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"
94 #define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
102 # if defined(BUGGY_MSC6)
103 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
104 # pragma optimize("a",off)
105 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
106 # pragma optimize("w",on )
107 # endif /* BUGGY_MSC6 */
111 #define STATIC static
115 typedef struct RExC_state_t {
116 U32 flags; /* RXf_* are we folding, multilining? */
117 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
118 char *precomp; /* uncompiled string. */
119 REGEXP *rx_sv; /* The SV that is the regexp. */
120 regexp *rx; /* perl core regexp structure */
121 regexp_internal *rxi; /* internal data for regexp object pprivate field */
122 char *start; /* Start of input for compile */
123 char *end; /* End of input for compile */
124 char *parse; /* Input-scan pointer. */
125 I32 whilem_seen; /* number of WHILEM in this expr */
126 regnode *emit_start; /* Start of emitted-code area */
127 regnode *emit_bound; /* First regnode outside of the allocated space */
128 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
129 I32 naughty; /* How bad is this pattern? */
130 I32 sawback; /* Did we see \1, ...? */
132 I32 size; /* Code size. */
133 I32 npar; /* Capture buffer count, (OPEN). */
134 I32 cpar; /* Capture buffer count, (CLOSE). */
135 I32 nestroot; /* root parens we are in - used by accept */
138 regnode **open_parens; /* pointers to open parens */
139 regnode **close_parens; /* pointers to close parens */
140 regnode *opend; /* END node in program */
141 I32 utf8; /* whether the pattern is utf8 or not */
142 I32 orig_utf8; /* whether the pattern was originally in utf8 */
143 /* XXX use this for future optimisation of case
144 * where pattern must be upgraded to utf8. */
145 I32 uni_semantics; /* If a d charset modifier should use unicode
146 rules, even if the pattern is not in
148 HV *paren_names; /* Paren names */
150 regnode **recurse; /* Recurse regops */
151 I32 recurse_count; /* Number of recurse regops */
154 I32 override_recoding;
155 struct reg_code_block *code_blocks; /* positions of literal (?{})
157 int num_code_blocks; /* size of code_blocks[] */
158 int code_index; /* next code_blocks[] slot */
160 char *starttry; /* -Dr: where regtry was called. */
161 #define RExC_starttry (pRExC_state->starttry)
163 SV *runtime_code_qr; /* qr with the runtime code blocks */
165 const char *lastparse;
167 AV *paren_name_list; /* idx -> name */
168 #define RExC_lastparse (pRExC_state->lastparse)
169 #define RExC_lastnum (pRExC_state->lastnum)
170 #define RExC_paren_name_list (pRExC_state->paren_name_list)
174 #define RExC_flags (pRExC_state->flags)
175 #define RExC_pm_flags (pRExC_state->pm_flags)
176 #define RExC_precomp (pRExC_state->precomp)
177 #define RExC_rx_sv (pRExC_state->rx_sv)
178 #define RExC_rx (pRExC_state->rx)
179 #define RExC_rxi (pRExC_state->rxi)
180 #define RExC_start (pRExC_state->start)
181 #define RExC_end (pRExC_state->end)
182 #define RExC_parse (pRExC_state->parse)
183 #define RExC_whilem_seen (pRExC_state->whilem_seen)
184 #ifdef RE_TRACK_PATTERN_OFFSETS
185 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
187 #define RExC_emit (pRExC_state->emit)
188 #define RExC_emit_start (pRExC_state->emit_start)
189 #define RExC_emit_bound (pRExC_state->emit_bound)
190 #define RExC_naughty (pRExC_state->naughty)
191 #define RExC_sawback (pRExC_state->sawback)
192 #define RExC_seen (pRExC_state->seen)
193 #define RExC_size (pRExC_state->size)
194 #define RExC_npar (pRExC_state->npar)
195 #define RExC_nestroot (pRExC_state->nestroot)
196 #define RExC_extralen (pRExC_state->extralen)
197 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
198 #define RExC_utf8 (pRExC_state->utf8)
199 #define RExC_uni_semantics (pRExC_state->uni_semantics)
200 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
201 #define RExC_open_parens (pRExC_state->open_parens)
202 #define RExC_close_parens (pRExC_state->close_parens)
203 #define RExC_opend (pRExC_state->opend)
204 #define RExC_paren_names (pRExC_state->paren_names)
205 #define RExC_recurse (pRExC_state->recurse)
206 #define RExC_recurse_count (pRExC_state->recurse_count)
207 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
208 #define RExC_contains_locale (pRExC_state->contains_locale)
209 #define RExC_override_recoding (pRExC_state->override_recoding)
212 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
213 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
214 ((*s) == '{' && regcurly(s)))
217 #undef SPSTART /* dratted cpp namespace... */
220 * Flags to be passed up and down.
222 #define WORST 0 /* Worst case. */
223 #define HASWIDTH 0x01 /* Known to match non-null strings. */
225 /* Simple enough to be STAR/PLUS operand; in an EXACT node must be a single
226 * character, and if utf8, must be invariant. Note that this is not the same
227 * thing as REGNODE_SIMPLE */
229 #define SPSTART 0x04 /* Starts with * or +. */
230 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
231 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
233 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
235 /* whether trie related optimizations are enabled */
236 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
237 #define TRIE_STUDY_OPT
238 #define FULL_TRIE_STUDY
244 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
245 #define PBITVAL(paren) (1 << ((paren) & 7))
246 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
247 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
248 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
250 /* If not already in utf8, do a longjmp back to the beginning */
251 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
252 #define REQUIRE_UTF8 STMT_START { \
253 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
256 /* About scan_data_t.
258 During optimisation we recurse through the regexp program performing
259 various inplace (keyhole style) optimisations. In addition study_chunk
260 and scan_commit populate this data structure with information about
261 what strings MUST appear in the pattern. We look for the longest
262 string that must appear at a fixed location, and we look for the
263 longest string that may appear at a floating location. So for instance
268 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
269 strings (because they follow a .* construct). study_chunk will identify
270 both FOO and BAR as being the longest fixed and floating strings respectively.
272 The strings can be composites, for instance
276 will result in a composite fixed substring 'foo'.
278 For each string some basic information is maintained:
280 - offset or min_offset
281 This is the position the string must appear at, or not before.
282 It also implicitly (when combined with minlenp) tells us how many
283 characters must match before the string we are searching for.
284 Likewise when combined with minlenp and the length of the string it
285 tells us how many characters must appear after the string we have
289 Only used for floating strings. This is the rightmost point that
290 the string can appear at. If set to I32 max it indicates that the
291 string can occur infinitely far to the right.
294 A pointer to the minimum length of the pattern that the string
295 was found inside. This is important as in the case of positive
296 lookahead or positive lookbehind we can have multiple patterns
301 The minimum length of the pattern overall is 3, the minimum length
302 of the lookahead part is 3, but the minimum length of the part that
303 will actually match is 1. So 'FOO's minimum length is 3, but the
304 minimum length for the F is 1. This is important as the minimum length
305 is used to determine offsets in front of and behind the string being
306 looked for. Since strings can be composites this is the length of the
307 pattern at the time it was committed with a scan_commit. Note that
308 the length is calculated by study_chunk, so that the minimum lengths
309 are not known until the full pattern has been compiled, thus the
310 pointer to the value.
314 In the case of lookbehind the string being searched for can be
315 offset past the start point of the final matching string.
316 If this value was just blithely removed from the min_offset it would
317 invalidate some of the calculations for how many chars must match
318 before or after (as they are derived from min_offset and minlen and
319 the length of the string being searched for).
320 When the final pattern is compiled and the data is moved from the
321 scan_data_t structure into the regexp structure the information
322 about lookbehind is factored in, with the information that would
323 have been lost precalculated in the end_shift field for the
326 The fields pos_min and pos_delta are used to store the minimum offset
327 and the delta to the maximum offset at the current point in the pattern.
331 typedef struct scan_data_t {
332 /*I32 len_min; unused */
333 /*I32 len_delta; unused */
337 I32 last_end; /* min value, <0 unless valid. */
340 SV **longest; /* Either &l_fixed, or &l_float. */
341 SV *longest_fixed; /* longest fixed string found in pattern */
342 I32 offset_fixed; /* offset where it starts */
343 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
344 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
345 SV *longest_float; /* longest floating string found in pattern */
346 I32 offset_float_min; /* earliest point in string it can appear */
347 I32 offset_float_max; /* latest point in string it can appear */
348 I32 *minlen_float; /* pointer to the minlen relevant to the string */
349 I32 lookbehind_float; /* is the position of the string modified by LB */
353 struct regnode_charclass_class *start_class;
357 * Forward declarations for pregcomp()'s friends.
360 static const scan_data_t zero_scan_data =
361 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
363 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
364 #define SF_BEFORE_SEOL 0x0001
365 #define SF_BEFORE_MEOL 0x0002
366 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
367 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
370 # define SF_FIX_SHIFT_EOL (0+2)
371 # define SF_FL_SHIFT_EOL (0+4)
373 # define SF_FIX_SHIFT_EOL (+2)
374 # define SF_FL_SHIFT_EOL (+4)
377 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
378 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
380 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
381 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
382 #define SF_IS_INF 0x0040
383 #define SF_HAS_PAR 0x0080
384 #define SF_IN_PAR 0x0100
385 #define SF_HAS_EVAL 0x0200
386 #define SCF_DO_SUBSTR 0x0400
387 #define SCF_DO_STCLASS_AND 0x0800
388 #define SCF_DO_STCLASS_OR 0x1000
389 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
390 #define SCF_WHILEM_VISITED_POS 0x2000
392 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
393 #define SCF_SEEN_ACCEPT 0x8000
395 #define UTF cBOOL(RExC_utf8)
397 /* The enums for all these are ordered so things work out correctly */
398 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
399 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
400 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
401 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
402 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
403 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
404 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
406 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
408 #define OOB_NAMEDCLASS -1
410 /* There is no code point that is out-of-bounds, so this is problematic. But
411 * its only current use is to initialize a variable that is always set before
413 #define OOB_UNICODE 0xDEADBEEF
415 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
416 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
419 /* length of regex to show in messages that don't mark a position within */
420 #define RegexLengthToShowInErrorMessages 127
423 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
424 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
425 * op/pragma/warn/regcomp.
427 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
428 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
430 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
433 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
434 * arg. Show regex, up to a maximum length. If it's too long, chop and add
437 #define _FAIL(code) STMT_START { \
438 const char *ellipses = ""; \
439 IV len = RExC_end - RExC_precomp; \
442 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
443 if (len > RegexLengthToShowInErrorMessages) { \
444 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
445 len = RegexLengthToShowInErrorMessages - 10; \
451 #define FAIL(msg) _FAIL( \
452 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
453 msg, (int)len, RExC_precomp, ellipses))
455 #define FAIL2(msg,arg) _FAIL( \
456 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
457 arg, (int)len, RExC_precomp, ellipses))
460 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
462 #define Simple_vFAIL(m) STMT_START { \
463 const IV offset = RExC_parse - RExC_precomp; \
464 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
465 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
469 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
471 #define vFAIL(m) STMT_START { \
473 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
478 * Like Simple_vFAIL(), but accepts two arguments.
480 #define Simple_vFAIL2(m,a1) STMT_START { \
481 const IV offset = RExC_parse - RExC_precomp; \
482 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
483 (int)offset, RExC_precomp, RExC_precomp + offset); \
487 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
489 #define vFAIL2(m,a1) STMT_START { \
491 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
492 Simple_vFAIL2(m, a1); \
497 * Like Simple_vFAIL(), but accepts three arguments.
499 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
500 const IV offset = RExC_parse - RExC_precomp; \
501 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
502 (int)offset, RExC_precomp, RExC_precomp + offset); \
506 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
508 #define vFAIL3(m,a1,a2) STMT_START { \
510 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
511 Simple_vFAIL3(m, a1, a2); \
515 * Like Simple_vFAIL(), but accepts four arguments.
517 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
518 const IV offset = RExC_parse - RExC_precomp; \
519 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
520 (int)offset, RExC_precomp, RExC_precomp + offset); \
523 #define ckWARNreg(loc,m) STMT_START { \
524 const IV offset = loc - RExC_precomp; \
525 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
526 (int)offset, RExC_precomp, RExC_precomp + offset); \
529 #define ckWARNregdep(loc,m) STMT_START { \
530 const IV offset = loc - RExC_precomp; \
531 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
533 (int)offset, RExC_precomp, RExC_precomp + offset); \
536 #define ckWARN2regdep(loc,m, a1) STMT_START { \
537 const IV offset = loc - RExC_precomp; \
538 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
540 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
543 #define ckWARN2reg(loc, m, a1) STMT_START { \
544 const IV offset = loc - RExC_precomp; \
545 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
546 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
549 #define vWARN3(loc, m, a1, a2) STMT_START { \
550 const IV offset = loc - RExC_precomp; \
551 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
552 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
555 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
556 const IV offset = loc - RExC_precomp; \
557 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
558 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
561 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
562 const IV offset = loc - RExC_precomp; \
563 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
564 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
567 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
568 const IV offset = loc - RExC_precomp; \
569 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
570 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
573 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
574 const IV offset = loc - RExC_precomp; \
575 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
576 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
580 /* Allow for side effects in s */
581 #define REGC(c,s) STMT_START { \
582 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
585 /* Macros for recording node offsets. 20001227 mjd@plover.com
586 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
587 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
588 * Element 0 holds the number n.
589 * Position is 1 indexed.
591 #ifndef RE_TRACK_PATTERN_OFFSETS
592 #define Set_Node_Offset_To_R(node,byte)
593 #define Set_Node_Offset(node,byte)
594 #define Set_Cur_Node_Offset
595 #define Set_Node_Length_To_R(node,len)
596 #define Set_Node_Length(node,len)
597 #define Set_Node_Cur_Length(node)
598 #define Node_Offset(n)
599 #define Node_Length(n)
600 #define Set_Node_Offset_Length(node,offset,len)
601 #define ProgLen(ri) ri->u.proglen
602 #define SetProgLen(ri,x) ri->u.proglen = x
604 #define ProgLen(ri) ri->u.offsets[0]
605 #define SetProgLen(ri,x) ri->u.offsets[0] = x
606 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
608 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
609 __LINE__, (int)(node), (int)(byte))); \
611 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
613 RExC_offsets[2*(node)-1] = (byte); \
618 #define Set_Node_Offset(node,byte) \
619 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
620 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
622 #define Set_Node_Length_To_R(node,len) STMT_START { \
624 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
625 __LINE__, (int)(node), (int)(len))); \
627 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
629 RExC_offsets[2*(node)] = (len); \
634 #define Set_Node_Length(node,len) \
635 Set_Node_Length_To_R((node)-RExC_emit_start, len)
636 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
637 #define Set_Node_Cur_Length(node) \
638 Set_Node_Length(node, RExC_parse - parse_start)
640 /* Get offsets and lengths */
641 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
642 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
644 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
645 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
646 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
650 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
651 #define EXPERIMENTAL_INPLACESCAN
652 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
654 #define DEBUG_STUDYDATA(str,data,depth) \
655 DEBUG_OPTIMISE_MORE_r(if(data){ \
656 PerlIO_printf(Perl_debug_log, \
657 "%*s" str "Pos:%"IVdf"/%"IVdf \
658 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
659 (int)(depth)*2, "", \
660 (IV)((data)->pos_min), \
661 (IV)((data)->pos_delta), \
662 (UV)((data)->flags), \
663 (IV)((data)->whilem_c), \
664 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
665 is_inf ? "INF " : "" \
667 if ((data)->last_found) \
668 PerlIO_printf(Perl_debug_log, \
669 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
670 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
671 SvPVX_const((data)->last_found), \
672 (IV)((data)->last_end), \
673 (IV)((data)->last_start_min), \
674 (IV)((data)->last_start_max), \
675 ((data)->longest && \
676 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
677 SvPVX_const((data)->longest_fixed), \
678 (IV)((data)->offset_fixed), \
679 ((data)->longest && \
680 (data)->longest==&((data)->longest_float)) ? "*" : "", \
681 SvPVX_const((data)->longest_float), \
682 (IV)((data)->offset_float_min), \
683 (IV)((data)->offset_float_max) \
685 PerlIO_printf(Perl_debug_log,"\n"); \
688 static void clear_re(pTHX_ void *r);
690 /* Mark that we cannot extend a found fixed substring at this point.
691 Update the longest found anchored substring and the longest found
692 floating substrings if needed. */
695 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
697 const STRLEN l = CHR_SVLEN(data->last_found);
698 const STRLEN old_l = CHR_SVLEN(*data->longest);
699 GET_RE_DEBUG_FLAGS_DECL;
701 PERL_ARGS_ASSERT_SCAN_COMMIT;
703 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
704 SvSetMagicSV(*data->longest, data->last_found);
705 if (*data->longest == data->longest_fixed) {
706 data->offset_fixed = l ? data->last_start_min : data->pos_min;
707 if (data->flags & SF_BEFORE_EOL)
709 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
711 data->flags &= ~SF_FIX_BEFORE_EOL;
712 data->minlen_fixed=minlenp;
713 data->lookbehind_fixed=0;
715 else { /* *data->longest == data->longest_float */
716 data->offset_float_min = l ? data->last_start_min : data->pos_min;
717 data->offset_float_max = (l
718 ? data->last_start_max
719 : data->pos_min + data->pos_delta);
720 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
721 data->offset_float_max = I32_MAX;
722 if (data->flags & SF_BEFORE_EOL)
724 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
726 data->flags &= ~SF_FL_BEFORE_EOL;
727 data->minlen_float=minlenp;
728 data->lookbehind_float=0;
731 SvCUR_set(data->last_found, 0);
733 SV * const sv = data->last_found;
734 if (SvUTF8(sv) && SvMAGICAL(sv)) {
735 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
741 data->flags &= ~SF_BEFORE_EOL;
742 DEBUG_STUDYDATA("commit: ",data,0);
745 /* Can match anything (initialization) */
747 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
749 PERL_ARGS_ASSERT_CL_ANYTHING;
751 ANYOF_BITMAP_SETALL(cl);
752 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
753 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
755 /* If any portion of the regex is to operate under locale rules,
756 * initialization includes it. The reason this isn't done for all regexes
757 * is that the optimizer was written under the assumption that locale was
758 * all-or-nothing. Given the complexity and lack of documentation in the
759 * optimizer, and that there are inadequate test cases for locale, so many
760 * parts of it may not work properly, it is safest to avoid locale unless
762 if (RExC_contains_locale) {
763 ANYOF_CLASS_SETALL(cl); /* /l uses class */
764 cl->flags |= ANYOF_LOCALE;
767 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
771 /* Can match anything (initialization) */
773 S_cl_is_anything(const struct regnode_charclass_class *cl)
777 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
779 for (value = 0; value <= ANYOF_MAX; value += 2)
780 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
782 if (!(cl->flags & ANYOF_UNICODE_ALL))
784 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
789 /* Can match anything (initialization) */
791 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
793 PERL_ARGS_ASSERT_CL_INIT;
795 Zero(cl, 1, struct regnode_charclass_class);
797 cl_anything(pRExC_state, cl);
798 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
801 /* These two functions currently do the exact same thing */
802 #define cl_init_zero S_cl_init
804 /* 'AND' a given class with another one. Can create false positives. 'cl'
805 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
806 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
808 S_cl_and(struct regnode_charclass_class *cl,
809 const struct regnode_charclass_class *and_with)
811 PERL_ARGS_ASSERT_CL_AND;
813 assert(and_with->type == ANYOF);
815 /* I (khw) am not sure all these restrictions are necessary XXX */
816 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
817 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
818 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
819 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
820 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
823 if (and_with->flags & ANYOF_INVERT)
824 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
825 cl->bitmap[i] &= ~and_with->bitmap[i];
827 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
828 cl->bitmap[i] &= and_with->bitmap[i];
829 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
831 if (and_with->flags & ANYOF_INVERT) {
833 /* Here, the and'ed node is inverted. Get the AND of the flags that
834 * aren't affected by the inversion. Those that are affected are
835 * handled individually below */
836 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
837 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
838 cl->flags |= affected_flags;
840 /* We currently don't know how to deal with things that aren't in the
841 * bitmap, but we know that the intersection is no greater than what
842 * is already in cl, so let there be false positives that get sorted
843 * out after the synthetic start class succeeds, and the node is
844 * matched for real. */
846 /* The inversion of these two flags indicate that the resulting
847 * intersection doesn't have them */
848 if (and_with->flags & ANYOF_UNICODE_ALL) {
849 cl->flags &= ~ANYOF_UNICODE_ALL;
851 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
852 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
855 else { /* and'd node is not inverted */
856 U8 outside_bitmap_but_not_utf8; /* Temp variable */
858 if (! ANYOF_NONBITMAP(and_with)) {
860 /* Here 'and_with' doesn't match anything outside the bitmap
861 * (except possibly ANYOF_UNICODE_ALL), which means the
862 * intersection can't either, except for ANYOF_UNICODE_ALL, in
863 * which case we don't know what the intersection is, but it's no
864 * greater than what cl already has, so can just leave it alone,
865 * with possible false positives */
866 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
867 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
868 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
871 else if (! ANYOF_NONBITMAP(cl)) {
873 /* Here, 'and_with' does match something outside the bitmap, and cl
874 * doesn't have a list of things to match outside the bitmap. If
875 * cl can match all code points above 255, the intersection will
876 * be those above-255 code points that 'and_with' matches. If cl
877 * can't match all Unicode code points, it means that it can't
878 * match anything outside the bitmap (since the 'if' that got us
879 * into this block tested for that), so we leave the bitmap empty.
881 if (cl->flags & ANYOF_UNICODE_ALL) {
882 ARG_SET(cl, ARG(and_with));
884 /* and_with's ARG may match things that don't require UTF8.
885 * And now cl's will too, in spite of this being an 'and'. See
886 * the comments below about the kludge */
887 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
891 /* Here, both 'and_with' and cl match something outside the
892 * bitmap. Currently we do not do the intersection, so just match
893 * whatever cl had at the beginning. */
897 /* Take the intersection of the two sets of flags. However, the
898 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
899 * kludge around the fact that this flag is not treated like the others
900 * which are initialized in cl_anything(). The way the optimizer works
901 * is that the synthetic start class (SSC) is initialized to match
902 * anything, and then the first time a real node is encountered, its
903 * values are AND'd with the SSC's with the result being the values of
904 * the real node. However, there are paths through the optimizer where
905 * the AND never gets called, so those initialized bits are set
906 * inappropriately, which is not usually a big deal, as they just cause
907 * false positives in the SSC, which will just mean a probably
908 * imperceptible slow down in execution. However this bit has a
909 * higher false positive consequence in that it can cause utf8.pm,
910 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
911 * bigger slowdown and also causes significant extra memory to be used.
912 * In order to prevent this, the code now takes a different tack. The
913 * bit isn't set unless some part of the regular expression needs it,
914 * but once set it won't get cleared. This means that these extra
915 * modules won't get loaded unless there was some path through the
916 * pattern that would have required them anyway, and so any false
917 * positives that occur by not ANDing them out when they could be
918 * aren't as severe as they would be if we treated this bit like all
920 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
921 & ANYOF_NONBITMAP_NON_UTF8;
922 cl->flags &= and_with->flags;
923 cl->flags |= outside_bitmap_but_not_utf8;
927 /* 'OR' a given class with another one. Can create false positives. 'cl'
928 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
929 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
931 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
933 PERL_ARGS_ASSERT_CL_OR;
935 if (or_with->flags & ANYOF_INVERT) {
937 /* Here, the or'd node is to be inverted. This means we take the
938 * complement of everything not in the bitmap, but currently we don't
939 * know what that is, so give up and match anything */
940 if (ANYOF_NONBITMAP(or_with)) {
941 cl_anything(pRExC_state, cl);
944 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
945 * <= (B1 | !B2) | (CL1 | !CL2)
946 * which is wasteful if CL2 is small, but we ignore CL2:
947 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
948 * XXXX Can we handle case-fold? Unclear:
949 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
950 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
952 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
953 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
954 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
957 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
958 cl->bitmap[i] |= ~or_with->bitmap[i];
959 } /* XXXX: logic is complicated otherwise */
961 cl_anything(pRExC_state, cl);
964 /* And, we can just take the union of the flags that aren't affected
965 * by the inversion */
966 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
968 /* For the remaining flags:
969 ANYOF_UNICODE_ALL and inverted means to not match anything above
970 255, which means that the union with cl should just be
971 what cl has in it, so can ignore this flag
972 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
973 is 127-255 to match them, but then invert that, so the
974 union with cl should just be what cl has in it, so can
977 } else { /* 'or_with' is not inverted */
978 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
979 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
980 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
981 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
984 /* OR char bitmap and class bitmap separately */
985 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
986 cl->bitmap[i] |= or_with->bitmap[i];
987 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
988 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
989 cl->classflags[i] |= or_with->classflags[i];
990 cl->flags |= ANYOF_CLASS;
993 else { /* XXXX: logic is complicated, leave it along for a moment. */
994 cl_anything(pRExC_state, cl);
997 if (ANYOF_NONBITMAP(or_with)) {
999 /* Use the added node's outside-the-bit-map match if there isn't a
1000 * conflict. If there is a conflict (both nodes match something
1001 * outside the bitmap, but what they match outside is not the same
1002 * pointer, and hence not easily compared until XXX we extend
1003 * inversion lists this far), give up and allow the start class to
1004 * match everything outside the bitmap. If that stuff is all above
1005 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1006 if (! ANYOF_NONBITMAP(cl)) {
1007 ARG_SET(cl, ARG(or_with));
1009 else if (ARG(cl) != ARG(or_with)) {
1011 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1012 cl_anything(pRExC_state, cl);
1015 cl->flags |= ANYOF_UNICODE_ALL;
1020 /* Take the union */
1021 cl->flags |= or_with->flags;
1025 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1026 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1027 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1028 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1033 dump_trie(trie,widecharmap,revcharmap)
1034 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1035 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1037 These routines dump out a trie in a somewhat readable format.
1038 The _interim_ variants are used for debugging the interim
1039 tables that are used to generate the final compressed
1040 representation which is what dump_trie expects.
1042 Part of the reason for their existence is to provide a form
1043 of documentation as to how the different representations function.
1048 Dumps the final compressed table form of the trie to Perl_debug_log.
1049 Used for debugging make_trie().
1053 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1054 AV *revcharmap, U32 depth)
1057 SV *sv=sv_newmortal();
1058 int colwidth= widecharmap ? 6 : 4;
1060 GET_RE_DEBUG_FLAGS_DECL;
1062 PERL_ARGS_ASSERT_DUMP_TRIE;
1064 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1065 (int)depth * 2 + 2,"",
1066 "Match","Base","Ofs" );
1068 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1069 SV ** const tmp = av_fetch( revcharmap, state, 0);
1071 PerlIO_printf( Perl_debug_log, "%*s",
1073 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1074 PL_colors[0], PL_colors[1],
1075 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1076 PERL_PV_ESCAPE_FIRSTCHAR
1081 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1082 (int)depth * 2 + 2,"");
1084 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1085 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1086 PerlIO_printf( Perl_debug_log, "\n");
1088 for( state = 1 ; state < trie->statecount ; state++ ) {
1089 const U32 base = trie->states[ state ].trans.base;
1091 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1093 if ( trie->states[ state ].wordnum ) {
1094 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1096 PerlIO_printf( Perl_debug_log, "%6s", "" );
1099 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1104 while( ( base + ofs < trie->uniquecharcount ) ||
1105 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1106 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1109 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1111 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1112 if ( ( base + ofs >= trie->uniquecharcount ) &&
1113 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1114 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1116 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1118 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1120 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1124 PerlIO_printf( Perl_debug_log, "]");
1127 PerlIO_printf( Perl_debug_log, "\n" );
1129 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1130 for (word=1; word <= trie->wordcount; word++) {
1131 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1132 (int)word, (int)(trie->wordinfo[word].prev),
1133 (int)(trie->wordinfo[word].len));
1135 PerlIO_printf(Perl_debug_log, "\n" );
1138 Dumps a fully constructed but uncompressed trie in list form.
1139 List tries normally only are used for construction when the number of
1140 possible chars (trie->uniquecharcount) is very high.
1141 Used for debugging make_trie().
1144 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1145 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1149 SV *sv=sv_newmortal();
1150 int colwidth= widecharmap ? 6 : 4;
1151 GET_RE_DEBUG_FLAGS_DECL;
1153 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1155 /* print out the table precompression. */
1156 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1157 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1158 "------:-----+-----------------\n" );
1160 for( state=1 ; state < next_alloc ; state ++ ) {
1163 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1164 (int)depth * 2 + 2,"", (UV)state );
1165 if ( ! trie->states[ state ].wordnum ) {
1166 PerlIO_printf( Perl_debug_log, "%5s| ","");
1168 PerlIO_printf( Perl_debug_log, "W%4x| ",
1169 trie->states[ state ].wordnum
1172 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1173 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1175 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1177 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1178 PL_colors[0], PL_colors[1],
1179 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1180 PERL_PV_ESCAPE_FIRSTCHAR
1182 TRIE_LIST_ITEM(state,charid).forid,
1183 (UV)TRIE_LIST_ITEM(state,charid).newstate
1186 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1187 (int)((depth * 2) + 14), "");
1190 PerlIO_printf( Perl_debug_log, "\n");
1195 Dumps a fully constructed but uncompressed trie in table form.
1196 This is the normal DFA style state transition table, with a few
1197 twists to facilitate compression later.
1198 Used for debugging make_trie().
1201 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1202 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1207 SV *sv=sv_newmortal();
1208 int colwidth= widecharmap ? 6 : 4;
1209 GET_RE_DEBUG_FLAGS_DECL;
1211 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1214 print out the table precompression so that we can do a visual check
1215 that they are identical.
1218 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1220 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1221 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1223 PerlIO_printf( Perl_debug_log, "%*s",
1225 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1226 PL_colors[0], PL_colors[1],
1227 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1228 PERL_PV_ESCAPE_FIRSTCHAR
1234 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1236 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1237 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1240 PerlIO_printf( Perl_debug_log, "\n" );
1242 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1244 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1245 (int)depth * 2 + 2,"",
1246 (UV)TRIE_NODENUM( state ) );
1248 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1249 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1251 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1253 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1255 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1256 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1258 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1259 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1267 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1268 startbranch: the first branch in the whole branch sequence
1269 first : start branch of sequence of branch-exact nodes.
1270 May be the same as startbranch
1271 last : Thing following the last branch.
1272 May be the same as tail.
1273 tail : item following the branch sequence
1274 count : words in the sequence
1275 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1276 depth : indent depth
1278 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1280 A trie is an N'ary tree where the branches are determined by digital
1281 decomposition of the key. IE, at the root node you look up the 1st character and
1282 follow that branch repeat until you find the end of the branches. Nodes can be
1283 marked as "accepting" meaning they represent a complete word. Eg:
1287 would convert into the following structure. Numbers represent states, letters
1288 following numbers represent valid transitions on the letter from that state, if
1289 the number is in square brackets it represents an accepting state, otherwise it
1290 will be in parenthesis.
1292 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1296 (1) +-i->(6)-+-s->[7]
1298 +-s->(3)-+-h->(4)-+-e->[5]
1300 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1302 This shows that when matching against the string 'hers' we will begin at state 1
1303 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1304 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1305 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1306 single traverse. We store a mapping from accepting to state to which word was
1307 matched, and then when we have multiple possibilities we try to complete the
1308 rest of the regex in the order in which they occured in the alternation.
1310 The only prior NFA like behaviour that would be changed by the TRIE support is
1311 the silent ignoring of duplicate alternations which are of the form:
1313 / (DUPE|DUPE) X? (?{ ... }) Y /x
1315 Thus EVAL blocks following a trie may be called a different number of times with
1316 and without the optimisation. With the optimisations dupes will be silently
1317 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1318 the following demonstrates:
1320 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1322 which prints out 'word' three times, but
1324 'words'=~/(word|word|word)(?{ print $1 })S/
1326 which doesnt print it out at all. This is due to other optimisations kicking in.
1328 Example of what happens on a structural level:
1330 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1332 1: CURLYM[1] {1,32767}(18)
1343 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1344 and should turn into:
1346 1: CURLYM[1] {1,32767}(18)
1348 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1356 Cases where tail != last would be like /(?foo|bar)baz/:
1366 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1367 and would end up looking like:
1370 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1377 d = uvuni_to_utf8_flags(d, uv, 0);
1379 is the recommended Unicode-aware way of saying
1384 #define TRIE_STORE_REVCHAR(val) \
1387 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1388 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1389 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1390 SvCUR_set(zlopp, kapow - flrbbbbb); \
1393 av_push(revcharmap, zlopp); \
1395 char ooooff = (char)val; \
1396 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1400 #define TRIE_READ_CHAR STMT_START { \
1403 /* if it is UTF then it is either already folded, or does not need folding */ \
1404 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1406 else if (folder == PL_fold_latin1) { \
1407 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1408 if ( foldlen > 0 ) { \
1409 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1415 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1416 skiplen = UNISKIP(uvc); \
1417 foldlen -= skiplen; \
1418 scan = foldbuf + skiplen; \
1421 /* raw data, will be folded later if needed */ \
1429 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1430 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1431 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1432 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1434 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1435 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1436 TRIE_LIST_CUR( state )++; \
1439 #define TRIE_LIST_NEW(state) STMT_START { \
1440 Newxz( trie->states[ state ].trans.list, \
1441 4, reg_trie_trans_le ); \
1442 TRIE_LIST_CUR( state ) = 1; \
1443 TRIE_LIST_LEN( state ) = 4; \
1446 #define TRIE_HANDLE_WORD(state) STMT_START { \
1447 U16 dupe= trie->states[ state ].wordnum; \
1448 regnode * const noper_next = regnext( noper ); \
1451 /* store the word for dumping */ \
1453 if (OP(noper) != NOTHING) \
1454 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1456 tmp = newSVpvn_utf8( "", 0, UTF ); \
1457 av_push( trie_words, tmp ); \
1461 trie->wordinfo[curword].prev = 0; \
1462 trie->wordinfo[curword].len = wordlen; \
1463 trie->wordinfo[curword].accept = state; \
1465 if ( noper_next < tail ) { \
1467 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1468 trie->jump[curword] = (U16)(noper_next - convert); \
1470 jumper = noper_next; \
1472 nextbranch= regnext(cur); \
1476 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1477 /* chain, so that when the bits of chain are later */\
1478 /* linked together, the dups appear in the chain */\
1479 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1480 trie->wordinfo[dupe].prev = curword; \
1482 /* we haven't inserted this word yet. */ \
1483 trie->states[ state ].wordnum = curword; \
1488 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1489 ( ( base + charid >= ucharcount \
1490 && base + charid < ubound \
1491 && state == trie->trans[ base - ucharcount + charid ].check \
1492 && trie->trans[ base - ucharcount + charid ].next ) \
1493 ? trie->trans[ base - ucharcount + charid ].next \
1494 : ( state==1 ? special : 0 ) \
1498 #define MADE_JUMP_TRIE 2
1499 #define MADE_EXACT_TRIE 4
1502 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1505 /* first pass, loop through and scan words */
1506 reg_trie_data *trie;
1507 HV *widecharmap = NULL;
1508 AV *revcharmap = newAV();
1510 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1515 regnode *jumper = NULL;
1516 regnode *nextbranch = NULL;
1517 regnode *convert = NULL;
1518 U32 *prev_states; /* temp array mapping each state to previous one */
1519 /* we just use folder as a flag in utf8 */
1520 const U8 * folder = NULL;
1523 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1524 AV *trie_words = NULL;
1525 /* along with revcharmap, this only used during construction but both are
1526 * useful during debugging so we store them in the struct when debugging.
1529 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1530 STRLEN trie_charcount=0;
1532 SV *re_trie_maxbuff;
1533 GET_RE_DEBUG_FLAGS_DECL;
1535 PERL_ARGS_ASSERT_MAKE_TRIE;
1537 PERL_UNUSED_ARG(depth);
1544 case EXACTFU_TRICKYFOLD:
1545 case EXACTFU: folder = PL_fold_latin1; break;
1546 case EXACTF: folder = PL_fold; break;
1547 case EXACTFL: folder = PL_fold_locale; break;
1548 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1551 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1553 trie->startstate = 1;
1554 trie->wordcount = word_count;
1555 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1556 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1558 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1559 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1560 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1563 trie_words = newAV();
1566 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1567 if (!SvIOK(re_trie_maxbuff)) {
1568 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1570 DEBUG_TRIE_COMPILE_r({
1571 PerlIO_printf( Perl_debug_log,
1572 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1573 (int)depth * 2 + 2, "",
1574 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1575 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1579 /* Find the node we are going to overwrite */
1580 if ( first == startbranch && OP( last ) != BRANCH ) {
1581 /* whole branch chain */
1584 /* branch sub-chain */
1585 convert = NEXTOPER( first );
1588 /* -- First loop and Setup --
1590 We first traverse the branches and scan each word to determine if it
1591 contains widechars, and how many unique chars there are, this is
1592 important as we have to build a table with at least as many columns as we
1595 We use an array of integers to represent the character codes 0..255
1596 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1597 native representation of the character value as the key and IV's for the
1600 *TODO* If we keep track of how many times each character is used we can
1601 remap the columns so that the table compression later on is more
1602 efficient in terms of memory by ensuring the most common value is in the
1603 middle and the least common are on the outside. IMO this would be better
1604 than a most to least common mapping as theres a decent chance the most
1605 common letter will share a node with the least common, meaning the node
1606 will not be compressible. With a middle is most common approach the worst
1607 case is when we have the least common nodes twice.
1611 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1612 regnode *noper = NEXTOPER( cur );
1613 const U8 *uc = (U8*)STRING( noper );
1614 const U8 *e = uc + STR_LEN( noper );
1616 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1618 const U8 *scan = (U8*)NULL;
1619 U32 wordlen = 0; /* required init */
1621 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1623 if (OP(noper) == NOTHING) {
1624 regnode *noper_next= regnext(noper);
1625 if (noper_next != tail && OP(noper_next) == flags) {
1627 uc= (U8*)STRING(noper);
1628 e= uc + STR_LEN(noper);
1629 trie->minlen= STR_LEN(noper);
1636 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1637 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1638 regardless of encoding */
1639 if (OP( noper ) == EXACTFU_SS) {
1640 /* false positives are ok, so just set this */
1641 TRIE_BITMAP_SET(trie,0xDF);
1644 for ( ; uc < e ; uc += len ) {
1645 TRIE_CHARCOUNT(trie)++;
1650 U8 folded= folder[ (U8) uvc ];
1651 if ( !trie->charmap[ folded ] ) {
1652 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1653 TRIE_STORE_REVCHAR( folded );
1656 if ( !trie->charmap[ uvc ] ) {
1657 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1658 TRIE_STORE_REVCHAR( uvc );
1661 /* store the codepoint in the bitmap, and its folded
1663 TRIE_BITMAP_SET(trie, uvc);
1665 /* store the folded codepoint */
1666 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1669 /* store first byte of utf8 representation of
1670 variant codepoints */
1671 if (! UNI_IS_INVARIANT(uvc)) {
1672 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1675 set_bit = 0; /* We've done our bit :-) */
1680 widecharmap = newHV();
1682 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1685 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1687 if ( !SvTRUE( *svpp ) ) {
1688 sv_setiv( *svpp, ++trie->uniquecharcount );
1689 TRIE_STORE_REVCHAR(uvc);
1693 if( cur == first ) {
1694 trie->minlen = chars;
1695 trie->maxlen = chars;
1696 } else if (chars < trie->minlen) {
1697 trie->minlen = chars;
1698 } else if (chars > trie->maxlen) {
1699 trie->maxlen = chars;
1701 if (OP( noper ) == EXACTFU_SS) {
1702 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1703 if (trie->minlen > 1)
1706 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1707 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1708 * - We assume that any such sequence might match a 2 byte string */
1709 if (trie->minlen > 2 )
1713 } /* end first pass */
1714 DEBUG_TRIE_COMPILE_r(
1715 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1716 (int)depth * 2 + 2,"",
1717 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1718 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1719 (int)trie->minlen, (int)trie->maxlen )
1723 We now know what we are dealing with in terms of unique chars and
1724 string sizes so we can calculate how much memory a naive
1725 representation using a flat table will take. If it's over a reasonable
1726 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1727 conservative but potentially much slower representation using an array
1730 At the end we convert both representations into the same compressed
1731 form that will be used in regexec.c for matching with. The latter
1732 is a form that cannot be used to construct with but has memory
1733 properties similar to the list form and access properties similar
1734 to the table form making it both suitable for fast searches and
1735 small enough that its feasable to store for the duration of a program.
1737 See the comment in the code where the compressed table is produced
1738 inplace from the flat tabe representation for an explanation of how
1739 the compression works.
1744 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1747 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1749 Second Pass -- Array Of Lists Representation
1751 Each state will be represented by a list of charid:state records
1752 (reg_trie_trans_le) the first such element holds the CUR and LEN
1753 points of the allocated array. (See defines above).
1755 We build the initial structure using the lists, and then convert
1756 it into the compressed table form which allows faster lookups
1757 (but cant be modified once converted).
1760 STRLEN transcount = 1;
1762 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1763 "%*sCompiling trie using list compiler\n",
1764 (int)depth * 2 + 2, ""));
1766 trie->states = (reg_trie_state *)
1767 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1768 sizeof(reg_trie_state) );
1772 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1774 regnode *noper = NEXTOPER( cur );
1775 U8 *uc = (U8*)STRING( noper );
1776 const U8 *e = uc + STR_LEN( noper );
1777 U32 state = 1; /* required init */
1778 U16 charid = 0; /* sanity init */
1779 U8 *scan = (U8*)NULL; /* sanity init */
1780 STRLEN foldlen = 0; /* required init */
1781 U32 wordlen = 0; /* required init */
1782 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1785 if (OP(noper) == NOTHING) {
1786 regnode *noper_next= regnext(noper);
1787 if (noper_next != tail && OP(noper_next) == flags) {
1789 uc= (U8*)STRING(noper);
1790 e= uc + STR_LEN(noper);
1794 if (OP(noper) != NOTHING) {
1795 for ( ; uc < e ; uc += len ) {
1800 charid = trie->charmap[ uvc ];
1802 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1806 charid=(U16)SvIV( *svpp );
1809 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1816 if ( !trie->states[ state ].trans.list ) {
1817 TRIE_LIST_NEW( state );
1819 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1820 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1821 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1826 newstate = next_alloc++;
1827 prev_states[newstate] = state;
1828 TRIE_LIST_PUSH( state, charid, newstate );
1833 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1837 TRIE_HANDLE_WORD(state);
1839 } /* end second pass */
1841 /* next alloc is the NEXT state to be allocated */
1842 trie->statecount = next_alloc;
1843 trie->states = (reg_trie_state *)
1844 PerlMemShared_realloc( trie->states,
1846 * sizeof(reg_trie_state) );
1848 /* and now dump it out before we compress it */
1849 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1850 revcharmap, next_alloc,
1854 trie->trans = (reg_trie_trans *)
1855 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1862 for( state=1 ; state < next_alloc ; state ++ ) {
1866 DEBUG_TRIE_COMPILE_MORE_r(
1867 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1871 if (trie->states[state].trans.list) {
1872 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1876 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1877 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1878 if ( forid < minid ) {
1880 } else if ( forid > maxid ) {
1884 if ( transcount < tp + maxid - minid + 1) {
1886 trie->trans = (reg_trie_trans *)
1887 PerlMemShared_realloc( trie->trans,
1889 * sizeof(reg_trie_trans) );
1890 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1892 base = trie->uniquecharcount + tp - minid;
1893 if ( maxid == minid ) {
1895 for ( ; zp < tp ; zp++ ) {
1896 if ( ! trie->trans[ zp ].next ) {
1897 base = trie->uniquecharcount + zp - minid;
1898 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1899 trie->trans[ zp ].check = state;
1905 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1906 trie->trans[ tp ].check = state;
1911 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1912 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1913 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1914 trie->trans[ tid ].check = state;
1916 tp += ( maxid - minid + 1 );
1918 Safefree(trie->states[ state ].trans.list);
1921 DEBUG_TRIE_COMPILE_MORE_r(
1922 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1925 trie->states[ state ].trans.base=base;
1927 trie->lasttrans = tp + 1;
1931 Second Pass -- Flat Table Representation.
1933 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1934 We know that we will need Charcount+1 trans at most to store the data
1935 (one row per char at worst case) So we preallocate both structures
1936 assuming worst case.
1938 We then construct the trie using only the .next slots of the entry
1941 We use the .check field of the first entry of the node temporarily to
1942 make compression both faster and easier by keeping track of how many non
1943 zero fields are in the node.
1945 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1948 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1949 number representing the first entry of the node, and state as a
1950 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1951 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1952 are 2 entrys per node. eg:
1960 The table is internally in the right hand, idx form. However as we also
1961 have to deal with the states array which is indexed by nodenum we have to
1962 use TRIE_NODENUM() to convert.
1965 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1966 "%*sCompiling trie using table compiler\n",
1967 (int)depth * 2 + 2, ""));
1969 trie->trans = (reg_trie_trans *)
1970 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1971 * trie->uniquecharcount + 1,
1972 sizeof(reg_trie_trans) );
1973 trie->states = (reg_trie_state *)
1974 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1975 sizeof(reg_trie_state) );
1976 next_alloc = trie->uniquecharcount + 1;
1979 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1981 regnode *noper = NEXTOPER( cur );
1982 const U8 *uc = (U8*)STRING( noper );
1983 const U8 *e = uc + STR_LEN( noper );
1985 U32 state = 1; /* required init */
1987 U16 charid = 0; /* sanity init */
1988 U32 accept_state = 0; /* sanity init */
1989 U8 *scan = (U8*)NULL; /* sanity init */
1991 STRLEN foldlen = 0; /* required init */
1992 U32 wordlen = 0; /* required init */
1994 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1996 if (OP(noper) == NOTHING) {
1997 regnode *noper_next= regnext(noper);
1998 if (noper_next != tail && OP(noper_next) == flags) {
2000 uc= (U8*)STRING(noper);
2001 e= uc + STR_LEN(noper);
2005 if ( OP(noper) != NOTHING ) {
2006 for ( ; uc < e ; uc += len ) {
2011 charid = trie->charmap[ uvc ];
2013 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2014 charid = svpp ? (U16)SvIV(*svpp) : 0;
2018 if ( !trie->trans[ state + charid ].next ) {
2019 trie->trans[ state + charid ].next = next_alloc;
2020 trie->trans[ state ].check++;
2021 prev_states[TRIE_NODENUM(next_alloc)]
2022 = TRIE_NODENUM(state);
2023 next_alloc += trie->uniquecharcount;
2025 state = trie->trans[ state + charid ].next;
2027 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2029 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2032 accept_state = TRIE_NODENUM( state );
2033 TRIE_HANDLE_WORD(accept_state);
2035 } /* end second pass */
2037 /* and now dump it out before we compress it */
2038 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2040 next_alloc, depth+1));
2044 * Inplace compress the table.*
2046 For sparse data sets the table constructed by the trie algorithm will
2047 be mostly 0/FAIL transitions or to put it another way mostly empty.
2048 (Note that leaf nodes will not contain any transitions.)
2050 This algorithm compresses the tables by eliminating most such
2051 transitions, at the cost of a modest bit of extra work during lookup:
2053 - Each states[] entry contains a .base field which indicates the
2054 index in the state[] array wheres its transition data is stored.
2056 - If .base is 0 there are no valid transitions from that node.
2058 - If .base is nonzero then charid is added to it to find an entry in
2061 -If trans[states[state].base+charid].check!=state then the
2062 transition is taken to be a 0/Fail transition. Thus if there are fail
2063 transitions at the front of the node then the .base offset will point
2064 somewhere inside the previous nodes data (or maybe even into a node
2065 even earlier), but the .check field determines if the transition is
2069 The following process inplace converts the table to the compressed
2070 table: We first do not compress the root node 1,and mark all its
2071 .check pointers as 1 and set its .base pointer as 1 as well. This
2072 allows us to do a DFA construction from the compressed table later,
2073 and ensures that any .base pointers we calculate later are greater
2076 - We set 'pos' to indicate the first entry of the second node.
2078 - We then iterate over the columns of the node, finding the first and
2079 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2080 and set the .check pointers accordingly, and advance pos
2081 appropriately and repreat for the next node. Note that when we copy
2082 the next pointers we have to convert them from the original
2083 NODEIDX form to NODENUM form as the former is not valid post
2086 - If a node has no transitions used we mark its base as 0 and do not
2087 advance the pos pointer.
2089 - If a node only has one transition we use a second pointer into the
2090 structure to fill in allocated fail transitions from other states.
2091 This pointer is independent of the main pointer and scans forward
2092 looking for null transitions that are allocated to a state. When it
2093 finds one it writes the single transition into the "hole". If the
2094 pointer doesnt find one the single transition is appended as normal.
2096 - Once compressed we can Renew/realloc the structures to release the
2099 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2100 specifically Fig 3.47 and the associated pseudocode.
2104 const U32 laststate = TRIE_NODENUM( next_alloc );
2107 trie->statecount = laststate;
2109 for ( state = 1 ; state < laststate ; state++ ) {
2111 const U32 stateidx = TRIE_NODEIDX( state );
2112 const U32 o_used = trie->trans[ stateidx ].check;
2113 U32 used = trie->trans[ stateidx ].check;
2114 trie->trans[ stateidx ].check = 0;
2116 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2117 if ( flag || trie->trans[ stateidx + charid ].next ) {
2118 if ( trie->trans[ stateidx + charid ].next ) {
2120 for ( ; zp < pos ; zp++ ) {
2121 if ( ! trie->trans[ zp ].next ) {
2125 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2126 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2127 trie->trans[ zp ].check = state;
2128 if ( ++zp > pos ) pos = zp;
2135 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2137 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2138 trie->trans[ pos ].check = state;
2143 trie->lasttrans = pos + 1;
2144 trie->states = (reg_trie_state *)
2145 PerlMemShared_realloc( trie->states, laststate
2146 * sizeof(reg_trie_state) );
2147 DEBUG_TRIE_COMPILE_MORE_r(
2148 PerlIO_printf( Perl_debug_log,
2149 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2150 (int)depth * 2 + 2,"",
2151 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2154 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2157 } /* end table compress */
2159 DEBUG_TRIE_COMPILE_MORE_r(
2160 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2161 (int)depth * 2 + 2, "",
2162 (UV)trie->statecount,
2163 (UV)trie->lasttrans)
2165 /* resize the trans array to remove unused space */
2166 trie->trans = (reg_trie_trans *)
2167 PerlMemShared_realloc( trie->trans, trie->lasttrans
2168 * sizeof(reg_trie_trans) );
2170 { /* Modify the program and insert the new TRIE node */
2171 U8 nodetype =(U8)(flags & 0xFF);
2175 regnode *optimize = NULL;
2176 #ifdef RE_TRACK_PATTERN_OFFSETS
2179 U32 mjd_nodelen = 0;
2180 #endif /* RE_TRACK_PATTERN_OFFSETS */
2181 #endif /* DEBUGGING */
2183 This means we convert either the first branch or the first Exact,
2184 depending on whether the thing following (in 'last') is a branch
2185 or not and whther first is the startbranch (ie is it a sub part of
2186 the alternation or is it the whole thing.)
2187 Assuming its a sub part we convert the EXACT otherwise we convert
2188 the whole branch sequence, including the first.
2190 /* Find the node we are going to overwrite */
2191 if ( first != startbranch || OP( last ) == BRANCH ) {
2192 /* branch sub-chain */
2193 NEXT_OFF( first ) = (U16)(last - first);
2194 #ifdef RE_TRACK_PATTERN_OFFSETS
2196 mjd_offset= Node_Offset((convert));
2197 mjd_nodelen= Node_Length((convert));
2200 /* whole branch chain */
2202 #ifdef RE_TRACK_PATTERN_OFFSETS
2205 const regnode *nop = NEXTOPER( convert );
2206 mjd_offset= Node_Offset((nop));
2207 mjd_nodelen= Node_Length((nop));
2211 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2212 (int)depth * 2 + 2, "",
2213 (UV)mjd_offset, (UV)mjd_nodelen)
2216 /* But first we check to see if there is a common prefix we can
2217 split out as an EXACT and put in front of the TRIE node. */
2218 trie->startstate= 1;
2219 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2221 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2225 const U32 base = trie->states[ state ].trans.base;
2227 if ( trie->states[state].wordnum )
2230 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2231 if ( ( base + ofs >= trie->uniquecharcount ) &&
2232 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2233 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2235 if ( ++count > 1 ) {
2236 SV **tmp = av_fetch( revcharmap, ofs, 0);
2237 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2238 if ( state == 1 ) break;
2240 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2242 PerlIO_printf(Perl_debug_log,
2243 "%*sNew Start State=%"UVuf" Class: [",
2244 (int)depth * 2 + 2, "",
2247 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2248 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2250 TRIE_BITMAP_SET(trie,*ch);
2252 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2254 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2258 TRIE_BITMAP_SET(trie,*ch);
2260 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2261 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2267 SV **tmp = av_fetch( revcharmap, idx, 0);
2269 char *ch = SvPV( *tmp, len );
2271 SV *sv=sv_newmortal();
2272 PerlIO_printf( Perl_debug_log,
2273 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2274 (int)depth * 2 + 2, "",
2276 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2277 PL_colors[0], PL_colors[1],
2278 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2279 PERL_PV_ESCAPE_FIRSTCHAR
2284 OP( convert ) = nodetype;
2285 str=STRING(convert);
2288 STR_LEN(convert) += len;
2294 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2299 trie->prefixlen = (state-1);
2301 regnode *n = convert+NODE_SZ_STR(convert);
2302 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2303 trie->startstate = state;
2304 trie->minlen -= (state - 1);
2305 trie->maxlen -= (state - 1);
2307 /* At least the UNICOS C compiler choked on this
2308 * being argument to DEBUG_r(), so let's just have
2311 #ifdef PERL_EXT_RE_BUILD
2317 regnode *fix = convert;
2318 U32 word = trie->wordcount;
2320 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2321 while( ++fix < n ) {
2322 Set_Node_Offset_Length(fix, 0, 0);
2325 SV ** const tmp = av_fetch( trie_words, word, 0 );
2327 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2328 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2330 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2338 NEXT_OFF(convert) = (U16)(tail - convert);
2339 DEBUG_r(optimize= n);
2345 if ( trie->maxlen ) {
2346 NEXT_OFF( convert ) = (U16)(tail - convert);
2347 ARG_SET( convert, data_slot );
2348 /* Store the offset to the first unabsorbed branch in
2349 jump[0], which is otherwise unused by the jump logic.
2350 We use this when dumping a trie and during optimisation. */
2352 trie->jump[0] = (U16)(nextbranch - convert);
2354 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2355 * and there is a bitmap
2356 * and the first "jump target" node we found leaves enough room
2357 * then convert the TRIE node into a TRIEC node, with the bitmap
2358 * embedded inline in the opcode - this is hypothetically faster.
2360 if ( !trie->states[trie->startstate].wordnum
2362 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2364 OP( convert ) = TRIEC;
2365 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2366 PerlMemShared_free(trie->bitmap);
2369 OP( convert ) = TRIE;
2371 /* store the type in the flags */
2372 convert->flags = nodetype;
2376 + regarglen[ OP( convert ) ];
2378 /* XXX We really should free up the resource in trie now,
2379 as we won't use them - (which resources?) dmq */
2381 /* needed for dumping*/
2382 DEBUG_r(if (optimize) {
2383 regnode *opt = convert;
2385 while ( ++opt < optimize) {
2386 Set_Node_Offset_Length(opt,0,0);
2389 Try to clean up some of the debris left after the
2392 while( optimize < jumper ) {
2393 mjd_nodelen += Node_Length((optimize));
2394 OP( optimize ) = OPTIMIZED;
2395 Set_Node_Offset_Length(optimize,0,0);
2398 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2400 } /* end node insert */
2402 /* Finish populating the prev field of the wordinfo array. Walk back
2403 * from each accept state until we find another accept state, and if
2404 * so, point the first word's .prev field at the second word. If the
2405 * second already has a .prev field set, stop now. This will be the
2406 * case either if we've already processed that word's accept state,
2407 * or that state had multiple words, and the overspill words were
2408 * already linked up earlier.
2415 for (word=1; word <= trie->wordcount; word++) {
2417 if (trie->wordinfo[word].prev)
2419 state = trie->wordinfo[word].accept;
2421 state = prev_states[state];
2424 prev = trie->states[state].wordnum;
2428 trie->wordinfo[word].prev = prev;
2430 Safefree(prev_states);
2434 /* and now dump out the compressed format */
2435 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2437 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2439 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2440 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2442 SvREFCNT_dec(revcharmap);
2446 : trie->startstate>1
2452 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2454 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2456 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2457 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2460 We find the fail state for each state in the trie, this state is the longest proper
2461 suffix of the current state's 'word' that is also a proper prefix of another word in our
2462 trie. State 1 represents the word '' and is thus the default fail state. This allows
2463 the DFA not to have to restart after its tried and failed a word at a given point, it
2464 simply continues as though it had been matching the other word in the first place.
2466 'abcdgu'=~/abcdefg|cdgu/
2467 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2468 fail, which would bring us to the state representing 'd' in the second word where we would
2469 try 'g' and succeed, proceeding to match 'cdgu'.
2471 /* add a fail transition */
2472 const U32 trie_offset = ARG(source);
2473 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2475 const U32 ucharcount = trie->uniquecharcount;
2476 const U32 numstates = trie->statecount;
2477 const U32 ubound = trie->lasttrans + ucharcount;
2481 U32 base = trie->states[ 1 ].trans.base;
2484 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2485 GET_RE_DEBUG_FLAGS_DECL;
2487 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2489 PERL_UNUSED_ARG(depth);
2493 ARG_SET( stclass, data_slot );
2494 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2495 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2496 aho->trie=trie_offset;
2497 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2498 Copy( trie->states, aho->states, numstates, reg_trie_state );
2499 Newxz( q, numstates, U32);
2500 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2503 /* initialize fail[0..1] to be 1 so that we always have
2504 a valid final fail state */
2505 fail[ 0 ] = fail[ 1 ] = 1;
2507 for ( charid = 0; charid < ucharcount ; charid++ ) {
2508 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2510 q[ q_write ] = newstate;
2511 /* set to point at the root */
2512 fail[ q[ q_write++ ] ]=1;
2515 while ( q_read < q_write) {
2516 const U32 cur = q[ q_read++ % numstates ];
2517 base = trie->states[ cur ].trans.base;
2519 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2520 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2522 U32 fail_state = cur;
2525 fail_state = fail[ fail_state ];
2526 fail_base = aho->states[ fail_state ].trans.base;
2527 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2529 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2530 fail[ ch_state ] = fail_state;
2531 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2533 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2535 q[ q_write++ % numstates] = ch_state;
2539 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2540 when we fail in state 1, this allows us to use the
2541 charclass scan to find a valid start char. This is based on the principle
2542 that theres a good chance the string being searched contains lots of stuff
2543 that cant be a start char.
2545 fail[ 0 ] = fail[ 1 ] = 0;
2546 DEBUG_TRIE_COMPILE_r({
2547 PerlIO_printf(Perl_debug_log,
2548 "%*sStclass Failtable (%"UVuf" states): 0",
2549 (int)(depth * 2), "", (UV)numstates
2551 for( q_read=1; q_read<numstates; q_read++ ) {
2552 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2554 PerlIO_printf(Perl_debug_log, "\n");
2557 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2562 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2563 * These need to be revisited when a newer toolchain becomes available.
2565 #if defined(__sparc64__) && defined(__GNUC__)
2566 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2567 # undef SPARC64_GCC_WORKAROUND
2568 # define SPARC64_GCC_WORKAROUND 1
2572 #define DEBUG_PEEP(str,scan,depth) \
2573 DEBUG_OPTIMISE_r({if (scan){ \
2574 SV * const mysv=sv_newmortal(); \
2575 regnode *Next = regnext(scan); \
2576 regprop(RExC_rx, mysv, scan); \
2577 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2578 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2579 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2583 /* The below joins as many adjacent EXACTish nodes as possible into a single
2584 * one, and looks for problematic sequences of characters whose folds vs.
2585 * non-folds have sufficiently different lengths, that the optimizer would be
2586 * fooled into rejecting legitimate matches of them, and the trie construction
2587 * code needs to handle specially. The joining is only done if:
2588 * 1) there is room in the current conglomerated node to entirely contain the
2590 * 2) they are the exact same node type
2592 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2593 * these get optimized out
2595 * If there are problematic code sequences, *min_subtract is set to the delta
2596 * that the minimum size of the node can be less than its actual size. And,
2597 * the node type of the result is changed to reflect that it contains these
2600 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2601 * and contains LATIN SMALL LETTER SHARP S
2603 * This is as good a place as any to discuss the design of handling these
2604 * problematic sequences. It's been wrong in Perl for a very long time. There
2605 * are three code points currently in Unicode whose folded lengths differ so
2606 * much from the un-folded lengths that it causes problems for the optimizer
2607 * and trie construction. Why only these are problematic, and not others where
2608 * lengths also differ is something I (khw) do not understand. New versions of
2609 * Unicode might add more such code points. Hopefully the logic in
2610 * fold_grind.t that figures out what to test (in part by verifying that each
2611 * size-combination gets tested) will catch any that do come along, so they can
2612 * be added to the special handling below. The chances of new ones are
2613 * actually rather small, as most, if not all, of the world's scripts that have
2614 * casefolding have already been encoded by Unicode. Also, a number of
2615 * Unicode's decisions were made to allow compatibility with pre-existing
2616 * standards, and almost all of those have already been dealt with. These
2617 * would otherwise be the most likely candidates for generating further tricky
2618 * sequences. In other words, Unicode by itself is unlikely to add new ones
2619 * unless it is for compatibility with pre-existing standards, and there aren't
2620 * many of those left.
2622 * The previous designs for dealing with these involved assigning a special
2623 * node for them. This approach doesn't work, as evidenced by this example:
2624 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2625 * Both these fold to "sss", but if the pattern is parsed to create a node
2626 * that would match just the \xDF, it won't be able to handle the case where a
2627 * successful match would have to cross the node's boundary. The new approach
2628 * that hopefully generally solves the problem generates an EXACTFU_SS node
2631 * There are a number of components to the approach (a lot of work for just
2632 * three code points!):
2633 * 1) This routine examines each EXACTFish node that could contain the
2634 * problematic sequences. It returns in *min_subtract how much to
2635 * subtract from the the actual length of the string to get a real minimum
2636 * for one that could match it. This number is usually 0 except for the
2637 * problematic sequences. This delta is used by the caller to adjust the
2638 * min length of the match, and the delta between min and max, so that the
2639 * optimizer doesn't reject these possibilities based on size constraints.
2640 * 2) These sequences require special handling by the trie code, so this code
2641 * changes the joined node type to special ops: EXACTFU_TRICKYFOLD and
2643 * 3) This is sufficient for the two Greek sequences (described below), but
2644 * the one involving the Sharp s (\xDF) needs more. The node type
2645 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2646 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2647 * case where there is a possible fold length change. That means that a
2648 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2649 * itself with length changes, and so can be processed faster. regexec.c
2650 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2651 * is pre-folded by regcomp.c. This saves effort in regex matching.
2652 * However, the pre-folding isn't done for non-UTF8 patterns because the
2653 * fold of the MICRO SIGN requires UTF-8. Also what EXACTF and EXACTFL
2654 * nodes fold to isn't known until runtime. The fold possibilities for
2655 * the non-UTF8 patterns are quite simple, except for the sharp s. All
2656 * the ones that don't involve a UTF-8 target string are members of a
2657 * fold-pair, and arrays are set up for all of them so that the other
2658 * member of the pair can be found quickly. Code elsewhere in this file
2659 * makes sure that in EXACTFU nodes, the sharp s gets folded to 'ss', even
2660 * if the pattern isn't UTF-8. This avoids the issues described in the
2662 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2663 * 'ss' or not is not knowable at compile time. It will match iff the
2664 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2665 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2666 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
2667 * described in item 3). An assumption that the optimizer part of
2668 * regexec.c (probably unwittingly) makes is that a character in the
2669 * pattern corresponds to at most a single character in the target string.
2670 * (And I do mean character, and not byte here, unlike other parts of the
2671 * documentation that have never been updated to account for multibyte
2672 * Unicode.) This assumption is wrong only in this case, as all other
2673 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2674 * virtue of having this file pre-fold UTF-8 patterns. I'm
2675 * reluctant to try to change this assumption, so instead the code punts.
2676 * This routine examines EXACTF nodes for the sharp s, and returns a
2677 * boolean indicating whether or not the node is an EXACTF node that
2678 * contains a sharp s. When it is true, the caller sets a flag that later
2679 * causes the optimizer in this file to not set values for the floating
2680 * and fixed string lengths, and thus avoids the optimizer code in
2681 * regexec.c that makes the invalid assumption. Thus, there is no
2682 * optimization based on string lengths for EXACTF nodes that contain the
2683 * sharp s. This only happens for /id rules (which means the pattern
2687 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2688 if (PL_regkind[OP(scan)] == EXACT) \
2689 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2692 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) {
2693 /* Merge several consecutive EXACTish nodes into one. */
2694 regnode *n = regnext(scan);
2696 regnode *next = scan + NODE_SZ_STR(scan);
2700 regnode *stop = scan;
2701 GET_RE_DEBUG_FLAGS_DECL;
2703 PERL_UNUSED_ARG(depth);
2706 PERL_ARGS_ASSERT_JOIN_EXACT;
2707 #ifndef EXPERIMENTAL_INPLACESCAN
2708 PERL_UNUSED_ARG(flags);
2709 PERL_UNUSED_ARG(val);
2711 DEBUG_PEEP("join",scan,depth);
2713 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2714 * EXACT ones that are mergeable to the current one. */
2716 && (PL_regkind[OP(n)] == NOTHING
2717 || (stringok && OP(n) == OP(scan)))
2719 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2722 if (OP(n) == TAIL || n > next)
2724 if (PL_regkind[OP(n)] == NOTHING) {
2725 DEBUG_PEEP("skip:",n,depth);
2726 NEXT_OFF(scan) += NEXT_OFF(n);
2727 next = n + NODE_STEP_REGNODE;
2734 else if (stringok) {
2735 const unsigned int oldl = STR_LEN(scan);
2736 regnode * const nnext = regnext(n);
2738 /* XXX I (khw) kind of doubt that this works on platforms where
2739 * U8_MAX is above 255 because of lots of other assumptions */
2740 if (oldl + STR_LEN(n) > U8_MAX)
2743 DEBUG_PEEP("merg",n,depth);
2746 NEXT_OFF(scan) += NEXT_OFF(n);
2747 STR_LEN(scan) += STR_LEN(n);
2748 next = n + NODE_SZ_STR(n);
2749 /* Now we can overwrite *n : */
2750 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2758 #ifdef EXPERIMENTAL_INPLACESCAN
2759 if (flags && !NEXT_OFF(n)) {
2760 DEBUG_PEEP("atch", val, depth);
2761 if (reg_off_by_arg[OP(n)]) {
2762 ARG_SET(n, val - n);
2765 NEXT_OFF(n) = val - n;
2773 *has_exactf_sharp_s = FALSE;
2775 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2776 * can now analyze for sequences of problematic code points. (Prior to
2777 * this final joining, sequences could have been split over boundaries, and
2778 * hence missed). The sequences only happen in folding, hence for any
2779 * non-EXACT EXACTish node */
2780 if (OP(scan) != EXACT) {
2782 U8 * s0 = (U8*) STRING(scan);
2783 U8 * const s_end = s0 + STR_LEN(scan);
2785 /* The below is perhaps overboard, but this allows us to save a test
2786 * each time through the loop at the expense of a mask. This is
2787 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2788 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2789 * This uses an exclusive 'or' to find that bit and then inverts it to
2790 * form a mask, with just a single 0, in the bit position where 'S' and
2792 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2793 const U8 s_masked = 's' & S_or_s_mask;
2795 /* One pass is made over the node's string looking for all the
2796 * possibilities. to avoid some tests in the loop, there are two main
2797 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2801 /* There are two problematic Greek code points in Unicode
2804 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2805 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2811 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2812 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2814 * This means that in case-insensitive matching (or "loose
2815 * matching", as Unicode calls it), an EXACTF of length six (the
2816 * UTF-8 encoded byte length of the above casefolded versions) can
2817 * match a target string of length two (the byte length of UTF-8
2818 * encoded U+0390 or U+03B0). This would rather mess up the
2819 * minimum length computation. (there are other code points that
2820 * also fold to these two sequences, but the delta is smaller)
2822 * If these sequences are found, the minimum length is decreased by
2823 * four (six minus two).
2825 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2826 * LETTER SHARP S. We decrease the min length by 1 for each
2827 * occurrence of 'ss' found */
2829 #define U390_FIRST_BYTE GREEK_SMALL_LETTER_IOTA_UTF8_FIRST_BYTE
2830 #define U3B0_FIRST_BYTE GREEK_SMALL_LETTER_UPSILON_UTF8_FIRST_BYTE
2831 const U8 U390_tail[] = GREEK_SMALL_LETTER_IOTA_UTF8_TAIL
2832 COMBINING_DIAERESIS_UTF8
2833 COMBINING_ACUTE_ACCENT_UTF8;
2834 const U8 U3B0_tail[] = GREEK_SMALL_LETTER_UPSILON_UTF8_TAIL
2835 COMBINING_DIAERESIS_UTF8
2836 COMBINING_ACUTE_ACCENT_UTF8;
2837 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2838 yields a net of 0 */
2839 /* Examine the string for one of the problematic sequences */
2841 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2842 * sequence we are looking for is 2 */
2846 /* Look for the first byte in each problematic sequence */
2848 /* We don't have to worry about other things that fold to
2849 * 's' (such as the long s, U+017F), as all above-latin1
2850 * code points have been pre-folded */
2854 /* Current character is an 's' or 'S'. If next one is
2855 * as well, we have the dreaded sequence */
2856 if (((*(s+1) & S_or_s_mask) == s_masked)
2857 /* These two node types don't have special handling
2859 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2862 OP(scan) = EXACTFU_SS;
2863 s++; /* No need to look at this character again */
2867 case U390_FIRST_BYTE:
2868 if (s_end - s >= len
2870 /* The 1's are because are skipping comparing the
2872 && memEQ(s + 1, U390_tail, len - 1))
2874 goto greek_sequence;
2878 case U3B0_FIRST_BYTE:
2879 if (! (s_end - s >= len
2880 && memEQ(s + 1, U3B0_tail, len - 1)))
2887 /* This requires special handling by trie's, so change
2888 * the node type to indicate this. If EXACTFA and
2889 * EXACTFL were ever to be handled by trie's, this
2890 * would have to be changed. If this node has already
2891 * been changed to EXACTFU_SS in this loop, leave it as
2892 * is. (I (khw) think it doesn't matter in regexec.c
2893 * for UTF patterns, but no need to change it */
2894 if (OP(scan) == EXACTFU) {
2895 OP(scan) = EXACTFU_TRICKYFOLD;
2897 s += 6; /* We already know what this sequence is. Skip
2903 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2905 /* Here, the pattern is not UTF-8. We need to look only for the
2906 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2907 * in the final position. Otherwise we can stop looking 1 byte
2908 * earlier because have to find both the first and second 's' */
2909 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2911 for (s = s0; s < upper; s++) {
2916 && ((*(s+1) & S_or_s_mask) == s_masked))
2920 /* EXACTF nodes need to know that the minimum
2921 * length changed so that a sharp s in the string
2922 * can match this ss in the pattern, but they
2923 * remain EXACTF nodes, as they won't match this
2924 * unless the target string is is UTF-8, which we
2925 * don't know until runtime */
2926 if (OP(scan) != EXACTF) {
2927 OP(scan) = EXACTFU_SS;
2932 case LATIN_SMALL_LETTER_SHARP_S:
2933 if (OP(scan) == EXACTF) {
2934 *has_exactf_sharp_s = TRUE;
2943 /* Allow dumping but overwriting the collection of skipped
2944 * ops and/or strings with fake optimized ops */
2945 n = scan + NODE_SZ_STR(scan);
2953 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2957 /* REx optimizer. Converts nodes into quicker variants "in place".
2958 Finds fixed substrings. */
2960 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2961 to the position after last scanned or to NULL. */
2963 #define INIT_AND_WITHP \
2964 assert(!and_withp); \
2965 Newx(and_withp,1,struct regnode_charclass_class); \
2966 SAVEFREEPV(and_withp)
2968 /* this is a chain of data about sub patterns we are processing that
2969 need to be handled separately/specially in study_chunk. Its so
2970 we can simulate recursion without losing state. */
2972 typedef struct scan_frame {
2973 regnode *last; /* last node to process in this frame */
2974 regnode *next; /* next node to process when last is reached */
2975 struct scan_frame *prev; /*previous frame*/
2976 I32 stop; /* what stopparen do we use */
2980 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2982 #define CASE_SYNST_FNC(nAmE) \
2984 if (flags & SCF_DO_STCLASS_AND) { \
2985 for (value = 0; value < 256; value++) \
2986 if (!is_ ## nAmE ## _cp(value)) \
2987 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2990 for (value = 0; value < 256; value++) \
2991 if (is_ ## nAmE ## _cp(value)) \
2992 ANYOF_BITMAP_SET(data->start_class, value); \
2996 if (flags & SCF_DO_STCLASS_AND) { \
2997 for (value = 0; value < 256; value++) \
2998 if (is_ ## nAmE ## _cp(value)) \
2999 ANYOF_BITMAP_CLEAR(data->start_class, value); \
3002 for (value = 0; value < 256; value++) \
3003 if (!is_ ## nAmE ## _cp(value)) \
3004 ANYOF_BITMAP_SET(data->start_class, value); \
3011 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3012 I32 *minlenp, I32 *deltap,
3017 struct regnode_charclass_class *and_withp,
3018 U32 flags, U32 depth)
3019 /* scanp: Start here (read-write). */
3020 /* deltap: Write maxlen-minlen here. */
3021 /* last: Stop before this one. */
3022 /* data: string data about the pattern */
3023 /* stopparen: treat close N as END */
3024 /* recursed: which subroutines have we recursed into */
3025 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3028 I32 min = 0, pars = 0, code;
3029 regnode *scan = *scanp, *next;
3031 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3032 int is_inf_internal = 0; /* The studied chunk is infinite */
3033 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3034 scan_data_t data_fake;
3035 SV *re_trie_maxbuff = NULL;
3036 regnode *first_non_open = scan;
3037 I32 stopmin = I32_MAX;
3038 scan_frame *frame = NULL;
3039 GET_RE_DEBUG_FLAGS_DECL;
3041 PERL_ARGS_ASSERT_STUDY_CHUNK;
3044 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3048 while (first_non_open && OP(first_non_open) == OPEN)
3049 first_non_open=regnext(first_non_open);
3054 while ( scan && OP(scan) != END && scan < last ){
3055 UV min_subtract = 0; /* How much to subtract from the minimum node
3056 length to get a real minimum (because the
3057 folded version may be shorter) */
3058 bool has_exactf_sharp_s = FALSE;
3059 /* Peephole optimizer: */
3060 DEBUG_STUDYDATA("Peep:", data,depth);
3061 DEBUG_PEEP("Peep",scan,depth);
3063 /* Its not clear to khw or hv why this is done here, and not in the
3064 * clauses that deal with EXACT nodes. khw's guess is that it's
3065 * because of a previous design */
3066 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3068 /* Follow the next-chain of the current node and optimize
3069 away all the NOTHINGs from it. */
3070 if (OP(scan) != CURLYX) {
3071 const int max = (reg_off_by_arg[OP(scan)]
3073 /* I32 may be smaller than U16 on CRAYs! */
3074 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3075 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3079 /* Skip NOTHING and LONGJMP. */
3080 while ((n = regnext(n))
3081 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3082 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3083 && off + noff < max)
3085 if (reg_off_by_arg[OP(scan)])
3088 NEXT_OFF(scan) = off;
3093 /* The principal pseudo-switch. Cannot be a switch, since we
3094 look into several different things. */
3095 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3096 || OP(scan) == IFTHEN) {
3097 next = regnext(scan);
3099 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3101 if (OP(next) == code || code == IFTHEN) {
3102 /* NOTE - There is similar code to this block below for handling
3103 TRIE nodes on a re-study. If you change stuff here check there
3105 I32 max1 = 0, min1 = I32_MAX, num = 0;
3106 struct regnode_charclass_class accum;
3107 regnode * const startbranch=scan;
3109 if (flags & SCF_DO_SUBSTR)
3110 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3111 if (flags & SCF_DO_STCLASS)
3112 cl_init_zero(pRExC_state, &accum);
3114 while (OP(scan) == code) {
3115 I32 deltanext, minnext, f = 0, fake;
3116 struct regnode_charclass_class this_class;
3119 data_fake.flags = 0;
3121 data_fake.whilem_c = data->whilem_c;
3122 data_fake.last_closep = data->last_closep;
3125 data_fake.last_closep = &fake;
3127 data_fake.pos_delta = delta;
3128 next = regnext(scan);
3129 scan = NEXTOPER(scan);
3131 scan = NEXTOPER(scan);
3132 if (flags & SCF_DO_STCLASS) {
3133 cl_init(pRExC_state, &this_class);
3134 data_fake.start_class = &this_class;
3135 f = SCF_DO_STCLASS_AND;
3137 if (flags & SCF_WHILEM_VISITED_POS)
3138 f |= SCF_WHILEM_VISITED_POS;
3140 /* we suppose the run is continuous, last=next...*/
3141 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3143 stopparen, recursed, NULL, f,depth+1);
3146 if (max1 < minnext + deltanext)
3147 max1 = minnext + deltanext;
3148 if (deltanext == I32_MAX)
3149 is_inf = is_inf_internal = 1;
3151 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3153 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3154 if ( stopmin > minnext)
3155 stopmin = min + min1;
3156 flags &= ~SCF_DO_SUBSTR;
3158 data->flags |= SCF_SEEN_ACCEPT;
3161 if (data_fake.flags & SF_HAS_EVAL)
3162 data->flags |= SF_HAS_EVAL;
3163 data->whilem_c = data_fake.whilem_c;
3165 if (flags & SCF_DO_STCLASS)
3166 cl_or(pRExC_state, &accum, &this_class);
3168 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3170 if (flags & SCF_DO_SUBSTR) {
3171 data->pos_min += min1;
3172 data->pos_delta += max1 - min1;
3173 if (max1 != min1 || is_inf)
3174 data->longest = &(data->longest_float);
3177 delta += max1 - min1;
3178 if (flags & SCF_DO_STCLASS_OR) {
3179 cl_or(pRExC_state, data->start_class, &accum);
3181 cl_and(data->start_class, and_withp);
3182 flags &= ~SCF_DO_STCLASS;
3185 else if (flags & SCF_DO_STCLASS_AND) {
3187 cl_and(data->start_class, &accum);
3188 flags &= ~SCF_DO_STCLASS;
3191 /* Switch to OR mode: cache the old value of
3192 * data->start_class */
3194 StructCopy(data->start_class, and_withp,
3195 struct regnode_charclass_class);
3196 flags &= ~SCF_DO_STCLASS_AND;
3197 StructCopy(&accum, data->start_class,
3198 struct regnode_charclass_class);
3199 flags |= SCF_DO_STCLASS_OR;
3200 data->start_class->flags |= ANYOF_EOS;
3204 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3207 Assuming this was/is a branch we are dealing with: 'scan' now
3208 points at the item that follows the branch sequence, whatever
3209 it is. We now start at the beginning of the sequence and look
3216 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3218 If we can find such a subsequence we need to turn the first
3219 element into a trie and then add the subsequent branch exact
3220 strings to the trie.
3224 1. patterns where the whole set of branches can be converted.
3226 2. patterns where only a subset can be converted.
3228 In case 1 we can replace the whole set with a single regop
3229 for the trie. In case 2 we need to keep the start and end
3232 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3233 becomes BRANCH TRIE; BRANCH X;
3235 There is an additional case, that being where there is a
3236 common prefix, which gets split out into an EXACT like node
3237 preceding the TRIE node.
3239 If x(1..n)==tail then we can do a simple trie, if not we make
3240 a "jump" trie, such that when we match the appropriate word
3241 we "jump" to the appropriate tail node. Essentially we turn
3242 a nested if into a case structure of sorts.
3247 if (!re_trie_maxbuff) {
3248 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3249 if (!SvIOK(re_trie_maxbuff))
3250 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3252 if ( SvIV(re_trie_maxbuff)>=0 ) {
3254 regnode *first = (regnode *)NULL;
3255 regnode *last = (regnode *)NULL;
3256 regnode *tail = scan;
3261 SV * const mysv = sv_newmortal(); /* for dumping */
3263 /* var tail is used because there may be a TAIL
3264 regop in the way. Ie, the exacts will point to the
3265 thing following the TAIL, but the last branch will
3266 point at the TAIL. So we advance tail. If we
3267 have nested (?:) we may have to move through several
3271 while ( OP( tail ) == TAIL ) {
3272 /* this is the TAIL generated by (?:) */
3273 tail = regnext( tail );
3277 DEBUG_TRIE_COMPILE_r({
3278 regprop(RExC_rx, mysv, tail );
3279 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3280 (int)depth * 2 + 2, "",
3281 "Looking for TRIE'able sequences. Tail node is: ",
3282 SvPV_nolen_const( mysv )
3288 Step through the branches
3289 cur represents each branch,
3290 noper is the first thing to be matched as part of that branch
3291 noper_next is the regnext() of that node.
3293 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3294 via a "jump trie" but we also support building with NOJUMPTRIE,
3295 which restricts the trie logic to structures like /FOO|BAR/.
3297 If noper is a trieable nodetype then the branch is a possible optimization
3298 target. If we are building under NOJUMPTRIE then we require that noper_next
3299 is the same as scan (our current position in the regex program).
3301 Once we have two or more consecutive such branches we can create a
3302 trie of the EXACT's contents and stitch it in place into the program.
3304 If the sequence represents all of the branches in the alternation we
3305 replace the entire thing with a single TRIE node.
3307 Otherwise when it is a subsequence we need to stitch it in place and
3308 replace only the relevant branches. This means the first branch has
3309 to remain as it is used by the alternation logic, and its next pointer,
3310 and needs to be repointed at the item on the branch chain following
3311 the last branch we have optimized away.
3313 This could be either a BRANCH, in which case the subsequence is internal,
3314 or it could be the item following the branch sequence in which case the
3315 subsequence is at the end (which does not necessarily mean the first node
3316 is the start of the alternation).
3318 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3321 ----------------+-----------
3325 EXACTFU_SS | EXACTFU
3326 EXACTFU_TRICKYFOLD | EXACTFU
3331 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3332 ( EXACT == (X) ) ? EXACT : \
3333 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3336 /* dont use tail as the end marker for this traverse */
3337 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3338 regnode * const noper = NEXTOPER( cur );
3339 U8 noper_type = OP( noper );
3340 U8 noper_trietype = TRIE_TYPE( noper_type );
3341 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3342 regnode * const noper_next = regnext( noper );
3343 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3344 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3347 DEBUG_TRIE_COMPILE_r({
3348 regprop(RExC_rx, mysv, cur);
3349 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3350 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3352 regprop(RExC_rx, mysv, noper);
3353 PerlIO_printf( Perl_debug_log, " -> %s",
3354 SvPV_nolen_const(mysv));
3357 regprop(RExC_rx, mysv, noper_next );
3358 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3359 SvPV_nolen_const(mysv));
3361 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3362 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3363 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3367 /* Is noper a trieable nodetype that can be merged with the
3368 * current trie (if there is one)? */
3372 ( noper_trietype == NOTHING)
3373 || ( trietype == NOTHING )
3374 || ( trietype == noper_trietype )
3377 && noper_next == tail
3381 /* Handle mergable triable node
3382 * Either we are the first node in a new trieable sequence,
3383 * in which case we do some bookkeeping, otherwise we update
3384 * the end pointer. */
3387 if ( noper_trietype == NOTHING ) {
3388 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3389 regnode * const noper_next = regnext( noper );
3390 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3391 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3394 if ( noper_next_trietype ) {
3395 trietype = noper_next_trietype;
3396 } else if (noper_next_type) {
3397 /* a NOTHING regop is 1 regop wide. We need at least two
3398 * for a trie so we can't merge this in */
3402 trietype = noper_trietype;
3405 if ( trietype == NOTHING )
3406 trietype = noper_trietype;
3411 } /* end handle mergable triable node */
3413 /* handle unmergable node -
3414 * noper may either be a triable node which can not be tried
3415 * together with the current trie, or a non triable node */
3417 /* If last is set and trietype is not NOTHING then we have found
3418 * at least two triable branch sequences in a row of a similar
3419 * trietype so we can turn them into a trie. If/when we
3420 * allow NOTHING to start a trie sequence this condition will be
3421 * required, and it isn't expensive so we leave it in for now. */
3422 if ( trietype != NOTHING )
3423 make_trie( pRExC_state,
3424 startbranch, first, cur, tail, count,
3425 trietype, depth+1 );
3426 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3430 && noper_next == tail
3433 /* noper is triable, so we can start a new trie sequence */
3436 trietype = noper_trietype;
3438 /* if we already saw a first but the current node is not triable then we have
3439 * to reset the first information. */
3444 } /* end handle unmergable node */
3445 } /* loop over branches */
3446 DEBUG_TRIE_COMPILE_r({
3447 regprop(RExC_rx, mysv, cur);
3448 PerlIO_printf( Perl_debug_log,
3449 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3450 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3454 if ( trietype != NOTHING ) {
3455 /* the last branch of the sequence was part of a trie,
3456 * so we have to construct it here outside of the loop
3458 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3459 #ifdef TRIE_STUDY_OPT
3460 if ( ((made == MADE_EXACT_TRIE &&
3461 startbranch == first)
3462 || ( first_non_open == first )) &&
3464 flags |= SCF_TRIE_RESTUDY;
3465 if ( startbranch == first
3468 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3473 /* at this point we know whatever we have is a NOTHING sequence/branch
3474 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3476 if ( startbranch == first ) {
3478 /* the entire thing is a NOTHING sequence, something like this:
3479 * (?:|) So we can turn it into a plain NOTHING op. */
3480 DEBUG_TRIE_COMPILE_r({
3481 regprop(RExC_rx, mysv, cur);
3482 PerlIO_printf( Perl_debug_log,
3483 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3484 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3487 OP(startbranch)= NOTHING;
3488 NEXT_OFF(startbranch)= tail - startbranch;
3489 for ( opt= startbranch + 1; opt < tail ; opt++ )
3493 } /* end if ( last) */
3494 } /* TRIE_MAXBUF is non zero */
3499 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3500 scan = NEXTOPER(NEXTOPER(scan));
3501 } else /* single branch is optimized. */
3502 scan = NEXTOPER(scan);
3504 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3505 scan_frame *newframe = NULL;
3510 if (OP(scan) != SUSPEND) {
3511 /* set the pointer */
3512 if (OP(scan) == GOSUB) {
3514 RExC_recurse[ARG2L(scan)] = scan;
3515 start = RExC_open_parens[paren-1];
3516 end = RExC_close_parens[paren-1];
3519 start = RExC_rxi->program + 1;
3523 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3524 SAVEFREEPV(recursed);
3526 if (!PAREN_TEST(recursed,paren+1)) {
3527 PAREN_SET(recursed,paren+1);
3528 Newx(newframe,1,scan_frame);
3530 if (flags & SCF_DO_SUBSTR) {
3531 SCAN_COMMIT(pRExC_state,data,minlenp);
3532 data->longest = &(data->longest_float);
3534 is_inf = is_inf_internal = 1;
3535 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3536 cl_anything(pRExC_state, data->start_class);
3537 flags &= ~SCF_DO_STCLASS;
3540 Newx(newframe,1,scan_frame);
3543 end = regnext(scan);
3548 SAVEFREEPV(newframe);
3549 newframe->next = regnext(scan);
3550 newframe->last = last;
3551 newframe->stop = stopparen;
3552 newframe->prev = frame;
3562 else if (OP(scan) == EXACT) {
3563 I32 l = STR_LEN(scan);
3566 const U8 * const s = (U8*)STRING(scan);
3567 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3568 l = utf8_length(s, s + l);
3570 uc = *((U8*)STRING(scan));
3573 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3574 /* The code below prefers earlier match for fixed
3575 offset, later match for variable offset. */
3576 if (data->last_end == -1) { /* Update the start info. */
3577 data->last_start_min = data->pos_min;
3578 data->last_start_max = is_inf
3579 ? I32_MAX : data->pos_min + data->pos_delta;
3581 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3583 SvUTF8_on(data->last_found);
3585 SV * const sv = data->last_found;
3586 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3587 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3588 if (mg && mg->mg_len >= 0)
3589 mg->mg_len += utf8_length((U8*)STRING(scan),
3590 (U8*)STRING(scan)+STR_LEN(scan));
3592 data->last_end = data->pos_min + l;
3593 data->pos_min += l; /* As in the first entry. */
3594 data->flags &= ~SF_BEFORE_EOL;
3596 if (flags & SCF_DO_STCLASS_AND) {
3597 /* Check whether it is compatible with what we know already! */
3601 /* If compatible, we or it in below. It is compatible if is
3602 * in the bitmp and either 1) its bit or its fold is set, or 2)
3603 * it's for a locale. Even if there isn't unicode semantics
3604 * here, at runtime there may be because of matching against a
3605 * utf8 string, so accept a possible false positive for
3606 * latin1-range folds */
3608 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3609 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3610 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
3611 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3616 ANYOF_CLASS_ZERO(data->start_class);
3617 ANYOF_BITMAP_ZERO(data->start_class);
3619 ANYOF_BITMAP_SET(data->start_class, uc);
3620 else if (uc >= 0x100) {
3623 /* Some Unicode code points fold to the Latin1 range; as
3624 * XXX temporary code, instead of figuring out if this is
3625 * one, just assume it is and set all the start class bits
3626 * that could be some such above 255 code point's fold
3627 * which will generate fals positives. As the code
3628 * elsewhere that does compute the fold settles down, it
3629 * can be extracted out and re-used here */
3630 for (i = 0; i < 256; i++){
3631 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3632 ANYOF_BITMAP_SET(data->start_class, i);
3636 data->start_class->flags &= ~ANYOF_EOS;
3638 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3640 else if (flags & SCF_DO_STCLASS_OR) {
3641 /* false positive possible if the class is case-folded */
3643 ANYOF_BITMAP_SET(data->start_class, uc);
3645 data->start_class->flags |= ANYOF_UNICODE_ALL;
3646 data->start_class->flags &= ~ANYOF_EOS;
3647 cl_and(data->start_class, and_withp);
3649 flags &= ~SCF_DO_STCLASS;
3651 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3652 I32 l = STR_LEN(scan);
3653 UV uc = *((U8*)STRING(scan));
3655 /* Search for fixed substrings supports EXACT only. */
3656 if (flags & SCF_DO_SUBSTR) {
3658 SCAN_COMMIT(pRExC_state, data, minlenp);
3661 const U8 * const s = (U8 *)STRING(scan);
3662 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3663 l = utf8_length(s, s + l);
3665 if (has_exactf_sharp_s) {
3666 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3668 min += l - min_subtract;
3672 delta += min_subtract;
3673 if (flags & SCF_DO_SUBSTR) {
3674 data->pos_min += l - min_subtract;
3675 if (data->pos_min < 0) {
3678 data->pos_delta += min_subtract;
3680 data->longest = &(data->longest_float);
3683 if (flags & SCF_DO_STCLASS_AND) {
3684 /* Check whether it is compatible with what we know already! */
3687 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3688 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3689 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3693 ANYOF_CLASS_ZERO(data->start_class);
3694 ANYOF_BITMAP_ZERO(data->start_class);
3696 ANYOF_BITMAP_SET(data->start_class, uc);
3697 data->start_class->flags &= ~ANYOF_EOS;
3698 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
3699 if (OP(scan) == EXACTFL) {
3700 /* XXX This set is probably no longer necessary, and
3701 * probably wrong as LOCALE now is on in the initial
3703 data->start_class->flags |= ANYOF_LOCALE;
3707 /* Also set the other member of the fold pair. In case
3708 * that unicode semantics is called for at runtime, use
3709 * the full latin1 fold. (Can't do this for locale,
3710 * because not known until runtime) */
3711 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3713 /* All other (EXACTFL handled above) folds except under
3714 * /iaa that include s, S, and sharp_s also may include
3716 if (OP(scan) != EXACTFA) {
3717 if (uc == 's' || uc == 'S') {
3718 ANYOF_BITMAP_SET(data->start_class,
3719 LATIN_SMALL_LETTER_SHARP_S);
3721 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3722 ANYOF_BITMAP_SET(data->start_class, 's');
3723 ANYOF_BITMAP_SET(data->start_class, 'S');
3728 else if (uc >= 0x100) {
3730 for (i = 0; i < 256; i++){
3731 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3732 ANYOF_BITMAP_SET(data->start_class, i);
3737 else if (flags & SCF_DO_STCLASS_OR) {
3738 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
3739 /* false positive possible if the class is case-folded.
3740 Assume that the locale settings are the same... */
3742 ANYOF_BITMAP_SET(data->start_class, uc);
3743 if (OP(scan) != EXACTFL) {
3745 /* And set the other member of the fold pair, but
3746 * can't do that in locale because not known until
3748 ANYOF_BITMAP_SET(data->start_class,
3749 PL_fold_latin1[uc]);
3751 /* All folds except under /iaa that include s, S,
3752 * and sharp_s also may include the others */
3753 if (OP(scan) != EXACTFA) {
3754 if (uc == 's' || uc == 'S') {
3755 ANYOF_BITMAP_SET(data->start_class,
3756 LATIN_SMALL_LETTER_SHARP_S);
3758 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3759 ANYOF_BITMAP_SET(data->start_class, 's');
3760 ANYOF_BITMAP_SET(data->start_class, 'S');
3765 data->start_class->flags &= ~ANYOF_EOS;
3767 cl_and(data->start_class, and_withp);
3769 flags &= ~SCF_DO_STCLASS;
3771 else if (REGNODE_VARIES(OP(scan))) {
3772 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3773 I32 f = flags, pos_before = 0;
3774 regnode * const oscan = scan;
3775 struct regnode_charclass_class this_class;
3776 struct regnode_charclass_class *oclass = NULL;
3777 I32 next_is_eval = 0;
3779 switch (PL_regkind[OP(scan)]) {
3780 case WHILEM: /* End of (?:...)* . */
3781 scan = NEXTOPER(scan);
3784 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3785 next = NEXTOPER(scan);
3786 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3788 maxcount = REG_INFTY;
3789 next = regnext(scan);
3790 scan = NEXTOPER(scan);
3794 if (flags & SCF_DO_SUBSTR)
3799 if (flags & SCF_DO_STCLASS) {
3801 maxcount = REG_INFTY;
3802 next = regnext(scan);
3803 scan = NEXTOPER(scan);
3806 is_inf = is_inf_internal = 1;
3807 scan = regnext(scan);
3808 if (flags & SCF_DO_SUBSTR) {
3809 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3810 data->longest = &(data->longest_float);
3812 goto optimize_curly_tail;
3814 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3815 && (scan->flags == stopparen))
3820 mincount = ARG1(scan);
3821 maxcount = ARG2(scan);
3823 next = regnext(scan);
3824 if (OP(scan) == CURLYX) {
3825 I32 lp = (data ? *(data->last_closep) : 0);
3826 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3828 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3829 next_is_eval = (OP(scan) == EVAL);
https://perl5.git.perl.org / perl5.git / blob