == flags) {
1637 uc= (U8*)STRING(noper);
1638 e= uc + STR_LEN(noper);
1639 trie->minlen= STR_LEN(noper);
1646 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1647 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1648 regardless of encoding */
1649 if (OP( noper ) == EXACTFU_SS) {
1650 /* false positives are ok, so just set this */
1651 TRIE_BITMAP_SET(trie,0xDF);
1654 for ( ; uc < e ; uc += len ) {
1655 TRIE_CHARCOUNT(trie)++;
1660 U8 folded= folder[ (U8) uvc ];
1661 if ( !trie->charmap[ folded ] ) {
1662 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1663 TRIE_STORE_REVCHAR( folded );
1666 if ( !trie->charmap[ uvc ] ) {
1667 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1668 TRIE_STORE_REVCHAR( uvc );
1671 /* store the codepoint in the bitmap, and its folded
1673 TRIE_BITMAP_SET(trie, uvc);
1675 /* store the folded codepoint */
1676 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1679 /* store first byte of utf8 representation of
1680 variant codepoints */
1681 if (! UNI_IS_INVARIANT(uvc)) {
1682 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1685 set_bit = 0; /* We've done our bit :-) */
1690 widecharmap = newHV();
1692 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1695 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1697 if ( !SvTRUE( *svpp ) ) {
1698 sv_setiv( *svpp, ++trie->uniquecharcount );
1699 TRIE_STORE_REVCHAR(uvc);
1703 if( cur == first ) {
1704 trie->minlen = chars;
1705 trie->maxlen = chars;
1706 } else if (chars < trie->minlen) {
1707 trie->minlen = chars;
1708 } else if (chars > trie->maxlen) {
1709 trie->maxlen = chars;
1711 if (OP( noper ) == EXACTFU_SS) {
1712 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1713 if (trie->minlen > 1)
1716 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1717 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1718 * - We assume that any such sequence might match a 2 byte string */
1719 if (trie->minlen > 2 )
1723 } /* end first pass */
1724 DEBUG_TRIE_COMPILE_r(
1725 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1726 (int)depth * 2 + 2,"",
1727 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1728 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1729 (int)trie->minlen, (int)trie->maxlen )
1733 We now know what we are dealing with in terms of unique chars and
1734 string sizes so we can calculate how much memory a naive
1735 representation using a flat table will take. If it's over a reasonable
1736 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1737 conservative but potentially much slower representation using an array
1740 At the end we convert both representations into the same compressed
1741 form that will be used in regexec.c for matching with. The latter
1742 is a form that cannot be used to construct with but has memory
1743 properties similar to the list form and access properties similar
1744 to the table form making it both suitable for fast searches and
1745 small enough that its feasable to store for the duration of a program.
1747 See the comment in the code where the compressed table is produced
1748 inplace from the flat tabe representation for an explanation of how
1749 the compression works.
1754 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1757 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1759 Second Pass -- Array Of Lists Representation
1761 Each state will be represented by a list of charid:state records
1762 (reg_trie_trans_le) the first such element holds the CUR and LEN
1763 points of the allocated array. (See defines above).
1765 We build the initial structure using the lists, and then convert
1766 it into the compressed table form which allows faster lookups
1767 (but cant be modified once converted).
1770 STRLEN transcount = 1;
1772 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1773 "%*sCompiling trie using list compiler\n",
1774 (int)depth * 2 + 2, ""));
1776 trie->states = (reg_trie_state *)
1777 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1778 sizeof(reg_trie_state) );
1782 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1784 regnode *noper = NEXTOPER( cur );
1785 U8 *uc = (U8*)STRING( noper );
1786 const U8 *e = uc + STR_LEN( noper );
1787 U32 state = 1; /* required init */
1788 U16 charid = 0; /* sanity init */
1789 U8 *scan = (U8*)NULL; /* sanity init */
1790 STRLEN foldlen = 0; /* required init */
1791 U32 wordlen = 0; /* required init */
1792 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1795 if (OP(noper) == NOTHING) {
1796 regnode *noper_next= regnext(noper);
1797 if (noper_next != tail && OP(noper_next) == flags) {
1799 uc= (U8*)STRING(noper);
1800 e= uc + STR_LEN(noper);
1804 if (OP(noper) != NOTHING) {
1805 for ( ; uc < e ; uc += len ) {
1810 charid = trie->charmap[ uvc ];
1812 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1816 charid=(U16)SvIV( *svpp );
1819 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1826 if ( !trie->states[ state ].trans.list ) {
1827 TRIE_LIST_NEW( state );
1829 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1830 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1831 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1836 newstate = next_alloc++;
1837 prev_states[newstate] = state;
1838 TRIE_LIST_PUSH( state, charid, newstate );
1843 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1847 TRIE_HANDLE_WORD(state);
1849 } /* end second pass */
1851 /* next alloc is the NEXT state to be allocated */
1852 trie->statecount = next_alloc;
1853 trie->states = (reg_trie_state *)
1854 PerlMemShared_realloc( trie->states,
1856 * sizeof(reg_trie_state) );
1858 /* and now dump it out before we compress it */
1859 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1860 revcharmap, next_alloc,
1864 trie->trans = (reg_trie_trans *)
1865 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1872 for( state=1 ; state < next_alloc ; state ++ ) {
1876 DEBUG_TRIE_COMPILE_MORE_r(
1877 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1881 if (trie->states[state].trans.list) {
1882 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1886 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1887 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1888 if ( forid < minid ) {
1890 } else if ( forid > maxid ) {
1894 if ( transcount < tp + maxid - minid + 1) {
1896 trie->trans = (reg_trie_trans *)
1897 PerlMemShared_realloc( trie->trans,
1899 * sizeof(reg_trie_trans) );
1900 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1902 base = trie->uniquecharcount + tp - minid;
1903 if ( maxid == minid ) {
1905 for ( ; zp < tp ; zp++ ) {
1906 if ( ! trie->trans[ zp ].next ) {
1907 base = trie->uniquecharcount + zp - minid;
1908 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1909 trie->trans[ zp ].check = state;
1915 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1916 trie->trans[ tp ].check = state;
1921 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1922 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1923 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1924 trie->trans[ tid ].check = state;
1926 tp += ( maxid - minid + 1 );
1928 Safefree(trie->states[ state ].trans.list);
1931 DEBUG_TRIE_COMPILE_MORE_r(
1932 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1935 trie->states[ state ].trans.base=base;
1937 trie->lasttrans = tp + 1;
1941 Second Pass -- Flat Table Representation.
1943 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1944 We know that we will need Charcount+1 trans at most to store the data
1945 (one row per char at worst case) So we preallocate both structures
1946 assuming worst case.
1948 We then construct the trie using only the .next slots of the entry
1951 We use the .check field of the first entry of the node temporarily to
1952 make compression both faster and easier by keeping track of how many non
1953 zero fields are in the node.
1955 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1958 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1959 number representing the first entry of the node, and state as a
1960 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1961 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1962 are 2 entrys per node. eg:
1970 The table is internally in the right hand, idx form. However as we also
1971 have to deal with the states array which is indexed by nodenum we have to
1972 use TRIE_NODENUM() to convert.
1975 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1976 "%*sCompiling trie using table compiler\n",
1977 (int)depth * 2 + 2, ""));
1979 trie->trans = (reg_trie_trans *)
1980 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1981 * trie->uniquecharcount + 1,
1982 sizeof(reg_trie_trans) );
1983 trie->states = (reg_trie_state *)
1984 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1985 sizeof(reg_trie_state) );
1986 next_alloc = trie->uniquecharcount + 1;
1989 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1991 regnode *noper = NEXTOPER( cur );
1992 const U8 *uc = (U8*)STRING( noper );
1993 const U8 *e = uc + STR_LEN( noper );
1995 U32 state = 1; /* required init */
1997 U16 charid = 0; /* sanity init */
1998 U32 accept_state = 0; /* sanity init */
1999 U8 *scan = (U8*)NULL; /* sanity init */
2001 STRLEN foldlen = 0; /* required init */
2002 U32 wordlen = 0; /* required init */
2004 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2006 if (OP(noper) == NOTHING) {
2007 regnode *noper_next= regnext(noper);
2008 if (noper_next != tail && OP(noper_next) == flags) {
2010 uc= (U8*)STRING(noper);
2011 e= uc + STR_LEN(noper);
2015 if ( OP(noper) != NOTHING ) {
2016 for ( ; uc < e ; uc += len ) {
2021 charid = trie->charmap[ uvc ];
2023 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2024 charid = svpp ? (U16)SvIV(*svpp) : 0;
2028 if ( !trie->trans[ state + charid ].next ) {
2029 trie->trans[ state + charid ].next = next_alloc;
2030 trie->trans[ state ].check++;
2031 prev_states[TRIE_NODENUM(next_alloc)]
2032 = TRIE_NODENUM(state);
2033 next_alloc += trie->uniquecharcount;
2035 state = trie->trans[ state + charid ].next;
2037 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2039 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2042 accept_state = TRIE_NODENUM( state );
2043 TRIE_HANDLE_WORD(accept_state);
2045 } /* end second pass */
2047 /* and now dump it out before we compress it */
2048 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2050 next_alloc, depth+1));
2054 * Inplace compress the table.*
2056 For sparse data sets the table constructed by the trie algorithm will
2057 be mostly 0/FAIL transitions or to put it another way mostly empty.
2058 (Note that leaf nodes will not contain any transitions.)
2060 This algorithm compresses the tables by eliminating most such
2061 transitions, at the cost of a modest bit of extra work during lookup:
2063 - Each states[] entry contains a .base field which indicates the
2064 index in the state[] array wheres its transition data is stored.
2066 - If .base is 0 there are no valid transitions from that node.
2068 - If .base is nonzero then charid is added to it to find an entry in
2071 -If trans[states[state].base+charid].check!=state then the
2072 transition is taken to be a 0/Fail transition. Thus if there are fail
2073 transitions at the front of the node then the .base offset will point
2074 somewhere inside the previous nodes data (or maybe even into a node
2075 even earlier), but the .check field determines if the transition is
2079 The following process inplace converts the table to the compressed
2080 table: We first do not compress the root node 1,and mark all its
2081 .check pointers as 1 and set its .base pointer as 1 as well. This
2082 allows us to do a DFA construction from the compressed table later,
2083 and ensures that any .base pointers we calculate later are greater
2086 - We set 'pos' to indicate the first entry of the second node.
2088 - We then iterate over the columns of the node, finding the first and
2089 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2090 and set the .check pointers accordingly, and advance pos
2091 appropriately and repreat for the next node. Note that when we copy
2092 the next pointers we have to convert them from the original
2093 NODEIDX form to NODENUM form as the former is not valid post
2096 - If a node has no transitions used we mark its base as 0 and do not
2097 advance the pos pointer.
2099 - If a node only has one transition we use a second pointer into the
2100 structure to fill in allocated fail transitions from other states.
2101 This pointer is independent of the main pointer and scans forward
2102 looking for null transitions that are allocated to a state. When it
2103 finds one it writes the single transition into the "hole". If the
2104 pointer doesnt find one the single transition is appended as normal.
2106 - Once compressed we can Renew/realloc the structures to release the
2109 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2110 specifically Fig 3.47 and the associated pseudocode.
2114 const U32 laststate = TRIE_NODENUM( next_alloc );
2117 trie->statecount = laststate;
2119 for ( state = 1 ; state < laststate ; state++ ) {
2121 const U32 stateidx = TRIE_NODEIDX( state );
2122 const U32 o_used = trie->trans[ stateidx ].check;
2123 U32 used = trie->trans[ stateidx ].check;
2124 trie->trans[ stateidx ].check = 0;
2126 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2127 if ( flag || trie->trans[ stateidx + charid ].next ) {
2128 if ( trie->trans[ stateidx + charid ].next ) {
2130 for ( ; zp < pos ; zp++ ) {
2131 if ( ! trie->trans[ zp ].next ) {
2135 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2136 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2137 trie->trans[ zp ].check = state;
2138 if ( ++zp > pos ) pos = zp;
2145 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2147 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2148 trie->trans[ pos ].check = state;
2153 trie->lasttrans = pos + 1;
2154 trie->states = (reg_trie_state *)
2155 PerlMemShared_realloc( trie->states, laststate
2156 * sizeof(reg_trie_state) );
2157 DEBUG_TRIE_COMPILE_MORE_r(
2158 PerlIO_printf( Perl_debug_log,
2159 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2160 (int)depth * 2 + 2,"",
2161 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2164 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2167 } /* end table compress */
2169 DEBUG_TRIE_COMPILE_MORE_r(
2170 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2171 (int)depth * 2 + 2, "",
2172 (UV)trie->statecount,
2173 (UV)trie->lasttrans)
2175 /* resize the trans array to remove unused space */
2176 trie->trans = (reg_trie_trans *)
2177 PerlMemShared_realloc( trie->trans, trie->lasttrans
2178 * sizeof(reg_trie_trans) );
2180 { /* Modify the program and insert the new TRIE node */
2181 U8 nodetype =(U8)(flags & 0xFF);
2185 regnode *optimize = NULL;
2186 #ifdef RE_TRACK_PATTERN_OFFSETS
2189 U32 mjd_nodelen = 0;
2190 #endif /* RE_TRACK_PATTERN_OFFSETS */
2191 #endif /* DEBUGGING */
2193 This means we convert either the first branch or the first Exact,
2194 depending on whether the thing following (in 'last') is a branch
2195 or not and whther first is the startbranch (ie is it a sub part of
2196 the alternation or is it the whole thing.)
2197 Assuming its a sub part we convert the EXACT otherwise we convert
2198 the whole branch sequence, including the first.
2200 /* Find the node we are going to overwrite */
2201 if ( first != startbranch || OP( last ) == BRANCH ) {
2202 /* branch sub-chain */
2203 NEXT_OFF( first ) = (U16)(last - first);
2204 #ifdef RE_TRACK_PATTERN_OFFSETS
2206 mjd_offset= Node_Offset((convert));
2207 mjd_nodelen= Node_Length((convert));
2210 /* whole branch chain */
2212 #ifdef RE_TRACK_PATTERN_OFFSETS
2215 const regnode *nop = NEXTOPER( convert );
2216 mjd_offset= Node_Offset((nop));
2217 mjd_nodelen= Node_Length((nop));
2221 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2222 (int)depth * 2 + 2, "",
2223 (UV)mjd_offset, (UV)mjd_nodelen)
2226 /* But first we check to see if there is a common prefix we can
2227 split out as an EXACT and put in front of the TRIE node. */
2228 trie->startstate= 1;
2229 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2231 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2235 const U32 base = trie->states[ state ].trans.base;
2237 if ( trie->states[state].wordnum )
2240 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2241 if ( ( base + ofs >= trie->uniquecharcount ) &&
2242 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2243 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2245 if ( ++count > 1 ) {
2246 SV **tmp = av_fetch( revcharmap, ofs, 0);
2247 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2248 if ( state == 1 ) break;
2250 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2252 PerlIO_printf(Perl_debug_log,
2253 "%*sNew Start State=%"UVuf" Class: [",
2254 (int)depth * 2 + 2, "",
2257 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2258 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2260 TRIE_BITMAP_SET(trie,*ch);
2262 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2264 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2268 TRIE_BITMAP_SET(trie,*ch);
2270 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2271 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2277 SV **tmp = av_fetch( revcharmap, idx, 0);
2279 char *ch = SvPV( *tmp, len );
2281 SV *sv=sv_newmortal();
2282 PerlIO_printf( Perl_debug_log,
2283 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2284 (int)depth * 2 + 2, "",
2286 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2287 PL_colors[0], PL_colors[1],
2288 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2289 PERL_PV_ESCAPE_FIRSTCHAR
2294 OP( convert ) = nodetype;
2295 str=STRING(convert);
2298 STR_LEN(convert) += len;
2304 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2309 trie->prefixlen = (state-1);
2311 regnode *n = convert+NODE_SZ_STR(convert);
2312 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2313 trie->startstate = state;
2314 trie->minlen -= (state - 1);
2315 trie->maxlen -= (state - 1);
2317 /* At least the UNICOS C compiler choked on this
2318 * being argument to DEBUG_r(), so let's just have
2321 #ifdef PERL_EXT_RE_BUILD
2327 regnode *fix = convert;
2328 U32 word = trie->wordcount;
2330 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2331 while( ++fix < n ) {
2332 Set_Node_Offset_Length(fix, 0, 0);
2335 SV ** const tmp = av_fetch( trie_words, word, 0 );
2337 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2338 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2340 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2348 NEXT_OFF(convert) = (U16)(tail - convert);
2349 DEBUG_r(optimize= n);
2355 if ( trie->maxlen ) {
2356 NEXT_OFF( convert ) = (U16)(tail - convert);
2357 ARG_SET( convert, data_slot );
2358 /* Store the offset to the first unabsorbed branch in
2359 jump[0], which is otherwise unused by the jump logic.
2360 We use this when dumping a trie and during optimisation. */
2362 trie->jump[0] = (U16)(nextbranch - convert);
2364 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2365 * and there is a bitmap
2366 * and the first "jump target" node we found leaves enough room
2367 * then convert the TRIE node into a TRIEC node, with the bitmap
2368 * embedded inline in the opcode - this is hypothetically faster.
2370 if ( !trie->states[trie->startstate].wordnum
2372 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2374 OP( convert ) = TRIEC;
2375 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2376 PerlMemShared_free(trie->bitmap);
2379 OP( convert ) = TRIE;
2381 /* store the type in the flags */
2382 convert->flags = nodetype;
2386 + regarglen[ OP( convert ) ];
2388 /* XXX We really should free up the resource in trie now,
2389 as we won't use them - (which resources?) dmq */
2391 /* needed for dumping*/
2392 DEBUG_r(if (optimize) {
2393 regnode *opt = convert;
2395 while ( ++opt < optimize) {
2396 Set_Node_Offset_Length(opt,0,0);
2399 Try to clean up some of the debris left after the
2402 while( optimize < jumper ) {
2403 mjd_nodelen += Node_Length((optimize));
2404 OP( optimize ) = OPTIMIZED;
2405 Set_Node_Offset_Length(optimize,0,0);
2408 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2410 } /* end node insert */
2412 /* Finish populating the prev field of the wordinfo array. Walk back
2413 * from each accept state until we find another accept state, and if
2414 * so, point the first word's .prev field at the second word. If the
2415 * second already has a .prev field set, stop now. This will be the
2416 * case either if we've already processed that word's accept state,
2417 * or that state had multiple words, and the overspill words were
2418 * already linked up earlier.
2425 for (word=1; word <= trie->wordcount; word++) {
2427 if (trie->wordinfo[word].prev)
2429 state = trie->wordinfo[word].accept;
2431 state = prev_states[state];
2434 prev = trie->states[state].wordnum;
2438 trie->wordinfo[word].prev = prev;
2440 Safefree(prev_states);
2444 /* and now dump out the compressed format */
2445 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2447 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2449 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2450 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2452 SvREFCNT_dec(revcharmap);
2456 : trie->startstate>1
2462 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2464 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2466 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2467 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2470 We find the fail state for each state in the trie, this state is the longest proper
2471 suffix of the current state's 'word' that is also a proper prefix of another word in our
2472 trie. State 1 represents the word '' and is thus the default fail state. This allows
2473 the DFA not to have to restart after its tried and failed a word at a given point, it
2474 simply continues as though it had been matching the other word in the first place.
2476 'abcdgu'=~/abcdefg|cdgu/
2477 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2478 fail, which would bring us to the state representing 'd' in the second word where we would
2479 try 'g' and succeed, proceeding to match 'cdgu'.
2481 /* add a fail transition */
2482 const U32 trie_offset = ARG(source);
2483 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2485 const U32 ucharcount = trie->uniquecharcount;
2486 const U32 numstates = trie->statecount;
2487 const U32 ubound = trie->lasttrans + ucharcount;
2491 U32 base = trie->states[ 1 ].trans.base;
2494 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2495 GET_RE_DEBUG_FLAGS_DECL;
2497 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2499 PERL_UNUSED_ARG(depth);
2503 ARG_SET( stclass, data_slot );
2504 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2505 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2506 aho->trie=trie_offset;
2507 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2508 Copy( trie->states, aho->states, numstates, reg_trie_state );
2509 Newxz( q, numstates, U32);
2510 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2513 /* initialize fail[0..1] to be 1 so that we always have
2514 a valid final fail state */
2515 fail[ 0 ] = fail[ 1 ] = 1;
2517 for ( charid = 0; charid < ucharcount ; charid++ ) {
2518 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2520 q[ q_write ] = newstate;
2521 /* set to point at the root */
2522 fail[ q[ q_write++ ] ]=1;
2525 while ( q_read < q_write) {
2526 const U32 cur = q[ q_read++ % numstates ];
2527 base = trie->states[ cur ].trans.base;
2529 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2530 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2532 U32 fail_state = cur;
2535 fail_state = fail[ fail_state ];
2536 fail_base = aho->states[ fail_state ].trans.base;
2537 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2539 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2540 fail[ ch_state ] = fail_state;
2541 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2543 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2545 q[ q_write++ % numstates] = ch_state;
2549 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2550 when we fail in state 1, this allows us to use the
2551 charclass scan to find a valid start char. This is based on the principle
2552 that theres a good chance the string being searched contains lots of stuff
2553 that cant be a start char.
2555 fail[ 0 ] = fail[ 1 ] = 0;
2556 DEBUG_TRIE_COMPILE_r({
2557 PerlIO_printf(Perl_debug_log,
2558 "%*sStclass Failtable (%"UVuf" states): 0",
2559 (int)(depth * 2), "", (UV)numstates
2561 for( q_read=1; q_read<numstates; q_read++ ) {
2562 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2564 PerlIO_printf(Perl_debug_log, "\n");
2567 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2572 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2573 * These need to be revisited when a newer toolchain becomes available.
2575 #if defined(__sparc64__) && defined(__GNUC__)
2576 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2577 # undef SPARC64_GCC_WORKAROUND
2578 # define SPARC64_GCC_WORKAROUND 1
2582 #define DEBUG_PEEP(str,scan,depth) \
2583 DEBUG_OPTIMISE_r({if (scan){ \
2584 SV * const mysv=sv_newmortal(); \
2585 regnode *Next = regnext(scan); \
2586 regprop(RExC_rx, mysv, scan); \
2587 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2588 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2589 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2593 /* The below joins as many adjacent EXACTish nodes as possible into a single
2594 * one. The regop may be changed if the node(s) contain certain sequences that
2595 * require special handling. The joining is only done if:
2596 * 1) there is room in the current conglomerated node to entirely contain the
2598 * 2) they are the exact same node type
2600 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2601 * these get optimized out
2603 * If a node is to match under /i (folded), the number of characters it matches
2604 * can be different than its character length if it contains a multi-character
2605 * fold. *min_subtract is set to the total delta of the input nodes.
2607 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2608 * and contains LATIN SMALL LETTER SHARP S
2610 * This is as good a place as any to discuss the design of handling these
2611 * multi-character fold sequences. It's been wrong in Perl for a very long
2612 * time. There are three code points in Unicode whose multi-character folds
2613 * were long ago discovered to mess things up. The previous designs for
2614 * dealing with these involved assigning a special node for them. This
2615 * approach doesn't work, as evidenced by this example:
2616 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2617 * Both these fold to "sss", but if the pattern is parsed to create a node that
2618 * would match just the \xDF, it won't be able to handle the case where a
2619 * successful match would have to cross the node's boundary. The new approach
2620 * that hopefully generally solves the problem generates an EXACTFU_SS node
2623 * It turns out that there are problems with all multi-character folds, and not
2624 * just these three. Now the code is general, for all such cases, but the
2625 * three still have some special handling. The approach taken is:
2626 * 1) This routine examines each EXACTFish node that could contain multi-
2627 * character fold sequences. It returns in *min_subtract how much to
2628 * subtract from the the actual length of the string to get a real minimum
2629 * match length; it is 0 if there are no multi-char folds. This delta is
2630 * used by the caller to adjust the min length of the match, and the delta
2631 * between min and max, so that the optimizer doesn't reject these
2632 * possibilities based on size constraints.
2633 * 2) Certain of these sequences require special handling by the trie code,
2634 * so, if found, this code changes the joined node type to special ops:
2635 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2636 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2637 * is used for an EXACTFU node that contains at least one "ss" sequence in
2638 * it. For non-UTF-8 patterns and strings, this is the only case where
2639 * there is a possible fold length change. That means that a regular
2640 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2641 * with length changes, and so can be processed faster. regexec.c takes
2642 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2643 * pre-folded by regcomp.c. This saves effort in regex matching.
2644 * However, the pre-folding isn't done for non-UTF8 patterns because the
2645 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2646 * down by forcing the pattern into UTF8 unless necessary. Also what
2647 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2648 * possibilities for the non-UTF8 patterns are quite simple, except for
2649 * the sharp s. All the ones that don't involve a UTF-8 target string are
2650 * members of a fold-pair, and arrays are set up for all of them so that
2651 * the other member of the pair can be found quickly. Code elsewhere in
2652 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2653 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2654 * described in the next item.
2655 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2656 * 'ss' or not is not knowable at compile time. It will match iff the
2657 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2658 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2659 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
2660 * described in item 3). An assumption that the optimizer part of
2661 * regexec.c (probably unwittingly) makes is that a character in the
2662 * pattern corresponds to at most a single character in the target string.
2663 * (And I do mean character, and not byte here, unlike other parts of the
2664 * documentation that have never been updated to account for multibyte
2665 * Unicode.) This assumption is wrong only in this case, as all other
2666 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2667 * virtue of having this file pre-fold UTF-8 patterns. I'm
2668 * reluctant to try to change this assumption, so instead the code punts.
2669 * This routine examines EXACTF nodes for the sharp s, and returns a
2670 * boolean indicating whether or not the node is an EXACTF node that
2671 * contains a sharp s. When it is true, the caller sets a flag that later
2672 * causes the optimizer in this file to not set values for the floating
2673 * and fixed string lengths, and thus avoids the optimizer code in
2674 * regexec.c that makes the invalid assumption. Thus, there is no
2675 * optimization based on string lengths for EXACTF nodes that contain the
2676 * sharp s. This only happens for /id rules (which means the pattern
2680 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2681 if (PL_regkind[OP(scan)] == EXACT) \
2682 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2685 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, UV *min_subtract, bool *has_exactf_sharp_s, U32 flags,regnode *val, U32 depth) {
2686 /* Merge several consecutive EXACTish nodes into one. */
2687 regnode *n = regnext(scan);
2689 regnode *next = scan + NODE_SZ_STR(scan);
2693 regnode *stop = scan;
2694 GET_RE_DEBUG_FLAGS_DECL;
2696 PERL_UNUSED_ARG(depth);
2699 PERL_ARGS_ASSERT_JOIN_EXACT;
2700 #ifndef EXPERIMENTAL_INPLACESCAN
2701 PERL_UNUSED_ARG(flags);
2702 PERL_UNUSED_ARG(val);
2704 DEBUG_PEEP("join",scan,depth);
2706 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2707 * EXACT ones that are mergeable to the current one. */
2709 && (PL_regkind[OP(n)] == NOTHING
2710 || (stringok && OP(n) == OP(scan)))
2712 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2715 if (OP(n) == TAIL || n > next)
2717 if (PL_regkind[OP(n)] == NOTHING) {
2718 DEBUG_PEEP("skip:",n,depth);
2719 NEXT_OFF(scan) += NEXT_OFF(n);
2720 next = n + NODE_STEP_REGNODE;
2727 else if (stringok) {
2728 const unsigned int oldl = STR_LEN(scan);
2729 regnode * const nnext = regnext(n);
2731 /* XXX I (khw) kind of doubt that this works on platforms where
2732 * U8_MAX is above 255 because of lots of other assumptions */
2733 if (oldl + STR_LEN(n) > U8_MAX)
2736 DEBUG_PEEP("merg",n,depth);
2739 NEXT_OFF(scan) += NEXT_OFF(n);
2740 STR_LEN(scan) += STR_LEN(n);
2741 next = n + NODE_SZ_STR(n);
2742 /* Now we can overwrite *n : */
2743 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2751 #ifdef EXPERIMENTAL_INPLACESCAN
2752 if (flags && !NEXT_OFF(n)) {
2753 DEBUG_PEEP("atch", val, depth);
2754 if (reg_off_by_arg[OP(n)]) {
2755 ARG_SET(n, val - n);
2758 NEXT_OFF(n) = val - n;
2766 *has_exactf_sharp_s = FALSE;
2768 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2769 * can now analyze for sequences of problematic code points. (Prior to
2770 * this final joining, sequences could have been split over boundaries, and
2771 * hence missed). The sequences only happen in folding, hence for any
2772 * non-EXACT EXACTish node */
2773 if (OP(scan) != EXACT) {
2774 const U8 * const s0 = (U8*) STRING(scan);
2776 const U8 * const s_end = s0 + STR_LEN(scan);
2778 /* One pass is made over the node's string looking for all the
2779 * possibilities. to avoid some tests in the loop, there are two main
2780 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2784 /* Examine the string for a multi-character fold sequence. UTF-8
2785 * patterns have all characters pre-folded by the time this code is
2787 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2788 length sequence we are looking for is 2 */
2791 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2792 if (! len) { /* Not a multi-char fold: get next char */
2797 /* Nodes with 'ss' require special handling, except for EXACTFL
2798 * and EXACTFA for which there is no multi-char fold to this */
2799 if (len == 2 && *s == 's' && *(s+1) == 's'
2800 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2803 OP(scan) = EXACTFU_SS;
2806 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2807 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2808 COMBINING_DIAERESIS_UTF8
2809 COMBINING_ACUTE_ACCENT_UTF8,
2811 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2812 COMBINING_DIAERESIS_UTF8
2813 COMBINING_ACUTE_ACCENT_UTF8,
2818 /* These two folds require special handling by trie's, so
2819 * change the node type to indicate this. If EXACTFA and
2820 * EXACTFL were ever to be handled by trie's, this would
2821 * have to be changed. If this node has already been
2822 * changed to EXACTFU_SS in this loop, leave it as is. (I
2823 * (khw) think it doesn't matter in regexec.c for UTF
2824 * patterns, but no need to change it */
2825 if (OP(scan) == EXACTFU) {
2826 OP(scan) = EXACTFU_TRICKYFOLD;
2830 else { /* Here is a generic multi-char fold. */
2831 const U8* multi_end = s + len;
2833 /* Count how many characters in it. In the case of /l and
2834 * /aa, no folds which contain ASCII code points are
2835 * allowed, so check for those, and skip if found. (In
2836 * EXACTFL, no folds are allowed to any Latin1 code point,
2837 * not just ASCII. But there aren't any of these
2838 * currently, nor ever likely, so don't take the time to
2839 * test for them. The code that generates the
2840 * is_MULTI_foo() macros croaks should one actually get put
2841 * into Unicode .) */
2842 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2843 count = utf8_length(s, multi_end);
2847 while (s < multi_end) {
2850 goto next_iteration;
2860 /* The delta is how long the sequence is minus 1 (1 is how long
2861 * the character that folds to the sequence is) */
2862 *min_subtract += count - 1;
2866 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2868 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2869 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2870 * nodes can't have multi-char folds to this range (and there are
2871 * no existing ones in the upper latin1 range). In the EXACTF
2872 * case we look also for the sharp s, which can be in the final
2873 * position. Otherwise we can stop looking 1 byte earlier because
2874 * have to find at least two characters for a multi-fold */
2875 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2877 /* The below is perhaps overboard, but this allows us to save a
2878 * test each time through the loop at the expense of a mask. This
2879 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2880 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2881 * are 64. This uses an exclusive 'or' to find that bit and then
2882 * inverts it to form a mask, with just a single 0, in the bit
2883 * position where 'S' and 's' differ. */
2884 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2885 const U8 s_masked = 's' & S_or_s_mask;
2888 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2889 if (! len) { /* Not a multi-char fold. */
2890 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2892 *has_exactf_sharp_s = TRUE;
2899 && ((*s & S_or_s_mask) == s_masked)
2900 && ((*(s+1) & S_or_s_mask) == s_masked))
2903 /* EXACTF nodes need to know that the minimum length
2904 * changed so that a sharp s in the string can match this
2905 * ss in the pattern, but they remain EXACTF nodes, as they
2906 * won't match this unless the target string is is UTF-8,
2907 * which we don't know until runtime */
2908 if (OP(scan) != EXACTF) {
2909 OP(scan) = EXACTFU_SS;
2913 *min_subtract += len - 1;
2920 /* Allow dumping but overwriting the collection of skipped
2921 * ops and/or strings with fake optimized ops */
2922 n = scan + NODE_SZ_STR(scan);
2930 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2934 /* REx optimizer. Converts nodes into quicker variants "in place".
2935 Finds fixed substrings. */
2937 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2938 to the position after last scanned or to NULL. */
2940 #define INIT_AND_WITHP \
2941 assert(!and_withp); \
2942 Newx(and_withp,1,struct regnode_charclass_class); \
2943 SAVEFREEPV(and_withp)
2945 /* this is a chain of data about sub patterns we are processing that
2946 need to be handled separately/specially in study_chunk. Its so
2947 we can simulate recursion without losing state. */
2949 typedef struct scan_frame {
2950 regnode *last; /* last node to process in this frame */
2951 regnode *next; /* next node to process when last is reached */
2952 struct scan_frame *prev; /*previous frame*/
2953 I32 stop; /* what stopparen do we use */
2957 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2959 #define CASE_SYNST_FNC(nAmE) \
2961 if (flags & SCF_DO_STCLASS_AND) { \
2962 for (value = 0; value < 256; value++) \
2963 if (!is_ ## nAmE ## _cp(value)) \
2964 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2967 for (value = 0; value < 256; value++) \
2968 if (is_ ## nAmE ## _cp(value)) \
2969 ANYOF_BITMAP_SET(data->start_class, value); \
2973 if (flags & SCF_DO_STCLASS_AND) { \
2974 for (value = 0; value < 256; value++) \
2975 if (is_ ## nAmE ## _cp(value)) \
2976 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2979 for (value = 0; value < 256; value++) \
2980 if (!is_ ## nAmE ## _cp(value)) \
2981 ANYOF_BITMAP_SET(data->start_class, value); \
2988 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2989 I32 *minlenp, I32 *deltap,
2994 struct regnode_charclass_class *and_withp,
2995 U32 flags, U32 depth)
2996 /* scanp: Start here (read-write). */
2997 /* deltap: Write maxlen-minlen here. */
2998 /* last: Stop before this one. */
2999 /* data: string data about the pattern */
3000 /* stopparen: treat close N as END */
3001 /* recursed: which subroutines have we recursed into */
3002 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3005 I32 min = 0; /* There must be at least this number of characters to match */
3007 regnode *scan = *scanp, *next;
3009 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3010 int is_inf_internal = 0; /* The studied chunk is infinite */
3011 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3012 scan_data_t data_fake;
3013 SV *re_trie_maxbuff = NULL;
3014 regnode *first_non_open = scan;
3015 I32 stopmin = I32_MAX;
3016 scan_frame *frame = NULL;
3017 GET_RE_DEBUG_FLAGS_DECL;
3019 PERL_ARGS_ASSERT_STUDY_CHUNK;
3022 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3026 while (first_non_open && OP(first_non_open) == OPEN)
3027 first_non_open=regnext(first_non_open);
3032 while ( scan && OP(scan) != END && scan < last ){
3033 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3034 node length to get a real minimum (because
3035 the folded version may be shorter) */
3036 bool has_exactf_sharp_s = FALSE;
3037 /* Peephole optimizer: */
3038 DEBUG_STUDYDATA("Peep:", data,depth);
3039 DEBUG_PEEP("Peep",scan,depth);
3041 /* Its not clear to khw or hv why this is done here, and not in the
3042 * clauses that deal with EXACT nodes. khw's guess is that it's
3043 * because of a previous design */
3044 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3046 /* Follow the next-chain of the current node and optimize
3047 away all the NOTHINGs from it. */
3048 if (OP(scan) != CURLYX) {
3049 const int max = (reg_off_by_arg[OP(scan)]
3051 /* I32 may be smaller than U16 on CRAYs! */
3052 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3053 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3057 /* Skip NOTHING and LONGJMP. */
3058 while ((n = regnext(n))
3059 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3060 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3061 && off + noff < max)
3063 if (reg_off_by_arg[OP(scan)])
3066 NEXT_OFF(scan) = off;
3071 /* The principal pseudo-switch. Cannot be a switch, since we
3072 look into several different things. */
3073 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3074 || OP(scan) == IFTHEN) {
3075 next = regnext(scan);
3077 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3079 if (OP(next) == code || code == IFTHEN) {
3080 /* NOTE - There is similar code to this block below for handling
3081 TRIE nodes on a re-study. If you change stuff here check there
3083 I32 max1 = 0, min1 = I32_MAX, num = 0;
3084 struct regnode_charclass_class accum;
3085 regnode * const startbranch=scan;
3087 if (flags & SCF_DO_SUBSTR)
3088 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3089 if (flags & SCF_DO_STCLASS)
3090 cl_init_zero(pRExC_state, &accum);
3092 while (OP(scan) == code) {
3093 I32 deltanext, minnext, f = 0, fake;
3094 struct regnode_charclass_class this_class;
3097 data_fake.flags = 0;
3099 data_fake.whilem_c = data->whilem_c;
3100 data_fake.last_closep = data->last_closep;
3103 data_fake.last_closep = &fake;
3105 data_fake.pos_delta = delta;
3106 next = regnext(scan);
3107 scan = NEXTOPER(scan);
3109 scan = NEXTOPER(scan);
3110 if (flags & SCF_DO_STCLASS) {
3111 cl_init(pRExC_state, &this_class);
3112 data_fake.start_class = &this_class;
3113 f = SCF_DO_STCLASS_AND;
3115 if (flags & SCF_WHILEM_VISITED_POS)
3116 f |= SCF_WHILEM_VISITED_POS;
3118 /* we suppose the run is continuous, last=next...*/
3119 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3121 stopparen, recursed, NULL, f,depth+1);
3124 if (max1 < minnext + deltanext)
3125 max1 = minnext + deltanext;
3126 if (deltanext == I32_MAX)
3127 is_inf = is_inf_internal = 1;
3129 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3131 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3132 if ( stopmin > minnext)
3133 stopmin = min + min1;
3134 flags &= ~SCF_DO_SUBSTR;
3136 data->flags |= SCF_SEEN_ACCEPT;
3139 if (data_fake.flags & SF_HAS_EVAL)
3140 data->flags |= SF_HAS_EVAL;
3141 data->whilem_c = data_fake.whilem_c;
3143 if (flags & SCF_DO_STCLASS)
3144 cl_or(pRExC_state, &accum, &this_class);
3146 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3148 if (flags & SCF_DO_SUBSTR) {
3149 data->pos_min += min1;
3150 data->pos_delta += max1 - min1;
3151 if (max1 != min1 || is_inf)
3152 data->longest = &(data->longest_float);
3155 delta += max1 - min1;
3156 if (flags & SCF_DO_STCLASS_OR) {
3157 cl_or(pRExC_state, data->start_class, &accum);
3159 cl_and(data->start_class, and_withp);
3160 flags &= ~SCF_DO_STCLASS;
3163 else if (flags & SCF_DO_STCLASS_AND) {
3165 cl_and(data->start_class, &accum);
3166 flags &= ~SCF_DO_STCLASS;
3169 /* Switch to OR mode: cache the old value of
3170 * data->start_class */
3172 StructCopy(data->start_class, and_withp,
3173 struct regnode_charclass_class);
3174 flags &= ~SCF_DO_STCLASS_AND;
3175 StructCopy(&accum, data->start_class,
3176 struct regnode_charclass_class);
3177 flags |= SCF_DO_STCLASS_OR;
3178 data->start_class->flags |= ANYOF_EOS;
3182 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3185 Assuming this was/is a branch we are dealing with: 'scan' now
3186 points at the item that follows the branch sequence, whatever
3187 it is. We now start at the beginning of the sequence and look
3194 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3196 If we can find such a subsequence we need to turn the first
3197 element into a trie and then add the subsequent branch exact
3198 strings to the trie.
3202 1. patterns where the whole set of branches can be converted.
3204 2. patterns where only a subset can be converted.
3206 In case 1 we can replace the whole set with a single regop
3207 for the trie. In case 2 we need to keep the start and end
3210 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3211 becomes BRANCH TRIE; BRANCH X;
3213 There is an additional case, that being where there is a
3214 common prefix, which gets split out into an EXACT like node
3215 preceding the TRIE node.
3217 If x(1..n)==tail then we can do a simple trie, if not we make
3218 a "jump" trie, such that when we match the appropriate word
3219 we "jump" to the appropriate tail node. Essentially we turn
3220 a nested if into a case structure of sorts.
3225 if (!re_trie_maxbuff) {
3226 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3227 if (!SvIOK(re_trie_maxbuff))
3228 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3230 if ( SvIV(re_trie_maxbuff)>=0 ) {
3232 regnode *first = (regnode *)NULL;
3233 regnode *last = (regnode *)NULL;
3234 regnode *tail = scan;
3239 SV * const mysv = sv_newmortal(); /* for dumping */
3241 /* var tail is used because there may be a TAIL
3242 regop in the way. Ie, the exacts will point to the
3243 thing following the TAIL, but the last branch will
3244 point at the TAIL. So we advance tail. If we
3245 have nested (?:) we may have to move through several
3249 while ( OP( tail ) == TAIL ) {
3250 /* this is the TAIL generated by (?:) */
3251 tail = regnext( tail );
3255 DEBUG_TRIE_COMPILE_r({
3256 regprop(RExC_rx, mysv, tail );
3257 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3258 (int)depth * 2 + 2, "",
3259 "Looking for TRIE'able sequences. Tail node is: ",
3260 SvPV_nolen_const( mysv )
3266 Step through the branches
3267 cur represents each branch,
3268 noper is the first thing to be matched as part of that branch
3269 noper_next is the regnext() of that node.
3271 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3272 via a "jump trie" but we also support building with NOJUMPTRIE,
3273 which restricts the trie logic to structures like /FOO|BAR/.
3275 If noper is a trieable nodetype then the branch is a possible optimization
3276 target. If we are building under NOJUMPTRIE then we require that noper_next
3277 is the same as scan (our current position in the regex program).
3279 Once we have two or more consecutive such branches we can create a
3280 trie of the EXACT's contents and stitch it in place into the program.
3282 If the sequence represents all of the branches in the alternation we
3283 replace the entire thing with a single TRIE node.
3285 Otherwise when it is a subsequence we need to stitch it in place and
3286 replace only the relevant branches. This means the first branch has
3287 to remain as it is used by the alternation logic, and its next pointer,
3288 and needs to be repointed at the item on the branch chain following
3289 the last branch we have optimized away.
3291 This could be either a BRANCH, in which case the subsequence is internal,
3292 or it could be the item following the branch sequence in which case the
3293 subsequence is at the end (which does not necessarily mean the first node
3294 is the start of the alternation).
3296 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3299 ----------------+-----------
3303 EXACTFU_SS | EXACTFU
3304 EXACTFU_TRICKYFOLD | EXACTFU
3309 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3310 ( EXACT == (X) ) ? EXACT : \
3311 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3314 /* dont use tail as the end marker for this traverse */
3315 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3316 regnode * const noper = NEXTOPER( cur );
3317 U8 noper_type = OP( noper );
3318 U8 noper_trietype = TRIE_TYPE( noper_type );
3319 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3320 regnode * const noper_next = regnext( noper );
3321 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3322 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3325 DEBUG_TRIE_COMPILE_r({
3326 regprop(RExC_rx, mysv, cur);
3327 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3328 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3330 regprop(RExC_rx, mysv, noper);
3331 PerlIO_printf( Perl_debug_log, " -> %s",
3332 SvPV_nolen_const(mysv));
3335 regprop(RExC_rx, mysv, noper_next );
3336 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3337 SvPV_nolen_const(mysv));
3339 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3340 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3341 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3345 /* Is noper a trieable nodetype that can be merged with the
3346 * current trie (if there is one)? */
3350 ( noper_trietype == NOTHING)
3351 || ( trietype == NOTHING )
3352 || ( trietype == noper_trietype )
3355 && noper_next == tail
3359 /* Handle mergable triable node
3360 * Either we are the first node in a new trieable sequence,
3361 * in which case we do some bookkeeping, otherwise we update
3362 * the end pointer. */
3365 if ( noper_trietype == NOTHING ) {
3366 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3367 regnode * const noper_next = regnext( noper );
3368 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3369 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3372 if ( noper_next_trietype ) {
3373 trietype = noper_next_trietype;
3374 } else if (noper_next_type) {
3375 /* a NOTHING regop is 1 regop wide. We need at least two
3376 * for a trie so we can't merge this in */
3380 trietype = noper_trietype;
3383 if ( trietype == NOTHING )
3384 trietype = noper_trietype;
3389 } /* end handle mergable triable node */
3391 /* handle unmergable node -
3392 * noper may either be a triable node which can not be tried
3393 * together with the current trie, or a non triable node */
3395 /* If last is set and trietype is not NOTHING then we have found
3396 * at least two triable branch sequences in a row of a similar
3397 * trietype so we can turn them into a trie. If/when we
3398 * allow NOTHING to start a trie sequence this condition will be
3399 * required, and it isn't expensive so we leave it in for now. */
3400 if ( trietype && trietype != NOTHING )
3401 make_trie( pRExC_state,
3402 startbranch, first, cur, tail, count,
3403 trietype, depth+1 );
3404 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3408 && noper_next == tail
3411 /* noper is triable, so we can start a new trie sequence */
3414 trietype = noper_trietype;
3416 /* if we already saw a first but the current node is not triable then we have
3417 * to reset the first information. */
3422 } /* end handle unmergable node */
3423 } /* loop over branches */
3424 DEBUG_TRIE_COMPILE_r({
3425 regprop(RExC_rx, mysv, cur);
3426 PerlIO_printf( Perl_debug_log,
3427 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3428 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3431 if ( last && trietype ) {
3432 if ( trietype != NOTHING ) {
3433 /* the last branch of the sequence was part of a trie,
3434 * so we have to construct it here outside of the loop
3436 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3437 #ifdef TRIE_STUDY_OPT
3438 if ( ((made == MADE_EXACT_TRIE &&
3439 startbranch == first)
3440 || ( first_non_open == first )) &&
3442 flags |= SCF_TRIE_RESTUDY;
3443 if ( startbranch == first
3446 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3451 /* at this point we know whatever we have is a NOTHING sequence/branch
3452 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3454 if ( startbranch == first ) {
3456 /* the entire thing is a NOTHING sequence, something like this:
3457 * (?:|) So we can turn it into a plain NOTHING op. */
3458 DEBUG_TRIE_COMPILE_r({
3459 regprop(RExC_rx, mysv, cur);
3460 PerlIO_printf( Perl_debug_log,
3461 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3462 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3465 OP(startbranch)= NOTHING;
3466 NEXT_OFF(startbranch)= tail - startbranch;
3467 for ( opt= startbranch + 1; opt < tail ; opt++ )
3471 } /* end if ( last) */
3472 } /* TRIE_MAXBUF is non zero */
3477 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3478 scan = NEXTOPER(NEXTOPER(scan));
3479 } else /* single branch is optimized. */
3480 scan = NEXTOPER(scan);
3482 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3483 scan_frame *newframe = NULL;
3488 if (OP(scan) != SUSPEND) {
3489 /* set the pointer */
3490 if (OP(scan) == GOSUB) {
3492 RExC_recurse[ARG2L(scan)] = scan;
3493 start = RExC_open_parens[paren-1];
3494 end = RExC_close_parens[paren-1];
3497 start = RExC_rxi->program + 1;
3501 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3502 SAVEFREEPV(recursed);
3504 if (!PAREN_TEST(recursed,paren+1)) {
3505 PAREN_SET(recursed,paren+1);
3506 Newx(newframe,1,scan_frame);
3508 if (flags & SCF_DO_SUBSTR) {
3509 SCAN_COMMIT(pRExC_state,data,minlenp);
3510 data->longest = &(data->longest_float);
3512 is_inf = is_inf_internal = 1;
3513 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3514 cl_anything(pRExC_state, data->start_class);
3515 flags &= ~SCF_DO_STCLASS;
3518 Newx(newframe,1,scan_frame);
3521 end = regnext(scan);
3526 SAVEFREEPV(newframe);
3527 newframe->next = regnext(scan);
3528 newframe->last = last;
3529 newframe->stop = stopparen;
3530 newframe->prev = frame;
3540 else if (OP(scan) == EXACT) {
3541 I32 l = STR_LEN(scan);
3544 const U8 * const s = (U8*)STRING(scan);
3545 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3546 l = utf8_length(s, s + l);
3548 uc = *((U8*)STRING(scan));
3551 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3552 /* The code below prefers earlier match for fixed
3553 offset, later match for variable offset. */
3554 if (data->last_end == -1) { /* Update the start info. */
3555 data->last_start_min = data->pos_min;
3556 data->last_start_max = is_inf
3557 ? I32_MAX : data->pos_min + data->pos_delta;
3559 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3561 SvUTF8_on(data->last_found);
3563 SV * const sv = data->last_found;
3564 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3565 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3566 if (mg && mg->mg_len >= 0)
3567 mg->mg_len += utf8_length((U8*)STRING(scan),
3568 (U8*)STRING(scan)+STR_LEN(scan));
3570 data->last_end = data->pos_min + l;
3571 data->pos_min += l; /* As in the first entry. */
3572 data->flags &= ~SF_BEFORE_EOL;
3574 if (flags & SCF_DO_STCLASS_AND) {
3575 /* Check whether it is compatible with what we know already! */
3579 /* If compatible, we or it in below. It is compatible if is
3580 * in the bitmp and either 1) its bit or its fold is set, or 2)
3581 * it's for a locale. Even if there isn't unicode semantics
3582 * here, at runtime there may be because of matching against a
3583 * utf8 string, so accept a possible false positive for
3584 * latin1-range folds */
3586 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3587 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3588 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3589 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3594 ANYOF_CLASS_ZERO(data->start_class);
3595 ANYOF_BITMAP_ZERO(data->start_class);
3597 ANYOF_BITMAP_SET(data->start_class, uc);
3598 else if (uc >= 0x100) {
3601 /* Some Unicode code points fold to the Latin1 range; as
3602 * XXX temporary code, instead of figuring out if this is
3603 * one, just assume it is and set all the start class bits
3604 * that could be some such above 255 code point's fold
3605 * which will generate fals positives. As the code
3606 * elsewhere that does compute the fold settles down, it
3607 * can be extracted out and re-used here */
3608 for (i = 0; i < 256; i++){
3609 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3610 ANYOF_BITMAP_SET(data->start_class, i);
3614 data->start_class->flags &= ~ANYOF_EOS;
3616 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3618 else if (flags & SCF_DO_STCLASS_OR) {
3619 /* false positive possible if the class is case-folded */
3621 ANYOF_BITMAP_SET(data->start_class, uc);
3623 data->start_class->flags |= ANYOF_UNICODE_ALL;
3624 data->start_class->flags &= ~ANYOF_EOS;
3625 cl_and(data->start_class, and_withp);
3627 flags &= ~SCF_DO_STCLASS;
3629 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3630 I32 l = STR_LEN(scan);
3631 UV uc = *((U8*)STRING(scan));
3633 /* Search for fixed substrings supports EXACT only. */
3634 if (flags & SCF_DO_SUBSTR) {
3636 SCAN_COMMIT(pRExC_state, data, minlenp);
3639 const U8 * const s = (U8 *)STRING(scan);
3640 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3641 l = utf8_length(s, s + l);
3643 if (has_exactf_sharp_s) {
3644 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3646 min += l - min_subtract;
3648 delta += min_subtract;
3649 if (flags & SCF_DO_SUBSTR) {
3650 data->pos_min += l - min_subtract;
3651 if (data->pos_min < 0) {
3654 data->pos_delta += min_subtract;
3656 data->longest = &(data->longest_float);
3659 if (flags & SCF_DO_STCLASS_AND) {
3660 /* Check whether it is compatible with what we know already! */
3663 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3664 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3665 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3669 ANYOF_CLASS_ZERO(data->start_class);
3670 ANYOF_BITMAP_ZERO(data->start_class);
3672 ANYOF_BITMAP_SET(data->start_class, uc);
3673 data->start_class->flags &= ~ANYOF_EOS;
3674 if (OP(scan) == EXACTFL) {
3675 /* XXX This set is probably no longer necessary, and
3676 * probably wrong as LOCALE now is on in the initial
3678 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3682 /* Also set the other member of the fold pair. In case
3683 * that unicode semantics is called for at runtime, use
3684 * the full latin1 fold. (Can't do this for locale,
3685 * because not known until runtime) */
3686 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3688 /* All other (EXACTFL handled above) folds except under
3689 * /iaa that include s, S, and sharp_s also may include
3691 if (OP(scan) != EXACTFA) {
3692 if (uc == 's' || uc == 'S') {
3693 ANYOF_BITMAP_SET(data->start_class,
3694 LATIN_SMALL_LETTER_SHARP_S);
3696 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3697 ANYOF_BITMAP_SET(data->start_class, 's');
3698 ANYOF_BITMAP_SET(data->start_class, 'S');
3703 else if (uc >= 0x100) {
3705 for (i = 0; i < 256; i++){
3706 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3707 ANYOF_BITMAP_SET(data->start_class, i);
3712 else if (flags & SCF_DO_STCLASS_OR) {
3713 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3714 /* false positive possible if the class is case-folded.
3715 Assume that the locale settings are the same... */
3717 ANYOF_BITMAP_SET(data->start_class, uc);
3718 if (OP(scan) != EXACTFL) {
3720 /* And set the other member of the fold pair, but
3721 * can't do that in locale because not known until
3723 ANYOF_BITMAP_SET(data->start_class,
3724 PL_fold_latin1[uc]);
3726 /* All folds except under /iaa that include s, S,
3727 * and sharp_s also may include the others */
3728 if (OP(scan) != EXACTFA) {
3729 if (uc == 's' || uc == 'S') {
3730 ANYOF_BITMAP_SET(data->start_class,
3731 LATIN_SMALL_LETTER_SHARP_S);
3733 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3734 ANYOF_BITMAP_SET(data->start_class, 's');
3735 ANYOF_BITMAP_SET(data->start_class, 'S');
3740 data->start_class->flags &= ~ANYOF_EOS;
3742 cl_and(data->start_class, and_withp);
3744 flags &= ~SCF_DO_STCLASS;
3746 else if (REGNODE_VARIES(OP(scan))) {
3747 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3748 I32 f = flags, pos_before = 0;
3749 regnode * const oscan = scan;
3750 struct regnode_charclass_class this_class;
3751 struct regnode_charclass_class *oclass = NULL;
3752 I32 next_is_eval = 0;
3754 switch (PL_regkind[OP(scan)]) {
3755 case WHILEM: /* End of (?:...)* . */
3756 scan = NEXTOPER(scan);
3759 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3760 next = NEXTOPER(scan);
3761 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3763 maxcount = REG_INFTY;
3764 next = regnext(scan);
3765 scan = NEXTOPER(scan);
3769 if (flags & SCF_DO_SUBSTR)
3774 if (flags & SCF_DO_STCLASS) {
3776 maxcount = REG_INFTY;
3777 next = regnext(scan);
3778 scan = NEXTOPER(scan);
3781 is_inf = is_inf_internal = 1;
3782 scan = regnext(scan);
3783 if (flags & SCF_DO_SUBSTR) {
3784 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3785 data->longest = &(data->longest_float);
3787 goto optimize_curly_tail;
3789 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3790 && (scan->flags == stopparen))
3795 mincount = ARG1(scan);
3796 maxcount = ARG2(scan);
3798 next = regnext(scan);
3799 if (OP(scan) == CURLYX) {
3800 I32 lp = (data ? *(data->last_closep) : 0);
3801 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3803 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3804 next_is_eval = (OP(scan) == EVAL);
3806 if (flags & SCF_DO_SUBSTR) {
3807 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3808 pos_before = data->pos_min;
3812 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3814 data->flags |= SF_IS_INF;
3816 if (flags & SCF_DO_STCLASS) {
3817 cl_init(pRExC_state, &this_class);
3818 oclass = data->start_class;
3819 data->start_class = &this_class;
3820 f |= SCF_DO_STCLASS_AND;
3821 f &= ~SCF_DO_STCLASS_OR;
3823 /* Exclude from super-linear cache processing any {n,m}
3824 regops for which the combination of input pos and regex
3825 pos is not enough information to determine if a match
3828 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3829 regex pos at the \s*, the prospects for a match depend not
3830 only on the input position but also on how many (bar\s*)
3831 repeats into the {4,8} we are. */