5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
88 #include "dquote_static.c"
89 #ifndef PERL_IN_XSUB_RE
90 # include "charclass_invlists.h"
98 # if defined(BUGGY_MSC6)
99 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
100 # pragma optimize("a",off)
101 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
102 # pragma optimize("w",on )
103 # endif /* BUGGY_MSC6 */
107 #define STATIC static
110 typedef struct RExC_state_t {
111 U32 flags; /* are we folding, multilining? */
112 char *precomp; /* uncompiled string. */
113 REGEXP *rx_sv; /* The SV that is the regexp. */
114 regexp *rx; /* perl core regexp structure */
115 regexp_internal *rxi; /* internal data for regexp object pprivate field */
116 char *start; /* Start of input for compile */
117 char *end; /* End of input for compile */
118 char *parse; /* Input-scan pointer. */
119 I32 whilem_seen; /* number of WHILEM in this expr */
120 regnode *emit_start; /* Start of emitted-code area */
121 regnode *emit_bound; /* First regnode outside of the allocated space */
122 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
123 I32 naughty; /* How bad is this pattern? */
124 I32 sawback; /* Did we see \1, ...? */
126 I32 size; /* Code size. */
127 I32 npar; /* Capture buffer count, (OPEN). */
128 I32 cpar; /* Capture buffer count, (CLOSE). */
129 I32 nestroot; /* root parens we are in - used by accept */
133 regnode **open_parens; /* pointers to open parens */
134 regnode **close_parens; /* pointers to close parens */
135 regnode *opend; /* END node in program */
136 I32 utf8; /* whether the pattern is utf8 or not */
137 I32 orig_utf8; /* whether the pattern was originally in utf8 */
138 /* XXX use this for future optimisation of case
139 * where pattern must be upgraded to utf8. */
140 I32 uni_semantics; /* If a d charset modifier should use unicode
141 rules, even if the pattern is not in
143 HV *paren_names; /* Paren names */
145 regnode **recurse; /* Recurse regops */
146 I32 recurse_count; /* Number of recurse regops */
149 I32 override_recoding;
151 char *starttry; /* -Dr: where regtry was called. */
152 #define RExC_starttry (pRExC_state->starttry)
155 const char *lastparse;
157 AV *paren_name_list; /* idx -> name */
158 #define RExC_lastparse (pRExC_state->lastparse)
159 #define RExC_lastnum (pRExC_state->lastnum)
160 #define RExC_paren_name_list (pRExC_state->paren_name_list)
164 #define RExC_flags (pRExC_state->flags)
165 #define RExC_precomp (pRExC_state->precomp)
166 #define RExC_rx_sv (pRExC_state->rx_sv)
167 #define RExC_rx (pRExC_state->rx)
168 #define RExC_rxi (pRExC_state->rxi)
169 #define RExC_start (pRExC_state->start)
170 #define RExC_end (pRExC_state->end)
171 #define RExC_parse (pRExC_state->parse)
172 #define RExC_whilem_seen (pRExC_state->whilem_seen)
173 #ifdef RE_TRACK_PATTERN_OFFSETS
174 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
176 #define RExC_emit (pRExC_state->emit)
177 #define RExC_emit_start (pRExC_state->emit_start)
178 #define RExC_emit_bound (pRExC_state->emit_bound)
179 #define RExC_naughty (pRExC_state->naughty)
180 #define RExC_sawback (pRExC_state->sawback)
181 #define RExC_seen (pRExC_state->seen)
182 #define RExC_size (pRExC_state->size)
183 #define RExC_npar (pRExC_state->npar)
184 #define RExC_nestroot (pRExC_state->nestroot)
185 #define RExC_extralen (pRExC_state->extralen)
186 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
187 #define RExC_seen_evals (pRExC_state->seen_evals)
188 #define RExC_utf8 (pRExC_state->utf8)
189 #define RExC_uni_semantics (pRExC_state->uni_semantics)
190 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
191 #define RExC_open_parens (pRExC_state->open_parens)
192 #define RExC_close_parens (pRExC_state->close_parens)
193 #define RExC_opend (pRExC_state->opend)
194 #define RExC_paren_names (pRExC_state->paren_names)
195 #define RExC_recurse (pRExC_state->recurse)
196 #define RExC_recurse_count (pRExC_state->recurse_count)
197 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
198 #define RExC_contains_locale (pRExC_state->contains_locale)
199 #define RExC_override_recoding (pRExC_state->override_recoding)
202 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
203 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
204 ((*s) == '{' && regcurly(s)))
207 #undef SPSTART /* dratted cpp namespace... */
210 * Flags to be passed up and down.
212 #define WORST 0 /* Worst case. */
213 #define HASWIDTH 0x01 /* Known to match non-null strings. */
215 /* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
216 * character, and if utf8, must be invariant. Note that this is not the same thing as REGNODE_SIMPLE */
218 #define SPSTART 0x04 /* Starts with * or +. */
219 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
220 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
222 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
224 /* whether trie related optimizations are enabled */
225 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
226 #define TRIE_STUDY_OPT
227 #define FULL_TRIE_STUDY
233 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
234 #define PBITVAL(paren) (1 << ((paren) & 7))
235 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
236 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
237 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
239 /* If not already in utf8, do a longjmp back to the beginning */
240 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
241 #define REQUIRE_UTF8 STMT_START { \
242 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
245 /* About scan_data_t.
247 During optimisation we recurse through the regexp program performing
248 various inplace (keyhole style) optimisations. In addition study_chunk
249 and scan_commit populate this data structure with information about
250 what strings MUST appear in the pattern. We look for the longest
251 string that must appear at a fixed location, and we look for the
252 longest string that may appear at a floating location. So for instance
257 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
258 strings (because they follow a .* construct). study_chunk will identify
259 both FOO and BAR as being the longest fixed and floating strings respectively.
261 The strings can be composites, for instance
265 will result in a composite fixed substring 'foo'.
267 For each string some basic information is maintained:
269 - offset or min_offset
270 This is the position the string must appear at, or not before.
271 It also implicitly (when combined with minlenp) tells us how many
272 characters must match before the string we are searching for.
273 Likewise when combined with minlenp and the length of the string it
274 tells us how many characters must appear after the string we have
278 Only used for floating strings. This is the rightmost point that
279 the string can appear at. If set to I32 max it indicates that the
280 string can occur infinitely far to the right.
283 A pointer to the minimum length of the pattern that the string
284 was found inside. This is important as in the case of positive
285 lookahead or positive lookbehind we can have multiple patterns
290 The minimum length of the pattern overall is 3, the minimum length
291 of the lookahead part is 3, but the minimum length of the part that
292 will actually match is 1. So 'FOO's minimum length is 3, but the
293 minimum length for the F is 1. This is important as the minimum length
294 is used to determine offsets in front of and behind the string being
295 looked for. Since strings can be composites this is the length of the
296 pattern at the time it was committed with a scan_commit. Note that
297 the length is calculated by study_chunk, so that the minimum lengths
298 are not known until the full pattern has been compiled, thus the
299 pointer to the value.
303 In the case of lookbehind the string being searched for can be
304 offset past the start point of the final matching string.
305 If this value was just blithely removed from the min_offset it would
306 invalidate some of the calculations for how many chars must match
307 before or after (as they are derived from min_offset and minlen and
308 the length of the string being searched for).
309 When the final pattern is compiled and the data is moved from the
310 scan_data_t structure into the regexp structure the information
311 about lookbehind is factored in, with the information that would
312 have been lost precalculated in the end_shift field for the
315 The fields pos_min and pos_delta are used to store the minimum offset
316 and the delta to the maximum offset at the current point in the pattern.
320 typedef struct scan_data_t {
321 /*I32 len_min; unused */
322 /*I32 len_delta; unused */
326 I32 last_end; /* min value, <0 unless valid. */
329 SV **longest; /* Either &l_fixed, or &l_float. */
330 SV *longest_fixed; /* longest fixed string found in pattern */
331 I32 offset_fixed; /* offset where it starts */
332 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
333 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
334 SV *longest_float; /* longest floating string found in pattern */
335 I32 offset_float_min; /* earliest point in string it can appear */
336 I32 offset_float_max; /* latest point in string it can appear */
337 I32 *minlen_float; /* pointer to the minlen relevant to the string */
338 I32 lookbehind_float; /* is the position of the string modified by LB */
342 struct regnode_charclass_class *start_class;
346 * Forward declarations for pregcomp()'s friends.
349 static const scan_data_t zero_scan_data =
350 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
352 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
353 #define SF_BEFORE_SEOL 0x0001
354 #define SF_BEFORE_MEOL 0x0002
355 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
356 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
359 # define SF_FIX_SHIFT_EOL (0+2)
360 # define SF_FL_SHIFT_EOL (0+4)
362 # define SF_FIX_SHIFT_EOL (+2)
363 # define SF_FL_SHIFT_EOL (+4)
366 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
367 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
369 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
370 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
371 #define SF_IS_INF 0x0040
372 #define SF_HAS_PAR 0x0080
373 #define SF_IN_PAR 0x0100
374 #define SF_HAS_EVAL 0x0200
375 #define SCF_DO_SUBSTR 0x0400
376 #define SCF_DO_STCLASS_AND 0x0800
377 #define SCF_DO_STCLASS_OR 0x1000
378 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
379 #define SCF_WHILEM_VISITED_POS 0x2000
381 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
382 #define SCF_SEEN_ACCEPT 0x8000
384 #define UTF cBOOL(RExC_utf8)
386 /* The enums for all these are ordered so things work out correctly */
387 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
388 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
389 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
390 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
391 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
392 #define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
393 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
395 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
397 #define OOB_UNICODE 12345678
398 #define OOB_NAMEDCLASS -1
400 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
401 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
404 /* length of regex to show in messages that don't mark a position within */
405 #define RegexLengthToShowInErrorMessages 127
408 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
409 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
410 * op/pragma/warn/regcomp.
412 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
413 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
415 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
418 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
419 * arg. Show regex, up to a maximum length. If it's too long, chop and add
422 #define _FAIL(code) STMT_START { \
423 const char *ellipses = ""; \
424 IV len = RExC_end - RExC_precomp; \
427 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
428 if (len > RegexLengthToShowInErrorMessages) { \
429 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
430 len = RegexLengthToShowInErrorMessages - 10; \
436 #define FAIL(msg) _FAIL( \
437 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
438 msg, (int)len, RExC_precomp, ellipses))
440 #define FAIL2(msg,arg) _FAIL( \
441 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
442 arg, (int)len, RExC_precomp, ellipses))
445 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
447 #define Simple_vFAIL(m) STMT_START { \
448 const IV offset = RExC_parse - RExC_precomp; \
449 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
450 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
454 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
456 #define vFAIL(m) STMT_START { \
458 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
463 * Like Simple_vFAIL(), but accepts two arguments.
465 #define Simple_vFAIL2(m,a1) STMT_START { \
466 const IV offset = RExC_parse - RExC_precomp; \
467 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
468 (int)offset, RExC_precomp, RExC_precomp + offset); \
472 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
474 #define vFAIL2(m,a1) STMT_START { \
476 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
477 Simple_vFAIL2(m, a1); \
482 * Like Simple_vFAIL(), but accepts three arguments.
484 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
485 const IV offset = RExC_parse - RExC_precomp; \
486 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
487 (int)offset, RExC_precomp, RExC_precomp + offset); \
491 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
493 #define vFAIL3(m,a1,a2) STMT_START { \
495 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
496 Simple_vFAIL3(m, a1, a2); \
500 * Like Simple_vFAIL(), but accepts four arguments.
502 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
503 const IV offset = RExC_parse - RExC_precomp; \
504 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
505 (int)offset, RExC_precomp, RExC_precomp + offset); \
508 #define ckWARNreg(loc,m) STMT_START { \
509 const IV offset = loc - RExC_precomp; \
510 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
511 (int)offset, RExC_precomp, RExC_precomp + offset); \
514 #define ckWARNregdep(loc,m) STMT_START { \
515 const IV offset = loc - RExC_precomp; \
516 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
518 (int)offset, RExC_precomp, RExC_precomp + offset); \
521 #define ckWARN2regdep(loc,m, a1) STMT_START { \
522 const IV offset = loc - RExC_precomp; \
523 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
525 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
528 #define ckWARN2reg(loc, m, a1) STMT_START { \
529 const IV offset = loc - RExC_precomp; \
530 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
531 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
534 #define vWARN3(loc, m, a1, a2) STMT_START { \
535 const IV offset = loc - RExC_precomp; \
536 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
537 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
540 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
541 const IV offset = loc - RExC_precomp; \
542 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
543 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
546 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
547 const IV offset = loc - RExC_precomp; \
548 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
549 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
552 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
553 const IV offset = loc - RExC_precomp; \
554 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
555 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
558 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
559 const IV offset = loc - RExC_precomp; \
560 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
561 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
565 /* Allow for side effects in s */
566 #define REGC(c,s) STMT_START { \
567 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
570 /* Macros for recording node offsets. 20001227 mjd@plover.com
571 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
572 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
573 * Element 0 holds the number n.
574 * Position is 1 indexed.
576 #ifndef RE_TRACK_PATTERN_OFFSETS
577 #define Set_Node_Offset_To_R(node,byte)
578 #define Set_Node_Offset(node,byte)
579 #define Set_Cur_Node_Offset
580 #define Set_Node_Length_To_R(node,len)
581 #define Set_Node_Length(node,len)
582 #define Set_Node_Cur_Length(node)
583 #define Node_Offset(n)
584 #define Node_Length(n)
585 #define Set_Node_Offset_Length(node,offset,len)
586 #define ProgLen(ri) ri->u.proglen
587 #define SetProgLen(ri,x) ri->u.proglen = x
589 #define ProgLen(ri) ri->u.offsets[0]
590 #define SetProgLen(ri,x) ri->u.offsets[0] = x
591 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
593 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
594 __LINE__, (int)(node), (int)(byte))); \
596 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
598 RExC_offsets[2*(node)-1] = (byte); \
603 #define Set_Node_Offset(node,byte) \
604 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
605 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
607 #define Set_Node_Length_To_R(node,len) STMT_START { \
609 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
610 __LINE__, (int)(node), (int)(len))); \
612 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
614 RExC_offsets[2*(node)] = (len); \
619 #define Set_Node_Length(node,len) \
620 Set_Node_Length_To_R((node)-RExC_emit_start, len)
621 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
622 #define Set_Node_Cur_Length(node) \
623 Set_Node_Length(node, RExC_parse - parse_start)
625 /* Get offsets and lengths */
626 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
627 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
629 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
630 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
631 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
635 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
636 #define EXPERIMENTAL_INPLACESCAN
637 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
639 #define DEBUG_STUDYDATA(str,data,depth) \
640 DEBUG_OPTIMISE_MORE_r(if(data){ \
641 PerlIO_printf(Perl_debug_log, \
642 "%*s" str "Pos:%"IVdf"/%"IVdf \
643 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
644 (int)(depth)*2, "", \
645 (IV)((data)->pos_min), \
646 (IV)((data)->pos_delta), \
647 (UV)((data)->flags), \
648 (IV)((data)->whilem_c), \
649 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
650 is_inf ? "INF " : "" \
652 if ((data)->last_found) \
653 PerlIO_printf(Perl_debug_log, \
654 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
655 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
656 SvPVX_const((data)->last_found), \
657 (IV)((data)->last_end), \
658 (IV)((data)->last_start_min), \
659 (IV)((data)->last_start_max), \
660 ((data)->longest && \
661 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
662 SvPVX_const((data)->longest_fixed), \
663 (IV)((data)->offset_fixed), \
664 ((data)->longest && \
665 (data)->longest==&((data)->longest_float)) ? "*" : "", \
666 SvPVX_const((data)->longest_float), \
667 (IV)((data)->offset_float_min), \
668 (IV)((data)->offset_float_max) \
670 PerlIO_printf(Perl_debug_log,"\n"); \
673 static void clear_re(pTHX_ void *r);
675 /* Mark that we cannot extend a found fixed substring at this point.
676 Update the longest found anchored substring and the longest found
677 floating substrings if needed. */
680 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
682 const STRLEN l = CHR_SVLEN(data->last_found);
683 const STRLEN old_l = CHR_SVLEN(*data->longest);
684 GET_RE_DEBUG_FLAGS_DECL;
686 PERL_ARGS_ASSERT_SCAN_COMMIT;
688 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
689 SvSetMagicSV(*data->longest, data->last_found);
690 if (*data->longest == data->longest_fixed) {
691 data->offset_fixed = l ? data->last_start_min : data->pos_min;
692 if (data->flags & SF_BEFORE_EOL)
694 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
696 data->flags &= ~SF_FIX_BEFORE_EOL;
697 data->minlen_fixed=minlenp;
698 data->lookbehind_fixed=0;
700 else { /* *data->longest == data->longest_float */
701 data->offset_float_min = l ? data->last_start_min : data->pos_min;
702 data->offset_float_max = (l
703 ? data->last_start_max
704 : data->pos_min + data->pos_delta);
705 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
706 data->offset_float_max = I32_MAX;
707 if (data->flags & SF_BEFORE_EOL)
709 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
711 data->flags &= ~SF_FL_BEFORE_EOL;
712 data->minlen_float=minlenp;
713 data->lookbehind_float=0;
716 SvCUR_set(data->last_found, 0);
718 SV * const sv = data->last_found;
719 if (SvUTF8(sv) && SvMAGICAL(sv)) {
720 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
726 data->flags &= ~SF_BEFORE_EOL;
727 DEBUG_STUDYDATA("commit: ",data,0);
730 /* Can match anything (initialization) */
732 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
734 PERL_ARGS_ASSERT_CL_ANYTHING;
736 ANYOF_BITMAP_SETALL(cl);
737 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
738 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
740 /* If any portion of the regex is to operate under locale rules,
741 * initialization includes it. The reason this isn't done for all regexes
742 * is that the optimizer was written under the assumption that locale was
743 * all-or-nothing. Given the complexity and lack of documentation in the
744 * optimizer, and that there are inadequate test cases for locale, so many
745 * parts of it may not work properly, it is safest to avoid locale unless
747 if (RExC_contains_locale) {
748 ANYOF_CLASS_SETALL(cl); /* /l uses class */
749 cl->flags |= ANYOF_LOCALE;
752 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
756 /* Can match anything (initialization) */
758 S_cl_is_anything(const struct regnode_charclass_class *cl)
762 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
764 for (value = 0; value <= ANYOF_MAX; value += 2)
765 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
767 if (!(cl->flags & ANYOF_UNICODE_ALL))
769 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
774 /* Can match anything (initialization) */
776 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
778 PERL_ARGS_ASSERT_CL_INIT;
780 Zero(cl, 1, struct regnode_charclass_class);
782 cl_anything(pRExC_state, cl);
783 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
786 /* These two functions currently do the exact same thing */
787 #define cl_init_zero S_cl_init
789 /* 'AND' a given class with another one. Can create false positives. 'cl'
790 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
791 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
793 S_cl_and(struct regnode_charclass_class *cl,
794 const struct regnode_charclass_class *and_with)
796 PERL_ARGS_ASSERT_CL_AND;
798 assert(and_with->type == ANYOF);
800 /* I (khw) am not sure all these restrictions are necessary XXX */
801 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
802 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
803 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
804 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
805 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
808 if (and_with->flags & ANYOF_INVERT)
809 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
810 cl->bitmap[i] &= ~and_with->bitmap[i];
812 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
813 cl->bitmap[i] &= and_with->bitmap[i];
814 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
816 if (and_with->flags & ANYOF_INVERT) {
818 /* Here, the and'ed node is inverted. Get the AND of the flags that
819 * aren't affected by the inversion. Those that are affected are
820 * handled individually below */
821 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
822 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
823 cl->flags |= affected_flags;
825 /* We currently don't know how to deal with things that aren't in the
826 * bitmap, but we know that the intersection is no greater than what
827 * is already in cl, so let there be false positives that get sorted
828 * out after the synthetic start class succeeds, and the node is
829 * matched for real. */
831 /* The inversion of these two flags indicate that the resulting
832 * intersection doesn't have them */
833 if (and_with->flags & ANYOF_UNICODE_ALL) {
834 cl->flags &= ~ANYOF_UNICODE_ALL;
836 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
837 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
840 else { /* and'd node is not inverted */
841 U8 outside_bitmap_but_not_utf8; /* Temp variable */
843 if (! ANYOF_NONBITMAP(and_with)) {
845 /* Here 'and_with' doesn't match anything outside the bitmap
846 * (except possibly ANYOF_UNICODE_ALL), which means the
847 * intersection can't either, except for ANYOF_UNICODE_ALL, in
848 * which case we don't know what the intersection is, but it's no
849 * greater than what cl already has, so can just leave it alone,
850 * with possible false positives */
851 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
852 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
853 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
856 else if (! ANYOF_NONBITMAP(cl)) {
858 /* Here, 'and_with' does match something outside the bitmap, and cl
859 * doesn't have a list of things to match outside the bitmap. If
860 * cl can match all code points above 255, the intersection will
861 * be those above-255 code points that 'and_with' matches. If cl
862 * can't match all Unicode code points, it means that it can't
863 * match anything outside the bitmap (since the 'if' that got us
864 * into this block tested for that), so we leave the bitmap empty.
866 if (cl->flags & ANYOF_UNICODE_ALL) {
867 ARG_SET(cl, ARG(and_with));
869 /* and_with's ARG may match things that don't require UTF8.
870 * And now cl's will too, in spite of this being an 'and'. See
871 * the comments below about the kludge */
872 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
876 /* Here, both 'and_with' and cl match something outside the
877 * bitmap. Currently we do not do the intersection, so just match
878 * whatever cl had at the beginning. */
882 /* Take the intersection of the two sets of flags. However, the
883 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
884 * kludge around the fact that this flag is not treated like the others
885 * which are initialized in cl_anything(). The way the optimizer works
886 * is that the synthetic start class (SSC) is initialized to match
887 * anything, and then the first time a real node is encountered, its
888 * values are AND'd with the SSC's with the result being the values of
889 * the real node. However, there are paths through the optimizer where
890 * the AND never gets called, so those initialized bits are set
891 * inappropriately, which is not usually a big deal, as they just cause
892 * false positives in the SSC, which will just mean a probably
893 * imperceptible slow down in execution. However this bit has a
894 * higher false positive consequence in that it can cause utf8.pm,
895 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
896 * bigger slowdown and also causes significant extra memory to be used.
897 * In order to prevent this, the code now takes a different tack. The
898 * bit isn't set unless some part of the regular expression needs it,
899 * but once set it won't get cleared. This means that these extra
900 * modules won't get loaded unless there was some path through the
901 * pattern that would have required them anyway, and so any false
902 * positives that occur by not ANDing them out when they could be
903 * aren't as severe as they would be if we treated this bit like all
905 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
906 & ANYOF_NONBITMAP_NON_UTF8;
907 cl->flags &= and_with->flags;
908 cl->flags |= outside_bitmap_but_not_utf8;
912 /* 'OR' a given class with another one. Can create false positives. 'cl'
913 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
914 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
916 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
918 PERL_ARGS_ASSERT_CL_OR;
920 if (or_with->flags & ANYOF_INVERT) {
922 /* Here, the or'd node is to be inverted. This means we take the
923 * complement of everything not in the bitmap, but currently we don't
924 * know what that is, so give up and match anything */
925 if (ANYOF_NONBITMAP(or_with)) {
926 cl_anything(pRExC_state, cl);
929 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
930 * <= (B1 | !B2) | (CL1 | !CL2)
931 * which is wasteful if CL2 is small, but we ignore CL2:
932 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
933 * XXXX Can we handle case-fold? Unclear:
934 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
935 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
937 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
938 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
939 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
942 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
943 cl->bitmap[i] |= ~or_with->bitmap[i];
944 } /* XXXX: logic is complicated otherwise */
946 cl_anything(pRExC_state, cl);
949 /* And, we can just take the union of the flags that aren't affected
950 * by the inversion */
951 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
953 /* For the remaining flags:
954 ANYOF_UNICODE_ALL and inverted means to not match anything above
955 255, which means that the union with cl should just be
956 what cl has in it, so can ignore this flag
957 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
958 is 127-255 to match them, but then invert that, so the
959 union with cl should just be what cl has in it, so can
962 } else { /* 'or_with' is not inverted */
963 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
964 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
965 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
966 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
969 /* OR char bitmap and class bitmap separately */
970 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
971 cl->bitmap[i] |= or_with->bitmap[i];
972 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
973 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
974 cl->classflags[i] |= or_with->classflags[i];
975 cl->flags |= ANYOF_CLASS;
978 else { /* XXXX: logic is complicated, leave it along for a moment. */
979 cl_anything(pRExC_state, cl);
982 if (ANYOF_NONBITMAP(or_with)) {
984 /* Use the added node's outside-the-bit-map match if there isn't a
985 * conflict. If there is a conflict (both nodes match something
986 * outside the bitmap, but what they match outside is not the same
987 * pointer, and hence not easily compared until XXX we extend
988 * inversion lists this far), give up and allow the start class to
989 * match everything outside the bitmap. If that stuff is all above
990 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
991 if (! ANYOF_NONBITMAP(cl)) {
992 ARG_SET(cl, ARG(or_with));
994 else if (ARG(cl) != ARG(or_with)) {
996 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
997 cl_anything(pRExC_state, cl);
1000 cl->flags |= ANYOF_UNICODE_ALL;
1005 /* Take the union */
1006 cl->flags |= or_with->flags;
1010 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1011 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1012 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1013 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1018 dump_trie(trie,widecharmap,revcharmap)
1019 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1020 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1022 These routines dump out a trie in a somewhat readable format.
1023 The _interim_ variants are used for debugging the interim
1024 tables that are used to generate the final compressed
1025 representation which is what dump_trie expects.
1027 Part of the reason for their existence is to provide a form
1028 of documentation as to how the different representations function.
1033 Dumps the final compressed table form of the trie to Perl_debug_log.
1034 Used for debugging make_trie().
1038 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1039 AV *revcharmap, U32 depth)
1042 SV *sv=sv_newmortal();
1043 int colwidth= widecharmap ? 6 : 4;
1045 GET_RE_DEBUG_FLAGS_DECL;
1047 PERL_ARGS_ASSERT_DUMP_TRIE;
1049 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1050 (int)depth * 2 + 2,"",
1051 "Match","Base","Ofs" );
1053 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1054 SV ** const tmp = av_fetch( revcharmap, state, 0);
1056 PerlIO_printf( Perl_debug_log, "%*s",
1058 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1059 PL_colors[0], PL_colors[1],
1060 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1061 PERL_PV_ESCAPE_FIRSTCHAR
1066 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1067 (int)depth * 2 + 2,"");
1069 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1070 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1071 PerlIO_printf( Perl_debug_log, "\n");
1073 for( state = 1 ; state < trie->statecount ; state++ ) {
1074 const U32 base = trie->states[ state ].trans.base;
1076 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1078 if ( trie->states[ state ].wordnum ) {
1079 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1081 PerlIO_printf( Perl_debug_log, "%6s", "" );
1084 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1089 while( ( base + ofs < trie->uniquecharcount ) ||
1090 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1091 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1094 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1096 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1097 if ( ( base + ofs >= trie->uniquecharcount ) &&
1098 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1099 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1101 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1103 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1105 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1109 PerlIO_printf( Perl_debug_log, "]");
1112 PerlIO_printf( Perl_debug_log, "\n" );
1114 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1115 for (word=1; word <= trie->wordcount; word++) {
1116 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1117 (int)word, (int)(trie->wordinfo[word].prev),
1118 (int)(trie->wordinfo[word].len));
1120 PerlIO_printf(Perl_debug_log, "\n" );
1123 Dumps a fully constructed but uncompressed trie in list form.
1124 List tries normally only are used for construction when the number of
1125 possible chars (trie->uniquecharcount) is very high.
1126 Used for debugging make_trie().
1129 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1130 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1134 SV *sv=sv_newmortal();
1135 int colwidth= widecharmap ? 6 : 4;
1136 GET_RE_DEBUG_FLAGS_DECL;
1138 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1140 /* print out the table precompression. */
1141 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1142 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1143 "------:-----+-----------------\n" );
1145 for( state=1 ; state < next_alloc ; state ++ ) {
1148 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1149 (int)depth * 2 + 2,"", (UV)state );
1150 if ( ! trie->states[ state ].wordnum ) {
1151 PerlIO_printf( Perl_debug_log, "%5s| ","");
1153 PerlIO_printf( Perl_debug_log, "W%4x| ",
1154 trie->states[ state ].wordnum
1157 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1158 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1160 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1162 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1163 PL_colors[0], PL_colors[1],
1164 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1165 PERL_PV_ESCAPE_FIRSTCHAR
1167 TRIE_LIST_ITEM(state,charid).forid,
1168 (UV)TRIE_LIST_ITEM(state,charid).newstate
1171 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1172 (int)((depth * 2) + 14), "");
1175 PerlIO_printf( Perl_debug_log, "\n");
1180 Dumps a fully constructed but uncompressed trie in table form.
1181 This is the normal DFA style state transition table, with a few
1182 twists to facilitate compression later.
1183 Used for debugging make_trie().
1186 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1187 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1192 SV *sv=sv_newmortal();
1193 int colwidth= widecharmap ? 6 : 4;
1194 GET_RE_DEBUG_FLAGS_DECL;
1196 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1199 print out the table precompression so that we can do a visual check
1200 that they are identical.
1203 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1205 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1206 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1208 PerlIO_printf( Perl_debug_log, "%*s",
1210 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1211 PL_colors[0], PL_colors[1],
1212 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1213 PERL_PV_ESCAPE_FIRSTCHAR
1219 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1221 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1222 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1225 PerlIO_printf( Perl_debug_log, "\n" );
1227 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1229 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1230 (int)depth * 2 + 2,"",
1231 (UV)TRIE_NODENUM( state ) );
1233 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1234 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1236 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1238 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1240 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1241 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1243 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1244 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1252 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1253 startbranch: the first branch in the whole branch sequence
1254 first : start branch of sequence of branch-exact nodes.
1255 May be the same as startbranch
1256 last : Thing following the last branch.
1257 May be the same as tail.
1258 tail : item following the branch sequence
1259 count : words in the sequence
1260 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1261 depth : indent depth
1263 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1265 A trie is an N'ary tree where the branches are determined by digital
1266 decomposition of the key. IE, at the root node you look up the 1st character and
1267 follow that branch repeat until you find the end of the branches. Nodes can be
1268 marked as "accepting" meaning they represent a complete word. Eg:
1272 would convert into the following structure. Numbers represent states, letters
1273 following numbers represent valid transitions on the letter from that state, if
1274 the number is in square brackets it represents an accepting state, otherwise it
1275 will be in parenthesis.
1277 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1281 (1) +-i->(6)-+-s->[7]
1283 +-s->(3)-+-h->(4)-+-e->[5]
1285 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1287 This shows that when matching against the string 'hers' we will begin at state 1
1288 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1289 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1290 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1291 single traverse. We store a mapping from accepting to state to which word was
1292 matched, and then when we have multiple possibilities we try to complete the
1293 rest of the regex in the order in which they occured in the alternation.
1295 The only prior NFA like behaviour that would be changed by the TRIE support is
1296 the silent ignoring of duplicate alternations which are of the form:
1298 / (DUPE|DUPE) X? (?{ ... }) Y /x
1300 Thus EVAL blocks following a trie may be called a different number of times with
1301 and without the optimisation. With the optimisations dupes will be silently
1302 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1303 the following demonstrates:
1305 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1307 which prints out 'word' three times, but
1309 'words'=~/(word|word|word)(?{ print $1 })S/
1311 which doesnt print it out at all. This is due to other optimisations kicking in.
1313 Example of what happens on a structural level:
1315 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1317 1: CURLYM[1] {1,32767}(18)
1328 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1329 and should turn into:
1331 1: CURLYM[1] {1,32767}(18)
1333 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1341 Cases where tail != last would be like /(?foo|bar)baz/:
1351 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1352 and would end up looking like:
1355 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1362 d = uvuni_to_utf8_flags(d, uv, 0);
1364 is the recommended Unicode-aware way of saying
1369 #define TRIE_STORE_REVCHAR \
1372 SV *zlopp = newSV(2); \
1373 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1374 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, uvc & 0xFF); \
1375 SvCUR_set(zlopp, kapow - flrbbbbb); \
1378 av_push(revcharmap, zlopp); \
1380 char ooooff = (char)uvc; \
1381 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1385 #define TRIE_READ_CHAR STMT_START { \
1389 if ( foldlen > 0 ) { \
1390 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1395 len = UTF8SKIP(uc);\
1396 uvc = to_utf8_fold( uc, foldbuf, &foldlen); \
1397 foldlen -= UNISKIP( uvc ); \
1398 scan = foldbuf + UNISKIP( uvc ); \
1401 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1411 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1412 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1413 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1414 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1416 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1417 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1418 TRIE_LIST_CUR( state )++; \
1421 #define TRIE_LIST_NEW(state) STMT_START { \
1422 Newxz( trie->states[ state ].trans.list, \
1423 4, reg_trie_trans_le ); \
1424 TRIE_LIST_CUR( state ) = 1; \
1425 TRIE_LIST_LEN( state ) = 4; \
1428 #define TRIE_HANDLE_WORD(state) STMT_START { \
1429 U16 dupe= trie->states[ state ].wordnum; \
1430 regnode * const noper_next = regnext( noper ); \
1433 /* store the word for dumping */ \
1435 if (OP(noper) != NOTHING) \
1436 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1438 tmp = newSVpvn_utf8( "", 0, UTF ); \
1439 av_push( trie_words, tmp ); \
1443 trie->wordinfo[curword].prev = 0; \
1444 trie->wordinfo[curword].len = wordlen; \
1445 trie->wordinfo[curword].accept = state; \
1447 if ( noper_next < tail ) { \
1449 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1450 trie->jump[curword] = (U16)(noper_next - convert); \
1452 jumper = noper_next; \
1454 nextbranch= regnext(cur); \
1458 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1459 /* chain, so that when the bits of chain are later */\
1460 /* linked together, the dups appear in the chain */\
1461 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1462 trie->wordinfo[dupe].prev = curword; \
1464 /* we haven't inserted this word yet. */ \
1465 trie->states[ state ].wordnum = curword; \
1470 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1471 ( ( base + charid >= ucharcount \
1472 && base + charid < ubound \
1473 && state == trie->trans[ base - ucharcount + charid ].check \
1474 && trie->trans[ base - ucharcount + charid ].next ) \
1475 ? trie->trans[ base - ucharcount + charid ].next \
1476 : ( state==1 ? special : 0 ) \
1480 #define MADE_JUMP_TRIE 2
1481 #define MADE_EXACT_TRIE 4
1484 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1487 /* first pass, loop through and scan words */
1488 reg_trie_data *trie;
1489 HV *widecharmap = NULL;
1490 AV *revcharmap = newAV();
1492 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1497 regnode *jumper = NULL;
1498 regnode *nextbranch = NULL;
1499 regnode *convert = NULL;
1500 U32 *prev_states; /* temp array mapping each state to previous one */
1501 /* we just use folder as a flag in utf8 */
1502 const U8 * folder = NULL;
1505 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1506 AV *trie_words = NULL;
1507 /* along with revcharmap, this only used during construction but both are
1508 * useful during debugging so we store them in the struct when debugging.
1511 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1512 STRLEN trie_charcount=0;
1514 SV *re_trie_maxbuff;
1515 GET_RE_DEBUG_FLAGS_DECL;
1517 PERL_ARGS_ASSERT_MAKE_TRIE;
1519 PERL_UNUSED_ARG(depth);
1525 case EXACTFU: folder = PL_fold_latin1; break;
1526 case EXACTF: folder = PL_fold; break;
1527 case EXACTFL: folder = PL_fold_locale; break;
1528 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u", (unsigned) flags );
1531 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1533 trie->startstate = 1;
1534 trie->wordcount = word_count;
1535 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1536 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1537 if (!(UTF && folder))
1538 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1539 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1540 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1543 trie_words = newAV();
1546 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1547 if (!SvIOK(re_trie_maxbuff)) {
1548 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1551 PerlIO_printf( Perl_debug_log,
1552 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1553 (int)depth * 2 + 2, "",
1554 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1555 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1559 /* Find the node we are going to overwrite */
1560 if ( first == startbranch && OP( last ) != BRANCH ) {
1561 /* whole branch chain */
1564 /* branch sub-chain */
1565 convert = NEXTOPER( first );
1568 /* -- First loop and Setup --
1570 We first traverse the branches and scan each word to determine if it
1571 contains widechars, and how many unique chars there are, this is
1572 important as we have to build a table with at least as many columns as we
1575 We use an array of integers to represent the character codes 0..255
1576 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1577 native representation of the character value as the key and IV's for the
1580 *TODO* If we keep track of how many times each character is used we can
1581 remap the columns so that the table compression later on is more
1582 efficient in terms of memory by ensuring the most common value is in the
1583 middle and the least common are on the outside. IMO this would be better
1584 than a most to least common mapping as theres a decent chance the most
1585 common letter will share a node with the least common, meaning the node
1586 will not be compressible. With a middle is most common approach the worst
1587 case is when we have the least common nodes twice.
1591 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1592 regnode * const noper = NEXTOPER( cur );
1593 const U8 *uc = (U8*)STRING( noper );
1594 const U8 * const e = uc + STR_LEN( noper );
1596 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1597 const U8 *scan = (U8*)NULL;
1598 U32 wordlen = 0; /* required init */
1600 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1602 if (OP(noper) == NOTHING) {
1606 if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
1607 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1608 regardless of encoding */
1610 for ( ; uc < e ; uc += len ) {
1611 TRIE_CHARCOUNT(trie)++;
1615 if ( !trie->charmap[ uvc ] ) {
1616 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1618 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
1622 /* store the codepoint in the bitmap, and its folded
1624 TRIE_BITMAP_SET(trie,uvc);
1626 /* store the folded codepoint */
1627 if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);
1630 /* store first byte of utf8 representation of
1631 variant codepoints */
1632 if (! UNI_IS_INVARIANT(uvc)) {
1633 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1636 set_bit = 0; /* We've done our bit :-) */
1641 widecharmap = newHV();
1643 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1646 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1648 if ( !SvTRUE( *svpp ) ) {
1649 sv_setiv( *svpp, ++trie->uniquecharcount );
1654 if( cur == first ) {
1657 } else if (chars < trie->minlen) {
1659 } else if (chars > trie->maxlen) {
1663 } /* end first pass */
1664 DEBUG_TRIE_COMPILE_r(
1665 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1666 (int)depth * 2 + 2,"",
1667 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1668 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1669 (int)trie->minlen, (int)trie->maxlen )
1673 We now know what we are dealing with in terms of unique chars and
1674 string sizes so we can calculate how much memory a naive
1675 representation using a flat table will take. If it's over a reasonable
1676 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1677 conservative but potentially much slower representation using an array
1680 At the end we convert both representations into the same compressed
1681 form that will be used in regexec.c for matching with. The latter
1682 is a form that cannot be used to construct with but has memory
1683 properties similar to the list form and access properties similar
1684 to the table form making it both suitable for fast searches and
1685 small enough that its feasable to store for the duration of a program.
1687 See the comment in the code where the compressed table is produced
1688 inplace from the flat tabe representation for an explanation of how
1689 the compression works.
1694 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1697 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1699 Second Pass -- Array Of Lists Representation
1701 Each state will be represented by a list of charid:state records
1702 (reg_trie_trans_le) the first such element holds the CUR and LEN
1703 points of the allocated array. (See defines above).
1705 We build the initial structure using the lists, and then convert
1706 it into the compressed table form which allows faster lookups
1707 (but cant be modified once converted).
1710 STRLEN transcount = 1;
1712 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1713 "%*sCompiling trie using list compiler\n",
1714 (int)depth * 2 + 2, ""));
1716 trie->states = (reg_trie_state *)
1717 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1718 sizeof(reg_trie_state) );
1722 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1724 regnode * const noper = NEXTOPER( cur );
1725 U8 *uc = (U8*)STRING( noper );
1726 const U8 * const e = uc + STR_LEN( noper );
1727 U32 state = 1; /* required init */
1728 U16 charid = 0; /* sanity init */
1729 U8 *scan = (U8*)NULL; /* sanity init */
1730 STRLEN foldlen = 0; /* required init */
1731 U32 wordlen = 0; /* required init */
1732 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1734 if (OP(noper) != NOTHING) {
1735 for ( ; uc < e ; uc += len ) {
1740 charid = trie->charmap[ uvc ];
1742 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1746 charid=(U16)SvIV( *svpp );
1749 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1756 if ( !trie->states[ state ].trans.list ) {
1757 TRIE_LIST_NEW( state );
1759 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1760 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1761 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1766 newstate = next_alloc++;
1767 prev_states[newstate] = state;
1768 TRIE_LIST_PUSH( state, charid, newstate );
1773 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1777 TRIE_HANDLE_WORD(state);
1779 } /* end second pass */
1781 /* next alloc is the NEXT state to be allocated */
1782 trie->statecount = next_alloc;
1783 trie->states = (reg_trie_state *)
1784 PerlMemShared_realloc( trie->states,
1786 * sizeof(reg_trie_state) );
1788 /* and now dump it out before we compress it */
1789 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1790 revcharmap, next_alloc,
1794 trie->trans = (reg_trie_trans *)
1795 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1802 for( state=1 ; state < next_alloc ; state ++ ) {
1806 DEBUG_TRIE_COMPILE_MORE_r(
1807 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1811 if (trie->states[state].trans.list) {
1812 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1816 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1817 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1818 if ( forid < minid ) {
1820 } else if ( forid > maxid ) {
1824 if ( transcount < tp + maxid - minid + 1) {
1826 trie->trans = (reg_trie_trans *)
1827 PerlMemShared_realloc( trie->trans,
1829 * sizeof(reg_trie_trans) );
1830 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1832 base = trie->uniquecharcount + tp - minid;
1833 if ( maxid == minid ) {
1835 for ( ; zp < tp ; zp++ ) {
1836 if ( ! trie->trans[ zp ].next ) {
1837 base = trie->uniquecharcount + zp - minid;
1838 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1839 trie->trans[ zp ].check = state;
1845 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1846 trie->trans[ tp ].check = state;
1851 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1852 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1853 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1854 trie->trans[ tid ].check = state;
1856 tp += ( maxid - minid + 1 );
1858 Safefree(trie->states[ state ].trans.list);
1861 DEBUG_TRIE_COMPILE_MORE_r(
1862 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1865 trie->states[ state ].trans.base=base;
1867 trie->lasttrans = tp + 1;
1871 Second Pass -- Flat Table Representation.
1873 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1874 We know that we will need Charcount+1 trans at most to store the data
1875 (one row per char at worst case) So we preallocate both structures
1876 assuming worst case.
1878 We then construct the trie using only the .next slots of the entry
1881 We use the .check field of the first entry of the node temporarily to
1882 make compression both faster and easier by keeping track of how many non
1883 zero fields are in the node.
1885 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1888 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1889 number representing the first entry of the node, and state as a
1890 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1891 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1892 are 2 entrys per node. eg:
1900 The table is internally in the right hand, idx form. However as we also
1901 have to deal with the states array which is indexed by nodenum we have to
1902 use TRIE_NODENUM() to convert.
1905 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1906 "%*sCompiling trie using table compiler\n",
1907 (int)depth * 2 + 2, ""));
1909 trie->trans = (reg_trie_trans *)
1910 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1911 * trie->uniquecharcount + 1,
1912 sizeof(reg_trie_trans) );
1913 trie->states = (reg_trie_state *)
1914 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1915 sizeof(reg_trie_state) );
1916 next_alloc = trie->uniquecharcount + 1;
1919 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1921 regnode * const noper = NEXTOPER( cur );
1922 const U8 *uc = (U8*)STRING( noper );
1923 const U8 * const e = uc + STR_LEN( noper );
1925 U32 state = 1; /* required init */
1927 U16 charid = 0; /* sanity init */
1928 U32 accept_state = 0; /* sanity init */
1929 U8 *scan = (U8*)NULL; /* sanity init */
1931 STRLEN foldlen = 0; /* required init */
1932 U32 wordlen = 0; /* required init */
1933 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1935 if ( OP(noper) != NOTHING ) {
1936 for ( ; uc < e ; uc += len ) {
1941 charid = trie->charmap[ uvc ];
1943 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1944 charid = svpp ? (U16)SvIV(*svpp) : 0;
1948 if ( !trie->trans[ state + charid ].next ) {
1949 trie->trans[ state + charid ].next = next_alloc;
1950 trie->trans[ state ].check++;
1951 prev_states[TRIE_NODENUM(next_alloc)]
1952 = TRIE_NODENUM(state);
1953 next_alloc += trie->uniquecharcount;
1955 state = trie->trans[ state + charid ].next;
1957 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1959 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1962 accept_state = TRIE_NODENUM( state );
1963 TRIE_HANDLE_WORD(accept_state);
1965 } /* end second pass */
1967 /* and now dump it out before we compress it */
1968 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1970 next_alloc, depth+1));
1974 * Inplace compress the table.*
1976 For sparse data sets the table constructed by the trie algorithm will
1977 be mostly 0/FAIL transitions or to put it another way mostly empty.
1978 (Note that leaf nodes will not contain any transitions.)
1980 This algorithm compresses the tables by eliminating most such
1981 transitions, at the cost of a modest bit of extra work during lookup:
1983 - Each states[] entry contains a .base field which indicates the
1984 index in the state[] array wheres its transition data is stored.
1986 - If .base is 0 there are no valid transitions from that node.
1988 - If .base is nonzero then charid is added to it to find an entry in
1991 -If trans[states[state].base+charid].check!=state then the
1992 transition is taken to be a 0/Fail transition. Thus if there are fail
1993 transitions at the front of the node then the .base offset will point
1994 somewhere inside the previous nodes data (or maybe even into a node
1995 even earlier), but the .check field determines if the transition is
1999 The following process inplace converts the table to the compressed
2000 table: We first do not compress the root node 1,and mark all its
2001 .check pointers as 1 and set its .base pointer as 1 as well. This
2002 allows us to do a DFA construction from the compressed table later,
2003 and ensures that any .base pointers we calculate later are greater
2006 - We set 'pos' to indicate the first entry of the second node.
2008 - We then iterate over the columns of the node, finding the first and
2009 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2010 and set the .check pointers accordingly, and advance pos
2011 appropriately and repreat for the next node. Note that when we copy
2012 the next pointers we have to convert them from the original
2013 NODEIDX form to NODENUM form as the former is not valid post
2016 - If a node has no transitions used we mark its base as 0 and do not
2017 advance the pos pointer.
2019 - If a node only has one transition we use a second pointer into the
2020 structure to fill in allocated fail transitions from other states.
2021 This pointer is independent of the main pointer and scans forward
2022 looking for null transitions that are allocated to a state. When it
2023 finds one it writes the single transition into the "hole". If the
2024 pointer doesnt find one the single transition is appended as normal.
2026 - Once compressed we can Renew/realloc the structures to release the
2029 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2030 specifically Fig 3.47 and the associated pseudocode.
2034 const U32 laststate = TRIE_NODENUM( next_alloc );
2037 trie->statecount = laststate;
2039 for ( state = 1 ; state < laststate ; state++ ) {
2041 const U32 stateidx = TRIE_NODEIDX( state );
2042 const U32 o_used = trie->trans[ stateidx ].check;
2043 U32 used = trie->trans[ stateidx ].check;
2044 trie->trans[ stateidx ].check = 0;
2046 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2047 if ( flag || trie->trans[ stateidx + charid ].next ) {
2048 if ( trie->trans[ stateidx + charid ].next ) {
2050 for ( ; zp < pos ; zp++ ) {
2051 if ( ! trie->trans[ zp ].next ) {
2055 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2056 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2057 trie->trans[ zp ].check = state;
2058 if ( ++zp > pos ) pos = zp;
2065 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2067 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2068 trie->trans[ pos ].check = state;
2073 trie->lasttrans = pos + 1;
2074 trie->states = (reg_trie_state *)
2075 PerlMemShared_realloc( trie->states, laststate
2076 * sizeof(reg_trie_state) );
2077 DEBUG_TRIE_COMPILE_MORE_r(
2078 PerlIO_printf( Perl_debug_log,
2079 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2080 (int)depth * 2 + 2,"",
2081 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2084 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2087 } /* end table compress */
2089 DEBUG_TRIE_COMPILE_MORE_r(
2090 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2091 (int)depth * 2 + 2, "",
2092 (UV)trie->statecount,
2093 (UV)trie->lasttrans)
2095 /* resize the trans array to remove unused space */
2096 trie->trans = (reg_trie_trans *)
2097 PerlMemShared_realloc( trie->trans, trie->lasttrans
2098 * sizeof(reg_trie_trans) );
2100 { /* Modify the program and insert the new TRIE node */
2101 U8 nodetype =(U8)(flags & 0xFF);
2105 regnode *optimize = NULL;
2106 #ifdef RE_TRACK_PATTERN_OFFSETS
2109 U32 mjd_nodelen = 0;
2110 #endif /* RE_TRACK_PATTERN_OFFSETS */
2111 #endif /* DEBUGGING */
2113 This means we convert either the first branch or the first Exact,
2114 depending on whether the thing following (in 'last') is a branch
2115 or not and whther first is the startbranch (ie is it a sub part of
2116 the alternation or is it the whole thing.)
2117 Assuming its a sub part we convert the EXACT otherwise we convert
2118 the whole branch sequence, including the first.
2120 /* Find the node we are going to overwrite */
2121 if ( first != startbranch || OP( last ) == BRANCH ) {
2122 /* branch sub-chain */
2123 NEXT_OFF( first ) = (U16)(last - first);
2124 #ifdef RE_TRACK_PATTERN_OFFSETS
2126 mjd_offset= Node_Offset((convert));
2127 mjd_nodelen= Node_Length((convert));
2130 /* whole branch chain */
2132 #ifdef RE_TRACK_PATTERN_OFFSETS
2135 const regnode *nop = NEXTOPER( convert );
2136 mjd_offset= Node_Offset((nop));
2137 mjd_nodelen= Node_Length((nop));
2141 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2142 (int)depth * 2 + 2, "",
2143 (UV)mjd_offset, (UV)mjd_nodelen)
2146 /* But first we check to see if there is a common prefix we can
2147 split out as an EXACT and put in front of the TRIE node. */
2148 trie->startstate= 1;
2149 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2151 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2155 const U32 base = trie->states[ state ].trans.base;
2157 if ( trie->states[state].wordnum )
2160 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2161 if ( ( base + ofs >= trie->uniquecharcount ) &&
2162 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2163 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2165 if ( ++count > 1 ) {
2166 SV **tmp = av_fetch( revcharmap, ofs, 0);
2167 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2168 if ( state == 1 ) break;
2170 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2172 PerlIO_printf(Perl_debug_log,
2173 "%*sNew Start State=%"UVuf" Class: [",
2174 (int)depth * 2 + 2, "",
2177 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2178 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2180 TRIE_BITMAP_SET(trie,*ch);
2182 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2184 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2188 TRIE_BITMAP_SET(trie,*ch);
2190 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2191 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2197 SV **tmp = av_fetch( revcharmap, idx, 0);
2199 char *ch = SvPV( *tmp, len );
2201 SV *sv=sv_newmortal();
2202 PerlIO_printf( Perl_debug_log,
2203 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2204 (int)depth * 2 + 2, "",
2206 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2207 PL_colors[0], PL_colors[1],
2208 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2209 PERL_PV_ESCAPE_FIRSTCHAR
2214 OP( convert ) = nodetype;
2215 str=STRING(convert);
2218 STR_LEN(convert) += len;
2224 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2229 trie->prefixlen = (state-1);
2231 regnode *n = convert+NODE_SZ_STR(convert);
2232 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2233 trie->startstate = state;
2234 trie->minlen -= (state - 1);
2235 trie->maxlen -= (state - 1);
2237 /* At least the UNICOS C compiler choked on this
2238 * being argument to DEBUG_r(), so let's just have
2241 #ifdef PERL_EXT_RE_BUILD
2247 regnode *fix = convert;
2248 U32 word = trie->wordcount;
2250 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2251 while( ++fix < n ) {
2252 Set_Node_Offset_Length(fix, 0, 0);
2255 SV ** const tmp = av_fetch( trie_words, word, 0 );
2257 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2258 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2260 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2268 NEXT_OFF(convert) = (U16)(tail - convert);
2269 DEBUG_r(optimize= n);
2275 if ( trie->maxlen ) {
2276 NEXT_OFF( convert ) = (U16)(tail - convert);
2277 ARG_SET( convert, data_slot );
2278 /* Store the offset to the first unabsorbed branch in
2279 jump[0], which is otherwise unused by the jump logic.
2280 We use this when dumping a trie and during optimisation. */
2282 trie->jump[0] = (U16)(nextbranch - convert);
2284 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2285 * and there is a bitmap
2286 * and the first "jump target" node we found leaves enough room
2287 * then convert the TRIE node into a TRIEC node, with the bitmap
2288 * embedded inline in the opcode - this is hypothetically faster.
2290 if ( !trie->states[trie->startstate].wordnum
2292 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2294 OP( convert ) = TRIEC;
2295 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2296 PerlMemShared_free(trie->bitmap);
2299 OP( convert ) = TRIE;
2301 /* store the type in the flags */
2302 convert->flags = nodetype;
2306 + regarglen[ OP( convert ) ];
2308 /* XXX We really should free up the resource in trie now,
2309 as we won't use them - (which resources?) dmq */
2311 /* needed for dumping*/
2312 DEBUG_r(if (optimize) {
2313 regnode *opt = convert;
2315 while ( ++opt < optimize) {
2316 Set_Node_Offset_Length(opt,0,0);
2319 Try to clean up some of the debris left after the
2322 while( optimize < jumper ) {
2323 mjd_nodelen += Node_Length((optimize));
2324 OP( optimize ) = OPTIMIZED;
2325 Set_Node_Offset_Length(optimize,0,0);
2328 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2330 } /* end node insert */
2332 /* Finish populating the prev field of the wordinfo array. Walk back
2333 * from each accept state until we find another accept state, and if
2334 * so, point the first word's .prev field at the second word. If the
2335 * second already has a .prev field set, stop now. This will be the
2336 * case either if we've already processed that word's accept state,
2337 * or that state had multiple words, and the overspill words were
2338 * already linked up earlier.
2345 for (word=1; word <= trie->wordcount; word++) {
2347 if (trie->wordinfo[word].prev)
2349 state = trie->wordinfo[word].accept;
2351 state = prev_states[state];
2354 prev = trie->states[state].wordnum;
2358 trie->wordinfo[word].prev = prev;
2360 Safefree(prev_states);
2364 /* and now dump out the compressed format */
2365 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2367 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2369 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2370 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2372 SvREFCNT_dec(revcharmap);
2376 : trie->startstate>1
2382 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2384 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2386 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2387 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2390 We find the fail state for each state in the trie, this state is the longest proper
2391 suffix of the current state's 'word' that is also a proper prefix of another word in our
2392 trie. State 1 represents the word '' and is thus the default fail state. This allows
2393 the DFA not to have to restart after its tried and failed a word at a given point, it
2394 simply continues as though it had been matching the other word in the first place.
2396 'abcdgu'=~/abcdefg|cdgu/
2397 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2398 fail, which would bring us to the state representing 'd' in the second word where we would
2399 try 'g' and succeed, proceeding to match 'cdgu'.
2401 /* add a fail transition */
2402 const U32 trie_offset = ARG(source);
2403 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2405 const U32 ucharcount = trie->uniquecharcount;
2406 const U32 numstates = trie->statecount;
2407 const U32 ubound = trie->lasttrans + ucharcount;
2411 U32 base = trie->states[ 1 ].trans.base;
2414 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2415 GET_RE_DEBUG_FLAGS_DECL;
2417 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2419 PERL_UNUSED_ARG(depth);
2423 ARG_SET( stclass, data_slot );
2424 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2425 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2426 aho->trie=trie_offset;
2427 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2428 Copy( trie->states, aho->states, numstates, reg_trie_state );
2429 Newxz( q, numstates, U32);
2430 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2433 /* initialize fail[0..1] to be 1 so that we always have
2434 a valid final fail state */
2435 fail[ 0 ] = fail[ 1 ] = 1;
2437 for ( charid = 0; charid < ucharcount ; charid++ ) {
2438 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2440 q[ q_write ] = newstate;
2441 /* set to point at the root */
2442 fail[ q[ q_write++ ] ]=1;
2445 while ( q_read < q_write) {
2446 const U32 cur = q[ q_read++ % numstates ];
2447 base = trie->states[ cur ].trans.base;
2449 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2450 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2452 U32 fail_state = cur;
2455 fail_state = fail[ fail_state ];
2456 fail_base = aho->states[ fail_state ].trans.base;
2457 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2459 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2460 fail[ ch_state ] = fail_state;
2461 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2463 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2465 q[ q_write++ % numstates] = ch_state;
2469 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2470 when we fail in state 1, this allows us to use the
2471 charclass scan to find a valid start char. This is based on the principle
2472 that theres a good chance the string being searched contains lots of stuff
2473 that cant be a start char.
2475 fail[ 0 ] = fail[ 1 ] = 0;
2476 DEBUG_TRIE_COMPILE_r({
2477 PerlIO_printf(Perl_debug_log,
2478 "%*sStclass Failtable (%"UVuf" states): 0",
2479 (int)(depth * 2), "", (UV)numstates
2481 for( q_read=1; q_read<numstates; q_read++ ) {
2482 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2484 PerlIO_printf(Perl_debug_log, "\n");
2487 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2492 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2493 * These need to be revisited when a newer toolchain becomes available.
2495 #if defined(__sparc64__) && defined(__GNUC__)
2496 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2497 # undef SPARC64_GCC_WORKAROUND
2498 # define SPARC64_GCC_WORKAROUND 1
2502 #define DEBUG_PEEP(str,scan,depth) \
2503 DEBUG_OPTIMISE_r({if (scan){ \
2504 SV * const mysv=sv_newmortal(); \
2505 regnode *Next = regnext(scan); \
2506 regprop(RExC_rx, mysv, scan); \
2507 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2508 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2509 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2513 /* The below joins as many adjacent EXACTish nodes as possible into a single
2514 * one, and looks for problematic sequences of characters whose folds vs.
2515 * non-folds have sufficiently different lengths, that the optimizer would be
2516 * fooled into rejecting legitimate matches of them, and the trie construction
2517 * code can't cope with them. The joining is only done if:
2518 * 1) there is room in the current conglomerated node to entirely contain the
2520 * 2) they are the exact same node type
2522 * The adjacent nodes actually may be separated by NOTHING kind nodes, and
2523 * these get optimized out
2525 * If there are problematic code sequences, *min_subtract is set to the delta
2526 * that the minimum size of the node can be less than its actual size. And,
2527 * the node type of the result is changed to reflect that it contains these
2530 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2531 * and contains LATIN SMALL LETTER SHARP S
2533 * This is as good a place as any to discuss the design of handling these
2534 * problematic sequences. It's been wrong in Perl for a very long time. There
2535 * are three code points in Unicode whose folded lengths differ so much from
2536 * the un-folded lengths that it causes problems for the optimizer and trie
2537 * construction. Why only these are problematic, and not others where lengths
2538 * also differ is something I (khw) do not understand. New versions of Unicode
2539 * might add more such code points. Hopefully the logic in fold_grind.t that
2540 * figures out what to test (in part by verifying that each size-combination
2541 * gets tested) will catch any that do come along, so they can be added to the
2542 * special handling below. The chances of new ones are actually rather small,
2543 * as most, if not all, of the world's scripts that have casefolding have
2544 * already been encoded by Unicode. Also, a number of Unicode's decisions were
2545 * made to allow compatibility with pre-existing standards, and almost all of
2546 * those have already been dealt with. These would otherwise be the most
2547 * likely candidates for generating further tricky sequences. In other words,
2548 * Unicode by itself is unlikely to add new ones unless it is for compatibility
2549 * with pre-existing standards, and there aren't many of those left.
2551 * The previous designs for dealing with these involved assigning a special
2552 * node for them. This approach doesn't work, as evidenced by this example:
2553 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2554 * Both these fold to "sss", but if the pattern is parsed to create a node of
2555 * that would match just the \xDF, it won't be able to handle the case where a
2556 * successful match would have to cross the node's boundary. The new approach
2557 * that hopefully generally solves the problem generates an EXACTFU_SS node
2560 * There are a number of components to the approach (a lot of work for just
2561 * three code points!):
2562 * 1) This routine examines each EXACTFish node that could contain the
2563 * problematic sequences. It returns in *min_subtract how much to
2564 * subtract from the the actual length of the string to get a real minimum
2565 * for one that could match it. This number is usually 0 except for the
2566 * problematic sequences. This delta is used by the caller to adjust the
2567 * min length of the match, and the delta between min and max, so that the
2568 * optimizer doesn't reject these possibilities based on size constraints.
2569 * 2) These sequences are not currently correctly handled by the trie code
2570 * either, so it changes the joined node type to ops that are not handled
2571 * by trie's, those new ops being EXACTFU_SS and EXACTFU_NO_TRIE.
2572 * 3) This is sufficient for the two Greek sequences (described below), but
2573 * the one involving the Sharp s (\xDF) needs more. The node type
2574 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2575 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2576 * case where there is a possible fold length change. That means that a
2577 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2578 * itself with length changes, and so can be processed faster. regexec.c
2579 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2580 * is pre-folded by regcomp.c. This saves effort in regex matching.
2581 * However, probably mostly for historical reasons, the pre-folding isn't
2582 * done for non-UTF8 patterns (and it can't be for EXACTF and EXACTFL
2583 * nodes, as what they fold to isn't known until runtime.) The fold
2584 * possibilities for the non-UTF8 patterns are quite simple, except for
2585 * the sharp s. All the ones that don't involve a UTF-8 target string
2586 * are members of a fold-pair, and arrays are set up for all of them
2587 * that quickly find the other member of the pair. It might actually
2588 * be faster to pre-fold these, but it isn't currently done, except for
2589 * the sharp s. Code elsewhere in this file makes sure that it gets
2590 * folded to 'ss', even if the pattern isn't UTF-8. This avoids the
2591 * issues described in the next item.
2592 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2593 * 'ss' or not is not knowable at compile time. It will match iff the
2594 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2595 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2596 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2597 * described in item 3). An assumption that the optimizer part of
2598 * regexec.c (probably unwittingly) makes is that a character in the
2599 * pattern corresponds to at most a single character in the target string.
2600 * (And I do mean character, and not byte here, unlike other parts of the
2601 * documentation that have never been updated to account for multibyte
2602 * Unicode.) This assumption is wrong only in this case, as all other
2603 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2604 * virtue of having this file pre-fold UTF-8 patterns. I'm
2605 * reluctant to try to change this assumption, so instead the code punts.
2606 * This routine examines EXACTF nodes for the sharp s, and returns a
2607 * boolean indicating whether or not the node is an EXACTF node that
2608 * contains a sharp s. When it is true, the caller sets a flag that later
2609 * causes the optimizer in this file to not set values for the floating
2610 * and fixed string lengths, and thus avoids the optimizer code in
2611 * regexec.c that makes the invalid assumption. Thus, there is no
2612 * optimization based on string lengths for EXACTF nodes that contain the
2613 * sharp s. This only happens for /id rules (which means the pattern
2617 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2618 if (PL_regkind[OP(scan)] == EXACT) \
2619 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2622 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) {
2623 /* Merge several consecutive EXACTish nodes into one. */
2624 regnode *n = regnext(scan);
2626 regnode *next = scan + NODE_SZ_STR(scan);
2630 regnode *stop = scan;
2631 GET_RE_DEBUG_FLAGS_DECL;
2633 PERL_UNUSED_ARG(depth);
2636 PERL_ARGS_ASSERT_JOIN_EXACT;
2637 #ifndef EXPERIMENTAL_INPLACESCAN
2638 PERL_UNUSED_ARG(flags);
2639 PERL_UNUSED_ARG(val);
2641 DEBUG_PEEP("join",scan,depth);
2643 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2644 * EXACT ones that are mergeable to the current one. */
2646 && (PL_regkind[OP(n)] == NOTHING
2647 || (stringok && OP(n) == OP(scan)))
2649 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2652 if (OP(n) == TAIL || n > next)
2654 if (PL_regkind[OP(n)] == NOTHING) {
2655 DEBUG_PEEP("skip:",n,depth);
2656 NEXT_OFF(scan) += NEXT_OFF(n);
2657 next = n + NODE_STEP_REGNODE;
2664 else if (stringok) {
2665 const unsigned int oldl = STR_LEN(scan);
2666 regnode * const nnext = regnext(n);
2668 if (oldl + STR_LEN(n) > U8_MAX)
2671 DEBUG_PEEP("merg",n,depth);
2674 NEXT_OFF(scan) += NEXT_OFF(n);
2675 STR_LEN(scan) += STR_LEN(n);
2676 next = n + NODE_SZ_STR(n);
2677 /* Now we can overwrite *n : */
2678 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2686 #ifdef EXPERIMENTAL_INPLACESCAN
2687 if (flags && !NEXT_OFF(n)) {
2688 DEBUG_PEEP("atch", val, depth);
2689 if (reg_off_by_arg[OP(n)]) {
2690 ARG_SET(n, val - n);
2693 NEXT_OFF(n) = val - n;
2701 *has_exactf_sharp_s = FALSE;
2703 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2704 * can now analyze for sequences of problematic code points. (Prior to
2705 * this final joining, sequences could have been split over boundaries, and
2706 * hence missed). The sequences only happen in folding, hence for any
2707 * non-EXACT EXACTish node */
2708 if (OP(scan) != EXACT) {
2710 U8 * s0 = (U8*) STRING(scan);
2711 U8 * const s_end = s0 + STR_LEN(scan);
2713 /* The below is perhaps overboard, but this allows us to save a test
2714 * each time through the loop at the expense of a mask. This is
2715 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2716 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2717 * This uses an exclusive 'or' to find that bit and then inverts it to
2718 * form a mask, with just a single 0, in the bit position where 'S' and
2720 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2721 const U8 s_masked = 's' & S_or_s_mask;
2723 /* One pass is made over the node's string looking for all the
2724 * possibilities. to avoid some tests in the loop, there are two main
2725 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2729 /* There are two problematic Greek code points in Unicode
2732 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2733 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2739 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2740 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2742 * This means that in case-insensitive matching (or "loose
2743 * matching", as Unicode calls it), an EXACTF of length six (the
2744 * UTF-8 encoded byte length of the above casefolded versions) can
2745 * match a target string of length two (the byte length of UTF-8
2746 * encoded U+0390 or U+03B0). This would rather mess up the
2747 * minimum length computation. (there are other code points that
2748 * also fold to these two sequences, but the delta is smaller)
2750 * If these sequences are found, the minimum length is decreased by
2751 * four (six minus two).
2753 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2754 * LETTER SHARP S. We decrease the min length by 1 for each
2755 * occurrence of 'ss' found */
2757 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2758 # define U390_first_byte 0xb4
2759 const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
2760 # define U3B0_first_byte 0xb5
2761 const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
2763 # define U390_first_byte 0xce
2764 const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
2765 # define U3B0_first_byte 0xcf
2766 const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
2768 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2769 yields a net of 0 */
2770 /* Examine the string for one of the problematic sequences */
2772 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2773 * sequence we are looking for is 2 */
2777 /* Look for the first byte in each problematic sequence */
2779 /* We don't have to worry about other things that fold to
2780 * 's' (such as the long s, U+017F), as all above-latin1
2781 * code points have been pre-folded */
2785 /* Current character is an 's' or 'S'. If next one is
2786 * as well, we have the dreaded sequence */
2787 if (((*(s+1) & S_or_s_mask) == s_masked)
2788 /* These two node types don't have special handling
2790 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2793 OP(scan) = EXACTFU_SS;
2794 s++; /* No need to look at this character again */
2798 case U390_first_byte:
2799 if (s_end - s >= len
2801 /* The 1's are because are skipping comparing the
2803 && memEQ(s + 1, U390_tail, len - 1))
2805 goto greek_sequence;
2809 case U3B0_first_byte:
2810 if (! (s_end - s >= len
2811 && memEQ(s + 1, U3B0_tail, len - 1)))
2818 /* This can't currently be handled by trie's, so change
2819 * the node type to indicate this. If EXACTFA and
2820 * EXACTFL were ever to be handled by trie's, this
2821 * would have to be changed. If this node has already
2822 * been changed to EXACTFU_SS in this loop, leave it as
2823 * is. (I (khw) think it doesn't matter in regexec.c
2824 * for UTF patterns, but no need to change it */
2825 if (OP(scan) == EXACTFU) {
2826 OP(scan) = EXACTFU_NO_TRIE;
2828 s += 6; /* We already know what this sequence is. Skip
2834 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2836 /* Here, the pattern is not UTF-8. We need to look only for the
2837 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2838 * in the final position. Otherwise we can stop looking 1 byte
2839 * earlier because have to find both the first and second 's' */
2840 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2842 for (s = s0; s < upper; s++) {
2847 && ((*(s+1) & S_or_s_mask) == s_masked))
2851 /* EXACTF nodes need to know that the minimum
2852 * length changed so that a sharp s in the string
2853 * can match this ss in the pattern, but they
2854 * remain EXACTF nodes, as they are not trie'able,
2855 * so don't have to invent a new node type to
2856 * exclude them from the trie code */
2857 if (OP(scan) != EXACTF) {
2858 OP(scan) = EXACTFU_SS;
2863 case LATIN_SMALL_LETTER_SHARP_S:
2864 if (OP(scan) == EXACTF) {
2865 *has_exactf_sharp_s = TRUE;
2874 /* Allow dumping but overwriting the collection of skipped
2875 * ops and/or strings with fake optimized ops */
2876 n = scan + NODE_SZ_STR(scan);
2884 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2888 /* REx optimizer. Converts nodes into quicker variants "in place".
2889 Finds fixed substrings. */
2891 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2892 to the position after last scanned or to NULL. */
2894 #define INIT_AND_WITHP \
2895 assert(!and_withp); \
2896 Newx(and_withp,1,struct regnode_charclass_class); \
2897 SAVEFREEPV(and_withp)
2899 /* this is a chain of data about sub patterns we are processing that
2900 need to be handled separately/specially in study_chunk. Its so
2901 we can simulate recursion without losing state. */
2903 typedef struct scan_frame {
2904 regnode *last; /* last node to process in this frame */
2905 regnode *next; /* next node to process when last is reached */
2906 struct scan_frame *prev; /*previous frame*/
2907 I32 stop; /* what stopparen do we use */
2911 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2913 #define CASE_SYNST_FNC(nAmE) \
2915 if (flags & SCF_DO_STCLASS_AND) { \
2916 for (value = 0; value < 256; value++) \
2917 if (!is_ ## nAmE ## _cp(value)) \
2918 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2921 for (value = 0; value < 256; value++) \
2922 if (is_ ## nAmE ## _cp(value)) \
2923 ANYOF_BITMAP_SET(data->start_class, value); \
2927 if (flags & SCF_DO_STCLASS_AND) { \
2928 for (value = 0; value < 256; value++) \
2929 if (is_ ## nAmE ## _cp(value)) \
2930 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2933 for (value = 0; value < 256; value++) \
2934 if (!is_ ## nAmE ## _cp(value)) \
2935 ANYOF_BITMAP_SET(data->start_class, value); \
2942 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2943 I32 *minlenp, I32 *deltap,
2948 struct regnode_charclass_class *and_withp,
2949 U32 flags, U32 depth)
2950 /* scanp: Start here (read-write). */
2951 /* deltap: Write maxlen-minlen here. */
2952 /* last: Stop before this one. */
2953 /* data: string data about the pattern */
2954 /* stopparen: treat close N as END */
2955 /* recursed: which subroutines have we recursed into */
2956 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
2959 I32 min = 0, pars = 0, code;
2960 regnode *scan = *scanp, *next;
2962 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2963 int is_inf_internal = 0; /* The studied chunk is infinite */
2964 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2965 scan_data_t data_fake;
2966 SV *re_trie_maxbuff = NULL;
2967 regnode *first_non_open = scan;
2968 I32 stopmin = I32_MAX;
2969 scan_frame *frame = NULL;
2970 GET_RE_DEBUG_FLAGS_DECL;
2972 PERL_ARGS_ASSERT_STUDY_CHUNK;
2975 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
2979 while (first_non_open && OP(first_non_open) == OPEN)
2980 first_non_open=regnext(first_non_open);
2985 while ( scan && OP(scan) != END && scan < last ){
2986 UV min_subtract = 0; /* How much to subtract from the minimum node
2987 length to get a real minimum (because the
2988 folded version may be shorter) */
2989 bool has_exactf_sharp_s = FALSE;
2990 /* Peephole optimizer: */
2991 DEBUG_STUDYDATA("Peep:", data,depth);
2992 DEBUG_PEEP("Peep",scan,depth);
2994 /* Its not clear to khw or hv why this is done here, and not in the
2995 * clauses that deal with EXACT nodes. khw's guess is that it's
2996 * because of a previous design */
2997 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
2999 /* Follow the next-chain of the current node and optimize
3000 away all the NOTHINGs from it. */
3001 if (OP(scan) != CURLYX) {
3002 const int max = (reg_off_by_arg[OP(scan)]
3004 /* I32 may be smaller than U16 on CRAYs! */
3005 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3006 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3010 /* Skip NOTHING and LONGJMP. */
3011 while ((n = regnext(n))
3012 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3013 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3014 && off + noff < max)
3016 if (reg_off_by_arg[OP(scan)])
3019 NEXT_OFF(scan) = off;
3024 /* The principal pseudo-switch. Cannot be a switch, since we
3025 look into several different things. */
3026 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3027 || OP(scan) == IFTHEN) {
3028 next = regnext(scan);
3030 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3032 if (OP(next) == code || code == IFTHEN) {
3033 /* NOTE - There is similar code to this block below for handling
3034 TRIE nodes on a re-study. If you change stuff here check there
3036 I32 max1 = 0, min1 = I32_MAX, num = 0;
3037 struct regnode_charclass_class accum;
3038 regnode * const startbranch=scan;
3040 if (flags & SCF_DO_SUBSTR)
3041 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3042 if (flags & SCF_DO_STCLASS)
3043 cl_init_zero(pRExC_state, &accum);
3045 while (OP(scan) == code) {
3046 I32 deltanext, minnext, f = 0, fake;
3047 struct regnode_charclass_class this_class;
3050 data_fake.flags = 0;
3052 data_fake.whilem_c = data->whilem_c;
3053 data_fake.last_closep = data->last_closep;
3056 data_fake.last_closep = &fake;
3058 data_fake.pos_delta = delta;
3059 next = regnext(scan);
3060 scan = NEXTOPER(scan);
3062 scan = NEXTOPER(scan);
3063 if (flags & SCF_DO_STCLASS) {
3064 cl_init(pRExC_state, &this_class);
3065 data_fake.start_class = &this_class;
3066 f = SCF_DO_STCLASS_AND;
3068 if (flags & SCF_WHILEM_VISITED_POS)
3069 f |= SCF_WHILEM_VISITED_POS;
3071 /* we suppose the run is continuous, last=next...*/
3072 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3074 stopparen, recursed, NULL, f,depth+1);
3077 if (max1 < minnext + deltanext)
3078 max1 = minnext + deltanext;
3079 if (deltanext == I32_MAX)
3080 is_inf = is_inf_internal = 1;
3082 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3084 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3085 if ( stopmin > minnext)
3086 stopmin = min + min1;
3087 flags &= ~SCF_DO_SUBSTR;
3089 data->flags |= SCF_SEEN_ACCEPT;
3092 if (data_fake.flags & SF_HAS_EVAL)
3093 data->flags |= SF_HAS_EVAL;
3094 data->whilem_c = data_fake.whilem_c;
3096 if (flags & SCF_DO_STCLASS)
3097 cl_or(pRExC_state, &accum, &this_class);
3099 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3101 if (flags & SCF_DO_SUBSTR) {
3102 data->pos_min += min1;
3103 data->pos_delta += max1 - min1;
3104 if (max1 != min1 || is_inf)
3105 data->longest = &(data->longest_float);
3108 delta += max1 - min1;
3109 if (flags & SCF_DO_STCLASS_OR) {
3110 cl_or(pRExC_state, data->start_class, &accum);
3112 cl_and(data->start_class, and_withp);
3113 flags &= ~SCF_DO_STCLASS;
3116 else if (flags & SCF_DO_STCLASS_AND) {
3118 cl_and(data->start_class, &accum);
3119 flags &= ~SCF_DO_STCLASS;
3122 /* Switch to OR mode: cache the old value of
3123 * data->start_class */
3125 StructCopy(data->start_class, and_withp,
3126 struct regnode_charclass_class);
3127 flags &= ~SCF_DO_STCLASS_AND;
3128 StructCopy(&accum, data->start_class,
3129 struct regnode_charclass_class);
3130 flags |= SCF_DO_STCLASS_OR;
3131 data->start_class->flags |= ANYOF_EOS;
3135 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3138 Assuming this was/is a branch we are dealing with: 'scan' now
3139 points at the item that follows the branch sequence, whatever
3140 it is. We now start at the beginning of the sequence and look
3147 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3149 If we can find such a subsequence we need to turn the first
3150 element into a trie and then add the subsequent branch exact
3151 strings to the trie.
3155 1. patterns where the whole set of branches can be converted.
3157 2. patterns where only a subset can be converted.
3159 In case 1 we can replace the whole set with a single regop
3160 for the trie. In case 2 we need to keep the start and end
3163 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3164 becomes BRANCH TRIE; BRANCH X;
3166 There is an additional case, that being where there is a
3167 common prefix, which gets split out into an EXACT like node
3168 preceding the TRIE node.
3170 If x(1..n)==tail then we can do a simple trie, if not we make
3171 a "jump" trie, such that when we match the appropriate word
3172 we "jump" to the appropriate tail node. Essentially we turn
3173 a nested if into a case structure of sorts.
3178 if (!re_trie_maxbuff) {
3179 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3180 if (!SvIOK(re_trie_maxbuff))
3181 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3183 if ( SvIV(re_trie_maxbuff)>=0 ) {
3185 regnode *first = (regnode *)NULL;
3186 regnode *last = (regnode *)NULL;
3187 regnode *tail = scan;
3192 SV * const mysv = sv_newmortal(); /* for dumping */
3194 /* var tail is used because there may be a TAIL
3195 regop in the way. Ie, the exacts will point to the
3196 thing following the TAIL, but the last branch will
3197 point at the TAIL. So we advance tail. If we
3198 have nested (?:) we may have to move through several
3202 while ( OP( tail ) == TAIL ) {
3203 /* this is the TAIL generated by (?:) */
3204 tail = regnext( tail );
3209 regprop(RExC_rx, mysv, tail );
3210 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3211 (int)depth * 2 + 2, "",
3212 "Looking for TRIE'able sequences. Tail node is: ",
3213 SvPV_nolen_const( mysv )
3219 step through the branches, cur represents each
3220 branch, noper is the first thing to be matched
3221 as part of that branch and noper_next is the
3222 regnext() of that node. if noper is an EXACT
3223 and noper_next is the same as scan (our current
3224 position in the regex) then the EXACT branch is
3225 a possible optimization target. Once we have
3226 two or more consecutive such branches we can
3227 create a trie of the EXACT's contents and stich
3228 it in place. If the sequence represents all of
3229 the branches we eliminate the whole thing and
3230 replace it with a single TRIE. If it is a
3231 subsequence then we need to stitch it in. This
3232 means the first branch has to remain, and needs
3233 to be repointed at the item on the branch chain
3234 following the last branch optimized. This could
3235 be either a BRANCH, in which case the
3236 subsequence is internal, or it could be the
3237 item following the branch sequence in which
3238 case the subsequence is at the end.
3242 /* dont use tail as the end marker for this traverse */
3243 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3244 regnode * const noper = NEXTOPER( cur );
3245 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3246 regnode * const noper_next = regnext( noper );
3250 regprop(RExC_rx, mysv, cur);
3251 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3252 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3254 regprop(RExC_rx, mysv, noper);
3255 PerlIO_printf( Perl_debug_log, " -> %s",
3256 SvPV_nolen_const(mysv));
3259 regprop(RExC_rx, mysv, noper_next );
3260 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3261 SvPV_nolen_const(mysv));
3263 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
3264 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
3266 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
3267 : PL_regkind[ OP( noper ) ] == EXACT )
3268 || OP(noper) == NOTHING )
3270 && noper_next == tail
3275 if ( !first || optype == NOTHING ) {
3276 if (!first) first = cur;
3277 optype = OP( noper );
3283 Currently the trie logic handles case insensitive matching properly only
3284 when the pattern is UTF-8 and the node is EXACTFU (thus forcing unicode
3287 If/when this is fixed the following define can be swapped
3288 in below to fully enable trie logic.
3290 #define TRIE_TYPE_IS_SAFE 1
3292 Note that join_exact() assumes that the other types of EXACTFish nodes are not
3293 used in tries, so that would have to be updated if this changed
3296 #define TRIE_TYPE_IS_SAFE ((UTF && optype == EXACTFU) || optype==EXACT)
3298 if ( last && TRIE_TYPE_IS_SAFE ) {
3299 make_trie( pRExC_state,
3300 startbranch, first, cur, tail, count,
3303 if ( PL_regkind[ OP( noper ) ] == EXACT
3305 && noper_next == tail
3310 optype = OP( noper );
3320 regprop(RExC_rx, mysv, cur);
3321 PerlIO_printf( Perl_debug_log,
3322 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3323 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3327 if ( last && TRIE_TYPE_IS_SAFE ) {
3328 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
3329 #ifdef TRIE_STUDY_OPT
3330 if ( ((made == MADE_EXACT_TRIE &&
3331 startbranch == first)
3332 || ( first_non_open == first )) &&
3334 flags |= SCF_TRIE_RESTUDY;
3335 if ( startbranch == first
3338 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3348 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3349 scan = NEXTOPER(NEXTOPER(scan));
3350 } else /* single branch is optimized. */
3351 scan = NEXTOPER(scan);
3353 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3354 scan_frame *newframe = NULL;
3359 if (OP(scan) != SUSPEND) {
3360 /* set the pointer */
3361 if (OP(scan) == GOSUB) {
3363 RExC_recurse[ARG2L(scan)] = scan;
3364 start = RExC_open_parens[paren-1];
3365 end = RExC_close_parens[paren-1];
3368 start = RExC_rxi->program + 1;
3372 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3373 SAVEFREEPV(recursed);
3375 if (!PAREN_TEST(recursed,paren+1)) {
3376 PAREN_SET(recursed,paren+1);
3377 Newx(newframe,1,scan_frame);
3379 if (flags & SCF_DO_SUBSTR) {
3380 SCAN_COMMIT(pRExC_state,data,minlenp);
3381 data->longest = &(data->longest_float);
3383 is_inf = is_inf_internal = 1;
3384 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3385 cl_anything(pRExC_state, data->start_class);
3386 flags &= ~SCF_DO_STCLASS;
3389 Newx(newframe,1,scan_frame);
3392 end = regnext(scan);
3397 SAVEFREEPV(newframe);
3398 newframe->next = regnext(scan);
3399 newframe->last = last;
3400 newframe->stop = stopparen;
3401 newframe->prev = frame;
3411 else if (OP(scan) == EXACT) {
3412 I32 l = STR_LEN(scan);
3415 const U8 * const s = (U8*)STRING(scan);
3416 l = utf8_length(s, s + l);
3417 uc = utf8_to_uvchr(s, NULL);
3419 uc = *((U8*)STRING(scan));
3422 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3423 /* The code below prefers earlier match for fixed
3424 offset, later match for variable offset. */
3425 if (data->last_end == -1) { /* Update the start info. */
3426 data->last_start_min = data->pos_min;
3427 data->last_start_max = is_inf
3428 ? I32_MAX : data->pos_min + data->pos_delta;
3430 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3432 SvUTF8_on(data->last_found);
3434 SV * const sv = data->last_found;
3435 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3436 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3437 if (mg && mg->mg_len >= 0)
3438 mg->mg_len += utf8_length((U8*)STRING(scan),
3439 (U8*)STRING(scan)+STR_LEN(scan));
3441 data->last_end = data->pos_min + l;
3442 data->pos_min += l; /* As in the first entry. */
3443 data->flags &= ~SF_BEFORE_EOL;
3445 if (flags & SCF_DO_STCLASS_AND) {
3446 /* Check whether it is compatible with what we know already! */
3450 /* If compatible, we or it in below. It is compatible if is
3451 * in the bitmp and either 1) its bit or its fold is set, or 2)
3452 * it's for a locale. Even if there isn't unicode semantics
3453 * here, at runtime there may be because of matching against a
3454 * utf8 string, so accept a possible false positive for
3455 * latin1-range folds */
3457 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3458 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3459 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
3460 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3465 ANYOF_CLASS_ZERO(data->start_class);
3466 ANYOF_BITMAP_ZERO(data->start_class);
3468 ANYOF_BITMAP_SET(data->start_class, uc);
3469 else if (uc >= 0x100) {
3472 /* Some Unicode code points fold to the Latin1 range; as
3473 * XXX temporary code, instead of figuring out if this is
3474 * one, just assume it is and set all the start class bits
3475 * that could be some such above 255 code point's fold
3476 * which will generate fals positives. As the code
3477 * elsewhere that does compute the fold settles down, it
3478 * can be extracted out and re-used here */
3479 for (i = 0; i < 256; i++){
3480 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3481 ANYOF_BITMAP_SET(data->start_class, i);
3485 data->start_class->flags &= ~ANYOF_EOS;
3487 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3489 else if (flags & SCF_DO_STCLASS_OR) {
3490 /* false positive possible if the class is case-folded */
3492 ANYOF_BITMAP_SET(data->start_class, uc);
3494 data->start_class->flags |= ANYOF_UNICODE_ALL;
3495 data->start_class->flags &= ~ANYOF_EOS;
3496 cl_and(data->start_class, and_withp);
3498 flags &= ~SCF_DO_STCLASS;
3500 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3501 I32 l = STR_LEN(scan);
3502 UV uc = *((U8*)STRING(scan));
3504 /* Search for fixed substrings supports EXACT only. */
3505 if (flags & SCF_DO_SUBSTR) {
3507 SCAN_COMMIT(pRExC_state, data, minlenp);
3510 const U8 * const s = (U8 *)STRING(scan);
3511 l = utf8_length(s, s + l);
3512 uc = utf8_to_uvchr(s, NULL);
3514 else if (has_exactf_sharp_s) {
3515 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3517 min += l - min_subtract;
3521 delta += min_subtract;
3522 if (flags & SCF_DO_SUBSTR) {
3523 data->pos_min += l - min_subtract;
3524 if (data->pos_min < 0) {
3527 data->pos_delta += min_subtract;
3529 data->longest = &(data->longest_float);
3532 if (flags & SCF_DO_STCLASS_AND) {
3533 /* Check whether it is compatible with what we know already! */
3536 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3537 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3538 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3542 ANYOF_CLASS_ZERO(data->start_class);
3543 ANYOF_BITMAP_ZERO(data->start_class);
3545 ANYOF_BITMAP_SET(data->start_class, uc);
3546 data->start_class->flags &= ~ANYOF_EOS;
3547 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
3548 if (OP(scan) == EXACTFL) {
3549 /* XXX This set is probably no longer necessary, and
3550 * probably wrong as LOCALE now is on in the initial
3552 data->start_class->flags |= ANYOF_LOCALE;
3556 /* Also set the other member of the fold pair. In case
3557 * that unicode semantics is called for at runtime, use
3558 * the full latin1 fold. (Can't do this for locale,
3559 * because not known until runtime) */
3560 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3562 /* All other (EXACTFL handled above) folds except under
3563 * /iaa that include s, S, and sharp_s also may include
3565 if (OP(scan) != EXACTFA) {
3566 if (uc == 's' || uc == 'S') {
3567 ANYOF_BITMAP_SET(data->start_class,
3568 LATIN_SMALL_LETTER_SHARP_S);
3570 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3571 ANYOF_BITMAP_SET(data->start_class, 's');
3572 ANYOF_BITMAP_SET(data->start_class, 'S');
3577 else if (uc >= 0x100) {
3579 for (i = 0; i < 256; i++){
3580 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3581 ANYOF_BITMAP_SET(data->start_class, i);
3586 else if (flags & SCF_DO_STCLASS_OR) {
3587 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
3588 /* false positive possible if the class is case-folded.
3589 Assume that the locale settings are the same... */
3591 ANYOF_BITMAP_SET(data->start_class, uc);
3592 if (OP(scan) != EXACTFL) {
3594 /* And set the other member of the fold pair, but
3595 * can't do that in locale because not known until
3597 ANYOF_BITMAP_SET(data->start_class,
3598 PL_fold_latin1[uc]);
3600 /* All folds except under /iaa that include s, S,
3601 * and sharp_s also may include the others */
3602 if (OP(scan) != EXACTFA) {
3603 if (uc == 's' || uc == 'S') {
3604 ANYOF_BITMAP_SET(data->start_class,
3605 LATIN_SMALL_LETTER_SHARP_S);
3607 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3608 ANYOF_BITMAP_SET(data->start_class, 's');
3609 ANYOF_BITMAP_SET(data->start_class, 'S');
3614 data->start_class->flags &= ~ANYOF_EOS;
3616 cl_and(data->start_class, and_withp);
3618 flags &= ~SCF_DO_STCLASS;
3620 else if (REGNODE_VARIES(OP(scan))) {
3621 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3622 I32 f = flags, pos_before = 0;
3623 regnode * const oscan = scan;
3624 struct regnode_charclass_class this_class;
3625 struct regnode_charclass_class *oclass = NULL;
3626 I32 next_is_eval = 0;
3628 switch (PL_regkind[OP(scan)]) {
3629 case WHILEM: /* End of (?:...)* . */
3630 scan = NEXTOPER(scan);
3633 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3634 next = NEXTOPER(scan);
3635 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3637 maxcount = REG_INFTY;
3638 next = regnext(scan);
3639 scan = NEXTOPER(scan);
3643 if (flags & SCF_DO_SUBSTR)
3648 if (flags & SCF_DO_STCLASS) {
3650 maxcount = REG_INFTY;
3651 next = regnext(scan);
3652 scan = NEXTOPER(scan);
3655 is_inf = is_inf_internal = 1;
3656 scan = regnext(scan);
3657 if (flags & SCF_DO_SUBSTR) {
3658 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3659 data->longest = &(data->longest_float);
3661 goto optimize_curly_tail;
3663 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3664 && (scan->flags == stopparen))
3669 mincount = ARG1(scan);
3670 maxcount = ARG2(scan);
3672 next = regnext(scan);
3673 if (OP(scan) == CURLYX) {
3674 I32 lp = (data ? *(data->last_closep) : 0);
3675 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3677 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3678 next_is_eval = (OP(scan) == EVAL);
3680 if (flags & SCF_DO_SUBSTR) {
3681 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3682 pos_before = data->pos_min;
3686 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3688 data->flags |= SF_IS_INF;
3690 if (flags & SCF_DO_STCLASS) {
3691 cl_init(pRExC_state, &this_class);
3692 oclass = data->start_class;
3693 data->start_class = &this_class;
3694 f |= SCF_DO_STCLASS_AND;
3695 f &= ~SCF_DO_STCLASS_OR;
3697 /* Exclude from super-linear cache processing any {n,m}
3698 regops for which the combination of input pos and regex
3699 pos is not enough information to determine if a match
3702 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3703 regex pos at the \s*, the prospects for a match depend not
3704 only on the input position but also on how many (bar\s*)
3705 repeats into the {4,8} we are. */
3706 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3707 f &= ~SCF_WHILEM_VISITED_POS;
3709 /* This will finish on WHILEM, setting scan, or on NULL: */
3710 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3711 last, data, stopparen, recursed, NULL,
3713 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3715 if (flags & SCF_DO_STCLASS)
3716 data->start_class = oclass;
3717 if (mincount == 0 || minnext == 0) {
3718 if (flags & SCF_DO_STCLASS_OR) {
3719 cl_or(pRExC_state, data->start_class, &this_class);
3721 else if (flags & SCF_DO_STCLASS_AND) {
3722 /* Switch to OR mode: cache the old value of
3723 * data->start_class */
3725 StructCopy(data->start_class, and_withp,
3726 struct regnode_charclass_class);
3727 flags &= ~SCF_DO_STCLASS_AND;
3728 StructCopy(&this_class, data->start_class,
3729 struct regnode_charclass_class);
3730 flags |= SCF_DO_STCLASS_OR;
3731 data->start_class->flags |= ANYOF_EOS;
3733 } else { /* Non-zero len */
3734 if (flags & SCF_DO_STCLASS_OR) {
3735 cl_or(pRExC_state, data->start_class, &this_class);
3736 cl_and(data->start_class, and_withp);
3738 else if (flags & SCF_DO_STCLASS_AND)
3739 cl_and(data->start_class, &this_class);
3740 flags &= ~SCF_DO_STCLASS;
3742 if (!scan) /* It was not CURLYX, but CURLY. */
3744 if ( /* ? quantifier ok, except for (?{ ... }) */
3745 (next_is_eval || !(mincount == 0 && maxcount == 1))
3746 && (minnext == 0) && (deltanext == 0)
3747 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3748 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3750 ckWARNreg(RExC_parse,
3751 "Quantifier unexpected on zero-length expression");
3754 min += minnext * mincount;
3755 is_inf_internal |= ((maxcount == REG_INFTY
3756 && (minnext + deltanext) > 0)
3757 || deltanext == I32_MAX);
3758 is_inf |= is_inf_internal;
3759 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3761 /* Try powerful optimization CURLYX => CURLYN. */
3762 if ( OP(oscan) == CURLYX && data
3763 && data->flags & SF_IN_PAR
3764 && !(data->flags & SF_HAS_EVAL)
3765 && !deltanext && minnext == 1 ) {
3766 /* Try to optimize to CURLYN. */
3767 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3768 regnode * const nxt1 = nxt;
3775 if (!REGNODE_SIMPLE(OP(nxt))
3776 && !(PL_regkind[OP(nxt)] == EXACT
3777 && STR_LEN(nxt) == 1))
3783 if (OP(nxt) != CLOSE)
3785 if (RExC_open_parens) {
3786 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3787 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3789 /* Now we know that nxt2 is the only contents: */
3790 oscan->flags = (U8)ARG(nxt);
3792 OP(nxt1) = NOTHING; /* was OPEN. */
3795 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3796 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3797 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3798 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3799 OP(nxt + 1) = OPTIMIZED; /* was count. */
3800 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3805 /* Try optimization CURLYX => CURLYM. */
3806 if ( OP(oscan) == CURLYX && data
3807 && !(data->flags & SF_HAS_PAR)
3808 && !(data->flags & SF_HAS_EVAL)
3809 && !deltanext /* atom is fixed width */
3810 && minnext != 0 /* CURLYM can't handle zero width */
3812 /* XXXX How to optimize if data == 0? */
3813 /* Optimize to a simpler form. */
3814 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3818 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3819 && (OP(nxt2) != WHILEM))
3821 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3822 /* Need to optimize away parenths. */
3823 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
3824 /* Set the parenth number. */
3825 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3827 oscan->flags = (U8)ARG(nxt);
3828 if (RExC_open_parens) {
3829 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3830 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3832 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3833 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3836 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3837 OP(nxt + 1) = OPTIMIZED; /* was count. */
3838 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
3839 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
3842 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3843 regnode *nnxt = regnext(nxt1);
3845 if (reg_off_by_arg[OP(nxt1)])
3846 ARG_SET(nxt1, nxt2 - nxt1);
3847 else if (nxt2 - nxt1 < U16_MAX)
3848 NEXT_OFF(nxt1) = nxt2 - nxt1;
3850 OP(nxt) = NOTHING; /* Cannot beautify */
3855 /* Optimize again: */
3856 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
3857 NULL, stopparen, recursed, NULL, 0,depth+1);
https://perl5.git.perl.org / perl5.git / blob