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"
95 # if defined(BUGGY_MSC6)
96 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
97 # pragma optimize("a",off)
98 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
99 # pragma optimize("w",on )
100 # endif /* BUGGY_MSC6 */
104 #define STATIC static
107 typedef struct RExC_state_t {
108 U32 flags; /* are we folding, multilining? */
109 char *precomp; /* uncompiled string. */
110 REGEXP *rx_sv; /* The SV that is the regexp. */
111 regexp *rx; /* perl core regexp structure */
112 regexp_internal *rxi; /* internal data for regexp object pprivate field */
113 char *start; /* Start of input for compile */
114 char *end; /* End of input for compile */
115 char *parse; /* Input-scan pointer. */
116 I32 whilem_seen; /* number of WHILEM in this expr */
117 regnode *emit_start; /* Start of emitted-code area */
118 regnode *emit_bound; /* First regnode outside of the allocated space */
119 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
120 I32 naughty; /* How bad is this pattern? */
121 I32 sawback; /* Did we see \1, ...? */
123 I32 size; /* Code size. */
124 I32 npar; /* Capture buffer count, (OPEN). */
125 I32 cpar; /* Capture buffer count, (CLOSE). */
126 I32 nestroot; /* root parens we are in - used by accept */
130 regnode **open_parens; /* pointers to open parens */
131 regnode **close_parens; /* pointers to close parens */
132 regnode *opend; /* END node in program */
133 I32 utf8; /* whether the pattern is utf8 or not */
134 I32 orig_utf8; /* whether the pattern was originally in utf8 */
135 /* XXX use this for future optimisation of case
136 * where pattern must be upgraded to utf8. */
137 I32 uni_semantics; /* If a d charset modifier should use unicode
138 rules, even if the pattern is not in
140 HV *paren_names; /* Paren names */
142 regnode **recurse; /* Recurse regops */
143 I32 recurse_count; /* Number of recurse regops */
146 char *starttry; /* -Dr: where regtry was called. */
147 #define RExC_starttry (pRExC_state->starttry)
150 const char *lastparse;
152 AV *paren_name_list; /* idx -> name */
153 #define RExC_lastparse (pRExC_state->lastparse)
154 #define RExC_lastnum (pRExC_state->lastnum)
155 #define RExC_paren_name_list (pRExC_state->paren_name_list)
159 #define RExC_flags (pRExC_state->flags)
160 #define RExC_precomp (pRExC_state->precomp)
161 #define RExC_rx_sv (pRExC_state->rx_sv)
162 #define RExC_rx (pRExC_state->rx)
163 #define RExC_rxi (pRExC_state->rxi)
164 #define RExC_start (pRExC_state->start)
165 #define RExC_end (pRExC_state->end)
166 #define RExC_parse (pRExC_state->parse)
167 #define RExC_whilem_seen (pRExC_state->whilem_seen)
168 #ifdef RE_TRACK_PATTERN_OFFSETS
169 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
171 #define RExC_emit (pRExC_state->emit)
172 #define RExC_emit_start (pRExC_state->emit_start)
173 #define RExC_emit_bound (pRExC_state->emit_bound)
174 #define RExC_naughty (pRExC_state->naughty)
175 #define RExC_sawback (pRExC_state->sawback)
176 #define RExC_seen (pRExC_state->seen)
177 #define RExC_size (pRExC_state->size)
178 #define RExC_npar (pRExC_state->npar)
179 #define RExC_nestroot (pRExC_state->nestroot)
180 #define RExC_extralen (pRExC_state->extralen)
181 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
182 #define RExC_seen_evals (pRExC_state->seen_evals)
183 #define RExC_utf8 (pRExC_state->utf8)
184 #define RExC_uni_semantics (pRExC_state->uni_semantics)
185 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
186 #define RExC_open_parens (pRExC_state->open_parens)
187 #define RExC_close_parens (pRExC_state->close_parens)
188 #define RExC_opend (pRExC_state->opend)
189 #define RExC_paren_names (pRExC_state->paren_names)
190 #define RExC_recurse (pRExC_state->recurse)
191 #define RExC_recurse_count (pRExC_state->recurse_count)
192 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
195 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
196 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
197 ((*s) == '{' && regcurly(s)))
200 #undef SPSTART /* dratted cpp namespace... */
203 * Flags to be passed up and down.
205 #define WORST 0 /* Worst case. */
206 #define HASWIDTH 0x01 /* Known to match non-null strings. */
208 /* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
209 * character, and if utf8, must be invariant. Note that this is not the same thing as REGNODE_SIMPLE */
211 #define SPSTART 0x04 /* Starts with * or +. */
212 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
213 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
215 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
217 /* whether trie related optimizations are enabled */
218 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
219 #define TRIE_STUDY_OPT
220 #define FULL_TRIE_STUDY
226 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
227 #define PBITVAL(paren) (1 << ((paren) & 7))
228 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
229 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
230 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
232 /* If not already in utf8, do a longjmp back to the beginning */
233 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
234 #define REQUIRE_UTF8 STMT_START { \
235 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
238 /* About scan_data_t.
240 During optimisation we recurse through the regexp program performing
241 various inplace (keyhole style) optimisations. In addition study_chunk
242 and scan_commit populate this data structure with information about
243 what strings MUST appear in the pattern. We look for the longest
244 string that must appear at a fixed location, and we look for the
245 longest string that may appear at a floating location. So for instance
250 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
251 strings (because they follow a .* construct). study_chunk will identify
252 both FOO and BAR as being the longest fixed and floating strings respectively.
254 The strings can be composites, for instance
258 will result in a composite fixed substring 'foo'.
260 For each string some basic information is maintained:
262 - offset or min_offset
263 This is the position the string must appear at, or not before.
264 It also implicitly (when combined with minlenp) tells us how many
265 characters must match before the string we are searching for.
266 Likewise when combined with minlenp and the length of the string it
267 tells us how many characters must appear after the string we have
271 Only used for floating strings. This is the rightmost point that
272 the string can appear at. If set to I32 max it indicates that the
273 string can occur infinitely far to the right.
276 A pointer to the minimum length of the pattern that the string
277 was found inside. This is important as in the case of positive
278 lookahead or positive lookbehind we can have multiple patterns
283 The minimum length of the pattern overall is 3, the minimum length
284 of the lookahead part is 3, but the minimum length of the part that
285 will actually match is 1. So 'FOO's minimum length is 3, but the
286 minimum length for the F is 1. This is important as the minimum length
287 is used to determine offsets in front of and behind the string being
288 looked for. Since strings can be composites this is the length of the
289 pattern at the time it was committed with a scan_commit. Note that
290 the length is calculated by study_chunk, so that the minimum lengths
291 are not known until the full pattern has been compiled, thus the
292 pointer to the value.
296 In the case of lookbehind the string being searched for can be
297 offset past the start point of the final matching string.
298 If this value was just blithely removed from the min_offset it would
299 invalidate some of the calculations for how many chars must match
300 before or after (as they are derived from min_offset and minlen and
301 the length of the string being searched for).
302 When the final pattern is compiled and the data is moved from the
303 scan_data_t structure into the regexp structure the information
304 about lookbehind is factored in, with the information that would
305 have been lost precalculated in the end_shift field for the
308 The fields pos_min and pos_delta are used to store the minimum offset
309 and the delta to the maximum offset at the current point in the pattern.
313 typedef struct scan_data_t {
314 /*I32 len_min; unused */
315 /*I32 len_delta; unused */
319 I32 last_end; /* min value, <0 unless valid. */
322 SV **longest; /* Either &l_fixed, or &l_float. */
323 SV *longest_fixed; /* longest fixed string found in pattern */
324 I32 offset_fixed; /* offset where it starts */
325 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
326 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
327 SV *longest_float; /* longest floating string found in pattern */
328 I32 offset_float_min; /* earliest point in string it can appear */
329 I32 offset_float_max; /* latest point in string it can appear */
330 I32 *minlen_float; /* pointer to the minlen relevant to the string */
331 I32 lookbehind_float; /* is the position of the string modified by LB */
335 struct regnode_charclass_class *start_class;
339 * Forward declarations for pregcomp()'s friends.
342 static const scan_data_t zero_scan_data =
343 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
345 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
346 #define SF_BEFORE_SEOL 0x0001
347 #define SF_BEFORE_MEOL 0x0002
348 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
349 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
352 # define SF_FIX_SHIFT_EOL (0+2)
353 # define SF_FL_SHIFT_EOL (0+4)
355 # define SF_FIX_SHIFT_EOL (+2)
356 # define SF_FL_SHIFT_EOL (+4)
359 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
360 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
362 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
363 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
364 #define SF_IS_INF 0x0040
365 #define SF_HAS_PAR 0x0080
366 #define SF_IN_PAR 0x0100
367 #define SF_HAS_EVAL 0x0200
368 #define SCF_DO_SUBSTR 0x0400
369 #define SCF_DO_STCLASS_AND 0x0800
370 #define SCF_DO_STCLASS_OR 0x1000
371 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
372 #define SCF_WHILEM_VISITED_POS 0x2000
374 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
375 #define SCF_SEEN_ACCEPT 0x8000
377 #define UTF cBOOL(RExC_utf8)
378 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
379 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
380 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
381 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
382 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
384 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
386 #define OOB_UNICODE 12345678
387 #define OOB_NAMEDCLASS -1
389 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
390 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
393 /* length of regex to show in messages that don't mark a position within */
394 #define RegexLengthToShowInErrorMessages 127
397 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
398 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
399 * op/pragma/warn/regcomp.
401 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
402 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
404 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
407 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
408 * arg. Show regex, up to a maximum length. If it's too long, chop and add
411 #define _FAIL(code) STMT_START { \
412 const char *ellipses = ""; \
413 IV len = RExC_end - RExC_precomp; \
416 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
417 if (len > RegexLengthToShowInErrorMessages) { \
418 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
419 len = RegexLengthToShowInErrorMessages - 10; \
425 #define FAIL(msg) _FAIL( \
426 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
427 msg, (int)len, RExC_precomp, ellipses))
429 #define FAIL2(msg,arg) _FAIL( \
430 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
431 arg, (int)len, RExC_precomp, ellipses))
434 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
436 #define Simple_vFAIL(m) STMT_START { \
437 const IV offset = RExC_parse - RExC_precomp; \
438 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
439 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
443 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
445 #define vFAIL(m) STMT_START { \
447 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
452 * Like Simple_vFAIL(), but accepts two arguments.
454 #define Simple_vFAIL2(m,a1) STMT_START { \
455 const IV offset = RExC_parse - RExC_precomp; \
456 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
457 (int)offset, RExC_precomp, RExC_precomp + offset); \
461 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
463 #define vFAIL2(m,a1) STMT_START { \
465 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
466 Simple_vFAIL2(m, a1); \
471 * Like Simple_vFAIL(), but accepts three arguments.
473 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
474 const IV offset = RExC_parse - RExC_precomp; \
475 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
476 (int)offset, RExC_precomp, RExC_precomp + offset); \
480 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
482 #define vFAIL3(m,a1,a2) STMT_START { \
484 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
485 Simple_vFAIL3(m, a1, a2); \
489 * Like Simple_vFAIL(), but accepts four arguments.
491 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
492 const IV offset = RExC_parse - RExC_precomp; \
493 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
494 (int)offset, RExC_precomp, RExC_precomp + offset); \
497 #define ckWARNreg(loc,m) STMT_START { \
498 const IV offset = loc - RExC_precomp; \
499 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
500 (int)offset, RExC_precomp, RExC_precomp + offset); \
503 #define ckWARNregdep(loc,m) STMT_START { \
504 const IV offset = loc - RExC_precomp; \
505 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
507 (int)offset, RExC_precomp, RExC_precomp + offset); \
510 #define ckWARN2reg(loc, m, a1) STMT_START { \
511 const IV offset = loc - RExC_precomp; \
512 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
513 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
516 #define vWARN3(loc, m, a1, a2) STMT_START { \
517 const IV offset = loc - RExC_precomp; \
518 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
519 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
522 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
523 const IV offset = loc - RExC_precomp; \
524 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
525 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
528 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
529 const IV offset = loc - RExC_precomp; \
530 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
531 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
534 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
535 const IV offset = loc - RExC_precomp; \
536 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
537 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
540 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
541 const IV offset = loc - RExC_precomp; \
542 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
543 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
547 /* Allow for side effects in s */
548 #define REGC(c,s) STMT_START { \
549 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
552 /* Macros for recording node offsets. 20001227 mjd@plover.com
553 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
554 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
555 * Element 0 holds the number n.
556 * Position is 1 indexed.
558 #ifndef RE_TRACK_PATTERN_OFFSETS
559 #define Set_Node_Offset_To_R(node,byte)
560 #define Set_Node_Offset(node,byte)
561 #define Set_Cur_Node_Offset
562 #define Set_Node_Length_To_R(node,len)
563 #define Set_Node_Length(node,len)
564 #define Set_Node_Cur_Length(node)
565 #define Node_Offset(n)
566 #define Node_Length(n)
567 #define Set_Node_Offset_Length(node,offset,len)
568 #define ProgLen(ri) ri->u.proglen
569 #define SetProgLen(ri,x) ri->u.proglen = x
571 #define ProgLen(ri) ri->u.offsets[0]
572 #define SetProgLen(ri,x) ri->u.offsets[0] = x
573 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
575 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
576 __LINE__, (int)(node), (int)(byte))); \
578 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
580 RExC_offsets[2*(node)-1] = (byte); \
585 #define Set_Node_Offset(node,byte) \
586 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
587 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
589 #define Set_Node_Length_To_R(node,len) STMT_START { \
591 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
592 __LINE__, (int)(node), (int)(len))); \
594 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
596 RExC_offsets[2*(node)] = (len); \
601 #define Set_Node_Length(node,len) \
602 Set_Node_Length_To_R((node)-RExC_emit_start, len)
603 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
604 #define Set_Node_Cur_Length(node) \
605 Set_Node_Length(node, RExC_parse - parse_start)
607 /* Get offsets and lengths */
608 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
609 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
611 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
612 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
613 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
617 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
618 #define EXPERIMENTAL_INPLACESCAN
619 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
621 #define DEBUG_STUDYDATA(str,data,depth) \
622 DEBUG_OPTIMISE_MORE_r(if(data){ \
623 PerlIO_printf(Perl_debug_log, \
624 "%*s" str "Pos:%"IVdf"/%"IVdf \
625 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
626 (int)(depth)*2, "", \
627 (IV)((data)->pos_min), \
628 (IV)((data)->pos_delta), \
629 (UV)((data)->flags), \
630 (IV)((data)->whilem_c), \
631 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
632 is_inf ? "INF " : "" \
634 if ((data)->last_found) \
635 PerlIO_printf(Perl_debug_log, \
636 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
637 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
638 SvPVX_const((data)->last_found), \
639 (IV)((data)->last_end), \
640 (IV)((data)->last_start_min), \
641 (IV)((data)->last_start_max), \
642 ((data)->longest && \
643 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
644 SvPVX_const((data)->longest_fixed), \
645 (IV)((data)->offset_fixed), \
646 ((data)->longest && \
647 (data)->longest==&((data)->longest_float)) ? "*" : "", \
648 SvPVX_const((data)->longest_float), \
649 (IV)((data)->offset_float_min), \
650 (IV)((data)->offset_float_max) \
652 PerlIO_printf(Perl_debug_log,"\n"); \
655 static void clear_re(pTHX_ void *r);
657 /* Mark that we cannot extend a found fixed substring at this point.
658 Update the longest found anchored substring and the longest found
659 floating substrings if needed. */
662 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
664 const STRLEN l = CHR_SVLEN(data->last_found);
665 const STRLEN old_l = CHR_SVLEN(*data->longest);
666 GET_RE_DEBUG_FLAGS_DECL;
668 PERL_ARGS_ASSERT_SCAN_COMMIT;
670 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
671 SvSetMagicSV(*data->longest, data->last_found);
672 if (*data->longest == data->longest_fixed) {
673 data->offset_fixed = l ? data->last_start_min : data->pos_min;
674 if (data->flags & SF_BEFORE_EOL)
676 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
678 data->flags &= ~SF_FIX_BEFORE_EOL;
679 data->minlen_fixed=minlenp;
680 data->lookbehind_fixed=0;
682 else { /* *data->longest == data->longest_float */
683 data->offset_float_min = l ? data->last_start_min : data->pos_min;
684 data->offset_float_max = (l
685 ? data->last_start_max
686 : data->pos_min + data->pos_delta);
687 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
688 data->offset_float_max = I32_MAX;
689 if (data->flags & SF_BEFORE_EOL)
691 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
693 data->flags &= ~SF_FL_BEFORE_EOL;
694 data->minlen_float=minlenp;
695 data->lookbehind_float=0;
698 SvCUR_set(data->last_found, 0);
700 SV * const sv = data->last_found;
701 if (SvUTF8(sv) && SvMAGICAL(sv)) {
702 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
708 data->flags &= ~SF_BEFORE_EOL;
709 DEBUG_STUDYDATA("commit: ",data,0);
712 /* Can match anything (initialization) */
714 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
716 PERL_ARGS_ASSERT_CL_ANYTHING;
718 ANYOF_CLASS_ZERO(cl);
719 ANYOF_BITMAP_SETALL(cl);
720 cl->flags = ANYOF_EOS|ANYOF_UNICODE_ALL|ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
722 cl->flags |= ANYOF_LOCALE;
725 /* Can match anything (initialization) */
727 S_cl_is_anything(const struct regnode_charclass_class *cl)
731 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
733 for (value = 0; value <= ANYOF_MAX; value += 2)
734 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
736 if (!(cl->flags & ANYOF_UNICODE_ALL))
738 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
743 /* Can match anything (initialization) */
745 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
747 PERL_ARGS_ASSERT_CL_INIT;
749 Zero(cl, 1, struct regnode_charclass_class);
751 cl_anything(pRExC_state, cl);
755 S_cl_init_zero(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
757 PERL_ARGS_ASSERT_CL_INIT_ZERO;
759 Zero(cl, 1, struct regnode_charclass_class);
761 cl_anything(pRExC_state, cl);
763 cl->flags |= ANYOF_LOCALE;
766 /* 'And' a given class with another one. Can create false positives */
767 /* We assume that cl is not inverted */
769 S_cl_and(struct regnode_charclass_class *cl,
770 const struct regnode_charclass_class *and_with)
772 PERL_ARGS_ASSERT_CL_AND;
774 assert(and_with->type == ANYOF);
776 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
777 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
778 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
779 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
780 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
783 if (and_with->flags & ANYOF_INVERT)
784 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
785 cl->bitmap[i] &= ~and_with->bitmap[i];
787 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
788 cl->bitmap[i] &= and_with->bitmap[i];
789 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
790 if (!(and_with->flags & ANYOF_EOS))
791 cl->flags &= ~ANYOF_EOS;
793 if (!(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD))
794 cl->flags &= ~ANYOF_LOC_NONBITMAP_FOLD;
795 if (!(and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL))
796 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
798 if (cl->flags & ANYOF_UNICODE_ALL
799 && and_with->flags & ANYOF_NONBITMAP
800 && !(and_with->flags & ANYOF_INVERT))
802 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
803 cl->flags &= ~ANYOF_UNICODE_ALL;
805 cl->flags |= and_with->flags & ANYOF_NONBITMAP; /* field is 2 bits; use
808 ARG_SET(cl, ARG(and_with));
810 if (!(and_with->flags & ANYOF_UNICODE_ALL) &&
811 !(and_with->flags & ANYOF_INVERT))
812 cl->flags &= ~ANYOF_UNICODE_ALL;
813 if (!(and_with->flags & (ANYOF_NONBITMAP|ANYOF_UNICODE_ALL)) &&
814 !(and_with->flags & ANYOF_INVERT))
815 cl->flags &= ~ANYOF_NONBITMAP;
818 /* 'OR' a given class with another one. Can create false positives */
819 /* We assume that cl is not inverted */
821 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
823 PERL_ARGS_ASSERT_CL_OR;
825 if (or_with->flags & ANYOF_INVERT) {
827 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
828 * <= (B1 | !B2) | (CL1 | !CL2)
829 * which is wasteful if CL2 is small, but we ignore CL2:
830 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
831 * XXXX Can we handle case-fold? Unclear:
832 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
833 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
835 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
836 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
837 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
840 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
841 cl->bitmap[i] |= ~or_with->bitmap[i];
842 } /* XXXX: logic is complicated otherwise */
844 cl_anything(pRExC_state, cl);
847 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
848 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
849 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
850 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
853 /* OR char bitmap and class bitmap separately */
854 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
855 cl->bitmap[i] |= or_with->bitmap[i];
856 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
857 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
858 cl->classflags[i] |= or_with->classflags[i];
859 cl->flags |= ANYOF_CLASS;
862 else { /* XXXX: logic is complicated, leave it along for a moment. */
863 cl_anything(pRExC_state, cl);
866 if (or_with->flags & ANYOF_EOS)
867 cl->flags |= ANYOF_EOS;
868 if (!(or_with->flags & ANYOF_NON_UTF8_LATIN1_ALL))
869 cl->flags |= ANYOF_NON_UTF8_LATIN1_ALL;
871 if (or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
872 cl->flags |= ANYOF_LOC_NONBITMAP_FOLD;
874 /* If both nodes match something outside the bitmap, but what they match
875 * outside is not the same pointer, and hence not easily compared, give up
876 * and allow the start class to match everything outside the bitmap */
877 if (cl->flags & ANYOF_NONBITMAP && or_with->flags & ANYOF_NONBITMAP &&
878 ARG(cl) != ARG(or_with)) {
879 cl->flags |= ANYOF_UNICODE_ALL;
882 if (or_with->flags & ANYOF_UNICODE_ALL) {
883 cl->flags |= ANYOF_UNICODE_ALL;
887 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
888 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
889 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
890 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
895 dump_trie(trie,widecharmap,revcharmap)
896 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
897 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
899 These routines dump out a trie in a somewhat readable format.
900 The _interim_ variants are used for debugging the interim
901 tables that are used to generate the final compressed
902 representation which is what dump_trie expects.
904 Part of the reason for their existence is to provide a form
905 of documentation as to how the different representations function.
910 Dumps the final compressed table form of the trie to Perl_debug_log.
911 Used for debugging make_trie().
915 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
916 AV *revcharmap, U32 depth)
919 SV *sv=sv_newmortal();
920 int colwidth= widecharmap ? 6 : 4;
922 GET_RE_DEBUG_FLAGS_DECL;
924 PERL_ARGS_ASSERT_DUMP_TRIE;
926 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
927 (int)depth * 2 + 2,"",
928 "Match","Base","Ofs" );
930 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
931 SV ** const tmp = av_fetch( revcharmap, state, 0);
933 PerlIO_printf( Perl_debug_log, "%*s",
935 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
936 PL_colors[0], PL_colors[1],
937 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
938 PERL_PV_ESCAPE_FIRSTCHAR
943 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
944 (int)depth * 2 + 2,"");
946 for( state = 0 ; state < trie->uniquecharcount ; state++ )
947 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
948 PerlIO_printf( Perl_debug_log, "\n");
950 for( state = 1 ; state < trie->statecount ; state++ ) {
951 const U32 base = trie->states[ state ].trans.base;
953 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
955 if ( trie->states[ state ].wordnum ) {
956 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
958 PerlIO_printf( Perl_debug_log, "%6s", "" );
961 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
966 while( ( base + ofs < trie->uniquecharcount ) ||
967 ( base + ofs - trie->uniquecharcount < trie->lasttrans
968 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
971 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
973 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
974 if ( ( base + ofs >= trie->uniquecharcount ) &&
975 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
976 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
978 PerlIO_printf( Perl_debug_log, "%*"UVXf,
980 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
982 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
986 PerlIO_printf( Perl_debug_log, "]");
989 PerlIO_printf( Perl_debug_log, "\n" );
991 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
992 for (word=1; word <= trie->wordcount; word++) {
993 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
994 (int)word, (int)(trie->wordinfo[word].prev),
995 (int)(trie->wordinfo[word].len));
997 PerlIO_printf(Perl_debug_log, "\n" );
1000 Dumps a fully constructed but uncompressed trie in list form.
1001 List tries normally only are used for construction when the number of
1002 possible chars (trie->uniquecharcount) is very high.
1003 Used for debugging make_trie().
1006 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1007 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1011 SV *sv=sv_newmortal();
1012 int colwidth= widecharmap ? 6 : 4;
1013 GET_RE_DEBUG_FLAGS_DECL;
1015 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1017 /* print out the table precompression. */
1018 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1019 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1020 "------:-----+-----------------\n" );
1022 for( state=1 ; state < next_alloc ; state ++ ) {
1025 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1026 (int)depth * 2 + 2,"", (UV)state );
1027 if ( ! trie->states[ state ].wordnum ) {
1028 PerlIO_printf( Perl_debug_log, "%5s| ","");
1030 PerlIO_printf( Perl_debug_log, "W%4x| ",
1031 trie->states[ state ].wordnum
1034 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1035 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1037 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1039 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1040 PL_colors[0], PL_colors[1],
1041 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1042 PERL_PV_ESCAPE_FIRSTCHAR
1044 TRIE_LIST_ITEM(state,charid).forid,
1045 (UV)TRIE_LIST_ITEM(state,charid).newstate
1048 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1049 (int)((depth * 2) + 14), "");
1052 PerlIO_printf( Perl_debug_log, "\n");
1057 Dumps a fully constructed but uncompressed trie in table form.
1058 This is the normal DFA style state transition table, with a few
1059 twists to facilitate compression later.
1060 Used for debugging make_trie().
1063 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1064 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1069 SV *sv=sv_newmortal();
1070 int colwidth= widecharmap ? 6 : 4;
1071 GET_RE_DEBUG_FLAGS_DECL;
1073 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1076 print out the table precompression so that we can do a visual check
1077 that they are identical.
1080 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1082 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1083 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1085 PerlIO_printf( Perl_debug_log, "%*s",
1087 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1088 PL_colors[0], PL_colors[1],
1089 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1090 PERL_PV_ESCAPE_FIRSTCHAR
1096 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1098 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1099 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1102 PerlIO_printf( Perl_debug_log, "\n" );
1104 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1106 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1107 (int)depth * 2 + 2,"",
1108 (UV)TRIE_NODENUM( state ) );
1110 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1111 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1113 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1115 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1117 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1118 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1120 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1121 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1129 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1130 startbranch: the first branch in the whole branch sequence
1131 first : start branch of sequence of branch-exact nodes.
1132 May be the same as startbranch
1133 last : Thing following the last branch.
1134 May be the same as tail.
1135 tail : item following the branch sequence
1136 count : words in the sequence
1137 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1138 depth : indent depth
1140 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1142 A trie is an N'ary tree where the branches are determined by digital
1143 decomposition of the key. IE, at the root node you look up the 1st character and
1144 follow that branch repeat until you find the end of the branches. Nodes can be
1145 marked as "accepting" meaning they represent a complete word. Eg:
1149 would convert into the following structure. Numbers represent states, letters
1150 following numbers represent valid transitions on the letter from that state, if
1151 the number is in square brackets it represents an accepting state, otherwise it
1152 will be in parenthesis.
1154 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1158 (1) +-i->(6)-+-s->[7]
1160 +-s->(3)-+-h->(4)-+-e->[5]
1162 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1164 This shows that when matching against the string 'hers' we will begin at state 1
1165 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1166 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1167 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1168 single traverse. We store a mapping from accepting to state to which word was
1169 matched, and then when we have multiple possibilities we try to complete the
1170 rest of the regex in the order in which they occured in the alternation.
1172 The only prior NFA like behaviour that would be changed by the TRIE support is
1173 the silent ignoring of duplicate alternations which are of the form:
1175 / (DUPE|DUPE) X? (?{ ... }) Y /x
1177 Thus EVAL blocks following a trie may be called a different number of times with
1178 and without the optimisation. With the optimisations dupes will be silently
1179 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1180 the following demonstrates:
1182 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1184 which prints out 'word' three times, but
1186 'words'=~/(word|word|word)(?{ print $1 })S/
1188 which doesnt print it out at all. This is due to other optimisations kicking in.
1190 Example of what happens on a structural level:
1192 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1194 1: CURLYM[1] {1,32767}(18)
1205 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1206 and should turn into:
1208 1: CURLYM[1] {1,32767}(18)
1210 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1218 Cases where tail != last would be like /(?foo|bar)baz/:
1228 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1229 and would end up looking like:
1232 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1239 d = uvuni_to_utf8_flags(d, uv, 0);
1241 is the recommended Unicode-aware way of saying
1246 #define TRIE_STORE_REVCHAR \
1249 SV *zlopp = newSV(2); \
1250 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1251 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, uvc & 0xFF); \
1252 SvCUR_set(zlopp, kapow - flrbbbbb); \
1255 av_push(revcharmap, zlopp); \
1257 char ooooff = (char)uvc; \
1258 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1262 #define TRIE_READ_CHAR STMT_START { \
1266 if ( foldlen > 0 ) { \
1267 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1272 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1273 uvc = to_uni_fold( uvc, foldbuf, &foldlen ); \
1274 foldlen -= UNISKIP( uvc ); \
1275 scan = foldbuf + UNISKIP( uvc ); \
1278 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1288 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1289 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1290 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1291 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1293 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1294 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1295 TRIE_LIST_CUR( state )++; \
1298 #define TRIE_LIST_NEW(state) STMT_START { \
1299 Newxz( trie->states[ state ].trans.list, \
1300 4, reg_trie_trans_le ); \
1301 TRIE_LIST_CUR( state ) = 1; \
1302 TRIE_LIST_LEN( state ) = 4; \
1305 #define TRIE_HANDLE_WORD(state) STMT_START { \
1306 U16 dupe= trie->states[ state ].wordnum; \
1307 regnode * const noper_next = regnext( noper ); \
1310 /* store the word for dumping */ \
1312 if (OP(noper) != NOTHING) \
1313 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1315 tmp = newSVpvn_utf8( "", 0, UTF ); \
1316 av_push( trie_words, tmp ); \
1320 trie->wordinfo[curword].prev = 0; \
1321 trie->wordinfo[curword].len = wordlen; \
1322 trie->wordinfo[curword].accept = state; \
1324 if ( noper_next < tail ) { \
1326 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1327 trie->jump[curword] = (U16)(noper_next - convert); \
1329 jumper = noper_next; \
1331 nextbranch= regnext(cur); \
1335 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1336 /* chain, so that when the bits of chain are later */\
1337 /* linked together, the dups appear in the chain */\
1338 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1339 trie->wordinfo[dupe].prev = curword; \
1341 /* we haven't inserted this word yet. */ \
1342 trie->states[ state ].wordnum = curword; \
1347 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1348 ( ( base + charid >= ucharcount \
1349 && base + charid < ubound \
1350 && state == trie->trans[ base - ucharcount + charid ].check \
1351 && trie->trans[ base - ucharcount + charid ].next ) \
1352 ? trie->trans[ base - ucharcount + charid ].next \
1353 : ( state==1 ? special : 0 ) \
1357 #define MADE_JUMP_TRIE 2
1358 #define MADE_EXACT_TRIE 4
1361 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1364 /* first pass, loop through and scan words */
1365 reg_trie_data *trie;
1366 HV *widecharmap = NULL;
1367 AV *revcharmap = newAV();
1369 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1374 regnode *jumper = NULL;
1375 regnode *nextbranch = NULL;
1376 regnode *convert = NULL;
1377 U32 *prev_states; /* temp array mapping each state to previous one */
1378 /* we just use folder as a flag in utf8 */
1379 const U8 * folder = NULL;
1382 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1383 AV *trie_words = NULL;
1384 /* along with revcharmap, this only used during construction but both are
1385 * useful during debugging so we store them in the struct when debugging.
1388 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1389 STRLEN trie_charcount=0;
1391 SV *re_trie_maxbuff;
1392 GET_RE_DEBUG_FLAGS_DECL;
1394 PERL_ARGS_ASSERT_MAKE_TRIE;
1396 PERL_UNUSED_ARG(depth);
1400 case EXACTFU: folder = PL_fold_latin1; break;
1401 case EXACTF: folder = PL_fold; break;
1402 case EXACTFL: folder = PL_fold_locale; break;
1405 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1407 trie->startstate = 1;
1408 trie->wordcount = word_count;
1409 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1410 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1411 if (!(UTF && folder))
1412 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1413 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1414 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1417 trie_words = newAV();
1420 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1421 if (!SvIOK(re_trie_maxbuff)) {
1422 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1425 PerlIO_printf( Perl_debug_log,
1426 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1427 (int)depth * 2 + 2, "",
1428 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1429 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1433 /* Find the node we are going to overwrite */
1434 if ( first == startbranch && OP( last ) != BRANCH ) {
1435 /* whole branch chain */
1438 /* branch sub-chain */
1439 convert = NEXTOPER( first );
1442 /* -- First loop and Setup --
1444 We first traverse the branches and scan each word to determine if it
1445 contains widechars, and how many unique chars there are, this is
1446 important as we have to build a table with at least as many columns as we
1449 We use an array of integers to represent the character codes 0..255
1450 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1451 native representation of the character value as the key and IV's for the
1454 *TODO* If we keep track of how many times each character is used we can
1455 remap the columns so that the table compression later on is more
1456 efficient in terms of memory by ensuring the most common value is in the
1457 middle and the least common are on the outside. IMO this would be better
1458 than a most to least common mapping as theres a decent chance the most
1459 common letter will share a node with the least common, meaning the node
1460 will not be compressible. With a middle is most common approach the worst
1461 case is when we have the least common nodes twice.
1465 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1466 regnode * const noper = NEXTOPER( cur );
1467 const U8 *uc = (U8*)STRING( noper );
1468 const U8 * const e = uc + STR_LEN( noper );
1470 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1471 const U8 *scan = (U8*)NULL;
1472 U32 wordlen = 0; /* required init */
1474 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1476 if (OP(noper) == NOTHING) {
1480 if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
1481 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1482 regardless of encoding */
1484 for ( ; uc < e ; uc += len ) {
1485 TRIE_CHARCOUNT(trie)++;
1489 if ( !trie->charmap[ uvc ] ) {
1490 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1492 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
1496 /* store the codepoint in the bitmap, and its folded
1498 TRIE_BITMAP_SET(trie,uvc);
1500 /* store the folded codepoint */
1501 if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);
1504 /* store first byte of utf8 representation of
1505 variant codepoints */
1506 if (! UNI_IS_INVARIANT(uvc)) {
1507 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1510 set_bit = 0; /* We've done our bit :-) */
1515 widecharmap = newHV();
1517 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1520 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1522 if ( !SvTRUE( *svpp ) ) {
1523 sv_setiv( *svpp, ++trie->uniquecharcount );
1528 if( cur == first ) {
1531 } else if (chars < trie->minlen) {
1533 } else if (chars > trie->maxlen) {
1537 } /* end first pass */
1538 DEBUG_TRIE_COMPILE_r(
1539 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1540 (int)depth * 2 + 2,"",
1541 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1542 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1543 (int)trie->minlen, (int)trie->maxlen )
1547 We now know what we are dealing with in terms of unique chars and
1548 string sizes so we can calculate how much memory a naive
1549 representation using a flat table will take. If it's over a reasonable
1550 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1551 conservative but potentially much slower representation using an array
1554 At the end we convert both representations into the same compressed
1555 form that will be used in regexec.c for matching with. The latter
1556 is a form that cannot be used to construct with but has memory
1557 properties similar to the list form and access properties similar
1558 to the table form making it both suitable for fast searches and
1559 small enough that its feasable to store for the duration of a program.
1561 See the comment in the code where the compressed table is produced
1562 inplace from the flat tabe representation for an explanation of how
1563 the compression works.
1568 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1571 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1573 Second Pass -- Array Of Lists Representation
1575 Each state will be represented by a list of charid:state records
1576 (reg_trie_trans_le) the first such element holds the CUR and LEN
1577 points of the allocated array. (See defines above).
1579 We build the initial structure using the lists, and then convert
1580 it into the compressed table form which allows faster lookups
1581 (but cant be modified once converted).
1584 STRLEN transcount = 1;
1586 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1587 "%*sCompiling trie using list compiler\n",
1588 (int)depth * 2 + 2, ""));
1590 trie->states = (reg_trie_state *)
1591 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1592 sizeof(reg_trie_state) );
1596 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1598 regnode * const noper = NEXTOPER( cur );
1599 U8 *uc = (U8*)STRING( noper );
1600 const U8 * const e = uc + STR_LEN( noper );
1601 U32 state = 1; /* required init */
1602 U16 charid = 0; /* sanity init */
1603 U8 *scan = (U8*)NULL; /* sanity init */
1604 STRLEN foldlen = 0; /* required init */
1605 U32 wordlen = 0; /* required init */
1606 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1608 if (OP(noper) != NOTHING) {
1609 for ( ; uc < e ; uc += len ) {
1614 charid = trie->charmap[ uvc ];
1616 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1620 charid=(U16)SvIV( *svpp );
1623 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1630 if ( !trie->states[ state ].trans.list ) {
1631 TRIE_LIST_NEW( state );
1633 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1634 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1635 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1640 newstate = next_alloc++;
1641 prev_states[newstate] = state;
1642 TRIE_LIST_PUSH( state, charid, newstate );
1647 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1651 TRIE_HANDLE_WORD(state);
1653 } /* end second pass */
1655 /* next alloc is the NEXT state to be allocated */
1656 trie->statecount = next_alloc;
1657 trie->states = (reg_trie_state *)
1658 PerlMemShared_realloc( trie->states,
1660 * sizeof(reg_trie_state) );
1662 /* and now dump it out before we compress it */
1663 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1664 revcharmap, next_alloc,
1668 trie->trans = (reg_trie_trans *)
1669 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1676 for( state=1 ; state < next_alloc ; state ++ ) {
1680 DEBUG_TRIE_COMPILE_MORE_r(
1681 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1685 if (trie->states[state].trans.list) {
1686 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1690 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1691 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1692 if ( forid < minid ) {
1694 } else if ( forid > maxid ) {
1698 if ( transcount < tp + maxid - minid + 1) {
1700 trie->trans = (reg_trie_trans *)
1701 PerlMemShared_realloc( trie->trans,
1703 * sizeof(reg_trie_trans) );
1704 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1706 base = trie->uniquecharcount + tp - minid;
1707 if ( maxid == minid ) {
1709 for ( ; zp < tp ; zp++ ) {
1710 if ( ! trie->trans[ zp ].next ) {
1711 base = trie->uniquecharcount + zp - minid;
1712 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1713 trie->trans[ zp ].check = state;
1719 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1720 trie->trans[ tp ].check = state;
1725 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1726 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1727 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1728 trie->trans[ tid ].check = state;
1730 tp += ( maxid - minid + 1 );
1732 Safefree(trie->states[ state ].trans.list);
1735 DEBUG_TRIE_COMPILE_MORE_r(
1736 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1739 trie->states[ state ].trans.base=base;
1741 trie->lasttrans = tp + 1;
1745 Second Pass -- Flat Table Representation.
1747 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1748 We know that we will need Charcount+1 trans at most to store the data
1749 (one row per char at worst case) So we preallocate both structures
1750 assuming worst case.
1752 We then construct the trie using only the .next slots of the entry
1755 We use the .check field of the first entry of the node temporarily to
1756 make compression both faster and easier by keeping track of how many non
1757 zero fields are in the node.
1759 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1762 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1763 number representing the first entry of the node, and state as a
1764 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1765 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1766 are 2 entrys per node. eg:
1774 The table is internally in the right hand, idx form. However as we also
1775 have to deal with the states array which is indexed by nodenum we have to
1776 use TRIE_NODENUM() to convert.
1779 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1780 "%*sCompiling trie using table compiler\n",
1781 (int)depth * 2 + 2, ""));
1783 trie->trans = (reg_trie_trans *)
1784 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1785 * trie->uniquecharcount + 1,
1786 sizeof(reg_trie_trans) );
1787 trie->states = (reg_trie_state *)
1788 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1789 sizeof(reg_trie_state) );
1790 next_alloc = trie->uniquecharcount + 1;
1793 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1795 regnode * const noper = NEXTOPER( cur );
1796 const U8 *uc = (U8*)STRING( noper );
1797 const U8 * const e = uc + STR_LEN( noper );
1799 U32 state = 1; /* required init */
1801 U16 charid = 0; /* sanity init */
1802 U32 accept_state = 0; /* sanity init */
1803 U8 *scan = (U8*)NULL; /* sanity init */
1805 STRLEN foldlen = 0; /* required init */
1806 U32 wordlen = 0; /* required init */
1807 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1809 if ( OP(noper) != NOTHING ) {
1810 for ( ; uc < e ; uc += len ) {
1815 charid = trie->charmap[ uvc ];
1817 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1818 charid = svpp ? (U16)SvIV(*svpp) : 0;
1822 if ( !trie->trans[ state + charid ].next ) {
1823 trie->trans[ state + charid ].next = next_alloc;
1824 trie->trans[ state ].check++;
1825 prev_states[TRIE_NODENUM(next_alloc)]
1826 = TRIE_NODENUM(state);
1827 next_alloc += trie->uniquecharcount;
1829 state = trie->trans[ state + charid ].next;
1831 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1833 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1836 accept_state = TRIE_NODENUM( state );
1837 TRIE_HANDLE_WORD(accept_state);
1839 } /* end second pass */
1841 /* and now dump it out before we compress it */
1842 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1844 next_alloc, depth+1));
1848 * Inplace compress the table.*
1850 For sparse data sets the table constructed by the trie algorithm will
1851 be mostly 0/FAIL transitions or to put it another way mostly empty.
1852 (Note that leaf nodes will not contain any transitions.)
1854 This algorithm compresses the tables by eliminating most such
1855 transitions, at the cost of a modest bit of extra work during lookup:
1857 - Each states[] entry contains a .base field which indicates the
1858 index in the state[] array wheres its transition data is stored.
1860 - If .base is 0 there are no valid transitions from that node.
1862 - If .base is nonzero then charid is added to it to find an entry in
1865 -If trans[states[state].base+charid].check!=state then the
1866 transition is taken to be a 0/Fail transition. Thus if there are fail
1867 transitions at the front of the node then the .base offset will point
1868 somewhere inside the previous nodes data (or maybe even into a node
1869 even earlier), but the .check field determines if the transition is
1873 The following process inplace converts the table to the compressed
1874 table: We first do not compress the root node 1,and mark all its
1875 .check pointers as 1 and set its .base pointer as 1 as well. This
1876 allows us to do a DFA construction from the compressed table later,
1877 and ensures that any .base pointers we calculate later are greater
1880 - We set 'pos' to indicate the first entry of the second node.
1882 - We then iterate over the columns of the node, finding the first and
1883 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
1884 and set the .check pointers accordingly, and advance pos
1885 appropriately and repreat for the next node. Note that when we copy
1886 the next pointers we have to convert them from the original
1887 NODEIDX form to NODENUM form as the former is not valid post
1890 - If a node has no transitions used we mark its base as 0 and do not
1891 advance the pos pointer.
1893 - If a node only has one transition we use a second pointer into the
1894 structure to fill in allocated fail transitions from other states.
1895 This pointer is independent of the main pointer and scans forward
1896 looking for null transitions that are allocated to a state. When it
1897 finds one it writes the single transition into the "hole". If the
1898 pointer doesnt find one the single transition is appended as normal.
1900 - Once compressed we can Renew/realloc the structures to release the
1903 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
1904 specifically Fig 3.47 and the associated pseudocode.
1908 const U32 laststate = TRIE_NODENUM( next_alloc );
1911 trie->statecount = laststate;
1913 for ( state = 1 ; state < laststate ; state++ ) {
1915 const U32 stateidx = TRIE_NODEIDX( state );
1916 const U32 o_used = trie->trans[ stateidx ].check;
1917 U32 used = trie->trans[ stateidx ].check;
1918 trie->trans[ stateidx ].check = 0;
1920 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
1921 if ( flag || trie->trans[ stateidx + charid ].next ) {
1922 if ( trie->trans[ stateidx + charid ].next ) {
1924 for ( ; zp < pos ; zp++ ) {
1925 if ( ! trie->trans[ zp ].next ) {
1929 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
1930 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1931 trie->trans[ zp ].check = state;
1932 if ( ++zp > pos ) pos = zp;
1939 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
1941 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
1942 trie->trans[ pos ].check = state;
1947 trie->lasttrans = pos + 1;
1948 trie->states = (reg_trie_state *)
1949 PerlMemShared_realloc( trie->states, laststate
1950 * sizeof(reg_trie_state) );
1951 DEBUG_TRIE_COMPILE_MORE_r(
1952 PerlIO_printf( Perl_debug_log,
1953 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
1954 (int)depth * 2 + 2,"",
1955 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
1958 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
1961 } /* end table compress */
1963 DEBUG_TRIE_COMPILE_MORE_r(
1964 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
1965 (int)depth * 2 + 2, "",
1966 (UV)trie->statecount,
1967 (UV)trie->lasttrans)
1969 /* resize the trans array to remove unused space */
1970 trie->trans = (reg_trie_trans *)
1971 PerlMemShared_realloc( trie->trans, trie->lasttrans
1972 * sizeof(reg_trie_trans) );
1974 { /* Modify the program and insert the new TRIE node */
1975 U8 nodetype =(U8)(flags & 0xFF);
1979 regnode *optimize = NULL;
1980 #ifdef RE_TRACK_PATTERN_OFFSETS
1983 U32 mjd_nodelen = 0;
1984 #endif /* RE_TRACK_PATTERN_OFFSETS */
1985 #endif /* DEBUGGING */
1987 This means we convert either the first branch or the first Exact,
1988 depending on whether the thing following (in 'last') is a branch
1989 or not and whther first is the startbranch (ie is it a sub part of
1990 the alternation or is it the whole thing.)
1991 Assuming its a sub part we convert the EXACT otherwise we convert
1992 the whole branch sequence, including the first.
1994 /* Find the node we are going to overwrite */
1995 if ( first != startbranch || OP( last ) == BRANCH ) {
1996 /* branch sub-chain */
1997 NEXT_OFF( first ) = (U16)(last - first);
1998 #ifdef RE_TRACK_PATTERN_OFFSETS
2000 mjd_offset= Node_Offset((convert));
2001 mjd_nodelen= Node_Length((convert));
2004 /* whole branch chain */
2006 #ifdef RE_TRACK_PATTERN_OFFSETS
2009 const regnode *nop = NEXTOPER( convert );
2010 mjd_offset= Node_Offset((nop));
2011 mjd_nodelen= Node_Length((nop));
2015 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2016 (int)depth * 2 + 2, "",
2017 (UV)mjd_offset, (UV)mjd_nodelen)
2020 /* But first we check to see if there is a common prefix we can
2021 split out as an EXACT and put in front of the TRIE node. */
2022 trie->startstate= 1;
2023 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2025 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2029 const U32 base = trie->states[ state ].trans.base;
2031 if ( trie->states[state].wordnum )
2034 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2035 if ( ( base + ofs >= trie->uniquecharcount ) &&
2036 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2037 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2039 if ( ++count > 1 ) {
2040 SV **tmp = av_fetch( revcharmap, ofs, 0);
2041 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2042 if ( state == 1 ) break;
2044 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2046 PerlIO_printf(Perl_debug_log,
2047 "%*sNew Start State=%"UVuf" Class: [",
2048 (int)depth * 2 + 2, "",
2051 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2052 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2054 TRIE_BITMAP_SET(trie,*ch);
2056 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2058 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2062 TRIE_BITMAP_SET(trie,*ch);
2064 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2065 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2071 SV **tmp = av_fetch( revcharmap, idx, 0);
2073 char *ch = SvPV( *tmp, len );
2075 SV *sv=sv_newmortal();
2076 PerlIO_printf( Perl_debug_log,
2077 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2078 (int)depth * 2 + 2, "",
2080 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2081 PL_colors[0], PL_colors[1],
2082 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2083 PERL_PV_ESCAPE_FIRSTCHAR
2088 OP( convert ) = nodetype;
2089 str=STRING(convert);
2092 STR_LEN(convert) += len;
2098 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2103 trie->prefixlen = (state-1);
2105 regnode *n = convert+NODE_SZ_STR(convert);
2106 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2107 trie->startstate = state;
2108 trie->minlen -= (state - 1);
2109 trie->maxlen -= (state - 1);
2111 /* At least the UNICOS C compiler choked on this
2112 * being argument to DEBUG_r(), so let's just have
2115 #ifdef PERL_EXT_RE_BUILD
2121 regnode *fix = convert;
2122 U32 word = trie->wordcount;
2124 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2125 while( ++fix < n ) {
2126 Set_Node_Offset_Length(fix, 0, 0);
2129 SV ** const tmp = av_fetch( trie_words, word, 0 );
2131 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2132 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2134 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2142 NEXT_OFF(convert) = (U16)(tail - convert);
2143 DEBUG_r(optimize= n);
2149 if ( trie->maxlen ) {
2150 NEXT_OFF( convert ) = (U16)(tail - convert);
2151 ARG_SET( convert, data_slot );
2152 /* Store the offset to the first unabsorbed branch in
2153 jump[0], which is otherwise unused by the jump logic.
2154 We use this when dumping a trie and during optimisation. */
2156 trie->jump[0] = (U16)(nextbranch - convert);
2158 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2159 * and there is a bitmap
2160 * and the first "jump target" node we found leaves enough room
2161 * then convert the TRIE node into a TRIEC node, with the bitmap
2162 * embedded inline in the opcode - this is hypothetically faster.
2164 if ( !trie->states[trie->startstate].wordnum
2166 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2168 OP( convert ) = TRIEC;
2169 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2170 PerlMemShared_free(trie->bitmap);
2173 OP( convert ) = TRIE;
2175 /* store the type in the flags */
2176 convert->flags = nodetype;
2180 + regarglen[ OP( convert ) ];
2182 /* XXX We really should free up the resource in trie now,
2183 as we won't use them - (which resources?) dmq */
2185 /* needed for dumping*/
2186 DEBUG_r(if (optimize) {
2187 regnode *opt = convert;
2189 while ( ++opt < optimize) {
2190 Set_Node_Offset_Length(opt,0,0);
2193 Try to clean up some of the debris left after the
2196 while( optimize < jumper ) {
2197 mjd_nodelen += Node_Length((optimize));
2198 OP( optimize ) = OPTIMIZED;
2199 Set_Node_Offset_Length(optimize,0,0);
2202 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2204 } /* end node insert */
2206 /* Finish populating the prev field of the wordinfo array. Walk back
2207 * from each accept state until we find another accept state, and if
2208 * so, point the first word's .prev field at the second word. If the
2209 * second already has a .prev field set, stop now. This will be the
2210 * case either if we've already processed that word's accept state,
2211 * or that state had multiple words, and the overspill words were
2212 * already linked up earlier.
2219 for (word=1; word <= trie->wordcount; word++) {
2221 if (trie->wordinfo[word].prev)
2223 state = trie->wordinfo[word].accept;
2225 state = prev_states[state];
2228 prev = trie->states[state].wordnum;
2232 trie->wordinfo[word].prev = prev;
2234 Safefree(prev_states);
2238 /* and now dump out the compressed format */
2239 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2241 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2243 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2244 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2246 SvREFCNT_dec(revcharmap);
2250 : trie->startstate>1
2256 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2258 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2260 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2261 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2264 We find the fail state for each state in the trie, this state is the longest proper
2265 suffix of the current state's 'word' that is also a proper prefix of another word in our
2266 trie. State 1 represents the word '' and is thus the default fail state. This allows
2267 the DFA not to have to restart after its tried and failed a word at a given point, it
2268 simply continues as though it had been matching the other word in the first place.
2270 'abcdgu'=~/abcdefg|cdgu/
2271 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2272 fail, which would bring us to the state representing 'd' in the second word where we would
2273 try 'g' and succeed, proceeding to match 'cdgu'.
2275 /* add a fail transition */
2276 const U32 trie_offset = ARG(source);
2277 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2279 const U32 ucharcount = trie->uniquecharcount;
2280 const U32 numstates = trie->statecount;
2281 const U32 ubound = trie->lasttrans + ucharcount;
2285 U32 base = trie->states[ 1 ].trans.base;
2288 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2289 GET_RE_DEBUG_FLAGS_DECL;
2291 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2293 PERL_UNUSED_ARG(depth);
2297 ARG_SET( stclass, data_slot );
2298 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2299 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2300 aho->trie=trie_offset;
2301 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2302 Copy( trie->states, aho->states, numstates, reg_trie_state );
2303 Newxz( q, numstates, U32);
2304 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2307 /* initialize fail[0..1] to be 1 so that we always have
2308 a valid final fail state */
2309 fail[ 0 ] = fail[ 1 ] = 1;
2311 for ( charid = 0; charid < ucharcount ; charid++ ) {
2312 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2314 q[ q_write ] = newstate;
2315 /* set to point at the root */
2316 fail[ q[ q_write++ ] ]=1;
2319 while ( q_read < q_write) {
2320 const U32 cur = q[ q_read++ % numstates ];
2321 base = trie->states[ cur ].trans.base;
2323 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2324 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2326 U32 fail_state = cur;
2329 fail_state = fail[ fail_state ];
2330 fail_base = aho->states[ fail_state ].trans.base;
2331 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2333 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2334 fail[ ch_state ] = fail_state;
2335 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2337 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2339 q[ q_write++ % numstates] = ch_state;
2343 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2344 when we fail in state 1, this allows us to use the
2345 charclass scan to find a valid start char. This is based on the principle
2346 that theres a good chance the string being searched contains lots of stuff
2347 that cant be a start char.
2349 fail[ 0 ] = fail[ 1 ] = 0;
2350 DEBUG_TRIE_COMPILE_r({
2351 PerlIO_printf(Perl_debug_log,
2352 "%*sStclass Failtable (%"UVuf" states): 0",
2353 (int)(depth * 2), "", (UV)numstates
2355 for( q_read=1; q_read<numstates; q_read++ ) {
2356 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2358 PerlIO_printf(Perl_debug_log, "\n");
2361 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2366 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2367 * These need to be revisited when a newer toolchain becomes available.
2369 #if defined(__sparc64__) && defined(__GNUC__)
2370 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2371 # undef SPARC64_GCC_WORKAROUND
2372 # define SPARC64_GCC_WORKAROUND 1
2376 #define DEBUG_PEEP(str,scan,depth) \
2377 DEBUG_OPTIMISE_r({if (scan){ \
2378 SV * const mysv=sv_newmortal(); \
2379 regnode *Next = regnext(scan); \
2380 regprop(RExC_rx, mysv, scan); \
2381 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2382 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2383 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2390 #define JOIN_EXACT(scan,min,flags) \
2391 if (PL_regkind[OP(scan)] == EXACT) \
2392 join_exact(pRExC_state,(scan),(min),(flags),NULL,depth+1)
2395 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, I32 *min, U32 flags,regnode *val, U32 depth) {
2396 /* Merge several consecutive EXACTish nodes into one. */
2397 regnode *n = regnext(scan);
2399 regnode *next = scan + NODE_SZ_STR(scan);
2403 regnode *stop = scan;
2404 GET_RE_DEBUG_FLAGS_DECL;
2406 PERL_UNUSED_ARG(depth);
2409 PERL_ARGS_ASSERT_JOIN_EXACT;
2410 #ifndef EXPERIMENTAL_INPLACESCAN
2411 PERL_UNUSED_ARG(flags);
2412 PERL_UNUSED_ARG(val);
2414 DEBUG_PEEP("join",scan,depth);
2416 /* Skip NOTHING, merge EXACT*. */
2418 ( PL_regkind[OP(n)] == NOTHING ||
2419 (stringok && (OP(n) == OP(scan))))
2421 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX) {
2423 if (OP(n) == TAIL || n > next)
2425 if (PL_regkind[OP(n)] == NOTHING) {
2426 DEBUG_PEEP("skip:",n,depth);
2427 NEXT_OFF(scan) += NEXT_OFF(n);
2428 next = n + NODE_STEP_REGNODE;
2435 else if (stringok) {
2436 const unsigned int oldl = STR_LEN(scan);
2437 regnode * const nnext = regnext(n);
2439 DEBUG_PEEP("merg",n,depth);
2442 if (oldl + STR_LEN(n) > U8_MAX)
2444 NEXT_OFF(scan) += NEXT_OFF(n);
2445 STR_LEN(scan) += STR_LEN(n);
2446 next = n + NODE_SZ_STR(n);
2447 /* Now we can overwrite *n : */
2448 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2456 #ifdef EXPERIMENTAL_INPLACESCAN
2457 if (flags && !NEXT_OFF(n)) {
2458 DEBUG_PEEP("atch", val, depth);
2459 if (reg_off_by_arg[OP(n)]) {
2460 ARG_SET(n, val - n);
2463 NEXT_OFF(n) = val - n;
2469 #define GREEK_SMALL_LETTER_IOTA_WITH_DIALYTIKA_AND_TONOS 0x0390
2470 #define IOTA_D_T GREEK_SMALL_LETTER_IOTA_WITH_DIALYTIKA_AND_TONOS
2471 #define GREEK_SMALL_LETTER_UPSILON_WITH_DIALYTIKA_AND_TONOS 0x03B0
2472 #define UPSILON_D_T GREEK_SMALL_LETTER_UPSILON_WITH_DIALYTIKA_AND_TONOS
2475 && ( OP(scan) == EXACTF || OP(scan) == EXACTFU)
2476 && ( STR_LEN(scan) >= 6 ) )
2479 Two problematic code points in Unicode casefolding of EXACT nodes:
2481 U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2482 U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2488 U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2489 U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2491 This means that in case-insensitive matching (or "loose matching",
2492 as Unicode calls it), an EXACTF of length six (the UTF-8 encoded byte
2493 length of the above casefolded versions) can match a target string
2494 of length two (the byte length of UTF-8 encoded U+0390 or U+03B0).
2495 This would rather mess up the minimum length computation.
2497 What we'll do is to look for the tail four bytes, and then peek
2498 at the preceding two bytes to see whether we need to decrease
2499 the minimum length by four (six minus two).
2501 Thanks to the design of UTF-8, there cannot be false matches:
2502 A sequence of valid UTF-8 bytes cannot be a subsequence of
2503 another valid sequence of UTF-8 bytes.
2506 char * const s0 = STRING(scan), *s, *t;
2507 char * const s1 = s0 + STR_LEN(scan) - 1;
2508 char * const s2 = s1 - 4;
2509 #ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2510 const char t0[] = "\xaf\x49\xaf\x42";
2512 const char t0[] = "\xcc\x88\xcc\x81";
2514 const char * const t1 = t0 + 3;
2517 s < s2 && (t = ninstr(s, s1, t0, t1));
2520 if (((U8)t[-1] == 0x68 && (U8)t[-2] == 0xB4) ||
2521 ((U8)t[-1] == 0x46 && (U8)t[-2] == 0xB5))
2523 if (((U8)t[-1] == 0xB9 && (U8)t[-2] == 0xCE) ||
2524 ((U8)t[-1] == 0x85 && (U8)t[-2] == 0xCF))
2532 n = scan + NODE_SZ_STR(scan);
2534 if (PL_regkind[OP(n)] != NOTHING || OP(n) == NOTHING) {
2541 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2545 /* REx optimizer. Converts nodes into quicker variants "in place".
2546 Finds fixed substrings. */
2548 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2549 to the position after last scanned or to NULL. */
2551 #define INIT_AND_WITHP \
2552 assert(!and_withp); \
2553 Newx(and_withp,1,struct regnode_charclass_class); \
2554 SAVEFREEPV(and_withp)
2556 /* this is a chain of data about sub patterns we are processing that
2557 need to be handled separately/specially in study_chunk. Its so
2558 we can simulate recursion without losing state. */
2560 typedef struct scan_frame {
2561 regnode *last; /* last node to process in this frame */
2562 regnode *next; /* next node to process when last is reached */
2563 struct scan_frame *prev; /*previous frame*/
2564 I32 stop; /* what stopparen do we use */
2568 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2570 #define CASE_SYNST_FNC(nAmE) \
2572 if (flags & SCF_DO_STCLASS_AND) { \
2573 for (value = 0; value < 256; value++) \
2574 if (!is_ ## nAmE ## _cp(value)) \
2575 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2578 for (value = 0; value < 256; value++) \
2579 if (is_ ## nAmE ## _cp(value)) \
2580 ANYOF_BITMAP_SET(data->start_class, value); \
2584 if (flags & SCF_DO_STCLASS_AND) { \
2585 for (value = 0; value < 256; value++) \
2586 if (is_ ## nAmE ## _cp(value)) \
2587 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2590 for (value = 0; value < 256; value++) \
2591 if (!is_ ## nAmE ## _cp(value)) \
2592 ANYOF_BITMAP_SET(data->start_class, value); \
2599 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2600 I32 *minlenp, I32 *deltap,
2605 struct regnode_charclass_class *and_withp,
2606 U32 flags, U32 depth)
2607 /* scanp: Start here (read-write). */
2608 /* deltap: Write maxlen-minlen here. */
2609 /* last: Stop before this one. */
2610 /* data: string data about the pattern */
2611 /* stopparen: treat close N as END */
2612 /* recursed: which subroutines have we recursed into */
2613 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
2616 I32 min = 0, pars = 0, code;
2617 regnode *scan = *scanp, *next;
2619 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2620 int is_inf_internal = 0; /* The studied chunk is infinite */
2621 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2622 scan_data_t data_fake;
2623 SV *re_trie_maxbuff = NULL;
2624 regnode *first_non_open = scan;
2625 I32 stopmin = I32_MAX;
2626 scan_frame *frame = NULL;
2627 GET_RE_DEBUG_FLAGS_DECL;
2629 PERL_ARGS_ASSERT_STUDY_CHUNK;
2632 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
2636 while (first_non_open && OP(first_non_open) == OPEN)
2637 first_non_open=regnext(first_non_open);
2642 while ( scan && OP(scan) != END && scan < last ){
2643 /* Peephole optimizer: */
2644 DEBUG_STUDYDATA("Peep:", data,depth);
2645 DEBUG_PEEP("Peep",scan,depth);
2646 JOIN_EXACT(scan,&min,0);
2648 /* Follow the next-chain of the current node and optimize
2649 away all the NOTHINGs from it. */
2650 if (OP(scan) != CURLYX) {
2651 const int max = (reg_off_by_arg[OP(scan)]
2653 /* I32 may be smaller than U16 on CRAYs! */
2654 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
2655 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
2659 /* Skip NOTHING and LONGJMP. */
2660 while ((n = regnext(n))
2661 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
2662 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
2663 && off + noff < max)
2665 if (reg_off_by_arg[OP(scan)])
2668 NEXT_OFF(scan) = off;
2673 /* The principal pseudo-switch. Cannot be a switch, since we
2674 look into several different things. */
2675 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
2676 || OP(scan) == IFTHEN) {
2677 next = regnext(scan);
2679 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
2681 if (OP(next) == code || code == IFTHEN) {
2682 /* NOTE - There is similar code to this block below for handling
2683 TRIE nodes on a re-study. If you change stuff here check there
2685 I32 max1 = 0, min1 = I32_MAX, num = 0;
2686 struct regnode_charclass_class accum;
2687 regnode * const startbranch=scan;
2689 if (flags & SCF_DO_SUBSTR)
2690 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
2691 if (flags & SCF_DO_STCLASS)
2692 cl_init_zero(pRExC_state, &accum);
2694 while (OP(scan) == code) {
2695 I32 deltanext, minnext, f = 0, fake;
2696 struct regnode_charclass_class this_class;
2699 data_fake.flags = 0;
2701 data_fake.whilem_c = data->whilem_c;
2702 data_fake.last_closep = data->last_closep;
2705 data_fake.last_closep = &fake;
2707 data_fake.pos_delta = delta;
2708 next = regnext(scan);
2709 scan = NEXTOPER(scan);
2711 scan = NEXTOPER(scan);
2712 if (flags & SCF_DO_STCLASS) {
2713 cl_init(pRExC_state, &this_class);
2714 data_fake.start_class = &this_class;
2715 f = SCF_DO_STCLASS_AND;
2717 if (flags & SCF_WHILEM_VISITED_POS)
2718 f |= SCF_WHILEM_VISITED_POS;
2720 /* we suppose the run is continuous, last=next...*/
2721 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
2723 stopparen, recursed, NULL, f,depth+1);
2726 if (max1 < minnext + deltanext)
2727 max1 = minnext + deltanext;
2728 if (deltanext == I32_MAX)
2729 is_inf = is_inf_internal = 1;
2731 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
2733 if (data_fake.flags & SCF_SEEN_ACCEPT) {
2734 if ( stopmin > minnext)
2735 stopmin = min + min1;
2736 flags &= ~SCF_DO_SUBSTR;
2738 data->flags |= SCF_SEEN_ACCEPT;
2741 if (data_fake.flags & SF_HAS_EVAL)
2742 data->flags |= SF_HAS_EVAL;
2743 data->whilem_c = data_fake.whilem_c;
2745 if (flags & SCF_DO_STCLASS)
2746 cl_or(pRExC_state, &accum, &this_class);
2748 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
2750 if (flags & SCF_DO_SUBSTR) {
2751 data->pos_min += min1;
2752 data->pos_delta += max1 - min1;
2753 if (max1 != min1 || is_inf)
2754 data->longest = &(data->longest_float);
2757 delta += max1 - min1;
2758 if (flags & SCF_DO_STCLASS_OR) {
2759 cl_or(pRExC_state, data->start_class, &accum);
2761 cl_and(data->start_class, and_withp);
2762 flags &= ~SCF_DO_STCLASS;
2765 else if (flags & SCF_DO_STCLASS_AND) {
2767 cl_and(data->start_class, &accum);
2768 flags &= ~SCF_DO_STCLASS;
2771 /* Switch to OR mode: cache the old value of
2772 * data->start_class */
2774 StructCopy(data->start_class, and_withp,
2775 struct regnode_charclass_class);
2776 flags &= ~SCF_DO_STCLASS_AND;
2777 StructCopy(&accum, data->start_class,
2778 struct regnode_charclass_class);
2779 flags |= SCF_DO_STCLASS_OR;
2780 data->start_class->flags |= ANYOF_EOS;
2784 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
2787 Assuming this was/is a branch we are dealing with: 'scan' now
2788 points at the item that follows the branch sequence, whatever
2789 it is. We now start at the beginning of the sequence and look
2796 which would be constructed from a pattern like /A|LIST|OF|WORDS/
2798 If we can find such a subsequence we need to turn the first
2799 element into a trie and then add the subsequent branch exact
2800 strings to the trie.
2804 1. patterns where the whole set of branches can be converted.
2806 2. patterns where only a subset can be converted.
2808 In case 1 we can replace the whole set with a single regop
2809 for the trie. In case 2 we need to keep the start and end
2812 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
2813 becomes BRANCH TRIE; BRANCH X;
2815 There is an additional case, that being where there is a
2816 common prefix, which gets split out into an EXACT like node
2817 preceding the TRIE node.
2819 If x(1..n)==tail then we can do a simple trie, if not we make
2820 a "jump" trie, such that when we match the appropriate word
2821 we "jump" to the appropriate tail node. Essentially we turn
2822 a nested if into a case structure of sorts.
2827 if (!re_trie_maxbuff) {
2828 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2829 if (!SvIOK(re_trie_maxbuff))
2830 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2832 if ( SvIV(re_trie_maxbuff)>=0 ) {
2834 regnode *first = (regnode *)NULL;
2835 regnode *last = (regnode *)NULL;
2836 regnode *tail = scan;
2841 SV * const mysv = sv_newmortal(); /* for dumping */
2843 /* var tail is used because there may be a TAIL
2844 regop in the way. Ie, the exacts will point to the
2845 thing following the TAIL, but the last branch will
2846 point at the TAIL. So we advance tail. If we
2847 have nested (?:) we may have to move through several
2851 while ( OP( tail ) == TAIL ) {
2852 /* this is the TAIL generated by (?:) */
2853 tail = regnext( tail );
2858 regprop(RExC_rx, mysv, tail );
2859 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
2860 (int)depth * 2 + 2, "",
2861 "Looking for TRIE'able sequences. Tail node is: ",
2862 SvPV_nolen_const( mysv )
2868 step through the branches, cur represents each
2869 branch, noper is the first thing to be matched
2870 as part of that branch and noper_next is the
2871 regnext() of that node. if noper is an EXACT
2872 and noper_next is the same as scan (our current
2873 position in the regex) then the EXACT branch is
2874 a possible optimization target. Once we have
2875 two or more consecutive such branches we can
2876 create a trie of the EXACT's contents and stich
2877 it in place. If the sequence represents all of
2878 the branches we eliminate the whole thing and
2879 replace it with a single TRIE. If it is a
2880 subsequence then we need to stitch it in. This
2881 means the first branch has to remain, and needs
2882 to be repointed at the item on the branch chain
2883 following the last branch optimized. This could
2884 be either a BRANCH, in which case the
2885 subsequence is internal, or it could be the
2886 item following the branch sequence in which
2887 case the subsequence is at the end.
2891 /* dont use tail as the end marker for this traverse */
2892 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
2893 regnode * const noper = NEXTOPER( cur );
2894 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
2895 regnode * const noper_next = regnext( noper );
2899 regprop(RExC_rx, mysv, cur);
2900 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
2901 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
2903 regprop(RExC_rx, mysv, noper);
2904 PerlIO_printf( Perl_debug_log, " -> %s",
2905 SvPV_nolen_const(mysv));
2908 regprop(RExC_rx, mysv, noper_next );
2909 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
2910 SvPV_nolen_const(mysv));
2912 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
2913 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
2915 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
2916 : PL_regkind[ OP( noper ) ] == EXACT )
2917 || OP(noper) == NOTHING )
2919 && noper_next == tail
2924 if ( !first || optype == NOTHING ) {
2925 if (!first) first = cur;
2926 optype = OP( noper );
2932 Currently we do not believe that the trie logic can
2933 handle case insensitive matching properly when the
2934 pattern is not unicode (thus forcing unicode semantics).
2936 If/when this is fixed the following define can be swapped
2937 in below to fully enable trie logic.
2939 #define TRIE_TYPE_IS_SAFE 1
2942 #define TRIE_TYPE_IS_SAFE (UTF || optype==EXACT)
2944 if ( last && TRIE_TYPE_IS_SAFE ) {
2945 make_trie( pRExC_state,
2946 startbranch, first, cur, tail, count,
2949 if ( PL_regkind[ OP( noper ) ] == EXACT
2951 && noper_next == tail
2956 optype = OP( noper );
2966 regprop(RExC_rx, mysv, cur);
2967 PerlIO_printf( Perl_debug_log,
2968 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
2969 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
2973 if ( last && TRIE_TYPE_IS_SAFE ) {
2974 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
2975 #ifdef TRIE_STUDY_OPT
2976 if ( ((made == MADE_EXACT_TRIE &&
2977 startbranch == first)
2978 || ( first_non_open == first )) &&
2980 flags |= SCF_TRIE_RESTUDY;
2981 if ( startbranch == first
2984 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
2994 else if ( code == BRANCHJ ) { /* single branch is optimized. */
2995 scan = NEXTOPER(NEXTOPER(scan));
2996 } else /* single branch is optimized. */
2997 scan = NEXTOPER(scan);
2999 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3000 scan_frame *newframe = NULL;
3005 if (OP(scan) != SUSPEND) {
3006 /* set the pointer */
3007 if (OP(scan) == GOSUB) {
3009 RExC_recurse[ARG2L(scan)] = scan;
3010 start = RExC_open_parens[paren-1];
3011 end = RExC_close_parens[paren-1];
3014 start = RExC_rxi->program + 1;
3018 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3019 SAVEFREEPV(recursed);
3021 if (!PAREN_TEST(recursed,paren+1)) {
3022 PAREN_SET(recursed,paren+1);
3023 Newx(newframe,1,scan_frame);
3025 if (flags & SCF_DO_SUBSTR) {
3026 SCAN_COMMIT(pRExC_state,data,minlenp);
3027 data->longest = &(data->longest_float);
3029 is_inf = is_inf_internal = 1;
3030 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3031 cl_anything(pRExC_state, data->start_class);
3032 flags &= ~SCF_DO_STCLASS;
3035 Newx(newframe,1,scan_frame);
3038 end = regnext(scan);
3043 SAVEFREEPV(newframe);
3044 newframe->next = regnext(scan);
3045 newframe->last = last;
3046 newframe->stop = stopparen;
3047 newframe->prev = frame;
3057 else if (OP(scan) == EXACT) {
3058 I32 l = STR_LEN(scan);
3061 const U8 * const s = (U8*)STRING(scan);
3062 l = utf8_length(s, s + l);
3063 uc = utf8_to_uvchr(s, NULL);
3065 uc = *((U8*)STRING(scan));
3068 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3069 /* The code below prefers earlier match for fixed
3070 offset, later match for variable offset. */
3071 if (data->last_end == -1) { /* Update the start info. */
3072 data->last_start_min = data->pos_min;
3073 data->last_start_max = is_inf
3074 ? I32_MAX : data->pos_min + data->pos_delta;
3076 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3078 SvUTF8_on(data->last_found);
3080 SV * const sv = data->last_found;
3081 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3082 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3083 if (mg && mg->mg_len >= 0)
3084 mg->mg_len += utf8_length((U8*)STRING(scan),
3085 (U8*)STRING(scan)+STR_LEN(scan));
3087 data->last_end = data->pos_min + l;
3088 data->pos_min += l; /* As in the first entry. */
3089 data->flags &= ~SF_BEFORE_EOL;
3091 if (flags & SCF_DO_STCLASS_AND) {
3092 /* Check whether it is compatible with what we know already! */
3096 /* If compatible, we or it in below. It is compatible if is
3097 * in the bitmp and either 1) its bit or its fold is set, or 2)
3098 * it's for a locale. Even if there isn't unicode semantics
3099 * here, at runtime there may be because of matching against a
3100 * utf8 string, so accept a possible false positive for
3101 * latin1-range folds */
3103 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3104 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3105 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
3106 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3109 ANYOF_CLASS_ZERO(data->start_class);
3110 ANYOF_BITMAP_ZERO(data->start_class);
3112 ANYOF_BITMAP_SET(data->start_class, uc);
3113 data->start_class->flags &= ~ANYOF_EOS;
3115 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3117 else if (flags & SCF_DO_STCLASS_OR) {
3118 /* false positive possible if the class is case-folded */
3120 ANYOF_BITMAP_SET(data->start_class, uc);
3122 data->start_class->flags |= ANYOF_UNICODE_ALL;
3123 data->start_class->flags &= ~ANYOF_EOS;
3124 cl_and(data->start_class, and_withp);
3126 flags &= ~SCF_DO_STCLASS;
3128 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3129 I32 l = STR_LEN(scan);
3130 UV uc = *((U8*)STRING(scan));
3132 /* Search for fixed substrings supports EXACT only. */
3133 if (flags & SCF_DO_SUBSTR) {
3135 SCAN_COMMIT(pRExC_state, data, minlenp);
3138 const U8 * const s = (U8 *)STRING(scan);
3139 l = utf8_length(s, s + l);
3140 uc = utf8_to_uvchr(s, NULL);
3143 if (flags & SCF_DO_SUBSTR)
3145 if (flags & SCF_DO_STCLASS_AND) {
3146 /* Check whether it is compatible with what we know already! */
3149 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3150 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3151 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3155 ANYOF_CLASS_ZERO(data->start_class);
3156 ANYOF_BITMAP_ZERO(data->start_class);
3158 ANYOF_BITMAP_SET(data->start_class, uc);
3159 data->start_class->flags &= ~ANYOF_EOS;
3160 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
3161 if (OP(scan) == EXACTFL) {
3162 data->start_class->flags |= ANYOF_LOCALE;
3166 /* Also set the other member of the fold pair. In case
3167 * that unicode semantics is called for at runtime, use
3168 * the full latin1 fold. (Can't do this for locale,
3169 * because not known until runtime */
3170 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3174 else if (flags & SCF_DO_STCLASS_OR) {
3175 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
3176 /* false positive possible if the class is case-folded.
3177 Assume that the locale settings are the same... */
3179 ANYOF_BITMAP_SET(data->start_class, uc);
3180 if (OP(scan) != EXACTFL) {
3182 /* And set the other member of the fold pair, but
3183 * can't do that in locale because not known until
3185 ANYOF_BITMAP_SET(data->start_class,
3186 PL_fold_latin1[uc]);
3189 data->start_class->flags &= ~ANYOF_EOS;
3191 cl_and(data->start_class, and_withp);
3193 flags &= ~SCF_DO_STCLASS;
3195 else if (REGNODE_VARIES(OP(scan))) {
3196 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3197 I32 f = flags, pos_before = 0;
3198 regnode * const oscan = scan;
3199 struct regnode_charclass_class this_class;
3200 struct regnode_charclass_class *oclass = NULL;
3201 I32 next_is_eval = 0;
3203 switch (PL_regkind[OP(scan)]) {
3204 case WHILEM: /* End of (?:...)* . */
3205 scan = NEXTOPER(scan);
3208 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3209 next = NEXTOPER(scan);
3210 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3212 maxcount = REG_INFTY;
3213 next = regnext(scan);
3214 scan = NEXTOPER(scan);
3218 if (flags & SCF_DO_SUBSTR)
3223 if (flags & SCF_DO_STCLASS) {
3225 maxcount = REG_INFTY;
3226 next = regnext(scan);
3227 scan = NEXTOPER(scan);
3230 is_inf = is_inf_internal = 1;
3231 scan = regnext(scan);
3232 if (flags & SCF_DO_SUBSTR) {
3233 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3234 data->longest = &(data->longest_float);
3236 goto optimize_curly_tail;
3238 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3239 && (scan->flags == stopparen))
3244 mincount = ARG1(scan);
3245 maxcount = ARG2(scan);
3247 next = regnext(scan);
3248 if (OP(scan) == CURLYX) {
3249 I32 lp = (data ? *(data->last_closep) : 0);
3250 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3252 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3253 next_is_eval = (OP(scan) == EVAL);
3255 if (flags & SCF_DO_SUBSTR) {
3256 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3257 pos_before = data->pos_min;
3261 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3263 data->flags |= SF_IS_INF;
3265 if (flags & SCF_DO_STCLASS) {
3266 cl_init(pRExC_state, &this_class);
3267 oclass = data->start_class;
3268 data->start_class = &this_class;
3269 f |= SCF_DO_STCLASS_AND;
3270 f &= ~SCF_DO_STCLASS_OR;
3272 /* Exclude from super-linear cache processing any {n,m}
3273 regops for which the combination of input pos and regex
3274 pos is not enough information to determine if a match
3277 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3278 regex pos at the \s*, the prospects for a match depend not
3279 only on the input position but also on how many (bar\s*)
3280 repeats into the {4,8} we are. */
3281 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3282 f &= ~SCF_WHILEM_VISITED_POS;
3284 /* This will finish on WHILEM, setting scan, or on NULL: */
3285 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3286 last, data, stopparen, recursed, NULL,
3288 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3290 if (flags & SCF_DO_STCLASS)
3291 data->start_class = oclass;
3292 if (mincount == 0 || minnext == 0) {
3293 if (flags & SCF_DO_STCLASS_OR) {
3294 cl_or(pRExC_state, data->start_class, &this_class);
3296 else if (flags & SCF_DO_STCLASS_AND) {
3297 /* Switch to OR mode: cache the old value of
3298 * data->start_class */
3300 StructCopy(data->start_class, and_withp,
3301 struct regnode_charclass_class);
3302 flags &= ~SCF_DO_STCLASS_AND;
3303 StructCopy(&this_class, data->start_class,
3304 struct regnode_charclass_class);
3305 flags |= SCF_DO_STCLASS_OR;
3306 data->start_class->flags |= ANYOF_EOS;
3308 } else { /* Non-zero len */
3309 if (flags & SCF_DO_STCLASS_OR) {
3310 cl_or(pRExC_state, data->start_class, &this_class);
3311 cl_and(data->start_class, and_withp);
3313 else if (flags & SCF_DO_STCLASS_AND)
3314 cl_and(data->start_class, &this_class);
3315 flags &= ~SCF_DO_STCLASS;
3317 if (!scan) /* It was not CURLYX, but CURLY. */
3319 if ( /* ? quantifier ok, except for (?{ ... }) */
3320 (next_is_eval || !(mincount == 0 && maxcount == 1))
3321 && (minnext == 0) && (deltanext == 0)
3322 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3323 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3325 ckWARNreg(RExC_parse,
3326 "Quantifier unexpected on zero-length expression");
3329 min += minnext * mincount;
3330 is_inf_internal |= ((maxcount == REG_INFTY
3331 && (minnext + deltanext) > 0)
3332 || deltanext == I32_MAX);
3333 is_inf |= is_inf_internal;
3334 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3336 /* Try powerful optimization CURLYX => CURLYN. */
3337 if ( OP(oscan) == CURLYX && data
3338 && data->flags & SF_IN_PAR
3339 && !(data->flags & SF_HAS_EVAL)
3340 && !deltanext && minnext == 1 ) {
3341 /* Try to optimize to CURLYN. */
3342 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3343 regnode * const nxt1 = nxt;
3350 if (!REGNODE_SIMPLE(OP(nxt))
3351 && !(PL_regkind[OP(nxt)] == EXACT
3352 && STR_LEN(nxt) == 1))
3358 if (OP(nxt) != CLOSE)
3360 if (RExC_open_parens) {
3361 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3362 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3364 /* Now we know that nxt2 is the only contents: */
3365 oscan->flags = (U8)ARG(nxt);
3367 OP(nxt1) = NOTHING; /* was OPEN. */
3370 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3371 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3372 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3373 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3374 OP(nxt + 1) = OPTIMIZED; /* was count. */
3375 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3380 /* Try optimization CURLYX => CURLYM. */
3381 if ( OP(oscan) == CURLYX && data
3382 && !(data->flags & SF_HAS_PAR)
3383 && !(data->flags & SF_HAS_EVAL)
3384 && !deltanext /* atom is fixed width */
3385 && minnext != 0 /* CURLYM can't handle zero width */
3387 /* XXXX How to optimize if data == 0? */
3388 /* Optimize to a simpler form. */
3389 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3393 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3394 && (OP(nxt2) != WHILEM))
3396 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3397 /* Need to optimize away parenths. */
3398 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
3399 /* Set the parenth number. */
3400 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3402 oscan->flags = (U8)ARG(nxt);
3403 if (RExC_open_parens) {
3404 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3405 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3407 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3408 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3411 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3412 OP(nxt + 1) = OPTIMIZED; /* was count. */
3413 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
3414 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
3417 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3418 regnode *nnxt = regnext(nxt1);
3420 if (reg_off_by_arg[OP(nxt1)])
3421 ARG_SET(nxt1, nxt2 - nxt1);
3422 else if (nxt2 - nxt1 < U16_MAX)
3423 NEXT_OFF(nxt1) = nxt2 - nxt1;
3425 OP(nxt) = NOTHING; /* Cannot beautify */
3430 /* Optimize again: */
3431 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
3432 NULL, stopparen, recursed, NULL, 0,depth+1);
3437 else if ((OP(oscan) == CURLYX)
3438 && (flags & SCF_WHILEM_VISITED_POS)
3439 /* See the comment on a similar expression above.
3440 However, this time it's not a subexpression
3441 we care about, but the expression itself. */
3442 && (maxcount == REG_INFTY)
3443 && data && ++data->whilem_c < 16) {
3444 /* This stays as CURLYX, we can put the count/of pair. */
3445 /* Find WHILEM (as in regexec.c) */
3446 regnode *nxt = oscan + NEXT_OFF(oscan);
3448 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
3450 PREVOPER(nxt)->flags = (U8)(data->whilem_c
3451 | (RExC_whilem_seen << 4)); /* On WHILEM */
3453 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
3455 if (flags & SCF_DO_SUBSTR) {
3456 SV *last_str = NULL;
3457 int counted = mincount != 0;
3459 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
3460 #if defined(SPARC64_GCC_WORKAROUND)
3463 const char *s = NULL;
3466 if (pos_before >= data->last_start_min)
3469 b = data->last_start_min;
3472 s = SvPV_const(data->last_found, l);
3473 old = b - data->last_start_min;
3476 I32 b = pos_before >= data->last_start_min
3477 ? pos_before : data->last_start_min;
3479 const char * const s = SvPV_const(data->last_found, l);
3480 I32 old = b - data->last_start_min;
3484 old = utf8_hop((U8*)s, old) - (U8*)s;
3486 /* Get the added string: */
3487 last_str = newSVpvn_utf8(s + old, l, UTF);
3488 if (deltanext == 0 && pos_before == b) {
3489 /* What was added is a constant string */
3491 SvGROW(last_str, (mincount * l) + 1);
3492 repeatcpy(SvPVX(last_str) + l,
3493 SvPVX_const(last_str), l, mincount - 1);
3494 SvCUR_set(last_str, SvCUR(last_str) * mincount);
3495 /* Add additional parts. */
3496 SvCUR_set(data->last_found,
3497 SvCUR(data->last_found) - l);
3498 sv_catsv(data->last_found, last_str);
3500 SV * sv = data->last_found;
3502 SvUTF8(sv) && SvMAGICAL(sv) ?
3503 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3504 if (mg && mg->mg_len >= 0)
3505 mg->mg_len += CHR_SVLEN(last_str) - l;
3507 data->last_end += l * (mincount - 1);
3510 /* start offset must point into the last copy */
3511 data->last_start_min += minnext * (mincount - 1);
3512 data->last_start_max += is_inf ? I32_MAX
3513 : (maxcount - 1) * (minnext + data->pos_delta);
3516 /* It is counted once already... */
3517 data->pos_min += minnext * (mincount - counted);
3518 data->pos_delta += - counted * deltanext +
3519 (minnext + deltanext) * maxcount - minnext * mincount;
3520 if (mincount != maxcount) {
3521 /* Cannot extend fixed substrings found inside
3523 SCAN_COMMIT(pRExC_state,data,minlenp);
3524 if (mincount && last_str) {
3525 SV * const sv = data->last_found;
3526 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3527 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3531 sv_setsv(sv, last_str);
3532 data->last_end = data->pos_min;
3533 data->last_start_min =
3534 data->pos_min - CHR_SVLEN(last_str);
3535 data->last_start_max = is_inf
3537 : data->pos_min + data->pos_delta
3538 - CHR_SVLEN(last_str);
3540 data->longest = &(data->longest_float);
3542 SvREFCNT_dec(last_str);
3544 if (data && (fl & SF_HAS_EVAL))
3545 data->flags |= SF_HAS_EVAL;
3546 optimize_curly_tail:
3547 if (OP(oscan) != CURLYX) {
3548 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
3550 NEXT_OFF(oscan) += NEXT_OFF(next);
3553 default: /* REF, ANYOFV, and CLUMP only? */
3554 if (flags & SCF_DO_SUBSTR) {
3555 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3556 data->longest = &(data->longest_float);
3558 is_inf = is_inf_internal = 1;
3559 if (flags & SCF_DO_STCLASS_OR)
3560 cl_anything(pRExC_state, data->start_class);
3561 flags &= ~SCF_DO_STCLASS;
3565 else if (OP(scan) == LNBREAK) {
3566 if (flags & SCF_DO_STCLASS) {
3568 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3569 if (flags & SCF_DO_STCLASS_AND) {
3570 for (value = 0; value < 256; value++)
3571 if (!is_VERTWS_cp(value))
3572 ANYOF_BITMAP_CLEAR(data->start_class, value);
3575 for (value = 0; value < 256; value++)
3576 if (is_VERTWS_cp(value))
3577 ANYOF_BITMAP_SET(data->start_class, value);
3579 if (flags & SCF_DO_STCLASS_OR)
3580 cl_and(data->start_class, and_withp);
3581 flags &= ~SCF_DO_STCLASS;
3585 if (flags & SCF_DO_SUBSTR) {
3586 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3588 data->pos_delta += 1;
3589 data->longest = &(data->longest_float);
3592 else if (OP(scan) == FOLDCHAR) {
3593 int d = ARG(scan) == LATIN_SMALL_LETTER_SHARP_S ? 1 : 2;
3594 flags &= ~SCF_DO_STCLASS;
3597 if (flags & SCF_DO_SUBSTR) {
3598 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
3600 data->pos_delta += d;
3601 data->longest = &(data->longest_float);
3604 else if (REGNODE_SIMPLE(OP(scan))) {
3607 if (flags & SCF_DO_SUBSTR) {
3608 SCAN_COMMIT(pRExC_state,data,minlenp);
3612 if (flags & SCF_DO_STCLASS) {
3613 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
3615 /* Some of the logic below assumes that switching
3616 locale on will only add false positives. */
3617 switch (PL_regkind[OP(scan)]) {
3621 /* Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan)); */
3622 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3623 cl_anything(pRExC_state, data->start_class);
3626 if (OP(scan) == SANY)
3628 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
3629 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
3630 || ANYOF_CLASS_TEST_ANY_SET(data->start_class));
3631 cl_anything(pRExC_state, data->start_class);
3633 if (flags & SCF_DO_STCLASS_AND || !value)
3634 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
3637 if (flags & SCF_DO_STCLASS_AND)
3638 cl_and(data->start_class,
3639 (struct regnode_charclass_class*)scan);
3641 cl_or(pRExC_state, data->start_class,
3642 (struct regnode_charclass_class*)scan);
3645 if (flags & SCF_DO_STCLASS_AND) {
3646 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3647 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
3648 if (OP(scan) == ALNUMU) {
3649 for (value = 0; value < 256; value++) {
3650 if (!isWORDCHAR_L1(value)) {
3651 ANYOF_BITMAP_CLEAR(data->start_class, value);
3655 for (value = 0; value < 256; value++) {
3656 if (!isALNUM(value)) {
3657 ANYOF_BITMAP_CLEAR(data->start_class, value);
3664 if (data->start_class->flags & ANYOF_LOCALE)
3665 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
3666 else if (OP(scan) == ALNUMU) {
3667 for (value = 0; value < 256; value++) {
3668 if (isWORDCHAR_L1(value)) {
3669 ANYOF_BITMAP_SET(data->start_class, value);
3673 for (value = 0; value < 256; value++) {
3674 if (isALNUM(value)) {
3675 ANYOF_BITMAP_SET(data->start_class, value);
3682 if (flags & SCF_DO_STCLASS_AND) {
3683 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3684 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
3685 if (OP(scan) == NALNUMU) {
3686 for (value = 0; value < 256; value++) {
3687 if (isWORDCHAR_L1(value)) {
3688 ANYOF_BITMAP_CLEAR(data->start_class, value);
3692 for (value = 0; value < 256; value++) {
3693 if (isALNUM(value)) {
3694 ANYOF_BITMAP_CLEAR(data->start_class, value);
3701 if (data->start_class->flags & ANYOF_LOCALE)
3702 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
3704 if (OP(scan) == NALNUMU) {
3705 for (value = 0; value < 256; value++) {
3706 if (! isWORDCHAR_L1(value)) {
3707 ANYOF_BITMAP_SET(data->start_class, value);
3711 for (value = 0; value < 256; value++) {
3712 if (! isALNUM(value)) {
3713 ANYOF_BITMAP_SET(data->start_class, value);
3721 if (flags & SCF_DO_STCLASS_AND) {
3722 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3723 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
3724 if (OP(scan) == SPACEU) {
3725 for (value = 0; value < 256; value++) {
3726 if (!isSPACE_L1(value)) {
3727 ANYOF_BITMAP_CLEAR(data->start_class, value);
3731 for (value = 0; value < 256; value++) {
3732 if (!isSPACE(value)) {
3733 ANYOF_BITMAP_CLEAR(data->start_class, value);
3740 if (data->start_class->flags & ANYOF_LOCALE) {
3741 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
3743 else if (OP(scan) == SPACEU) {
3744 for (value = 0; value < 256; value++) {
3745 if (isSPACE_L1(value)) {
3746 ANYOF_BITMAP_SET(data->start_class, value);
3750 for (value = 0; value < 256; value++) {
3751 if (isSPACE(value)) {
3752 ANYOF_BITMAP_SET(data->start_class, value);
3759 if (flags & SCF_DO_STCLASS_AND) {
3760 if (!(data->start_class->flags & ANYOF_LOCALE)) {
3761 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
3762 if (OP(scan) == NSPACEU) {
3763 for (value = 0; value < 256; value++) {
3764 if (isSPACE_L1(value)) {
3765 ANYOF_BITMAP_CLEAR(data->start_class, value);
3769 for (value = 0; value < 256; value++) {
3770 if (isSPACE(value)) {
3771 ANYOF_BITMAP_CLEAR(data->start_class, value);
3778 if (data->start_class->flags & ANYOF_LOCALE)
3779 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
3780 else if (OP(scan) == NSPACEU) {
3781 for (value = 0; value < 256; value++) {
3782 if (!isSPACE_L1(value)) {
3783 ANYOF_BITMAP_SET(data->start_class, value);
3788 for (value = 0; value < 256; value++) {
3789 if (!isSPACE(value)) {
3790 ANYOF_BITMAP_SET(data->start_class, value);
3797 if (flags & SCF_DO_STCLASS_AND) {
3798 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NDIGIT);
3799 for (value = 0; value < 256; value++)
3800 if (!isDIGIT(value))
3801 ANYOF_BITMAP_CLEAR(data->start_class, value);
3804 if (data->start_class->flags & ANYOF_LOCALE)
3805 ANYOF_CLASS_SET(data->start_class,ANYOF_DIGIT);
3807 for (value = 0; value < 256; value++)
3809 ANYOF_BITMAP_SET(data->start_class, value);
3814 if (flags & SCF_DO_STCLASS_AND) {
3815 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_DIGIT);
3816 for (value = 0; value < 256; value++)
3818 ANYOF_BITMAP_CLEAR(data->start_class, value);