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 #include "charclass_invlists.h"
96 # if defined(BUGGY_MSC6)
97 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
98 # pragma optimize("a",off)
99 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
100 # pragma optimize("w",on )
101 # endif /* BUGGY_MSC6 */
105 #define STATIC static
108 typedef struct RExC_state_t {
109 U32 flags; /* are we folding, multilining? */
110 char *precomp; /* uncompiled string. */
111 REGEXP *rx_sv; /* The SV that is the regexp. */
112 regexp *rx; /* perl core regexp structure */
113 regexp_internal *rxi; /* internal data for regexp object pprivate field */
114 char *start; /* Start of input for compile */
115 char *end; /* End of input for compile */
116 char *parse; /* Input-scan pointer. */
117 I32 whilem_seen; /* number of WHILEM in this expr */
118 regnode *emit_start; /* Start of emitted-code area */
119 regnode *emit_bound; /* First regnode outside of the allocated space */
120 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
121 I32 naughty; /* How bad is this pattern? */
122 I32 sawback; /* Did we see \1, ...? */
124 I32 size; /* Code size. */
125 I32 npar; /* Capture buffer count, (OPEN). */
126 I32 cpar; /* Capture buffer count, (CLOSE). */
127 I32 nestroot; /* root parens we are in - used by accept */
131 regnode **open_parens; /* pointers to open parens */
132 regnode **close_parens; /* pointers to close parens */
133 regnode *opend; /* END node in program */
134 I32 utf8; /* whether the pattern is utf8 or not */
135 I32 orig_utf8; /* whether the pattern was originally in utf8 */
136 /* XXX use this for future optimisation of case
137 * where pattern must be upgraded to utf8. */
138 I32 uni_semantics; /* If a d charset modifier should use unicode
139 rules, even if the pattern is not in
141 HV *paren_names; /* Paren names */
143 regnode **recurse; /* Recurse regops */
144 I32 recurse_count; /* Number of recurse regops */
147 I32 override_recoding;
149 char *starttry; /* -Dr: where regtry was called. */
150 #define RExC_starttry (pRExC_state->starttry)
153 const char *lastparse;
155 AV *paren_name_list; /* idx -> name */
156 #define RExC_lastparse (pRExC_state->lastparse)
157 #define RExC_lastnum (pRExC_state->lastnum)
158 #define RExC_paren_name_list (pRExC_state->paren_name_list)
162 #define RExC_flags (pRExC_state->flags)
163 #define RExC_precomp (pRExC_state->precomp)
164 #define RExC_rx_sv (pRExC_state->rx_sv)
165 #define RExC_rx (pRExC_state->rx)
166 #define RExC_rxi (pRExC_state->rxi)
167 #define RExC_start (pRExC_state->start)
168 #define RExC_end (pRExC_state->end)
169 #define RExC_parse (pRExC_state->parse)
170 #define RExC_whilem_seen (pRExC_state->whilem_seen)
171 #ifdef RE_TRACK_PATTERN_OFFSETS
172 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
174 #define RExC_emit (pRExC_state->emit)
175 #define RExC_emit_start (pRExC_state->emit_start)
176 #define RExC_emit_bound (pRExC_state->emit_bound)
177 #define RExC_naughty (pRExC_state->naughty)
178 #define RExC_sawback (pRExC_state->sawback)
179 #define RExC_seen (pRExC_state->seen)
180 #define RExC_size (pRExC_state->size)
181 #define RExC_npar (pRExC_state->npar)
182 #define RExC_nestroot (pRExC_state->nestroot)
183 #define RExC_extralen (pRExC_state->extralen)
184 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
185 #define RExC_seen_evals (pRExC_state->seen_evals)
186 #define RExC_utf8 (pRExC_state->utf8)
187 #define RExC_uni_semantics (pRExC_state->uni_semantics)
188 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
189 #define RExC_open_parens (pRExC_state->open_parens)
190 #define RExC_close_parens (pRExC_state->close_parens)
191 #define RExC_opend (pRExC_state->opend)
192 #define RExC_paren_names (pRExC_state->paren_names)
193 #define RExC_recurse (pRExC_state->recurse)
194 #define RExC_recurse_count (pRExC_state->recurse_count)
195 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
196 #define RExC_contains_locale (pRExC_state->contains_locale)
197 #define RExC_override_recoding (pRExC_state->override_recoding)
200 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
201 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
202 ((*s) == '{' && regcurly(s)))
205 #undef SPSTART /* dratted cpp namespace... */
208 * Flags to be passed up and down.
210 #define WORST 0 /* Worst case. */
211 #define HASWIDTH 0x01 /* Known to match non-null strings. */
213 /* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
214 * character, and if utf8, must be invariant. Note that this is not the same thing as REGNODE_SIMPLE */
216 #define SPSTART 0x04 /* Starts with * or +. */
217 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
218 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
220 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
222 /* whether trie related optimizations are enabled */
223 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
224 #define TRIE_STUDY_OPT
225 #define FULL_TRIE_STUDY
231 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
232 #define PBITVAL(paren) (1 << ((paren) & 7))
233 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
234 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
235 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
237 /* If not already in utf8, do a longjmp back to the beginning */
238 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
239 #define REQUIRE_UTF8 STMT_START { \
240 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
243 /* About scan_data_t.
245 During optimisation we recurse through the regexp program performing
246 various inplace (keyhole style) optimisations. In addition study_chunk
247 and scan_commit populate this data structure with information about
248 what strings MUST appear in the pattern. We look for the longest
249 string that must appear at a fixed location, and we look for the
250 longest string that may appear at a floating location. So for instance
255 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
256 strings (because they follow a .* construct). study_chunk will identify
257 both FOO and BAR as being the longest fixed and floating strings respectively.
259 The strings can be composites, for instance
263 will result in a composite fixed substring 'foo'.
265 For each string some basic information is maintained:
267 - offset or min_offset
268 This is the position the string must appear at, or not before.
269 It also implicitly (when combined with minlenp) tells us how many
270 characters must match before the string we are searching for.
271 Likewise when combined with minlenp and the length of the string it
272 tells us how many characters must appear after the string we have
276 Only used for floating strings. This is the rightmost point that
277 the string can appear at. If set to I32 max it indicates that the
278 string can occur infinitely far to the right.
281 A pointer to the minimum length of the pattern that the string
282 was found inside. This is important as in the case of positive
283 lookahead or positive lookbehind we can have multiple patterns
288 The minimum length of the pattern overall is 3, the minimum length
289 of the lookahead part is 3, but the minimum length of the part that
290 will actually match is 1. So 'FOO's minimum length is 3, but the
291 minimum length for the F is 1. This is important as the minimum length
292 is used to determine offsets in front of and behind the string being
293 looked for. Since strings can be composites this is the length of the
294 pattern at the time it was committed with a scan_commit. Note that
295 the length is calculated by study_chunk, so that the minimum lengths
296 are not known until the full pattern has been compiled, thus the
297 pointer to the value.
301 In the case of lookbehind the string being searched for can be
302 offset past the start point of the final matching string.
303 If this value was just blithely removed from the min_offset it would
304 invalidate some of the calculations for how many chars must match
305 before or after (as they are derived from min_offset and minlen and
306 the length of the string being searched for).
307 When the final pattern is compiled and the data is moved from the
308 scan_data_t structure into the regexp structure the information
309 about lookbehind is factored in, with the information that would
310 have been lost precalculated in the end_shift field for the
313 The fields pos_min and pos_delta are used to store the minimum offset
314 and the delta to the maximum offset at the current point in the pattern.
318 typedef struct scan_data_t {
319 /*I32 len_min; unused */
320 /*I32 len_delta; unused */
324 I32 last_end; /* min value, <0 unless valid. */
327 SV **longest; /* Either &l_fixed, or &l_float. */
328 SV *longest_fixed; /* longest fixed string found in pattern */
329 I32 offset_fixed; /* offset where it starts */
330 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
331 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
332 SV *longest_float; /* longest floating string found in pattern */
333 I32 offset_float_min; /* earliest point in string it can appear */
334 I32 offset_float_max; /* latest point in string it can appear */
335 I32 *minlen_float; /* pointer to the minlen relevant to the string */
336 I32 lookbehind_float; /* is the position of the string modified by LB */
340 struct regnode_charclass_class *start_class;
344 * Forward declarations for pregcomp()'s friends.
347 static const scan_data_t zero_scan_data =
348 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
350 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
351 #define SF_BEFORE_SEOL 0x0001
352 #define SF_BEFORE_MEOL 0x0002
353 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
354 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
357 # define SF_FIX_SHIFT_EOL (0+2)
358 # define SF_FL_SHIFT_EOL (0+4)
360 # define SF_FIX_SHIFT_EOL (+2)
361 # define SF_FL_SHIFT_EOL (+4)
364 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
365 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
367 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
368 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
369 #define SF_IS_INF 0x0040
370 #define SF_HAS_PAR 0x0080
371 #define SF_IN_PAR 0x0100
372 #define SF_HAS_EVAL 0x0200
373 #define SCF_DO_SUBSTR 0x0400
374 #define SCF_DO_STCLASS_AND 0x0800
375 #define SCF_DO_STCLASS_OR 0x1000
376 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
377 #define SCF_WHILEM_VISITED_POS 0x2000
379 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
380 #define SCF_SEEN_ACCEPT 0x8000
382 #define UTF cBOOL(RExC_utf8)
383 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
384 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
385 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
386 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
387 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
388 #define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
389 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
391 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
393 #define OOB_UNICODE 12345678
394 #define OOB_NAMEDCLASS -1
396 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
397 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
400 /* length of regex to show in messages that don't mark a position within */
401 #define RegexLengthToShowInErrorMessages 127
404 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
405 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
406 * op/pragma/warn/regcomp.
408 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
409 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
411 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
414 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
415 * arg. Show regex, up to a maximum length. If it's too long, chop and add
418 #define _FAIL(code) STMT_START { \
419 const char *ellipses = ""; \
420 IV len = RExC_end - RExC_precomp; \
423 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
424 if (len > RegexLengthToShowInErrorMessages) { \
425 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
426 len = RegexLengthToShowInErrorMessages - 10; \
432 #define FAIL(msg) _FAIL( \
433 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
434 msg, (int)len, RExC_precomp, ellipses))
436 #define FAIL2(msg,arg) _FAIL( \
437 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
438 arg, (int)len, RExC_precomp, ellipses))
441 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
443 #define Simple_vFAIL(m) STMT_START { \
444 const IV offset = RExC_parse - RExC_precomp; \
445 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
446 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
450 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
452 #define vFAIL(m) STMT_START { \
454 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
459 * Like Simple_vFAIL(), but accepts two arguments.
461 #define Simple_vFAIL2(m,a1) STMT_START { \
462 const IV offset = RExC_parse - RExC_precomp; \
463 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
464 (int)offset, RExC_precomp, RExC_precomp + offset); \
468 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
470 #define vFAIL2(m,a1) STMT_START { \
472 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
473 Simple_vFAIL2(m, a1); \
478 * Like Simple_vFAIL(), but accepts three arguments.
480 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
481 const IV offset = RExC_parse - RExC_precomp; \
482 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
483 (int)offset, RExC_precomp, RExC_precomp + offset); \
487 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
489 #define vFAIL3(m,a1,a2) STMT_START { \
491 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
492 Simple_vFAIL3(m, a1, a2); \
496 * Like Simple_vFAIL(), but accepts four arguments.
498 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
499 const IV offset = RExC_parse - RExC_precomp; \
500 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
501 (int)offset, RExC_precomp, RExC_precomp + offset); \
504 #define ckWARNreg(loc,m) STMT_START { \
505 const IV offset = loc - RExC_precomp; \
506 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
507 (int)offset, RExC_precomp, RExC_precomp + offset); \
510 #define ckWARNregdep(loc,m) STMT_START { \
511 const IV offset = loc - RExC_precomp; \
512 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
514 (int)offset, RExC_precomp, RExC_precomp + offset); \
517 #define ckWARN2regdep(loc,m, a1) STMT_START { \
518 const IV offset = loc - RExC_precomp; \
519 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
521 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
524 #define ckWARN2reg(loc, m, a1) STMT_START { \
525 const IV offset = loc - RExC_precomp; \
526 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
527 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
530 #define vWARN3(loc, m, a1, a2) STMT_START { \
531 const IV offset = loc - RExC_precomp; \
532 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
533 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
536 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
537 const IV offset = loc - RExC_precomp; \
538 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
539 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
542 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
543 const IV offset = loc - RExC_precomp; \
544 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
545 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
548 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
549 const IV offset = loc - RExC_precomp; \
550 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
551 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
554 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
555 const IV offset = loc - RExC_precomp; \
556 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
557 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
561 /* Allow for side effects in s */
562 #define REGC(c,s) STMT_START { \
563 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
566 /* Macros for recording node offsets. 20001227 mjd@plover.com
567 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
568 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
569 * Element 0 holds the number n.
570 * Position is 1 indexed.
572 #ifndef RE_TRACK_PATTERN_OFFSETS
573 #define Set_Node_Offset_To_R(node,byte)
574 #define Set_Node_Offset(node,byte)
575 #define Set_Cur_Node_Offset
576 #define Set_Node_Length_To_R(node,len)
577 #define Set_Node_Length(node,len)
578 #define Set_Node_Cur_Length(node)
579 #define Node_Offset(n)
580 #define Node_Length(n)
581 #define Set_Node_Offset_Length(node,offset,len)
582 #define ProgLen(ri) ri->u.proglen
583 #define SetProgLen(ri,x) ri->u.proglen = x
585 #define ProgLen(ri) ri->u.offsets[0]
586 #define SetProgLen(ri,x) ri->u.offsets[0] = x
587 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
589 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
590 __LINE__, (int)(node), (int)(byte))); \
592 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
594 RExC_offsets[2*(node)-1] = (byte); \
599 #define Set_Node_Offset(node,byte) \
600 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
601 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
603 #define Set_Node_Length_To_R(node,len) STMT_START { \
605 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
606 __LINE__, (int)(node), (int)(len))); \
608 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
610 RExC_offsets[2*(node)] = (len); \
615 #define Set_Node_Length(node,len) \
616 Set_Node_Length_To_R((node)-RExC_emit_start, len)
617 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
618 #define Set_Node_Cur_Length(node) \
619 Set_Node_Length(node, RExC_parse - parse_start)
621 /* Get offsets and lengths */
622 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
623 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
625 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
626 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
627 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
631 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
632 #define EXPERIMENTAL_INPLACESCAN
633 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
635 #define DEBUG_STUDYDATA(str,data,depth) \
636 DEBUG_OPTIMISE_MORE_r(if(data){ \
637 PerlIO_printf(Perl_debug_log, \
638 "%*s" str "Pos:%"IVdf"/%"IVdf \
639 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
640 (int)(depth)*2, "", \
641 (IV)((data)->pos_min), \
642 (IV)((data)->pos_delta), \
643 (UV)((data)->flags), \
644 (IV)((data)->whilem_c), \
645 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
646 is_inf ? "INF " : "" \
648 if ((data)->last_found) \
649 PerlIO_printf(Perl_debug_log, \
650 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
651 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
652 SvPVX_const((data)->last_found), \
653 (IV)((data)->last_end), \
654 (IV)((data)->last_start_min), \
655 (IV)((data)->last_start_max), \
656 ((data)->longest && \
657 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
658 SvPVX_const((data)->longest_fixed), \
659 (IV)((data)->offset_fixed), \
660 ((data)->longest && \
661 (data)->longest==&((data)->longest_float)) ? "*" : "", \
662 SvPVX_const((data)->longest_float), \
663 (IV)((data)->offset_float_min), \
664 (IV)((data)->offset_float_max) \
666 PerlIO_printf(Perl_debug_log,"\n"); \
669 static void clear_re(pTHX_ void *r);
671 /* Mark that we cannot extend a found fixed substring at this point.
672 Update the longest found anchored substring and the longest found
673 floating substrings if needed. */
676 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
678 const STRLEN l = CHR_SVLEN(data->last_found);
679 const STRLEN old_l = CHR_SVLEN(*data->longest);
680 GET_RE_DEBUG_FLAGS_DECL;
682 PERL_ARGS_ASSERT_SCAN_COMMIT;
684 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
685 SvSetMagicSV(*data->longest, data->last_found);
686 if (*data->longest == data->longest_fixed) {
687 data->offset_fixed = l ? data->last_start_min : data->pos_min;
688 if (data->flags & SF_BEFORE_EOL)
690 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
692 data->flags &= ~SF_FIX_BEFORE_EOL;
693 data->minlen_fixed=minlenp;
694 data->lookbehind_fixed=0;
696 else { /* *data->longest == data->longest_float */
697 data->offset_float_min = l ? data->last_start_min : data->pos_min;
698 data->offset_float_max = (l
699 ? data->last_start_max
700 : data->pos_min + data->pos_delta);
701 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
702 data->offset_float_max = I32_MAX;
703 if (data->flags & SF_BEFORE_EOL)
705 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
707 data->flags &= ~SF_FL_BEFORE_EOL;
708 data->minlen_float=minlenp;
709 data->lookbehind_float=0;
712 SvCUR_set(data->last_found, 0);
714 SV * const sv = data->last_found;
715 if (SvUTF8(sv) && SvMAGICAL(sv)) {
716 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
722 data->flags &= ~SF_BEFORE_EOL;
723 DEBUG_STUDYDATA("commit: ",data,0);
726 /* Can match anything (initialization) */
728 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
730 PERL_ARGS_ASSERT_CL_ANYTHING;
732 ANYOF_BITMAP_SETALL(cl);
733 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
734 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
736 /* If any portion of the regex is to operate under locale rules,
737 * initialization includes it. The reason this isn't done for all regexes
738 * is that the optimizer was written under the assumption that locale was
739 * all-or-nothing. Given the complexity and lack of documentation in the
740 * optimizer, and that there are inadequate test cases for locale, so many
741 * parts of it may not work properly, it is safest to avoid locale unless
743 if (RExC_contains_locale) {
744 ANYOF_CLASS_SETALL(cl); /* /l uses class */
745 cl->flags |= ANYOF_LOCALE;
748 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
752 /* Can match anything (initialization) */
754 S_cl_is_anything(const struct regnode_charclass_class *cl)
758 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
760 for (value = 0; value <= ANYOF_MAX; value += 2)
761 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
763 if (!(cl->flags & ANYOF_UNICODE_ALL))
765 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
770 /* Can match anything (initialization) */
772 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
774 PERL_ARGS_ASSERT_CL_INIT;
776 Zero(cl, 1, struct regnode_charclass_class);
778 cl_anything(pRExC_state, cl);
779 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
782 /* These two functions currently do the exact same thing */
783 #define cl_init_zero S_cl_init
785 /* 'AND' a given class with another one. Can create false positives. 'cl'
786 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
787 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
789 S_cl_and(struct regnode_charclass_class *cl,
790 const struct regnode_charclass_class *and_with)
792 PERL_ARGS_ASSERT_CL_AND;
794 assert(and_with->type == ANYOF);
796 /* I (khw) am not sure all these restrictions are necessary XXX */
797 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
798 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
799 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
800 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
801 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
804 if (and_with->flags & ANYOF_INVERT)
805 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
806 cl->bitmap[i] &= ~and_with->bitmap[i];
808 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
809 cl->bitmap[i] &= and_with->bitmap[i];
810 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
812 if (and_with->flags & ANYOF_INVERT) {
814 /* Here, the and'ed node is inverted. Get the AND of the flags that
815 * aren't affected by the inversion. Those that are affected are
816 * handled individually below */
817 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
818 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
819 cl->flags |= affected_flags;
821 /* We currently don't know how to deal with things that aren't in the
822 * bitmap, but we know that the intersection is no greater than what
823 * is already in cl, so let there be false positives that get sorted
824 * out after the synthetic start class succeeds, and the node is
825 * matched for real. */
827 /* The inversion of these two flags indicate that the resulting
828 * intersection doesn't have them */
829 if (and_with->flags & ANYOF_UNICODE_ALL) {
830 cl->flags &= ~ANYOF_UNICODE_ALL;
832 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
833 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
836 else { /* and'd node is not inverted */
837 U8 outside_bitmap_but_not_utf8; /* Temp variable */
839 if (! ANYOF_NONBITMAP(and_with)) {
841 /* Here 'and_with' doesn't match anything outside the bitmap
842 * (except possibly ANYOF_UNICODE_ALL), which means the
843 * intersection can't either, except for ANYOF_UNICODE_ALL, in
844 * which case we don't know what the intersection is, but it's no
845 * greater than what cl already has, so can just leave it alone,
846 * with possible false positives */
847 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
848 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
849 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
852 else if (! ANYOF_NONBITMAP(cl)) {
854 /* Here, 'and_with' does match something outside the bitmap, and cl
855 * doesn't have a list of things to match outside the bitmap. If
856 * cl can match all code points above 255, the intersection will
857 * be those above-255 code points that 'and_with' matches. If cl
858 * can't match all Unicode code points, it means that it can't
859 * match anything outside the bitmap (since the 'if' that got us
860 * into this block tested for that), so we leave the bitmap empty.
862 if (cl->flags & ANYOF_UNICODE_ALL) {
863 ARG_SET(cl, ARG(and_with));
865 /* and_with's ARG may match things that don't require UTF8.
866 * And now cl's will too, in spite of this being an 'and'. See
867 * the comments below about the kludge */
868 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
872 /* Here, both 'and_with' and cl match something outside the
873 * bitmap. Currently we do not do the intersection, so just match
874 * whatever cl had at the beginning. */
878 /* Take the intersection of the two sets of flags. However, the
879 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
880 * kludge around the fact that this flag is not treated like the others
881 * which are initialized in cl_anything(). The way the optimizer works
882 * is that the synthetic start class (SSC) is initialized to match
883 * anything, and then the first time a real node is encountered, its
884 * values are AND'd with the SSC's with the result being the values of
885 * the real node. However, there are paths through the optimizer where
886 * the AND never gets called, so those initialized bits are set
887 * inappropriately, which is not usually a big deal, as they just cause
888 * false positives in the SSC, which will just mean a probably
889 * imperceptible slow down in execution. However this bit has a
890 * higher false positive consequence in that it can cause utf8.pm,
891 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
892 * bigger slowdown and also causes significant extra memory to be used.
893 * In order to prevent this, the code now takes a different tack. The
894 * bit isn't set unless some part of the regular expression needs it,
895 * but once set it won't get cleared. This means that these extra
896 * modules won't get loaded unless there was some path through the
897 * pattern that would have required them anyway, and so any false
898 * positives that occur by not ANDing them out when they could be
899 * aren't as severe as they would be if we treated this bit like all
901 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
902 & ANYOF_NONBITMAP_NON_UTF8;
903 cl->flags &= and_with->flags;
904 cl->flags |= outside_bitmap_but_not_utf8;
908 /* 'OR' a given class with another one. Can create false positives. 'cl'
909 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
910 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
912 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
914 PERL_ARGS_ASSERT_CL_OR;
916 if (or_with->flags & ANYOF_INVERT) {
918 /* Here, the or'd node is to be inverted. This means we take the
919 * complement of everything not in the bitmap, but currently we don't
920 * know what that is, so give up and match anything */
921 if (ANYOF_NONBITMAP(or_with)) {
922 cl_anything(pRExC_state, cl);
925 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
926 * <= (B1 | !B2) | (CL1 | !CL2)
927 * which is wasteful if CL2 is small, but we ignore CL2:
928 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
929 * XXXX Can we handle case-fold? Unclear:
930 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
931 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
933 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
934 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
935 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
938 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
939 cl->bitmap[i] |= ~or_with->bitmap[i];
940 } /* XXXX: logic is complicated otherwise */
942 cl_anything(pRExC_state, cl);
945 /* And, we can just take the union of the flags that aren't affected
946 * by the inversion */
947 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
949 /* For the remaining flags:
950 ANYOF_UNICODE_ALL and inverted means to not match anything above
951 255, which means that the union with cl should just be
952 what cl has in it, so can ignore this flag
953 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
954 is 127-255 to match them, but then invert that, so the
955 union with cl should just be what cl has in it, so can
958 } else { /* 'or_with' is not inverted */
959 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
960 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
961 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
962 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
965 /* OR char bitmap and class bitmap separately */
966 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
967 cl->bitmap[i] |= or_with->bitmap[i];
968 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
969 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
970 cl->classflags[i] |= or_with->classflags[i];
971 cl->flags |= ANYOF_CLASS;
974 else { /* XXXX: logic is complicated, leave it along for a moment. */
975 cl_anything(pRExC_state, cl);
978 if (ANYOF_NONBITMAP(or_with)) {
980 /* Use the added node's outside-the-bit-map match if there isn't a
981 * conflict. If there is a conflict (both nodes match something
982 * outside the bitmap, but what they match outside is not the same
983 * pointer, and hence not easily compared until XXX we extend
984 * inversion lists this far), give up and allow the start class to
985 * match everything outside the bitmap. If that stuff is all above
986 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
987 if (! ANYOF_NONBITMAP(cl)) {
988 ARG_SET(cl, ARG(or_with));
990 else if (ARG(cl) != ARG(or_with)) {
992 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
993 cl_anything(pRExC_state, cl);
996 cl->flags |= ANYOF_UNICODE_ALL;
1001 /* Take the union */
1002 cl->flags |= or_with->flags;
1006 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1007 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1008 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1009 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1014 dump_trie(trie,widecharmap,revcharmap)
1015 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1016 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1018 These routines dump out a trie in a somewhat readable format.
1019 The _interim_ variants are used for debugging the interim
1020 tables that are used to generate the final compressed
1021 representation which is what dump_trie expects.
1023 Part of the reason for their existence is to provide a form
1024 of documentation as to how the different representations function.
1029 Dumps the final compressed table form of the trie to Perl_debug_log.
1030 Used for debugging make_trie().
1034 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1035 AV *revcharmap, U32 depth)
1038 SV *sv=sv_newmortal();
1039 int colwidth= widecharmap ? 6 : 4;
1041 GET_RE_DEBUG_FLAGS_DECL;
1043 PERL_ARGS_ASSERT_DUMP_TRIE;
1045 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1046 (int)depth * 2 + 2,"",
1047 "Match","Base","Ofs" );
1049 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1050 SV ** const tmp = av_fetch( revcharmap, state, 0);
1052 PerlIO_printf( Perl_debug_log, "%*s",
1054 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1055 PL_colors[0], PL_colors[1],
1056 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1057 PERL_PV_ESCAPE_FIRSTCHAR
1062 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1063 (int)depth * 2 + 2,"");
1065 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1066 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1067 PerlIO_printf( Perl_debug_log, "\n");
1069 for( state = 1 ; state < trie->statecount ; state++ ) {
1070 const U32 base = trie->states[ state ].trans.base;
1072 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1074 if ( trie->states[ state ].wordnum ) {
1075 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1077 PerlIO_printf( Perl_debug_log, "%6s", "" );
1080 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1085 while( ( base + ofs < trie->uniquecharcount ) ||
1086 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1087 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1090 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1092 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1093 if ( ( base + ofs >= trie->uniquecharcount ) &&
1094 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1095 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1097 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1099 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1101 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1105 PerlIO_printf( Perl_debug_log, "]");
1108 PerlIO_printf( Perl_debug_log, "\n" );
1110 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1111 for (word=1; word <= trie->wordcount; word++) {
1112 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1113 (int)word, (int)(trie->wordinfo[word].prev),
1114 (int)(trie->wordinfo[word].len));
1116 PerlIO_printf(Perl_debug_log, "\n" );
1119 Dumps a fully constructed but uncompressed trie in list form.
1120 List tries normally only are used for construction when the number of
1121 possible chars (trie->uniquecharcount) is very high.
1122 Used for debugging make_trie().
1125 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1126 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1130 SV *sv=sv_newmortal();
1131 int colwidth= widecharmap ? 6 : 4;
1132 GET_RE_DEBUG_FLAGS_DECL;
1134 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1136 /* print out the table precompression. */
1137 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1138 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1139 "------:-----+-----------------\n" );
1141 for( state=1 ; state < next_alloc ; state ++ ) {
1144 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1145 (int)depth * 2 + 2,"", (UV)state );
1146 if ( ! trie->states[ state ].wordnum ) {
1147 PerlIO_printf( Perl_debug_log, "%5s| ","");
1149 PerlIO_printf( Perl_debug_log, "W%4x| ",
1150 trie->states[ state ].wordnum
1153 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1154 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1156 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1158 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1159 PL_colors[0], PL_colors[1],
1160 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1161 PERL_PV_ESCAPE_FIRSTCHAR
1163 TRIE_LIST_ITEM(state,charid).forid,
1164 (UV)TRIE_LIST_ITEM(state,charid).newstate
1167 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1168 (int)((depth * 2) + 14), "");
1171 PerlIO_printf( Perl_debug_log, "\n");
1176 Dumps a fully constructed but uncompressed trie in table form.
1177 This is the normal DFA style state transition table, with a few
1178 twists to facilitate compression later.
1179 Used for debugging make_trie().
1182 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1183 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1188 SV *sv=sv_newmortal();
1189 int colwidth= widecharmap ? 6 : 4;
1190 GET_RE_DEBUG_FLAGS_DECL;
1192 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1195 print out the table precompression so that we can do a visual check
1196 that they are identical.
1199 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1201 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1202 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1204 PerlIO_printf( Perl_debug_log, "%*s",
1206 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1207 PL_colors[0], PL_colors[1],
1208 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1209 PERL_PV_ESCAPE_FIRSTCHAR
1215 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1217 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1218 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1221 PerlIO_printf( Perl_debug_log, "\n" );
1223 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1225 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1226 (int)depth * 2 + 2,"",
1227 (UV)TRIE_NODENUM( state ) );
1229 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1230 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1232 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1234 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1236 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1237 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1239 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1240 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1248 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1249 startbranch: the first branch in the whole branch sequence
1250 first : start branch of sequence of branch-exact nodes.
1251 May be the same as startbranch
1252 last : Thing following the last branch.
1253 May be the same as tail.
1254 tail : item following the branch sequence
1255 count : words in the sequence
1256 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1257 depth : indent depth
1259 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1261 A trie is an N'ary tree where the branches are determined by digital
1262 decomposition of the key. IE, at the root node you look up the 1st character and
1263 follow that branch repeat until you find the end of the branches. Nodes can be
1264 marked as "accepting" meaning they represent a complete word. Eg:
1268 would convert into the following structure. Numbers represent states, letters
1269 following numbers represent valid transitions on the letter from that state, if
1270 the number is in square brackets it represents an accepting state, otherwise it
1271 will be in parenthesis.
1273 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1277 (1) +-i->(6)-+-s->[7]
1279 +-s->(3)-+-h->(4)-+-e->[5]
1281 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1283 This shows that when matching against the string 'hers' we will begin at state 1
1284 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1285 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1286 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1287 single traverse. We store a mapping from accepting to state to which word was
1288 matched, and then when we have multiple possibilities we try to complete the
1289 rest of the regex in the order in which they occured in the alternation.
1291 The only prior NFA like behaviour that would be changed by the TRIE support is
1292 the silent ignoring of duplicate alternations which are of the form:
1294 / (DUPE|DUPE) X? (?{ ... }) Y /x
1296 Thus EVAL blocks following a trie may be called a different number of times with
1297 and without the optimisation. With the optimisations dupes will be silently
1298 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1299 the following demonstrates:
1301 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1303 which prints out 'word' three times, but
1305 'words'=~/(word|word|word)(?{ print $1 })S/
1307 which doesnt print it out at all. This is due to other optimisations kicking in.
1309 Example of what happens on a structural level:
1311 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1313 1: CURLYM[1] {1,32767}(18)
1324 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1325 and should turn into:
1327 1: CURLYM[1] {1,32767}(18)
1329 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1337 Cases where tail != last would be like /(?foo|bar)baz/:
1347 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1348 and would end up looking like:
1351 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1358 d = uvuni_to_utf8_flags(d, uv, 0);
1360 is the recommended Unicode-aware way of saying
1365 #define TRIE_STORE_REVCHAR \
1368 SV *zlopp = newSV(2); \
1369 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1370 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, uvc & 0xFF); \
1371 SvCUR_set(zlopp, kapow - flrbbbbb); \
1374 av_push(revcharmap, zlopp); \
1376 char ooooff = (char)uvc; \
1377 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1381 #define TRIE_READ_CHAR STMT_START { \
1385 if ( foldlen > 0 ) { \
1386 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1391 len = UTF8SKIP(uc);\
1392 uvc = to_utf8_fold( uc, foldbuf, &foldlen); \
1393 foldlen -= UNISKIP( uvc ); \
1394 scan = foldbuf + UNISKIP( uvc ); \
1397 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1407 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1408 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1409 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1410 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1412 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1413 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1414 TRIE_LIST_CUR( state )++; \
1417 #define TRIE_LIST_NEW(state) STMT_START { \
1418 Newxz( trie->states[ state ].trans.list, \
1419 4, reg_trie_trans_le ); \
1420 TRIE_LIST_CUR( state ) = 1; \
1421 TRIE_LIST_LEN( state ) = 4; \
1424 #define TRIE_HANDLE_WORD(state) STMT_START { \
1425 U16 dupe= trie->states[ state ].wordnum; \
1426 regnode * const noper_next = regnext( noper ); \
1429 /* store the word for dumping */ \
1431 if (OP(noper) != NOTHING) \
1432 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1434 tmp = newSVpvn_utf8( "", 0, UTF ); \
1435 av_push( trie_words, tmp ); \
1439 trie->wordinfo[curword].prev = 0; \
1440 trie->wordinfo[curword].len = wordlen; \
1441 trie->wordinfo[curword].accept = state; \
1443 if ( noper_next < tail ) { \
1445 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1446 trie->jump[curword] = (U16)(noper_next - convert); \
1448 jumper = noper_next; \
1450 nextbranch= regnext(cur); \
1454 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1455 /* chain, so that when the bits of chain are later */\
1456 /* linked together, the dups appear in the chain */\
1457 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1458 trie->wordinfo[dupe].prev = curword; \
1460 /* we haven't inserted this word yet. */ \
1461 trie->states[ state ].wordnum = curword; \
1466 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1467 ( ( base + charid >= ucharcount \
1468 && base + charid < ubound \
1469 && state == trie->trans[ base - ucharcount + charid ].check \
1470 && trie->trans[ base - ucharcount + charid ].next ) \
1471 ? trie->trans[ base - ucharcount + charid ].next \
1472 : ( state==1 ? special : 0 ) \
1476 #define MADE_JUMP_TRIE 2
1477 #define MADE_EXACT_TRIE 4
1480 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1483 /* first pass, loop through and scan words */
1484 reg_trie_data *trie;
1485 HV *widecharmap = NULL;
1486 AV *revcharmap = newAV();
1488 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1493 regnode *jumper = NULL;
1494 regnode *nextbranch = NULL;
1495 regnode *convert = NULL;
1496 U32 *prev_states; /* temp array mapping each state to previous one */
1497 /* we just use folder as a flag in utf8 */
1498 const U8 * folder = NULL;
1501 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1502 AV *trie_words = NULL;
1503 /* along with revcharmap, this only used during construction but both are
1504 * useful during debugging so we store them in the struct when debugging.
1507 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1508 STRLEN trie_charcount=0;
1510 SV *re_trie_maxbuff;
1511 GET_RE_DEBUG_FLAGS_DECL;
1513 PERL_ARGS_ASSERT_MAKE_TRIE;
1515 PERL_UNUSED_ARG(depth);
1521 case EXACTFU: folder = PL_fold_latin1; break;
1522 case EXACTF: folder = PL_fold; break;
1523 case EXACTFL: folder = PL_fold_locale; break;
1524 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u", (unsigned) flags );
1527 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1529 trie->startstate = 1;
1530 trie->wordcount = word_count;
1531 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1532 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1533 if (!(UTF && folder))
1534 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1535 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1536 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1539 trie_words = newAV();
1542 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1543 if (!SvIOK(re_trie_maxbuff)) {
1544 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1547 PerlIO_printf( Perl_debug_log,
1548 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1549 (int)depth * 2 + 2, "",
1550 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1551 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1555 /* Find the node we are going to overwrite */
1556 if ( first == startbranch && OP( last ) != BRANCH ) {
1557 /* whole branch chain */
1560 /* branch sub-chain */
1561 convert = NEXTOPER( first );
1564 /* -- First loop and Setup --
1566 We first traverse the branches and scan each word to determine if it
1567 contains widechars, and how many unique chars there are, this is
1568 important as we have to build a table with at least as many columns as we
1571 We use an array of integers to represent the character codes 0..255
1572 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1573 native representation of the character value as the key and IV's for the
1576 *TODO* If we keep track of how many times each character is used we can
1577 remap the columns so that the table compression later on is more
1578 efficient in terms of memory by ensuring the most common value is in the
1579 middle and the least common are on the outside. IMO this would be better
1580 than a most to least common mapping as theres a decent chance the most
1581 common letter will share a node with the least common, meaning the node
1582 will not be compressible. With a middle is most common approach the worst
1583 case is when we have the least common nodes twice.
1587 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1588 regnode * const noper = NEXTOPER( cur );
1589 const U8 *uc = (U8*)STRING( noper );
1590 const U8 * const e = uc + STR_LEN( noper );
1592 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1593 const U8 *scan = (U8*)NULL;
1594 U32 wordlen = 0; /* required init */
1596 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1598 if (OP(noper) == NOTHING) {
1602 if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
1603 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1604 regardless of encoding */
1606 for ( ; uc < e ; uc += len ) {
1607 TRIE_CHARCOUNT(trie)++;
1611 if ( !trie->charmap[ uvc ] ) {
1612 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1614 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
1618 /* store the codepoint in the bitmap, and its folded
1620 TRIE_BITMAP_SET(trie,uvc);
1622 /* store the folded codepoint */
1623 if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);
1626 /* store first byte of utf8 representation of
1627 variant codepoints */
1628 if (! UNI_IS_INVARIANT(uvc)) {
1629 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1632 set_bit = 0; /* We've done our bit :-) */
1637 widecharmap = newHV();
1639 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1642 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1644 if ( !SvTRUE( *svpp ) ) {
1645 sv_setiv( *svpp, ++trie->uniquecharcount );
1650 if( cur == first ) {
1653 } else if (chars < trie->minlen) {
1655 } else if (chars > trie->maxlen) {
1659 } /* end first pass */
1660 DEBUG_TRIE_COMPILE_r(
1661 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1662 (int)depth * 2 + 2,"",
1663 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1664 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1665 (int)trie->minlen, (int)trie->maxlen )
1669 We now know what we are dealing with in terms of unique chars and
1670 string sizes so we can calculate how much memory a naive
1671 representation using a flat table will take. If it's over a reasonable
1672 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1673 conservative but potentially much slower representation using an array
1676 At the end we convert both representations into the same compressed
1677 form that will be used in regexec.c for matching with. The latter
1678 is a form that cannot be used to construct with but has memory
1679 properties similar to the list form and access properties similar
1680 to the table form making it both suitable for fast searches and
1681 small enough that its feasable to store for the duration of a program.
1683 See the comment in the code where the compressed table is produced
1684 inplace from the flat tabe representation for an explanation of how
1685 the compression works.
1690 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1693 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1695 Second Pass -- Array Of Lists Representation
1697 Each state will be represented by a list of charid:state records
1698 (reg_trie_trans_le) the first such element holds the CUR and LEN
1699 points of the allocated array. (See defines above).
1701 We build the initial structure using the lists, and then convert
1702 it into the compressed table form which allows faster lookups
1703 (but cant be modified once converted).
1706 STRLEN transcount = 1;
1708 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1709 "%*sCompiling trie using list compiler\n",
1710 (int)depth * 2 + 2, ""));
1712 trie->states = (reg_trie_state *)
1713 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1714 sizeof(reg_trie_state) );
1718 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1720 regnode * const noper = NEXTOPER( cur );
1721 U8 *uc = (U8*)STRING( noper );
1722 const U8 * const e = uc + STR_LEN( noper );
1723 U32 state = 1; /* required init */
1724 U16 charid = 0; /* sanity init */
1725 U8 *scan = (U8*)NULL; /* sanity init */
1726 STRLEN foldlen = 0; /* required init */
1727 U32 wordlen = 0; /* required init */
1728 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1730 if (OP(noper) != NOTHING) {
1731 for ( ; uc < e ; uc += len ) {
1736 charid = trie->charmap[ uvc ];
1738 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1742 charid=(U16)SvIV( *svpp );
1745 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1752 if ( !trie->states[ state ].trans.list ) {
1753 TRIE_LIST_NEW( state );
1755 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1756 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1757 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1762 newstate = next_alloc++;
1763 prev_states[newstate] = state;
1764 TRIE_LIST_PUSH( state, charid, newstate );
1769 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1773 TRIE_HANDLE_WORD(state);
1775 } /* end second pass */
1777 /* next alloc is the NEXT state to be allocated */
1778 trie->statecount = next_alloc;
1779 trie->states = (reg_trie_state *)
1780 PerlMemShared_realloc( trie->states,
1782 * sizeof(reg_trie_state) );
1784 /* and now dump it out before we compress it */
1785 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1786 revcharmap, next_alloc,
1790 trie->trans = (reg_trie_trans *)
1791 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1798 for( state=1 ; state < next_alloc ; state ++ ) {
1802 DEBUG_TRIE_COMPILE_MORE_r(
1803 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1807 if (trie->states[state].trans.list) {
1808 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1812 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1813 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1814 if ( forid < minid ) {
1816 } else if ( forid > maxid ) {
1820 if ( transcount < tp + maxid - minid + 1) {
1822 trie->trans = (reg_trie_trans *)
1823 PerlMemShared_realloc( trie->trans,
1825 * sizeof(reg_trie_trans) );
1826 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1828 base = trie->uniquecharcount + tp - minid;
1829 if ( maxid == minid ) {
1831 for ( ; zp < tp ; zp++ ) {
1832 if ( ! trie->trans[ zp ].next ) {
1833 base = trie->uniquecharcount + zp - minid;
1834 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1835 trie->trans[ zp ].check = state;
1841 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1842 trie->trans[ tp ].check = state;
1847 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1848 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1849 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1850 trie->trans[ tid ].check = state;
1852 tp += ( maxid - minid + 1 );
1854 Safefree(trie->states[ state ].trans.list);
1857 DEBUG_TRIE_COMPILE_MORE_r(
1858 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1861 trie->states[ state ].trans.base=base;
1863 trie->lasttrans = tp + 1;
1867 Second Pass -- Flat Table Representation.
1869 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1870 We know that we will need Charcount+1 trans at most to store the data
1871 (one row per char at worst case) So we preallocate both structures
1872 assuming worst case.
1874 We then construct the trie using only the .next slots of the entry
1877 We use the .check field of the first entry of the node temporarily to
1878 make compression both faster and easier by keeping track of how many non
1879 zero fields are in the node.
1881 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1884 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1885 number representing the first entry of the node, and state as a
1886 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1887 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1888 are 2 entrys per node. eg:
1896 The table is internally in the right hand, idx form. However as we also
1897 have to deal with the states array which is indexed by nodenum we have to
1898 use TRIE_NODENUM() to convert.
1901 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1902 "%*sCompiling trie using table compiler\n",
1903 (int)depth * 2 + 2, ""));
1905 trie->trans = (reg_trie_trans *)
1906 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1907 * trie->uniquecharcount + 1,
1908 sizeof(reg_trie_trans) );
1909 trie->states = (reg_trie_state *)
1910 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1911 sizeof(reg_trie_state) );
1912 next_alloc = trie->uniquecharcount + 1;
1915 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1917 regnode * const noper = NEXTOPER( cur );
1918 const U8 *uc = (U8*)STRING( noper );
1919 const U8 * const e = uc + STR_LEN( noper );
1921 U32 state = 1; /* required init */
1923 U16 charid = 0; /* sanity init */
1924 U32 accept_state = 0; /* sanity init */
1925 U8 *scan = (U8*)NULL; /* sanity init */
1927 STRLEN foldlen = 0; /* required init */
1928 U32 wordlen = 0; /* required init */
1929 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1931 if ( OP(noper) != NOTHING ) {
1932 for ( ; uc < e ; uc += len ) {
1937 charid = trie->charmap[ uvc ];
1939 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1940 charid = svpp ? (U16)SvIV(*svpp) : 0;
1944 if ( !trie->trans[ state + charid ].next ) {
1945 trie->trans[ state + charid ].next = next_alloc;
1946 trie->trans[ state ].check++;
1947 prev_states[TRIE_NODENUM(next_alloc)]
1948 = TRIE_NODENUM(state);
1949 next_alloc += trie->uniquecharcount;
1951 state = trie->trans[ state + charid ].next;
1953 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1955 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1958 accept_state = TRIE_NODENUM( state );
1959 TRIE_HANDLE_WORD(accept_state);
1961 } /* end second pass */
1963 /* and now dump it out before we compress it */
1964 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1966 next_alloc, depth+1));
1970 * Inplace compress the table.*
1972 For sparse data sets the table constructed by the trie algorithm will
1973 be mostly 0/FAIL transitions or to put it another way mostly empty.
1974 (Note that leaf nodes will not contain any transitions.)
1976 This algorithm compresses the tables by eliminating most such
1977 transitions, at the cost of a modest bit of extra work during lookup:
1979 - Each states[] entry contains a .base field which indicates the
1980 index in the state[] array wheres its transition data is stored.
1982 - If .base is 0 there are no valid transitions from that node.
1984 - If .base is nonzero then charid is added to it to find an entry in
1987 -If trans[states[state].base+charid].check!=state then the
1988 transition is taken to be a 0/Fail transition. Thus if there are fail
1989 transitions at the front of the node then the .base offset will point
1990 somewhere inside the previous nodes data (or maybe even into a node
1991 even earlier), but the .check field determines if the transition is
1995 The following process inplace converts the table to the compressed
1996 table: We first do not compress the root node 1,and mark all its
1997 .check pointers as 1 and set its .base pointer as 1 as well. This
1998 allows us to do a DFA construction from the compressed table later,
1999 and ensures that any .base pointers we calculate later are greater
2002 - We set 'pos' to indicate the first entry of the second node.
2004 - We then iterate over the columns of the node, finding the first and
2005 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2006 and set the .check pointers accordingly, and advance pos
2007 appropriately and repreat for the next node. Note that when we copy
2008 the next pointers we have to convert them from the original
2009 NODEIDX form to NODENUM form as the former is not valid post
2012 - If a node has no transitions used we mark its base as 0 and do not
2013 advance the pos pointer.
2015 - If a node only has one transition we use a second pointer into the
2016 structure to fill in allocated fail transitions from other states.
2017 This pointer is independent of the main pointer and scans forward
2018 looking for null transitions that are allocated to a state. When it
2019 finds one it writes the single transition into the "hole". If the
2020 pointer doesnt find one the single transition is appended as normal.
2022 - Once compressed we can Renew/realloc the structures to release the
2025 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2026 specifically Fig 3.47 and the associated pseudocode.
2030 const U32 laststate = TRIE_NODENUM( next_alloc );
2033 trie->statecount = laststate;
2035 for ( state = 1 ; state < laststate ; state++ ) {
2037 const U32 stateidx = TRIE_NODEIDX( state );
2038 const U32 o_used = trie->trans[ stateidx ].check;
2039 U32 used = trie->trans[ stateidx ].check;
2040 trie->trans[ stateidx ].check = 0;
2042 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2043 if ( flag || trie->trans[ stateidx + charid ].next ) {
2044 if ( trie->trans[ stateidx + charid ].next ) {
2046 for ( ; zp < pos ; zp++ ) {
2047 if ( ! trie->trans[ zp ].next ) {
2051 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2052 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2053 trie->trans[ zp ].check = state;
2054 if ( ++zp > pos ) pos = zp;
2061 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2063 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2064 trie->trans[ pos ].check = state;
2069 trie->lasttrans = pos + 1;
2070 trie->states = (reg_trie_state *)
2071 PerlMemShared_realloc( trie->states, laststate
2072 * sizeof(reg_trie_state) );
2073 DEBUG_TRIE_COMPILE_MORE_r(
2074 PerlIO_printf( Perl_debug_log,
2075 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2076 (int)depth * 2 + 2,"",
2077 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2080 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2083 } /* end table compress */
2085 DEBUG_TRIE_COMPILE_MORE_r(
2086 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2087 (int)depth * 2 + 2, "",
2088 (UV)trie->statecount,
2089 (UV)trie->lasttrans)
2091 /* resize the trans array to remove unused space */
2092 trie->trans = (reg_trie_trans *)
2093 PerlMemShared_realloc( trie->trans, trie->lasttrans
2094 * sizeof(reg_trie_trans) );
2096 { /* Modify the program and insert the new TRIE node */
2097 U8 nodetype =(U8)(flags & 0xFF);
2101 regnode *optimize = NULL;
2102 #ifdef RE_TRACK_PATTERN_OFFSETS
2105 U32 mjd_nodelen = 0;
2106 #endif /* RE_TRACK_PATTERN_OFFSETS */
2107 #endif /* DEBUGGING */
2109 This means we convert either the first branch or the first Exact,
2110 depending on whether the thing following (in 'last') is a branch
2111 or not and whther first is the startbranch (ie is it a sub part of
2112 the alternation or is it the whole thing.)
2113 Assuming its a sub part we convert the EXACT otherwise we convert
2114 the whole branch sequence, including the first.
2116 /* Find the node we are going to overwrite */
2117 if ( first != startbranch || OP( last ) == BRANCH ) {
2118 /* branch sub-chain */
2119 NEXT_OFF( first ) = (U16)(last - first);
2120 #ifdef RE_TRACK_PATTERN_OFFSETS
2122 mjd_offset= Node_Offset((convert));
2123 mjd_nodelen= Node_Length((convert));
2126 /* whole branch chain */
2128 #ifdef RE_TRACK_PATTERN_OFFSETS
2131 const regnode *nop = NEXTOPER( convert );
2132 mjd_offset= Node_Offset((nop));
2133 mjd_nodelen= Node_Length((nop));
2137 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2138 (int)depth * 2 + 2, "",
2139 (UV)mjd_offset, (UV)mjd_nodelen)
2142 /* But first we check to see if there is a common prefix we can
2143 split out as an EXACT and put in front of the TRIE node. */
2144 trie->startstate= 1;
2145 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2147 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2151 const U32 base = trie->states[ state ].trans.base;
2153 if ( trie->states[state].wordnum )
2156 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2157 if ( ( base + ofs >= trie->uniquecharcount ) &&
2158 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2159 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2161 if ( ++count > 1 ) {
2162 SV **tmp = av_fetch( revcharmap, ofs, 0);
2163 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2164 if ( state == 1 ) break;
2166 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2168 PerlIO_printf(Perl_debug_log,
2169 "%*sNew Start State=%"UVuf" Class: [",
2170 (int)depth * 2 + 2, "",
2173 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2174 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2176 TRIE_BITMAP_SET(trie,*ch);
2178 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2180 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2184 TRIE_BITMAP_SET(trie,*ch);
2186 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2187 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2193 SV **tmp = av_fetch( revcharmap, idx, 0);
2195 char *ch = SvPV( *tmp, len );
2197 SV *sv=sv_newmortal();
2198 PerlIO_printf( Perl_debug_log,
2199 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2200 (int)depth * 2 + 2, "",
2202 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2203 PL_colors[0], PL_colors[1],
2204 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2205 PERL_PV_ESCAPE_FIRSTCHAR
2210 OP( convert ) = nodetype;
2211 str=STRING(convert);
2214 STR_LEN(convert) += len;
2220 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2225 trie->prefixlen = (state-1);
2227 regnode *n = convert+NODE_SZ_STR(convert);
2228 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2229 trie->startstate = state;
2230 trie->minlen -= (state - 1);
2231 trie->maxlen -= (state - 1);
2233 /* At least the UNICOS C compiler choked on this
2234 * being argument to DEBUG_r(), so let's just have
2237 #ifdef PERL_EXT_RE_BUILD
2243 regnode *fix = convert;
2244 U32 word = trie->wordcount;
2246 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2247 while( ++fix < n ) {
2248 Set_Node_Offset_Length(fix, 0, 0);
2251 SV ** const tmp = av_fetch( trie_words, word, 0 );
2253 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2254 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2256 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2264 NEXT_OFF(convert) = (U16)(tail - convert);
2265 DEBUG_r(optimize= n);
2271 if ( trie->maxlen ) {
2272 NEXT_OFF( convert ) = (U16)(tail - convert);
2273 ARG_SET( convert, data_slot );
2274 /* Store the offset to the first unabsorbed branch in
2275 jump[0], which is otherwise unused by the jump logic.
2276 We use this when dumping a trie and during optimisation. */
2278 trie->jump[0] = (U16)(nextbranch - convert);
2280 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2281 * and there is a bitmap
2282 * and the first "jump target" node we found leaves enough room
2283 * then convert the TRIE node into a TRIEC node, with the bitmap
2284 * embedded inline in the opcode - this is hypothetically faster.
2286 if ( !trie->states[trie->startstate].wordnum
2288 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2290 OP( convert ) = TRIEC;
2291 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2292 PerlMemShared_free(trie->bitmap);
2295 OP( convert ) = TRIE;
2297 /* store the type in the flags */
2298 convert->flags = nodetype;
2302 + regarglen[ OP( convert ) ];
2304 /* XXX We really should free up the resource in trie now,
2305 as we won't use them - (which resources?) dmq */
2307 /* needed for dumping*/
2308 DEBUG_r(if (optimize) {
2309 regnode *opt = convert;
2311 while ( ++opt < optimize) {
2312 Set_Node_Offset_Length(opt,0,0);
2315 Try to clean up some of the debris left after the
2318 while( optimize < jumper ) {
2319 mjd_nodelen += Node_Length((optimize));
2320 OP( optimize ) = OPTIMIZED;
2321 Set_Node_Offset_Length(optimize,0,0);
2324 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2326 } /* end node insert */
2328 /* Finish populating the prev field of the wordinfo array. Walk back
2329 * from each accept state until we find another accept state, and if
2330 * so, point the first word's .prev field at the second word. If the
2331 * second already has a .prev field set, stop now. This will be the
2332 * case either if we've already processed that word's accept state,
2333 * or that state had multiple words, and the overspill words were
2334 * already linked up earlier.
2341 for (word=1; word <= trie->wordcount; word++) {
2343 if (trie->wordinfo[word].prev)
2345 state = trie->wordinfo[word].accept;
2347 state = prev_states[state];
2350 prev = trie->states[state].wordnum;
2354 trie->wordinfo[word].prev = prev;
2356 Safefree(prev_states);
2360 /* and now dump out the compressed format */
2361 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2363 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2365 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2366 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2368 SvREFCNT_dec(revcharmap);
2372 : trie->startstate>1
2378 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2380 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2382 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2383 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2386 We find the fail state for each state in the trie, this state is the longest proper
2387 suffix of the current state's 'word' that is also a proper prefix of another word in our
2388 trie. State 1 represents the word '' and is thus the default fail state. This allows
2389 the DFA not to have to restart after its tried and failed a word at a given point, it
2390 simply continues as though it had been matching the other word in the first place.
2392 'abcdgu'=~/abcdefg|cdgu/
2393 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2394 fail, which would bring us to the state representing 'd' in the second word where we would
2395 try 'g' and succeed, proceeding to match 'cdgu'.
2397 /* add a fail transition */
2398 const U32 trie_offset = ARG(source);
2399 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2401 const U32 ucharcount = trie->uniquecharcount;
2402 const U32 numstates = trie->statecount;
2403 const U32 ubound = trie->lasttrans + ucharcount;
2407 U32 base = trie->states[ 1 ].trans.base;
2410 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2411 GET_RE_DEBUG_FLAGS_DECL;
2413 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2415 PERL_UNUSED_ARG(depth);
2419 ARG_SET( stclass, data_slot );
2420 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2421 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2422 aho->trie=trie_offset;
2423 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2424 Copy( trie->states, aho->states, numstates, reg_trie_state );
2425 Newxz( q, numstates, U32);
2426 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2429 /* initialize fail[0..1] to be 1 so that we always have
2430 a valid final fail state */
2431 fail[ 0 ] = fail[ 1 ] = 1;
2433 for ( charid = 0; charid < ucharcount ; charid++ ) {
2434 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2436 q[ q_write ] = newstate;
2437 /* set to point at the root */
2438 fail[ q[ q_write++ ] ]=1;
2441 while ( q_read < q_write) {
2442 const U32 cur = q[ q_read++ % numstates ];
2443 base = trie->states[ cur ].trans.base;
2445 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2446 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2448 U32 fail_state = cur;
2451 fail_state = fail[ fail_state ];
2452 fail_base = aho->states[ fail_state ].trans.base;
2453 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2455 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2456 fail[ ch_state ] = fail_state;
2457 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2459 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2461 q[ q_write++ % numstates] = ch_state;
2465 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2466 when we fail in state 1, this allows us to use the
2467 charclass scan to find a valid start char. This is based on the principle
2468 that theres a good chance the string being searched contains lots of stuff
2469 that cant be a start char.
2471 fail[ 0 ] = fail[ 1 ] = 0;
2472 DEBUG_TRIE_COMPILE_r({
2473 PerlIO_printf(Perl_debug_log,
2474 "%*sStclass Failtable (%"UVuf" states): 0",
2475 (int)(depth * 2), "", (UV)numstates
2477 for( q_read=1; q_read<numstates; q_read++ ) {
2478 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2480 PerlIO_printf(Perl_debug_log, "\n");
2483 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2488 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2489 * These need to be revisited when a newer toolchain becomes available.
2491 #if defined(__sparc64__) && defined(__GNUC__)
2492 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2493 # undef SPARC64_GCC_WORKAROUND
2494 # define SPARC64_GCC_WORKAROUND 1
2498 #define DEBUG_PEEP(str,scan,depth) \
2499 DEBUG_OPTIMISE_r({if (scan){ \
2500 SV * const mysv=sv_newmortal(); \
2501 regnode *Next = regnext(scan); \
2502 regprop(RExC_rx, mysv, scan); \
2503 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2504 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2505 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2509 /* The below joins as many adjacent EXACTish nodes as possible into a single
2510 * one, and looks for problematic sequences of characters whose folds vs.
2511 * non-folds have sufficiently different lengths, that the optimizer would be
2512 * fooled into rejecting legitimate matches of them, and the trie construction
2513 * code can't cope with them. The joining is only done if:
2514 * 1) there is room in the current conglomerated node to entirely contain the
2516 * 2) they are the exact same node type
2518 * The adjacent nodes actually may be separated by NOTHING kind nodes, and
2519 * these get optimized out
2521 * If there are problematic code sequences, *min_subtract is set to the delta
2522 * that the minimum size of the node can be less than its actual size. And,
2523 * the node type of the result is changed to reflect that it contains these
2526 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2527 * and contains LATIN SMALL LETTER SHARP S
2529 * This is as good a place as any to discuss the design of handling these
2530 * problematic sequences. It's been wrong in Perl for a very long time. There
2531 * are three code points in Unicode whose folded lengths differ so much from
2532 * the un-folded lengths that it causes problems for the optimizer and trie
2533 * construction. Why only these are problematic, and not others where lengths
2534 * also differ is something I (khw) do not understand. New versions of Unicode
2535 * might add more such code points. Hopefully the logic in fold_grind.t that
2536 * figures out what to test (in part by verifying that each size-combination
2537 * gets tested) will catch any that do come along, so they can be added to the
2538 * special handling below. The chances of new ones are actually rather small,
2539 * as most, if not all, of the world's scripts that have casefolding have
2540 * already been encoded by Unicode. Also, a number of Unicode's decisions were
2541 * made to allow compatibility with pre-existing standards, and almost all of
2542 * those have already been dealt with. These would otherwise be the most
2543 * likely candidates for generating further tricky sequences. In other words,
2544 * Unicode by itself is unlikely to add new ones unless it is for compatibility
2545 * with pre-existing standards, and there aren't many of those left.
2547 * The previous designs for dealing with these involved assigning a special
2548 * node for them. This approach doesn't work, as evidenced by this example:
2549 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2550 * Both these fold to "sss", but if the pattern is parsed to create a node of
2551 * that would match just the \xDF, it won't be able to handle the case where a
2552 * successful match would have to cross the node's boundary. The new approach
2553 * that hopefully generally solves the problem generates an EXACTFU_SS node
2556 * There are a number of components to the approach (a lot of work for just
2557 * three code points!):
2558 * 1) This routine examines each EXACTFish node that could contain the
2559 * problematic sequences. It returns in *min_subtract how much to
2560 * subtract from the the actual length of the string to get a real minimum
2561 * for one that could match it. This number is usually 0 except for the
2562 * problematic sequences. This delta is used by the caller to adjust the
2563 * min length of the match, and the delta between min and max, so that the
2564 * optimizer doesn't reject these possibilities based on size constraints.
2565 * 2) These sequences are not currently correctly handled by the trie code
2566 * either, so it changes the joined node type to ops that are not handled
2567 * by trie's, those new ops being EXACTFU_SS and EXACTFU_NO_TRIE.
2568 * 3) This is sufficient for the two Greek sequences (described below), but
2569 * the one involving the Sharp s (\xDF) needs more. The node type
2570 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2571 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2572 * case where there is a possible fold length change. That means that a
2573 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2574 * itself with length changes, and so can be processed faster. regexec.c
2575 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2576 * is pre-folded by regcomp.c. This saves effort in regex matching.
2577 * However, probably mostly for historical reasons, the pre-folding isn't
2578 * done for non-UTF8 patterns (and it can't be for EXACTF and EXACTFL
2579 * nodes, as what they fold to isn't known until runtime.) The fold
2580 * possibilities for the non-UTF8 patterns are quite simple, except for
2581 * the sharp s. All the ones that don't involve a UTF-8 target string
2582 * are members of a fold-pair, and arrays are set up for all of them
2583 * that quickly find the other member of the pair. It might actually
2584 * be faster to pre-fold these, but it isn't currently done, except for
2585 * the sharp s. Code elsewhere in this file makes sure that it gets
2586 * folded to 'ss', even if the pattern isn't UTF-8. This avoids the
2587 * issues described in the next item.
2588 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2589 * 'ss' or not is not knowable at compile time. It will match iff the
2590 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2591 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2592 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2593 * described in item 3). An assumption that the optimizer part of
2594 * regexec.c (probably unwittingly) makes is that a character in the
2595 * pattern corresponds to at most a single character in the target string.
2596 * (And I do mean character, and not byte here, unlike other parts of the
2597 * documentation that have never been updated to account for multibyte
2598 * Unicode.) This assumption is wrong only in this case, as all other
2599 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2600 * virtue of having this file pre-fold UTF-8 patterns. I'm
2601 * reluctant to try to change this assumption, so instead the code punts.
2602 * This routine examines EXACTF nodes for the sharp s, and returns a
2603 * boolean indicating whether or not the node is an EXACTF node that
2604 * contains a sharp s. When it is true, the caller sets a flag that later
2605 * causes the optimizer in this file to not set values for the floating
2606 * and fixed string lengths, and thus avoids the optimizer code in
2607 * regexec.c that makes the invalid assumption. Thus, there is no
2608 * optimization based on string lengths for EXACTF nodes that contain the
2609 * sharp s. This only happens for /id rules (which means the pattern
2613 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2614 if (PL_regkind[OP(scan)] == EXACT) \
2615 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2618 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) {
2619 /* Merge several consecutive EXACTish nodes into one. */
2620 regnode *n = regnext(scan);
2622 regnode *next = scan + NODE_SZ_STR(scan);
2626 regnode *stop = scan;
2627 GET_RE_DEBUG_FLAGS_DECL;
2629 PERL_UNUSED_ARG(depth);
2632 PERL_ARGS_ASSERT_JOIN_EXACT;
2633 #ifndef EXPERIMENTAL_INPLACESCAN
2634 PERL_UNUSED_ARG(flags);
2635 PERL_UNUSED_ARG(val);
2637 DEBUG_PEEP("join",scan,depth);
2639 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2640 * EXACT ones that are mergeable to the current one. */
2642 && (PL_regkind[OP(n)] == NOTHING
2643 || (stringok && OP(n) == OP(scan)))
2645 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2648 if (OP(n) == TAIL || n > next)
2650 if (PL_regkind[OP(n)] == NOTHING) {
2651 DEBUG_PEEP("skip:",n,depth);
2652 NEXT_OFF(scan) += NEXT_OFF(n);
2653 next = n + NODE_STEP_REGNODE;
2660 else if (stringok) {
2661 const unsigned int oldl = STR_LEN(scan);
2662 regnode * const nnext = regnext(n);
2664 if (oldl + STR_LEN(n) > U8_MAX)
2667 DEBUG_PEEP("merg",n,depth);
2670 NEXT_OFF(scan) += NEXT_OFF(n);
2671 STR_LEN(scan) += STR_LEN(n);
2672 next = n + NODE_SZ_STR(n);
2673 /* Now we can overwrite *n : */
2674 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2682 #ifdef EXPERIMENTAL_INPLACESCAN
2683 if (flags && !NEXT_OFF(n)) {
2684 DEBUG_PEEP("atch", val, depth);
2685 if (reg_off_by_arg[OP(n)]) {
2686 ARG_SET(n, val - n);
2689 NEXT_OFF(n) = val - n;
2697 *has_exactf_sharp_s = FALSE;
2699 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2700 * can now analyze for sequences of problematic code points. (Prior to
2701 * this final joining, sequences could have been split over boundaries, and
2702 * hence missed). The sequences only happen in folding, hence for any
2703 * non-EXACT EXACTish node */
2704 if (OP(scan) != EXACT) {
2706 U8 * s0 = (U8*) STRING(scan);
2707 U8 * const s_end = s0 + STR_LEN(scan);
2709 /* The below is perhaps overboard, but this allows us to save a test
2710 * each time through the loop at the expense of a mask. This is
2711 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2712 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2713 * This uses an exclusive 'or' to find that bit and then inverts it to
2714 * form a mask, with just a single 0, in the bit position where 'S' and
2716 const U8 S_or_s_mask = ~ ('S' ^ 's');
2717 const U8 s_masked = 's' & S_or_s_mask;
2719 /* One pass is made over the node's string looking for all the
2720 * possibilities. to avoid some tests in the loop, there are two main
2721 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2725 /* There are two problematic Greek code points in Unicode
2728 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2729 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2735 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2736 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2738 * This means that in case-insensitive matching (or "loose
2739 * matching", as Unicode calls it), an EXACTF of length six (the
2740 * UTF-8 encoded byte length of the above casefolded versions) can
2741 * match a target string of length two (the byte length of UTF-8
2742 * encoded U+0390 or U+03B0). This would rather mess up the
2743 * minimum length computation. (there are other code points that
2744 * also fold to these two sequences, but the delta is smaller)
2746 * If these sequences are found, the minimum length is decreased by
2747 * four (six minus two).
2749 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2750 * LETTER SHARP S. We decrease the min length by 1 for each
2751 * occurrence of 'ss' found */
2753 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2754 # define U390_first_byte 0xb4
2755 const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
2756 # define U3B0_first_byte 0xb5
2757 const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
2759 # define U390_first_byte 0xce
2760 const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
2761 # define U3B0_first_byte 0xcf
2762 const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
2764 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2765 yields a net of 0 */
2766 /* Examine the string for one of the problematic sequences */
2768 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2769 * sequence we are looking for is 2 */
2773 /* Look for the first byte in each problematic sequence */
2775 /* We don't have to worry about other things that fold to
2776 * 's' (such as the long s, U+017F), as all above-latin1
2777 * code points have been pre-folded */
2781 /* Current character is an 's' or 'S'. If next one is
2782 * as well, we have the dreaded sequence */
2783 if (((*(s+1) & S_or_s_mask) == s_masked)
2784 /* These two node types don't have special handling
2786 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2789 OP(scan) = EXACTFU_SS;
2790 s++; /* No need to look at this character again */
2794 case U390_first_byte:
2795 if (s_end - s >= len
2797 /* The 1's are because are skipping comparing the
2799 && memEQ(s + 1, U390_tail, len - 1))
2801 goto greek_sequence;
2805 case U3B0_first_byte:
2806 if (! (s_end - s >= len
2807 && memEQ(s + 1, U3B0_tail, len - 1)))
2814 /* This can't currently be handled by trie's, so change
2815 * the node type to indicate this. If EXACTFA and
2816 * EXACTFL were ever to be handled by trie's, this
2817 * would have to be changed. If this node has already
2818 * been changed to EXACTFU_SS in this loop, leave it as
2819 * is. (I (khw) think it doesn't matter in regexec.c
2820 * for UTF patterns, but no need to change it */
2821 if (OP(scan) == EXACTFU) {
2822 OP(scan) = EXACTFU_NO_TRIE;
2824 s += 6; /* We already know what this sequence is. Skip
2830 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2832 /* Here, the pattern is not UTF-8. We need to look only for the
2833 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2834 * in the final position. Otherwise we can stop looking 1 byte
2835 * earlier because have to find both the first and second 's' */
2836 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2838 for (s = s0; s < upper; s++) {
2843 && ((*(s+1) & S_or_s_mask) == s_masked))
2847 /* EXACTF nodes need to know that the minimum
2848 * length changed so that a sharp s in the string
2849 * can match this ss in the pattern, but they
2850 * remain EXACTF nodes, as they are not trie'able,
2851 * so don't have to invent a new node type to
2852 * exclude them from the trie code */
2853 if (OP(scan) != EXACTF) {
2854 OP(scan) = EXACTFU_SS;
2859 case LATIN_SMALL_LETTER_SHARP_S:
2860 if (OP(scan) == EXACTF) {
2861 *has_exactf_sharp_s = TRUE;
2870 /* Allow dumping but overwriting the collection of skipped
2871 * ops and/or strings with fake optimized ops */
2872 n = scan + NODE_SZ_STR(scan);
2880 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2884 /* REx optimizer. Converts nodes into quicker variants "in place".
2885 Finds fixed substrings. */
2887 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2888 to the position after last scanned or to NULL. */
2890 #define INIT_AND_WITHP \
2891 assert(!and_withp); \
2892 Newx(and_withp,1,struct regnode_charclass_class); \
2893 SAVEFREEPV(and_withp)
2895 /* this is a chain of data about sub patterns we are processing that
2896 need to be handled separately/specially in study_chunk. Its so
2897 we can simulate recursion without losing state. */
2899 typedef struct scan_frame {
2900 regnode *last; /* last node to process in this frame */
2901 regnode *next; /* next node to process when last is reached */
2902 struct scan_frame *prev; /*previous frame*/
2903 I32 stop; /* what stopparen do we use */
2907 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2909 #define CASE_SYNST_FNC(nAmE) \
2911 if (flags & SCF_DO_STCLASS_AND) { \
2912 for (value = 0; value < 256; value++) \
2913 if (!is_ ## nAmE ## _cp(value)) \
2914 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2917 for (value = 0; value < 256; value++) \
2918 if (is_ ## nAmE ## _cp(value)) \
2919 ANYOF_BITMAP_SET(data->start_class, value); \
2923 if (flags & SCF_DO_STCLASS_AND) { \
2924 for (value = 0; value < 256; value++) \
2925 if (is_ ## nAmE ## _cp(value)) \
2926 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2929 for (value = 0; value < 256; value++) \
2930 if (!is_ ## nAmE ## _cp(value)) \
2931 ANYOF_BITMAP_SET(data->start_class, value); \
2938 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2939 I32 *minlenp, I32 *deltap,
2944 struct regnode_charclass_class *and_withp,
2945 U32 flags, U32 depth)
2946 /* scanp: Start here (read-write). */
2947 /* deltap: Write maxlen-minlen here. */
2948 /* last: Stop before this one. */
2949 /* data: string data about the pattern */
2950 /* stopparen: treat close N as END */
2951 /* recursed: which subroutines have we recursed into */
2952 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
2955 I32 min = 0, pars = 0, code;
2956 regnode *scan = *scanp, *next;
2958 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2959 int is_inf_internal = 0; /* The studied chunk is infinite */
2960 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2961 scan_data_t data_fake;
2962 SV *re_trie_maxbuff = NULL;
2963 regnode *first_non_open = scan;
2964 I32 stopmin = I32_MAX;
2965 scan_frame *frame = NULL;
2966 GET_RE_DEBUG_FLAGS_DECL;
2968 PERL_ARGS_ASSERT_STUDY_CHUNK;
2971 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
2975 while (first_non_open && OP(first_non_open) == OPEN)
2976 first_non_open=regnext(first_non_open);
2981 while ( scan && OP(scan) != END && scan < last ){
2982 UV min_subtract = 0; /* How much to subtract from the minimum node
2983 length to get a real minimum (because the
2984 folded version may be shorter) */
2985 bool has_exactf_sharp_s = FALSE;
2986 /* Peephole optimizer: */
2987 DEBUG_STUDYDATA("Peep:", data,depth);
2988 DEBUG_PEEP("Peep",scan,depth);
2990 /* Its not clear to khw or hv why this is done here, and not in the
2991 * clauses that deal with EXACT nodes. khw's guess is that it's
2992 * because of a previous design */
2993 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
2995 /* Follow the next-chain of the current node and optimize
2996 away all the NOTHINGs from it. */
2997 if (OP(scan) != CURLYX) {
2998 const int max = (reg_off_by_arg[OP(scan)]
3000 /* I32 may be smaller than U16 on CRAYs! */
3001 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3002 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3006 /* Skip NOTHING and LONGJMP. */
3007 while ((n = regnext(n))
3008 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3009 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3010 && off + noff < max)
3012 if (reg_off_by_arg[OP(scan)])
3015 NEXT_OFF(scan) = off;
3020 /* The principal pseudo-switch. Cannot be a switch, since we
3021 look into several different things. */
3022 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3023 || OP(scan) == IFTHEN) {
3024 next = regnext(scan);
3026 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3028 if (OP(next) == code || code == IFTHEN) {
3029 /* NOTE - There is similar code to this block below for handling
3030 TRIE nodes on a re-study. If you change stuff here check there
3032 I32 max1 = 0, min1 = I32_MAX, num = 0;
3033 struct regnode_charclass_class accum;
3034 regnode * const startbranch=scan;
3036 if (flags & SCF_DO_SUBSTR)
3037 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3038 if (flags & SCF_DO_STCLASS)
3039 cl_init_zero(pRExC_state, &accum);
3041 while (OP(scan) == code) {
3042 I32 deltanext, minnext, f = 0, fake;
3043 struct regnode_charclass_class this_class;
3046 data_fake.flags = 0;
3048 data_fake.whilem_c = data->whilem_c;
3049 data_fake.last_closep = data->last_closep;
3052 data_fake.last_closep = &fake;
3054 data_fake.pos_delta = delta;
3055 next = regnext(scan);
3056 scan = NEXTOPER(scan);
3058 scan = NEXTOPER(scan);
3059 if (flags & SCF_DO_STCLASS) {
3060 cl_init(pRExC_state, &this_class);
3061 data_fake.start_class = &this_class;
3062 f = SCF_DO_STCLASS_AND;
3064 if (flags & SCF_WHILEM_VISITED_POS)
3065 f |= SCF_WHILEM_VISITED_POS;
3067 /* we suppose the run is continuous, last=next...*/
3068 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3070 stopparen, recursed, NULL, f,depth+1);
3073 if (max1 < minnext + deltanext)
3074 max1 = minnext + deltanext;
3075 if (deltanext == I32_MAX)
3076 is_inf = is_inf_internal = 1;
3078 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3080 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3081 if ( stopmin > minnext)
3082 stopmin = min + min1;
3083 flags &= ~SCF_DO_SUBSTR;
3085 data->flags |= SCF_SEEN_ACCEPT;
3088 if (data_fake.flags & SF_HAS_EVAL)
3089 data->flags |= SF_HAS_EVAL;
3090 data->whilem_c = data_fake.whilem_c;
3092 if (flags & SCF_DO_STCLASS)
3093 cl_or(pRExC_state, &accum, &this_class);
3095 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3097 if (flags & SCF_DO_SUBSTR) {
3098 data->pos_min += min1;
3099 data->pos_delta += max1 - min1;
3100 if (max1 != min1 || is_inf)
3101 data->longest = &(data->longest_float);
3104 delta += max1 - min1;
3105 if (flags & SCF_DO_STCLASS_OR) {
3106 cl_or(pRExC_state, data->start_class, &accum);
3108 cl_and(data->start_class, and_withp);
3109 flags &= ~SCF_DO_STCLASS;
3112 else if (flags & SCF_DO_STCLASS_AND) {
3114 cl_and(data->start_class, &accum);
3115 flags &= ~SCF_DO_STCLASS;
3118 /* Switch to OR mode: cache the old value of
3119 * data->start_class */
3121 StructCopy(data->start_class, and_withp,
3122 struct regnode_charclass_class);
3123 flags &= ~SCF_DO_STCLASS_AND;
3124 StructCopy(&accum, data->start_class,
3125 struct regnode_charclass_class);
3126 flags |= SCF_DO_STCLASS_OR;
3127 data->start_class->flags |= ANYOF_EOS;
3131 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3134 Assuming this was/is a branch we are dealing with: 'scan' now
3135 points at the item that follows the branch sequence, whatever
3136 it is. We now start at the beginning of the sequence and look
3143 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3145 If we can find such a subsequence we need to turn the first
3146 element into a trie and then add the subsequent branch exact
3147 strings to the trie.
3151 1. patterns where the whole set of branches can be converted.
3153 2. patterns where only a subset can be converted.
3155 In case 1 we can replace the whole set with a single regop
3156 for the trie. In case 2 we need to keep the start and end
3159 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3160 becomes BRANCH TRIE; BRANCH X;
3162 There is an additional case, that being where there is a
3163 common prefix, which gets split out into an EXACT like node
3164 preceding the TRIE node.
3166 If x(1..n)==tail then we can do a simple trie, if not we make
3167 a "jump" trie, such that when we match the appropriate word
3168 we "jump" to the appropriate tail node. Essentially we turn
3169 a nested if into a case structure of sorts.
3174 if (!re_trie_maxbuff) {
3175 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3176 if (!SvIOK(re_trie_maxbuff))
3177 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3179 if ( SvIV(re_trie_maxbuff)>=0 ) {
3181 regnode *first = (regnode *)NULL;
3182 regnode *last = (regnode *)NULL;
3183 regnode *tail = scan;
3188 SV * const mysv = sv_newmortal(); /* for dumping */
3190 /* var tail is used because there may be a TAIL
3191 regop in the way. Ie, the exacts will point to the
3192 thing following the TAIL, but the last branch will
3193 point at the TAIL. So we advance tail. If we
3194 have nested (?:) we may have to move through several
3198 while ( OP( tail ) == TAIL ) {
3199 /* this is the TAIL generated by (?:) */
3200 tail = regnext( tail );
3205 regprop(RExC_rx, mysv, tail );
3206 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3207 (int)depth * 2 + 2, "",
3208 "Looking for TRIE'able sequences. Tail node is: ",
3209 SvPV_nolen_const( mysv )
3215 step through the branches, cur represents each
3216 branch, noper is the first thing to be matched
3217 as part of that branch and noper_next is the
3218 regnext() of that node. if noper is an EXACT
3219 and noper_next is the same as scan (our current
3220 position in the regex) then the EXACT branch is
3221 a possible optimization target. Once we have
3222 two or more consecutive such branches we can
3223 create a trie of the EXACT's contents and stich
3224 it in place. If the sequence represents all of
3225 the branches we eliminate the whole thing and
3226 replace it with a single TRIE. If it is a
3227 subsequence then we need to stitch it in. This
3228 means the first branch has to remain, and needs
3229 to be repointed at the item on the branch chain
3230 following the last branch optimized. This could
3231 be either a BRANCH, in which case the
3232 subsequence is internal, or it could be the
3233 item following the branch sequence in which
3234 case the subsequence is at the end.
3238 /* dont use tail as the end marker for this traverse */
3239 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3240 regnode * const noper = NEXTOPER( cur );
3241 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3242 regnode * const noper_next = regnext( noper );
3246 regprop(RExC_rx, mysv, cur);
3247 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3248 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3250 regprop(RExC_rx, mysv, noper);
3251 PerlIO_printf( Perl_debug_log, " -> %s",
3252 SvPV_nolen_const(mysv));
3255 regprop(RExC_rx, mysv, noper_next );
3256 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3257 SvPV_nolen_const(mysv));
3259 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
3260 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
3262 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
3263 : PL_regkind[ OP( noper ) ] == EXACT )
3264 || OP(noper) == NOTHING )
3266 && noper_next == tail
3271 if ( !first || optype == NOTHING ) {
3272 if (!first) first = cur;
3273 optype = OP( noper );
3279 Currently the trie logic handles case insensitive matching properly only
3280 when the pattern is UTF-8 and the node is EXACTFU (thus forcing unicode
3283 If/when this is fixed the following define can be swapped
3284 in below to fully enable trie logic.
3286 #define TRIE_TYPE_IS_SAFE 1
3288 Note that join_exact() assumes that the other types of EXACTFish nodes are not
3289 used in tries, so that would have to be updated if this changed
3292 #define TRIE_TYPE_IS_SAFE ((UTF && optype == EXACTFU) || optype==EXACT)
3294 if ( last && TRIE_TYPE_IS_SAFE ) {
3295 make_trie( pRExC_state,
3296 startbranch, first, cur, tail, count,
3299 if ( PL_regkind[ OP( noper ) ] == EXACT
3301 && noper_next == tail
3306 optype = OP( noper );
3316 regprop(RExC_rx, mysv, cur);
3317 PerlIO_printf( Perl_debug_log,
3318 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3319 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3323 if ( last && TRIE_TYPE_IS_SAFE ) {
3324 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
3325 #ifdef TRIE_STUDY_OPT
3326 if ( ((made == MADE_EXACT_TRIE &&
3327 startbranch == first)
3328 || ( first_non_open == first )) &&
3330 flags |= SCF_TRIE_RESTUDY;
3331 if ( startbranch == first
3334 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3344 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3345 scan = NEXTOPER(NEXTOPER(scan));
3346 } else /* single branch is optimized. */
3347 scan = NEXTOPER(scan);
3349 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3350 scan_frame *newframe = NULL;
3355 if (OP(scan) != SUSPEND) {
3356 /* set the pointer */
3357 if (OP(scan) == GOSUB) {
3359 RExC_recurse[ARG2L(scan)] = scan;
3360 start = RExC_open_parens[paren-1];
3361 end = RExC_close_parens[paren-1];
3364 start = RExC_rxi->program + 1;
3368 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3369 SAVEFREEPV(recursed);
3371 if (!PAREN_TEST(recursed,paren+1)) {
3372 PAREN_SET(recursed,paren+1);
3373 Newx(newframe,1,scan_frame);
3375 if (flags & SCF_DO_SUBSTR) {
3376 SCAN_COMMIT(pRExC_state,data,minlenp);
3377 data->longest = &(data->longest_float);
3379 is_inf = is_inf_internal = 1;
3380 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3381 cl_anything(pRExC_state, data->start_class);
3382 flags &= ~SCF_DO_STCLASS;
3385 Newx(newframe,1,scan_frame);
3388 end = regnext(scan);
3393 SAVEFREEPV(newframe);
3394 newframe->next = regnext(scan);
3395 newframe->last = last;
3396 newframe->stop = stopparen;
3397 newframe->prev = frame;
3407 else if (OP(scan) == EXACT) {
3408 I32 l = STR_LEN(scan);
3411 const U8 * const s = (U8*)STRING(scan);
3412 l = utf8_length(s, s + l);
3413 uc = utf8_to_uvchr(s, NULL);
3415 uc = *((U8*)STRING(scan));
3418 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3419 /* The code below prefers earlier match for fixed
3420 offset, later match for variable offset. */
3421 if (data->last_end == -1) { /* Update the start info. */
3422 data->last_start_min = data->pos_min;
3423 data->last_start_max = is_inf
3424 ? I32_MAX : data->pos_min + data->pos_delta;
3426 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3428 SvUTF8_on(data->last_found);
3430 SV * const sv = data->last_found;
3431 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3432 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3433 if (mg && mg->mg_len >= 0)
3434 mg->mg_len += utf8_length((U8*)STRING(scan),
3435 (U8*)STRING(scan)+STR_LEN(scan));
3437 data->last_end = data->pos_min + l;
3438 data->pos_min += l; /* As in the first entry. */
3439 data->flags &= ~SF_BEFORE_EOL;
3441 if (flags & SCF_DO_STCLASS_AND) {
3442 /* Check whether it is compatible with what we know already! */
3446 /* If compatible, we or it in below. It is compatible if is
3447 * in the bitmp and either 1) its bit or its fold is set, or 2)
3448 * it's for a locale. Even if there isn't unicode semantics
3449 * here, at runtime there may be because of matching against a
3450 * utf8 string, so accept a possible false positive for
3451 * latin1-range folds */
3453 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3454 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3455 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
3456 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3461 ANYOF_CLASS_ZERO(data->start_class);
3462 ANYOF_BITMAP_ZERO(data->start_class);
3464 ANYOF_BITMAP_SET(data->start_class, uc);
3465 else if (uc >= 0x100) {
3468 /* Some Unicode code points fold to the Latin1 range; as
3469 * XXX temporary code, instead of figuring out if this is
3470 * one, just assume it is and set all the start class bits
3471 * that could be some such above 255 code point's fold
3472 * which will generate fals positives. As the code
3473 * elsewhere that does compute the fold settles down, it
3474 * can be extracted out and re-used here */
3475 for (i = 0; i < 256; i++){
3476 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3477 ANYOF_BITMAP_SET(data->start_class, i);
3481 data->start_class->flags &= ~ANYOF_EOS;
3483 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3485 else if (flags & SCF_DO_STCLASS_OR) {
3486 /* false positive possible if the class is case-folded */
3488 ANYOF_BITMAP_SET(data->start_class, uc);
3490 data->start_class->flags |= ANYOF_UNICODE_ALL;
3491 data->start_class->flags &= ~ANYOF_EOS;
3492 cl_and(data->start_class, and_withp);
3494 flags &= ~SCF_DO_STCLASS;
3496 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3497 I32 l = STR_LEN(scan);
3498 UV uc = *((U8*)STRING(scan));
3500 /* Search for fixed substrings supports EXACT only. */
3501 if (flags & SCF_DO_SUBSTR) {
3503 SCAN_COMMIT(pRExC_state, data, minlenp);
3506 const U8 * const s = (U8 *)STRING(scan);
3507 l = utf8_length(s, s + l);
3508 uc = utf8_to_uvchr(s, NULL);
3510 else if (has_exactf_sharp_s) {
3511 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3513 min += l - min_subtract;
3517 delta += min_subtract;
3518 if (flags & SCF_DO_SUBSTR) {
3519 data->pos_min += l - min_subtract;
3520 if (data->pos_min < 0) {
3523 data->pos_delta += min_subtract;
3525 data->longest = &(data->longest_float);
3528 if (flags & SCF_DO_STCLASS_AND) {
3529 /* Check whether it is compatible with what we know already! */
3532 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3533 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3534 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3538 ANYOF_CLASS_ZERO(data->start_class);
3539 ANYOF_BITMAP_ZERO(data->start_class);
3541 ANYOF_BITMAP_SET(data->start_class, uc);
3542 data->start_class->flags &= ~ANYOF_EOS;
3543 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
3544 if (OP(scan) == EXACTFL) {
3545 /* XXX This set is probably no longer necessary, and
3546 * probably wrong as LOCALE now is on in the initial
3548 data->start_class->flags |= ANYOF_LOCALE;
3552 /* Also set the other member of the fold pair. In case
3553 * that unicode semantics is called for at runtime, use
3554 * the full latin1 fold. (Can't do this for locale,
3555 * because not known until runtime) */
3556 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3558 /* All other (EXACTFL handled above) folds except under
3559 * /iaa that include s, S, and sharp_s also may include
3561 if (OP(scan) != EXACTFA) {
3562 if (uc == 's' || uc == 'S') {
3563 ANYOF_BITMAP_SET(data->start_class,
3564 LATIN_SMALL_LETTER_SHARP_S);
3566 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3567 ANYOF_BITMAP_SET(data->start_class, 's');
3568 ANYOF_BITMAP_SET(data->start_class, 'S');
3573 else if (uc >= 0x100) {
3575 for (i = 0; i < 256; i++){
3576 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3577 ANYOF_BITMAP_SET(data->start_class, i);
3582 else if (flags & SCF_DO_STCLASS_OR) {
3583 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
3584 /* false positive possible if the class is case-folded.
3585 Assume that the locale settings are the same... */
3587 ANYOF_BITMAP_SET(data->start_class, uc);
3588 if (OP(scan) != EXACTFL) {
3590 /* And set the other member of the fold pair, but
3591 * can't do that in locale because not known until
3593 ANYOF_BITMAP_SET(data->start_class,
3594 PL_fold_latin1[uc]);
3596 /* All folds except under /iaa that include s, S,
3597 * and sharp_s also may include the others */
3598 if (OP(scan) != EXACTFA) {
3599 if (uc == 's' || uc == 'S') {
3600 ANYOF_BITMAP_SET(data->start_class,
3601 LATIN_SMALL_LETTER_SHARP_S);
3603 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3604 ANYOF_BITMAP_SET(data->start_class, 's');
3605 ANYOF_BITMAP_SET(data->start_class, 'S');
3610 data->start_class->flags &= ~ANYOF_EOS;
3612 cl_and(data->start_class, and_withp);
3614 flags &= ~SCF_DO_STCLASS;
3616 else if (REGNODE_VARIES(OP(scan))) {
3617 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3618 I32 f = flags, pos_before = 0;
3619 regnode * const oscan = scan;
3620 struct regnode_charclass_class this_class;
3621 struct regnode_charclass_class *oclass = NULL;
3622 I32 next_is_eval = 0;
3624 switch (PL_regkind[OP(scan)]) {
3625 case WHILEM: /* End of (?:...)* . */
3626 scan = NEXTOPER(scan);
3629 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3630 next = NEXTOPER(scan);
3631 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3633 maxcount = REG_INFTY;
3634 next = regnext(scan);
3635 scan = NEXTOPER(scan);
3639 if (flags & SCF_DO_SUBSTR)
3644 if (flags & SCF_DO_STCLASS) {
3646 maxcount = REG_INFTY;
3647 next = regnext(scan);
3648 scan = NEXTOPER(scan);
3651 is_inf = is_inf_internal = 1;
3652 scan = regnext(scan);
3653 if (flags & SCF_DO_SUBSTR) {
3654 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3655 data->longest = &(data->longest_float);
3657 goto optimize_curly_tail;
3659 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3660 && (scan->flags == stopparen))
3665 mincount = ARG1(scan);
3666 maxcount = ARG2(scan);
3668 next = regnext(scan);
3669 if (OP(scan) == CURLYX) {
3670 I32 lp = (data ? *(data->last_closep) : 0);
3671 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3673 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3674 next_is_eval = (OP(scan) == EVAL);
3676 if (flags & SCF_DO_SUBSTR) {
3677 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3678 pos_before = data->pos_min;
3682 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3684 data->flags |= SF_IS_INF;
3686 if (flags & SCF_DO_STCLASS) {
3687 cl_init(pRExC_state, &this_class);
3688 oclass = data->start_class;
3689 data->start_class = &this_class;
3690 f |= SCF_DO_STCLASS_AND;
3691 f &= ~SCF_DO_STCLASS_OR;
3693 /* Exclude from super-linear cache processing any {n,m}
3694 regops for which the combination of input pos and regex
3695 pos is not enough information to determine if a match
3698 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3699 regex pos at the \s*, the prospects for a match depend not
3700 only on the input position but also on how many (bar\s*)
3701 repeats into the {4,8} we are. */
3702 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3703 f &= ~SCF_WHILEM_VISITED_POS;
3705 /* This will finish on WHILEM, setting scan, or on NULL: */
3706 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3707 last, data, stopparen, recursed, NULL,
3709 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3711 if (flags & SCF_DO_STCLASS)
3712 data->start_class = oclass;
3713 if (mincount == 0 || minnext == 0) {
3714 if (flags & SCF_DO_STCLASS_OR) {
3715 cl_or(pRExC_state, data->start_class, &this_class);
3717 else if (flags & SCF_DO_STCLASS_AND) {
3718 /* Switch to OR mode: cache the old value of
3719 * data->start_class */
3721 StructCopy(data->start_class, and_withp,
3722 struct regnode_charclass_class);
3723 flags &= ~SCF_DO_STCLASS_AND;
3724 StructCopy(&this_class, data->start_class,
3725 struct regnode_charclass_class);
3726 flags |= SCF_DO_STCLASS_OR;
3727 data->start_class->flags |= ANYOF_EOS;
3729 } else { /* Non-zero len */
3730 if (flags & SCF_DO_STCLASS_OR) {
3731 cl_or(pRExC_state, data->start_class, &this_class);
3732 cl_and(data->start_class, and_withp);
3734 else if (flags & SCF_DO_STCLASS_AND)
3735 cl_and(data->start_class, &this_class);
3736 flags &= ~SCF_DO_STCLASS;
3738 if (!scan) /* It was not CURLYX, but CURLY. */
3740 if ( /* ? quantifier ok, except for (?{ ... }) */
3741 (next_is_eval || !(mincount == 0 && maxcount == 1))
3742 && (minnext == 0) && (deltanext == 0)
3743 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3744 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3746 ckWARNreg(RExC_parse,
3747 "Quantifier unexpected on zero-length expression");
3750 min += minnext * mincount;
3751 is_inf_internal |= ((maxcount == REG_INFTY
3752 && (minnext + deltanext) > 0)
3753 || deltanext == I32_MAX);
3754 is_inf |= is_inf_internal;
3755 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3757 /* Try powerful optimization CURLYX => CURLYN. */
3758 if ( OP(oscan) == CURLYX && data
3759 && data->flags & SF_IN_PAR
3760 && !(data->flags & SF_HAS_EVAL)
3761 && !deltanext && minnext == 1 ) {
3762 /* Try to optimize to CURLYN. */
3763 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3764 regnode * const nxt1 = nxt;
3771 if (!REGNODE_SIMPLE(OP(nxt))
3772 && !(PL_regkind[OP(nxt)] == EXACT
3773 && STR_LEN(nxt) == 1))
3779 if (OP(nxt) != CLOSE)
3781 if (RExC_open_parens) {
3782 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3783 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3785 /* Now we know that nxt2 is the only contents: */
3786 oscan->flags = (U8)ARG(nxt);
3788 OP(nxt1) = NOTHING; /* was OPEN. */
3791 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3792 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3793 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3794 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3795 OP(nxt + 1) = OPTIMIZED; /* was count. */
3796 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3801 /* Try optimization CURLYX => CURLYM. */
3802 if ( OP(oscan) == CURLYX && data
3803 && !(data->flags & SF_HAS_PAR)
3804 && !(data->flags & SF_HAS_EVAL)
3805 && !deltanext /* atom is fixed width */
3806 && minnext != 0 /* CURLYM can't handle zero width */
3808 /* XXXX How to optimize if data == 0? */
3809 /* Optimize to a simpler form. */
3810 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3814 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3815 && (OP(nxt2) != WHILEM))
3817 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3818 /* Need to optimize away parenths. */
3819 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
3820 /* Set the parenth number. */
3821 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3823 oscan->flags = (U8)ARG(nxt);
3824 if (RExC_open_parens) {
3825 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3826 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3828 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3829 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3832 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3833 OP(nxt + 1) = OPTIMIZED; /* was count. */
3834 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
3835 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
3838 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3839 regnode *nnxt = regnext(nxt1);
3841 if (reg_off_by_arg[OP(nxt1)])
3842 ARG_SET(nxt1, nxt2 - nxt1);
3843 else if (nxt2 - nxt1 < U16_MAX)
3844 NEXT_OFF(nxt1) = nxt2 - nxt1;
3846 OP(nxt) = NOTHING; /* Cannot beautify */
3851 /* Optimize again: */
3852 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
3853 NULL, stopparen, recursed, NULL, 0,depth+1);
3858 else if ((OP(oscan) == CURLYX)
https://perl5.git.perl.org / perl5.git / blob