This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
fix segv in regcomp.c:S_join_exact()
[perl5.git] / regcomp.c
CommitLineData
a0d0e21e
LW
1/* regcomp.c
2 */
3
4/*
4ac71550
TC
5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
6 *
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
a0d0e21e
LW
8 */
9
61296642
DM
10/* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
166f8a29 12 * a regular expression.
e4a054ea
DM
13 *
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.
166f8a29
DM
18 */
19
a687059c
LW
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!
22 */
23
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.
27 */
28
e50aee73 29/* The names of the functions have been changed from regcomp and
3b753521 30 * regexec to pregcomp and pregexec in order to avoid conflicts
e50aee73
AD
31 * with the POSIX routines of the same names.
32*/
33
b9d5759e 34#ifdef PERL_EXT_RE_BUILD
54df2634 35#include "re_top.h"
b81d288d 36#endif
56953603 37
a687059c 38/*
e50aee73 39 * pregcomp and pregexec -- regsub and regerror are not used in perl
a687059c
LW
40 *
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
43 *
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:
47 *
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
50 * from defects in it.
51 *
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
54 *
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
57 *
58 *
59 **** Alterations to Henry's code are...
60 ****
4bb101f2 61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
1129b882
NC
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
a687059c 64 ****
9ef589d8
LW
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.
67
a687059c
LW
68 *
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.
72 */
73#include "EXTERN.h"
864dbfa3 74#define PERL_IN_REGCOMP_C
a687059c 75#include "perl.h"
d06ea78c 76
acfe0abc 77#ifndef PERL_IN_XSUB_RE
d06ea78c
GS
78# include "INTERN.h"
79#endif
c277df42
IZ
80
81#define REG_COMP_C
54df2634
NC
82#ifdef PERL_IN_XSUB_RE
83# include "re_comp.h"
84#else
85# include "regcomp.h"
86#endif
a687059c 87
04e98a4d
AD
88#include "dquote_static.c"
89
d4cce5f1 90#ifdef op
11343788 91#undef op
d4cce5f1 92#endif /* op */
11343788 93
fe14fcc3 94#ifdef MSDOS
7e4e8c89 95# if defined(BUGGY_MSC6)
fe14fcc3 96 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
7e4e8c89 97# pragma optimize("a",off)
fe14fcc3 98 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
7e4e8c89
NC
99# pragma optimize("w",on )
100# endif /* BUGGY_MSC6 */
fe14fcc3
LW
101#endif /* MSDOS */
102
a687059c
LW
103#ifndef STATIC
104#define STATIC static
105#endif
106
830247a4 107typedef struct RExC_state_t {
e2509266 108 U32 flags; /* are we folding, multilining? */
830247a4 109 char *precomp; /* uncompiled string. */
288b8c02 110 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf
YO
111 regexp *rx; /* perl core regexp structure */
112 regexp_internal *rxi; /* internal data for regexp object pprivate field */
fac92740 113 char *start; /* Start of input for compile */
830247a4
IZ
114 char *end; /* End of input for compile */
115 char *parse; /* Input-scan pointer. */
116 I32 whilem_seen; /* number of WHILEM in this expr */
fac92740 117 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 118 regnode *emit_bound; /* First regnode outside of the allocated space */
ffc61ed2 119 regnode *emit; /* Code-emit pointer; &regdummy = don't = compiling */
830247a4
IZ
120 I32 naughty; /* How bad is this pattern? */
121 I32 sawback; /* Did we see \1, ...? */
122 U32 seen;
123 I32 size; /* Code size. */
c74340f9
YO
124 I32 npar; /* Capture buffer count, (OPEN). */
125 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 126 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
127 I32 extralen;
128 I32 seen_zerolen;
129 I32 seen_evals;
40d049e4
YO
130 regnode **open_parens; /* pointers to open parens */
131 regnode **close_parens; /* pointers to close parens */
132 regnode *opend; /* END node in program */
02daf0ab
YO
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. */
e40e74fe
KW
137 I32 uni_semantics; /* If a d charset modifier should use unicode
138 rules, even if the pattern is not in
139 utf8 */
81714fb9 140 HV *paren_names; /* Paren names */
1f1031fe 141
40d049e4
YO
142 regnode **recurse; /* Recurse regops */
143 I32 recurse_count; /* Number of recurse regops */
b57e4118 144 I32 in_lookbehind;
4624b182 145 I32 contains_locale;
bb3f3ed2 146 I32 override_recoding;
830247a4
IZ
147#if ADD_TO_REGEXEC
148 char *starttry; /* -Dr: where regtry was called. */
149#define RExC_starttry (pRExC_state->starttry)
150#endif
3dab1dad 151#ifdef DEBUGGING
be8e71aa 152 const char *lastparse;
3dab1dad 153 I32 lastnum;
1f1031fe 154 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
155#define RExC_lastparse (pRExC_state->lastparse)
156#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 157#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 158#endif
830247a4
IZ
159} RExC_state_t;
160
e2509266 161#define RExC_flags (pRExC_state->flags)
830247a4 162#define RExC_precomp (pRExC_state->precomp)
288b8c02 163#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 164#define RExC_rx (pRExC_state->rx)
f8fc2ecf 165#define RExC_rxi (pRExC_state->rxi)
fac92740 166#define RExC_start (pRExC_state->start)
830247a4
IZ
167#define RExC_end (pRExC_state->end)
168#define RExC_parse (pRExC_state->parse)
169#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
170#ifdef RE_TRACK_PATTERN_OFFSETS
171#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
172#endif
830247a4 173#define RExC_emit (pRExC_state->emit)
fac92740 174#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 175#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
176#define RExC_naughty (pRExC_state->naughty)
177#define RExC_sawback (pRExC_state->sawback)
178#define RExC_seen (pRExC_state->seen)
179#define RExC_size (pRExC_state->size)
180#define RExC_npar (pRExC_state->npar)
e2e6a0f1 181#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
182#define RExC_extralen (pRExC_state->extralen)
183#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
184#define RExC_seen_evals (pRExC_state->seen_evals)
1aa99e6b 185#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 186#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 187#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
188#define RExC_open_parens (pRExC_state->open_parens)
189#define RExC_close_parens (pRExC_state->close_parens)
190#define RExC_opend (pRExC_state->opend)
81714fb9 191#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
192#define RExC_recurse (pRExC_state->recurse)
193#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 194#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 195#define RExC_contains_locale (pRExC_state->contains_locale)
bb3f3ed2 196#define RExC_override_recoding (pRExC_state->override_recoding)
830247a4 197
cde0cee5 198
a687059c
LW
199#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
200#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
201 ((*s) == '{' && regcurly(s)))
a687059c 202
35c8bce7
LW
203#ifdef SPSTART
204#undef SPSTART /* dratted cpp namespace... */
205#endif
a687059c
LW
206/*
207 * Flags to be passed up and down.
208 */
a687059c 209#define WORST 0 /* Worst case. */
a3b492c3 210#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee
KW
211
212/* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
d7b56a3c 213 * character, and if utf8, must be invariant. Note that this is not the same thing as REGNODE_SIMPLE */
fda99bee 214#define SIMPLE 0x02
a3b492c3
YO
215#define SPSTART 0x04 /* Starts with * or +. */
216#define TRYAGAIN 0x08 /* Weeded out a declaration. */
217#define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
a687059c 218
3dab1dad
YO
219#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
220
07be1b83
YO
221/* whether trie related optimizations are enabled */
222#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
223#define TRIE_STUDY_OPT
786e8c11 224#define FULL_TRIE_STUDY
07be1b83
YO
225#define TRIE_STCLASS
226#endif
1de06328
YO
227
228
40d049e4
YO
229
230#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
231#define PBITVAL(paren) (1 << ((paren) & 7))
232#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
233#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
234#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
235
bbd61b5f
KW
236/* If not already in utf8, do a longjmp back to the beginning */
237#define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
238#define REQUIRE_UTF8 STMT_START { \
239 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
240 } STMT_END
40d049e4 241
1de06328
YO
242/* About scan_data_t.
243
244 During optimisation we recurse through the regexp program performing
245 various inplace (keyhole style) optimisations. In addition study_chunk
246 and scan_commit populate this data structure with information about
247 what strings MUST appear in the pattern. We look for the longest
3b753521 248 string that must appear at a fixed location, and we look for the
1de06328
YO
249 longest string that may appear at a floating location. So for instance
250 in the pattern:
251
252 /FOO[xX]A.*B[xX]BAR/
253
254 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
255 strings (because they follow a .* construct). study_chunk will identify
256 both FOO and BAR as being the longest fixed and floating strings respectively.
257
258 The strings can be composites, for instance
259
260 /(f)(o)(o)/
261
262 will result in a composite fixed substring 'foo'.
263
264 For each string some basic information is maintained:
265
266 - offset or min_offset
267 This is the position the string must appear at, or not before.
268 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
269 characters must match before the string we are searching for.
270 Likewise when combined with minlenp and the length of the string it
1de06328
YO
271 tells us how many characters must appear after the string we have
272 found.
273
274 - max_offset
275 Only used for floating strings. This is the rightmost point that
3b753521 276 the string can appear at. If set to I32 max it indicates that the
1de06328
YO
277 string can occur infinitely far to the right.
278
279 - minlenp
280 A pointer to the minimum length of the pattern that the string
281 was found inside. This is important as in the case of positive
282 lookahead or positive lookbehind we can have multiple patterns
283 involved. Consider
284
285 /(?=FOO).*F/
286
287 The minimum length of the pattern overall is 3, the minimum length
288 of the lookahead part is 3, but the minimum length of the part that
289 will actually match is 1. So 'FOO's minimum length is 3, but the
290 minimum length for the F is 1. This is important as the minimum length
291 is used to determine offsets in front of and behind the string being
292 looked for. Since strings can be composites this is the length of the
486ec47a 293 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
294 the length is calculated by study_chunk, so that the minimum lengths
295 are not known until the full pattern has been compiled, thus the
296 pointer to the value.
297
298 - lookbehind
299
300 In the case of lookbehind the string being searched for can be
301 offset past the start point of the final matching string.
302 If this value was just blithely removed from the min_offset it would
303 invalidate some of the calculations for how many chars must match
304 before or after (as they are derived from min_offset and minlen and
305 the length of the string being searched for).
306 When the final pattern is compiled and the data is moved from the
307 scan_data_t structure into the regexp structure the information
308 about lookbehind is factored in, with the information that would
309 have been lost precalculated in the end_shift field for the
310 associated string.
311
312 The fields pos_min and pos_delta are used to store the minimum offset
313 and the delta to the maximum offset at the current point in the pattern.
314
315*/
2c2d71f5
JH
316
317typedef struct scan_data_t {
1de06328
YO
318 /*I32 len_min; unused */
319 /*I32 len_delta; unused */
2c2d71f5
JH
320 I32 pos_min;
321 I32 pos_delta;
322 SV *last_found;
1de06328 323 I32 last_end; /* min value, <0 unless valid. */
2c2d71f5
JH
324 I32 last_start_min;
325 I32 last_start_max;
1de06328
YO
326 SV **longest; /* Either &l_fixed, or &l_float. */
327 SV *longest_fixed; /* longest fixed string found in pattern */
328 I32 offset_fixed; /* offset where it starts */
486ec47a 329 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
330 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
331 SV *longest_float; /* longest floating string found in pattern */
332 I32 offset_float_min; /* earliest point in string it can appear */
333 I32 offset_float_max; /* latest point in string it can appear */
486ec47a 334 I32 *minlen_float; /* pointer to the minlen relevant to the string */
1de06328 335 I32 lookbehind_float; /* is the position of the string modified by LB */
2c2d71f5
JH
336 I32 flags;
337 I32 whilem_c;
cb434fcc 338 I32 *last_closep;
653099ff 339 struct regnode_charclass_class *start_class;
2c2d71f5
JH
340} scan_data_t;
341
a687059c 342/*
e50aee73 343 * Forward declarations for pregcomp()'s friends.
a687059c 344 */
a0d0e21e 345
27da23d5 346static const scan_data_t zero_scan_data =
1de06328 347 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
348
349#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
350#define SF_BEFORE_SEOL 0x0001
351#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
352#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
353#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
354
09b7f37c
CB
355#ifdef NO_UNARY_PLUS
356# define SF_FIX_SHIFT_EOL (0+2)
357# define SF_FL_SHIFT_EOL (0+4)
358#else
359# define SF_FIX_SHIFT_EOL (+2)
360# define SF_FL_SHIFT_EOL (+4)
361#endif
c277df42
IZ
362
363#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
364#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
365
366#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
367#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
368#define SF_IS_INF 0x0040
369#define SF_HAS_PAR 0x0080
370#define SF_IN_PAR 0x0100
371#define SF_HAS_EVAL 0x0200
372#define SCF_DO_SUBSTR 0x0400
653099ff
GS
373#define SCF_DO_STCLASS_AND 0x0800
374#define SCF_DO_STCLASS_OR 0x1000
375#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 376#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 377
786e8c11 378#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 379#define SCF_SEEN_ACCEPT 0x8000
07be1b83 380
43fead97 381#define UTF cBOOL(RExC_utf8)
a62b1201
KW
382#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
383#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
384#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
385#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
386#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1
KW
387#define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
388#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a62b1201 389
43fead97 390#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 391
ffc61ed2 392#define OOB_UNICODE 12345678
93733859 393#define OOB_NAMEDCLASS -1
b8c5462f 394
a0ed51b3
LW
395#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
396#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
397
8615cb43 398
b45f050a
JF
399/* length of regex to show in messages that don't mark a position within */
400#define RegexLengthToShowInErrorMessages 127
401
402/*
403 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
404 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
405 * op/pragma/warn/regcomp.
406 */
7253e4e3
RK
407#define MARKER1 "<-- HERE" /* marker as it appears in the description */
408#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 409
7253e4e3 410#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
411
412/*
413 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
414 * arg. Show regex, up to a maximum length. If it's too long, chop and add
415 * "...".
416 */
58e23c8d 417#define _FAIL(code) STMT_START { \
bfed75c6 418 const char *ellipses = ""; \
ccb2c380
MP
419 IV len = RExC_end - RExC_precomp; \
420 \
421 if (!SIZE_ONLY) \
288b8c02 422 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
423 if (len > RegexLengthToShowInErrorMessages) { \
424 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
425 len = RegexLengthToShowInErrorMessages - 10; \
426 ellipses = "..."; \
427 } \
58e23c8d 428 code; \
ccb2c380 429} STMT_END
8615cb43 430
58e23c8d
YO
431#define FAIL(msg) _FAIL( \
432 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
433 msg, (int)len, RExC_precomp, ellipses))
434
435#define FAIL2(msg,arg) _FAIL( \
436 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
437 arg, (int)len, RExC_precomp, ellipses))
438
b45f050a 439/*
b45f050a
JF
440 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
441 */
ccb2c380 442#define Simple_vFAIL(m) STMT_START { \
a28509cc 443 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
444 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
445 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
446} STMT_END
b45f050a
JF
447
448/*
449 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
450 */
ccb2c380
MP
451#define vFAIL(m) STMT_START { \
452 if (!SIZE_ONLY) \
288b8c02 453 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
454 Simple_vFAIL(m); \
455} STMT_END
b45f050a
JF
456
457/*
458 * Like Simple_vFAIL(), but accepts two arguments.
459 */
ccb2c380 460#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 461 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
462 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
463 (int)offset, RExC_precomp, RExC_precomp + offset); \
464} STMT_END
b45f050a
JF
465
466/*
467 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
468 */
ccb2c380
MP
469#define vFAIL2(m,a1) STMT_START { \
470 if (!SIZE_ONLY) \
288b8c02 471 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
472 Simple_vFAIL2(m, a1); \
473} STMT_END
b45f050a
JF
474
475
476/*
477 * Like Simple_vFAIL(), but accepts three arguments.
478 */
ccb2c380 479#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 480 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
481 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
482 (int)offset, RExC_precomp, RExC_precomp + offset); \
483} STMT_END
b45f050a
JF
484
485/*
486 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
487 */
ccb2c380
MP
488#define vFAIL3(m,a1,a2) STMT_START { \
489 if (!SIZE_ONLY) \
288b8c02 490 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
491 Simple_vFAIL3(m, a1, a2); \
492} STMT_END
b45f050a
JF
493
494/*
495 * Like Simple_vFAIL(), but accepts four arguments.
496 */
ccb2c380 497#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 498 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
499 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
500 (int)offset, RExC_precomp, RExC_precomp + offset); \
501} STMT_END
b45f050a 502
668c081a 503#define ckWARNreg(loc,m) STMT_START { \
a28509cc 504 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
505 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
506 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
507} STMT_END
508
668c081a 509#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 510 const IV offset = loc - RExC_precomp; \
d1d15184 511 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
512 m REPORT_LOCATION, \
513 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
514} STMT_END
515
2335b3d3
KW
516#define ckWARN2regdep(loc,m, a1) STMT_START { \
517 const IV offset = loc - RExC_precomp; \
518 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
519 m REPORT_LOCATION, \
520 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
521} STMT_END
522
668c081a 523#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 524 const IV offset = loc - RExC_precomp; \
668c081a 525 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
526 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
527} STMT_END
528
529#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 530 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
531 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
532 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
533} STMT_END
534
668c081a
NC
535#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
536 const IV offset = loc - RExC_precomp; \
537 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
538 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
539} STMT_END
540
ccb2c380 541#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 542 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
543 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
544 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
545} STMT_END
546
668c081a
NC
547#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
548 const IV offset = loc - RExC_precomp; \
549 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
550 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
551} STMT_END
552
ccb2c380 553#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 554 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
555 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
556 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
557} STMT_END
9d1d55b5 558
8615cb43 559
cd439c50 560/* Allow for side effects in s */
ccb2c380
MP
561#define REGC(c,s) STMT_START { \
562 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
563} STMT_END
cd439c50 564
fac92740
MJD
565/* Macros for recording node offsets. 20001227 mjd@plover.com
566 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
567 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
568 * Element 0 holds the number n.
07be1b83 569 * Position is 1 indexed.
fac92740 570 */
7122b237
YO
571#ifndef RE_TRACK_PATTERN_OFFSETS
572#define Set_Node_Offset_To_R(node,byte)
573#define Set_Node_Offset(node,byte)
574#define Set_Cur_Node_Offset
575#define Set_Node_Length_To_R(node,len)
576#define Set_Node_Length(node,len)
577#define Set_Node_Cur_Length(node)
578#define Node_Offset(n)
579#define Node_Length(n)
580#define Set_Node_Offset_Length(node,offset,len)
581#define ProgLen(ri) ri->u.proglen
582#define SetProgLen(ri,x) ri->u.proglen = x
583#else
584#define ProgLen(ri) ri->u.offsets[0]
585#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
586#define Set_Node_Offset_To_R(node,byte) STMT_START { \
587 if (! SIZE_ONLY) { \
588 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 589 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 590 if((node) < 0) { \
551405c4 591 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
592 } else { \
593 RExC_offsets[2*(node)-1] = (byte); \
594 } \
595 } \
596} STMT_END
597
598#define Set_Node_Offset(node,byte) \
599 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
600#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
601
602#define Set_Node_Length_To_R(node,len) STMT_START { \
603 if (! SIZE_ONLY) { \
604 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 605 __LINE__, (int)(node), (int)(len))); \
ccb2c380 606 if((node) < 0) { \
551405c4 607 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
608 } else { \
609 RExC_offsets[2*(node)] = (len); \
610 } \
611 } \
612} STMT_END
613
614#define Set_Node_Length(node,len) \
615 Set_Node_Length_To_R((node)-RExC_emit_start, len)
616#define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
617#define Set_Node_Cur_Length(node) \
618 Set_Node_Length(node, RExC_parse - parse_start)
fac92740
MJD
619
620/* Get offsets and lengths */
621#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
622#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
623
07be1b83
YO
624#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
625 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
626 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
627} STMT_END
7122b237 628#endif
07be1b83
YO
629
630#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
631#define EXPERIMENTAL_INPLACESCAN
f427392e 632#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 633
304ee84b
YO
634#define DEBUG_STUDYDATA(str,data,depth) \
635DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 636 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
637 "%*s" str "Pos:%"IVdf"/%"IVdf \
638 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
639 (int)(depth)*2, "", \
640 (IV)((data)->pos_min), \
641 (IV)((data)->pos_delta), \
304ee84b 642 (UV)((data)->flags), \
1de06328 643 (IV)((data)->whilem_c), \
304ee84b
YO
644 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
645 is_inf ? "INF " : "" \
1de06328
YO
646 ); \
647 if ((data)->last_found) \
648 PerlIO_printf(Perl_debug_log, \
649 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
650 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
651 SvPVX_const((data)->last_found), \
652 (IV)((data)->last_end), \
653 (IV)((data)->last_start_min), \
654 (IV)((data)->last_start_max), \
655 ((data)->longest && \
656 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
657 SvPVX_const((data)->longest_fixed), \
658 (IV)((data)->offset_fixed), \
659 ((data)->longest && \
660 (data)->longest==&((data)->longest_float)) ? "*" : "", \
661 SvPVX_const((data)->longest_float), \
662 (IV)((data)->offset_float_min), \
663 (IV)((data)->offset_float_max) \
664 ); \
665 PerlIO_printf(Perl_debug_log,"\n"); \
666});
667
acfe0abc 668static void clear_re(pTHX_ void *r);
4327152a 669
653099ff 670/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 671 Update the longest found anchored substring and the longest found
653099ff
GS
672 floating substrings if needed. */
673
4327152a 674STATIC void
304ee84b 675S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 676{
e1ec3a88
AL
677 const STRLEN l = CHR_SVLEN(data->last_found);
678 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 679 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 680
7918f24d
NC
681 PERL_ARGS_ASSERT_SCAN_COMMIT;
682
c277df42 683 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 684 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
685 if (*data->longest == data->longest_fixed) {
686 data->offset_fixed = l ? data->last_start_min : data->pos_min;
687 if (data->flags & SF_BEFORE_EOL)
b81d288d 688 data->flags
c277df42
IZ
689 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
690 else
691 data->flags &= ~SF_FIX_BEFORE_EOL;
1de06328
YO
692 data->minlen_fixed=minlenp;
693 data->lookbehind_fixed=0;
a0ed51b3 694 }
304ee84b 695 else { /* *data->longest == data->longest_float */
c277df42 696 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
697 data->offset_float_max = (l
698 ? data->last_start_max
c277df42 699 : data->pos_min + data->pos_delta);
304ee84b 700 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 701 data->offset_float_max = I32_MAX;
c277df42 702 if (data->flags & SF_BEFORE_EOL)
b81d288d 703 data->flags
c277df42
IZ
704 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
705 else
706 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
707 data->minlen_float=minlenp;
708 data->lookbehind_float=0;
c277df42
IZ
709 }
710 }
711 SvCUR_set(data->last_found, 0);
0eda9292 712 {
a28509cc 713 SV * const sv = data->last_found;
097eb12c
AL
714 if (SvUTF8(sv) && SvMAGICAL(sv)) {
715 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
716 if (mg)
717 mg->mg_len = 0;
718 }
0eda9292 719 }
c277df42
IZ
720 data->last_end = -1;
721 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 722 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
723}
724
653099ff
GS
725/* Can match anything (initialization) */
726STATIC void
3fffb88a 727S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 728{
7918f24d
NC
729 PERL_ARGS_ASSERT_CL_ANYTHING;
730
f8bef550 731 ANYOF_BITMAP_SETALL(cl);
dd58aee1 732 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
3ad98780 733 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
3fffb88a
KW
734
735 /* If any portion of the regex is to operate under locale rules,
736 * initialization includes it. The reason this isn't done for all regexes
737 * is that the optimizer was written under the assumption that locale was
738 * all-or-nothing. Given the complexity and lack of documentation in the
739 * optimizer, and that there are inadequate test cases for locale, so many
740 * parts of it may not work properly, it is safest to avoid locale unless
741 * necessary. */
742 if (RExC_contains_locale) {
9d7a1e63 743 ANYOF_CLASS_SETALL(cl); /* /l uses class */
3fffb88a
KW
744 cl->flags |= ANYOF_LOCALE;
745 }
9d7a1e63
KW
746 else {
747 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
748 }
653099ff
GS
749}
750
751/* Can match anything (initialization) */
752STATIC int
5f66b61c 753S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
754{
755 int value;
756
7918f24d
NC
757 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
758
aaa51d5e 759 for (value = 0; value <= ANYOF_MAX; value += 2)
653099ff
GS
760 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
761 return 1;
1aa99e6b
IH
762 if (!(cl->flags & ANYOF_UNICODE_ALL))
763 return 0;
10edeb5d 764 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 765 return 0;
653099ff
GS
766 return 1;
767}
768
769/* Can match anything (initialization) */
770STATIC void
e755fd73 771S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 772{
7918f24d
NC
773 PERL_ARGS_ASSERT_CL_INIT;
774
8ecf7187 775 Zero(cl, 1, struct regnode_charclass_class);
653099ff 776 cl->type = ANYOF;
3fffb88a 777 cl_anything(pRExC_state, cl);
1411dba4 778 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
779}
780
1051e1c4
KW
781/* These two functions currently do the exact same thing */
782#define cl_init_zero S_cl_init
653099ff 783
dd58aee1
KW
784/* 'AND' a given class with another one. Can create false positives. 'cl'
785 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
786 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 787STATIC void
5f66b61c 788S_cl_and(struct regnode_charclass_class *cl,
a28509cc 789 const struct regnode_charclass_class *and_with)
653099ff 790{
7918f24d 791 PERL_ARGS_ASSERT_CL_AND;
40d049e4
YO
792
793 assert(and_with->type == ANYOF);
1e6ade67 794
c6b76537 795 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
796 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
797 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 798 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
799 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
800 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
653099ff
GS
801 int i;
802
803 if (and_with->flags & ANYOF_INVERT)
804 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
805 cl->bitmap[i] &= ~and_with->bitmap[i];
806 else
807 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
808 cl->bitmap[i] &= and_with->bitmap[i];
809 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 810
c6b76537 811 if (and_with->flags & ANYOF_INVERT) {
8951c461 812
c6b76537
KW
813 /* Here, the and'ed node is inverted. Get the AND of the flags that
814 * aren't affected by the inversion. Those that are affected are
815 * handled individually below */
816 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
817 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
818 cl->flags |= affected_flags;
819
820 /* We currently don't know how to deal with things that aren't in the
821 * bitmap, but we know that the intersection is no greater than what
822 * is already in cl, so let there be false positives that get sorted
823 * out after the synthetic start class succeeds, and the node is
824 * matched for real. */
825
826 /* The inversion of these two flags indicate that the resulting
827 * intersection doesn't have them */
828 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
829 cl->flags &= ~ANYOF_UNICODE_ALL;
830 }
c6b76537
KW
831 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
832 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 833 }
1aa99e6b 834 }
c6b76537 835 else { /* and'd node is not inverted */
3ad98780
KW
836 U8 outside_bitmap_but_not_utf8; /* Temp variable */
837
137165a6 838 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
839
840 /* Here 'and_with' doesn't match anything outside the bitmap
841 * (except possibly ANYOF_UNICODE_ALL), which means the
842 * intersection can't either, except for ANYOF_UNICODE_ALL, in
843 * which case we don't know what the intersection is, but it's no
844 * greater than what cl already has, so can just leave it alone,
845 * with possible false positives */
846 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
847 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 848 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 849 }
137165a6 850 }
c6b76537
KW
851 else if (! ANYOF_NONBITMAP(cl)) {
852
853 /* Here, 'and_with' does match something outside the bitmap, and cl
854 * doesn't have a list of things to match outside the bitmap. If
855 * cl can match all code points above 255, the intersection will
3ad98780
KW
856 * be those above-255 code points that 'and_with' matches. If cl
857 * can't match all Unicode code points, it means that it can't
858 * match anything outside the bitmap (since the 'if' that got us
859 * into this block tested for that), so we leave the bitmap empty.
860 */
c6b76537
KW
861 if (cl->flags & ANYOF_UNICODE_ALL) {
862 ARG_SET(cl, ARG(and_with));
3ad98780
KW
863
864 /* and_with's ARG may match things that don't require UTF8.
865 * And now cl's will too, in spite of this being an 'and'. See
866 * the comments below about the kludge */
867 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
868 }
869 }
870 else {
871 /* Here, both 'and_with' and cl match something outside the
872 * bitmap. Currently we do not do the intersection, so just match
873 * whatever cl had at the beginning. */
874 }
875
876
3ad98780
KW
877 /* Take the intersection of the two sets of flags. However, the
878 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
879 * kludge around the fact that this flag is not treated like the others
880 * which are initialized in cl_anything(). The way the optimizer works
881 * is that the synthetic start class (SSC) is initialized to match
882 * anything, and then the first time a real node is encountered, its
883 * values are AND'd with the SSC's with the result being the values of
884 * the real node. However, there are paths through the optimizer where
885 * the AND never gets called, so those initialized bits are set
886 * inappropriately, which is not usually a big deal, as they just cause
887 * false positives in the SSC, which will just mean a probably
888 * imperceptible slow down in execution. However this bit has a
889 * higher false positive consequence in that it can cause utf8.pm,
890 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
891 * bigger slowdown and also causes significant extra memory to be used.
892 * In order to prevent this, the code now takes a different tack. The
893 * bit isn't set unless some part of the regular expression needs it,
894 * but once set it won't get cleared. This means that these extra
895 * modules won't get loaded unless there was some path through the
896 * pattern that would have required them anyway, and so any false
897 * positives that occur by not ANDing them out when they could be
898 * aren't as severe as they would be if we treated this bit like all
899 * the others */
900 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
901 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 902 cl->flags &= and_with->flags;
3ad98780 903 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 904 }
653099ff
GS
905}
906
dd58aee1
KW
907/* 'OR' a given class with another one. Can create false positives. 'cl'
908 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
909 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 910STATIC void
3fffb88a 911S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 912{
7918f24d
NC
913 PERL_ARGS_ASSERT_CL_OR;
914
653099ff 915 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
916
917 /* Here, the or'd node is to be inverted. This means we take the
918 * complement of everything not in the bitmap, but currently we don't
919 * know what that is, so give up and match anything */
920 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 921 cl_anything(pRExC_state, cl);
c6b76537 922 }
653099ff
GS
923 /* We do not use
924 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
925 * <= (B1 | !B2) | (CL1 | !CL2)
926 * which is wasteful if CL2 is small, but we ignore CL2:
927 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
928 * XXXX Can we handle case-fold? Unclear:
929 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
930 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
931 */
c6b76537 932 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
933 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
934 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
653099ff
GS
935 int i;
936
937 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
938 cl->bitmap[i] |= ~or_with->bitmap[i];
939 } /* XXXX: logic is complicated otherwise */
940 else {
3fffb88a 941 cl_anything(pRExC_state, cl);
653099ff 942 }
c6b76537
KW
943
944 /* And, we can just take the union of the flags that aren't affected
945 * by the inversion */
946 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
947
948 /* For the remaining flags:
949 ANYOF_UNICODE_ALL and inverted means to not match anything above
950 255, which means that the union with cl should just be
951 what cl has in it, so can ignore this flag
952 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
953 is 127-255 to match them, but then invert that, so the
954 union with cl should just be what cl has in it, so can
955 ignore this flag
956 */
957 } else { /* 'or_with' is not inverted */
653099ff
GS
958 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
959 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
960 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
961 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
653099ff
GS
962 int i;
963
964 /* OR char bitmap and class bitmap separately */
965 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
966 cl->bitmap[i] |= or_with->bitmap[i];
1e6ade67 967 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
653099ff
GS
968 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
969 cl->classflags[i] |= or_with->classflags[i];
970 cl->flags |= ANYOF_CLASS;
971 }
972 }
973 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 974 cl_anything(pRExC_state, cl);
653099ff 975 }
9826f543 976
c6b76537
KW
977 if (ANYOF_NONBITMAP(or_with)) {
978
979 /* Use the added node's outside-the-bit-map match if there isn't a
980 * conflict. If there is a conflict (both nodes match something
981 * outside the bitmap, but what they match outside is not the same
982 * pointer, and hence not easily compared until XXX we extend
983 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
984 * match everything outside the bitmap. If that stuff is all above
985 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
986 if (! ANYOF_NONBITMAP(cl)) {
987 ARG_SET(cl, ARG(or_with));
988 }
989 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
990
991 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
992 cl_anything(pRExC_state, cl);
993 }
994 else {
995 cl->flags |= ANYOF_UNICODE_ALL;
996 }
c6b76537 997 }
4c34a693 998 }
0b9668ee
KW
999
1000 /* Take the union */
1001 cl->flags |= or_with->flags;
1aa99e6b 1002 }
653099ff
GS
1003}
1004
a3621e74
YO
1005#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1006#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1007#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1008#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1009
3dab1dad
YO
1010
1011#ifdef DEBUGGING
07be1b83 1012/*
2b8b4781
NC
1013 dump_trie(trie,widecharmap,revcharmap)
1014 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1015 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1016
1017 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1018 The _interim_ variants are used for debugging the interim
1019 tables that are used to generate the final compressed
1020 representation which is what dump_trie expects.
1021
486ec47a 1022 Part of the reason for their existence is to provide a form
3dab1dad 1023 of documentation as to how the different representations function.
07be1b83
YO
1024
1025*/
3dab1dad
YO
1026
1027/*
3dab1dad
YO
1028 Dumps the final compressed table form of the trie to Perl_debug_log.
1029 Used for debugging make_trie().
1030*/
b9a59e08 1031
3dab1dad 1032STATIC void
2b8b4781
NC
1033S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1034 AV *revcharmap, U32 depth)
3dab1dad
YO
1035{
1036 U32 state;
ab3bbdeb 1037 SV *sv=sv_newmortal();
55eed653 1038 int colwidth= widecharmap ? 6 : 4;
2e64971a 1039 U16 word;
3dab1dad
YO
1040 GET_RE_DEBUG_FLAGS_DECL;
1041
7918f24d 1042 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1043
3dab1dad
YO
1044 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1045 (int)depth * 2 + 2,"",
1046 "Match","Base","Ofs" );
1047
1048 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1049 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1050 if ( tmp ) {
ab3bbdeb
YO
1051 PerlIO_printf( Perl_debug_log, "%*s",
1052 colwidth,
ddc5bc0f 1053 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1054 PL_colors[0], PL_colors[1],
1055 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1056 PERL_PV_ESCAPE_FIRSTCHAR
1057 )
1058 );
3dab1dad
YO
1059 }
1060 }
1061 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1062 (int)depth * 2 + 2,"");
1063
1064 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1065 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1066 PerlIO_printf( Perl_debug_log, "\n");
1067
1e2e3d02 1068 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1069 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1070
1071 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1072
1073 if ( trie->states[ state ].wordnum ) {
1074 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1075 } else {
1076 PerlIO_printf( Perl_debug_log, "%6s", "" );
1077 }
1078
1079 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1080
1081 if ( base ) {
1082 U32 ofs = 0;
1083
1084 while( ( base + ofs < trie->uniquecharcount ) ||
1085 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1086 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1087 ofs++;
1088
1089 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1090
1091 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1092 if ( ( base + ofs >= trie->uniquecharcount ) &&
1093 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1094 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1095 {
ab3bbdeb
YO
1096 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1097 colwidth,
3dab1dad
YO
1098 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1099 } else {
ab3bbdeb 1100 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1101 }
1102 }
1103
1104 PerlIO_printf( Perl_debug_log, "]");
1105
1106 }
1107 PerlIO_printf( Perl_debug_log, "\n" );
1108 }
2e64971a
DM
1109 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1110 for (word=1; word <= trie->wordcount; word++) {
1111 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1112 (int)word, (int)(trie->wordinfo[word].prev),
1113 (int)(trie->wordinfo[word].len));
1114 }
1115 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1116}
1117/*
3dab1dad
YO
1118 Dumps a fully constructed but uncompressed trie in list form.
1119 List tries normally only are used for construction when the number of
1120 possible chars (trie->uniquecharcount) is very high.
1121 Used for debugging make_trie().
1122*/
1123STATIC void
55eed653 1124S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1125 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1126 U32 depth)
3dab1dad
YO
1127{
1128 U32 state;
ab3bbdeb 1129 SV *sv=sv_newmortal();
55eed653 1130 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1131 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1132
1133 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1134
3dab1dad 1135 /* print out the table precompression. */
ab3bbdeb
YO
1136 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1137 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1138 "------:-----+-----------------\n" );
3dab1dad
YO
1139
1140 for( state=1 ; state < next_alloc ; state ++ ) {
1141 U16 charid;
1142
ab3bbdeb 1143 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1144 (int)depth * 2 + 2,"", (UV)state );
1145 if ( ! trie->states[ state ].wordnum ) {
1146 PerlIO_printf( Perl_debug_log, "%5s| ","");
1147 } else {
1148 PerlIO_printf( Perl_debug_log, "W%4x| ",
1149 trie->states[ state ].wordnum
1150 );
1151 }
1152 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1153 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1154 if ( tmp ) {
1155 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1156 colwidth,
ddc5bc0f 1157 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1158 PL_colors[0], PL_colors[1],
1159 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1160 PERL_PV_ESCAPE_FIRSTCHAR
1161 ) ,
1e2e3d02
YO
1162 TRIE_LIST_ITEM(state,charid).forid,
1163 (UV)TRIE_LIST_ITEM(state,charid).newstate
1164 );
1165 if (!(charid % 10))
664e119d
RGS
1166 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1167 (int)((depth * 2) + 14), "");
1e2e3d02 1168 }
ab3bbdeb
YO
1169 }
1170 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1171 }
1172}
1173
1174/*
3dab1dad
YO
1175 Dumps a fully constructed but uncompressed trie in table form.
1176 This is the normal DFA style state transition table, with a few
1177 twists to facilitate compression later.
1178 Used for debugging make_trie().
1179*/
1180STATIC void
55eed653 1181S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1182 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1183 U32 depth)
3dab1dad
YO
1184{
1185 U32 state;
1186 U16 charid;
ab3bbdeb 1187 SV *sv=sv_newmortal();
55eed653 1188 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1189 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1190
1191 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1192
1193 /*
1194 print out the table precompression so that we can do a visual check
1195 that they are identical.
1196 */
1197
1198 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1199
1200 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1201 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1202 if ( tmp ) {
ab3bbdeb
YO
1203 PerlIO_printf( Perl_debug_log, "%*s",
1204 colwidth,
ddc5bc0f 1205 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1206 PL_colors[0], PL_colors[1],
1207 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1208 PERL_PV_ESCAPE_FIRSTCHAR
1209 )
1210 );
3dab1dad
YO
1211 }
1212 }
1213
1214 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1215
1216 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1217 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1218 }
1219
1220 PerlIO_printf( Perl_debug_log, "\n" );
1221
1222 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1223
1224 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1225 (int)depth * 2 + 2,"",
1226 (UV)TRIE_NODENUM( state ) );
1227
1228 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1229 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1230 if (v)
1231 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1232 else
1233 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1234 }
1235 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1236 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1237 } else {
1238 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1239 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1240 }
1241 }
07be1b83 1242}
3dab1dad
YO
1243
1244#endif
1245
2e64971a 1246
786e8c11
YO
1247/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1248 startbranch: the first branch in the whole branch sequence
1249 first : start branch of sequence of branch-exact nodes.
1250 May be the same as startbranch
1251 last : Thing following the last branch.
1252 May be the same as tail.
1253 tail : item following the branch sequence
1254 count : words in the sequence
1255 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1256 depth : indent depth
3dab1dad 1257
786e8c11 1258Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1259
786e8c11
YO
1260A trie is an N'ary tree where the branches are determined by digital
1261decomposition of the key. IE, at the root node you look up the 1st character and
1262follow that branch repeat until you find the end of the branches. Nodes can be
1263marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1264
786e8c11 1265 /he|she|his|hers/
72f13be8 1266
786e8c11
YO
1267would convert into the following structure. Numbers represent states, letters
1268following numbers represent valid transitions on the letter from that state, if
1269the number is in square brackets it represents an accepting state, otherwise it
1270will be in parenthesis.
07be1b83 1271
786e8c11
YO
1272 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1273 | |
1274 | (2)
1275 | |
1276 (1) +-i->(6)-+-s->[7]
1277 |
1278 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1279
786e8c11
YO
1280 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1281
1282This shows that when matching against the string 'hers' we will begin at state 1
1283read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1284then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1285is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1286single traverse. We store a mapping from accepting to state to which word was
1287matched, and then when we have multiple possibilities we try to complete the
1288rest of the regex in the order in which they occured in the alternation.
1289
1290The only prior NFA like behaviour that would be changed by the TRIE support is
1291the silent ignoring of duplicate alternations which are of the form:
1292
1293 / (DUPE|DUPE) X? (?{ ... }) Y /x
1294
4b714af6 1295Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1296and without the optimisation. With the optimisations dupes will be silently
486ec47a 1297ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1298the following demonstrates:
1299
1300 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1301
1302which prints out 'word' three times, but
1303
1304 'words'=~/(word|word|word)(?{ print $1 })S/
1305
1306which doesnt print it out at all. This is due to other optimisations kicking in.
1307
1308Example of what happens on a structural level:
1309
486ec47a 1310The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1311
1312 1: CURLYM[1] {1,32767}(18)
1313 5: BRANCH(8)
1314 6: EXACT <ac>(16)
1315 8: BRANCH(11)
1316 9: EXACT <ad>(16)
1317 11: BRANCH(14)
1318 12: EXACT <ab>(16)
1319 16: SUCCEED(0)
1320 17: NOTHING(18)
1321 18: END(0)
1322
1323This would be optimizable with startbranch=5, first=5, last=16, tail=16
1324and should turn into:
1325
1326 1: CURLYM[1] {1,32767}(18)
1327 5: TRIE(16)
1328 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1329 <ac>
1330 <ad>
1331 <ab>
1332 16: SUCCEED(0)
1333 17: NOTHING(18)
1334 18: END(0)
1335
1336Cases where tail != last would be like /(?foo|bar)baz/:
1337
1338 1: BRANCH(4)
1339 2: EXACT <foo>(8)
1340 4: BRANCH(7)
1341 5: EXACT <bar>(8)
1342 7: TAIL(8)
1343 8: EXACT <baz>(10)
1344 10: END(0)
1345
1346which would be optimizable with startbranch=1, first=1, last=7, tail=8
1347and would end up looking like:
1348
1349 1: TRIE(8)
1350 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1351 <foo>
1352 <bar>
1353 7: TAIL(8)
1354 8: EXACT <baz>(10)
1355 10: END(0)
1356
1357 d = uvuni_to_utf8_flags(d, uv, 0);
1358
1359is the recommended Unicode-aware way of saying
1360
1361 *(d++) = uv;
1362*/
1363
1e2e3d02 1364#define TRIE_STORE_REVCHAR \
786e8c11 1365 STMT_START { \
73031816
NC
1366 if (UTF) { \
1367 SV *zlopp = newSV(2); \
88c9ea1e
CB
1368 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1369 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, uvc & 0xFF); \
73031816
NC
1370 SvCUR_set(zlopp, kapow - flrbbbbb); \
1371 SvPOK_on(zlopp); \
1372 SvUTF8_on(zlopp); \
1373 av_push(revcharmap, zlopp); \
1374 } else { \
6bdeddd2 1375 char ooooff = (char)uvc; \
73031816
NC
1376 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1377 } \
1378 } STMT_END
786e8c11
YO
1379
1380#define TRIE_READ_CHAR STMT_START { \
1381 wordlen++; \
1382 if ( UTF ) { \
1383 if ( folder ) { \
1384 if ( foldlen > 0 ) { \
1385 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1386 foldlen -= len; \
1387 scan += len; \
1388 len = 0; \
1389 } else { \
1390 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1391 uvc = to_uni_fold( uvc, foldbuf, &foldlen ); \
1392 foldlen -= UNISKIP( uvc ); \
1393 scan = foldbuf + UNISKIP( uvc ); \
1394 } \
1395 } else { \
1396 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1397 } \
1398 } else { \
1399 uvc = (U32)*uc; \
1400 len = 1; \
1401 } \
1402} STMT_END
1403
1404
1405
1406#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1407 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1408 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1409 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1410 } \
1411 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1412 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1413 TRIE_LIST_CUR( state )++; \
1414} STMT_END
07be1b83 1415
786e8c11
YO
1416#define TRIE_LIST_NEW(state) STMT_START { \
1417 Newxz( trie->states[ state ].trans.list, \
1418 4, reg_trie_trans_le ); \
1419 TRIE_LIST_CUR( state ) = 1; \
1420 TRIE_LIST_LEN( state ) = 4; \
1421} STMT_END
07be1b83 1422
786e8c11
YO
1423#define TRIE_HANDLE_WORD(state) STMT_START { \
1424 U16 dupe= trie->states[ state ].wordnum; \
1425 regnode * const noper_next = regnext( noper ); \
1426 \
786e8c11
YO
1427 DEBUG_r({ \
1428 /* store the word for dumping */ \
1429 SV* tmp; \
1430 if (OP(noper) != NOTHING) \
740cce10 1431 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1432 else \
740cce10 1433 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1434 av_push( trie_words, tmp ); \
786e8c11
YO
1435 }); \
1436 \
1437 curword++; \
2e64971a
DM
1438 trie->wordinfo[curword].prev = 0; \
1439 trie->wordinfo[curword].len = wordlen; \
1440 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1441 \
1442 if ( noper_next < tail ) { \
1443 if (!trie->jump) \
c944940b 1444 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1445 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1446 if (!jumper) \
1447 jumper = noper_next; \
1448 if (!nextbranch) \
1449 nextbranch= regnext(cur); \
1450 } \
1451 \
1452 if ( dupe ) { \
2e64971a
DM
1453 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1454 /* chain, so that when the bits of chain are later */\
1455 /* linked together, the dups appear in the chain */\
1456 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1457 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1458 } else { \
1459 /* we haven't inserted this word yet. */ \
1460 trie->states[ state ].wordnum = curword; \
1461 } \
1462} STMT_END
07be1b83 1463
3dab1dad 1464
786e8c11
YO
1465#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1466 ( ( base + charid >= ucharcount \
1467 && base + charid < ubound \
1468 && state == trie->trans[ base - ucharcount + charid ].check \
1469 && trie->trans[ base - ucharcount + charid ].next ) \
1470 ? trie->trans[ base - ucharcount + charid ].next \
1471 : ( state==1 ? special : 0 ) \
1472 )
3dab1dad 1473
786e8c11
YO
1474#define MADE_TRIE 1
1475#define MADE_JUMP_TRIE 2
1476#define MADE_EXACT_TRIE 4
3dab1dad 1477
a3621e74 1478STATIC I32
786e8c11 1479S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1480{
27da23d5 1481 dVAR;
a3621e74
YO
1482 /* first pass, loop through and scan words */
1483 reg_trie_data *trie;
55eed653 1484 HV *widecharmap = NULL;
2b8b4781 1485 AV *revcharmap = newAV();
a3621e74 1486 regnode *cur;
9f7f3913 1487 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1488 STRLEN len = 0;
1489 UV uvc = 0;
1490 U16 curword = 0;
1491 U32 next_alloc = 0;
786e8c11
YO
1492 regnode *jumper = NULL;
1493 regnode *nextbranch = NULL;
7f69552c 1494 regnode *convert = NULL;
2e64971a 1495 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1496 /* we just use folder as a flag in utf8 */
1e696034 1497 const U8 * folder = NULL;
a3621e74 1498
2b8b4781
NC
1499#ifdef DEBUGGING
1500 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1501 AV *trie_words = NULL;
1502 /* along with revcharmap, this only used during construction but both are
1503 * useful during debugging so we store them in the struct when debugging.
8e11feef 1504 */
2b8b4781
NC
1505#else
1506 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1507 STRLEN trie_charcount=0;
3dab1dad 1508#endif
2b8b4781 1509 SV *re_trie_maxbuff;
a3621e74 1510 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1511
1512 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1513#ifndef DEBUGGING
1514 PERL_UNUSED_ARG(depth);
1515#endif
a3621e74 1516
1e696034 1517 switch (flags) {
2f7f8cb1 1518 case EXACTFA:
1e696034
KW
1519 case EXACTFU: folder = PL_fold_latin1; break;
1520 case EXACTF: folder = PL_fold; break;
1521 case EXACTFL: folder = PL_fold_locale; break;
1522 }
1523
c944940b 1524 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1525 trie->refcount = 1;
3dab1dad 1526 trie->startstate = 1;
786e8c11 1527 trie->wordcount = word_count;
f8fc2ecf 1528 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1529 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
3dab1dad 1530 if (!(UTF && folder))
c944940b 1531 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1532 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1533 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1534
a3621e74 1535 DEBUG_r({
2b8b4781 1536 trie_words = newAV();
a3621e74 1537 });
a3621e74 1538
0111c4fd 1539 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1540 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1541 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1542 }
3dab1dad
YO
1543 DEBUG_OPTIMISE_r({
1544 PerlIO_printf( Perl_debug_log,
786e8c11 1545 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1546 (int)depth * 2 + 2, "",
1547 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1548 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1549 (int)depth);
3dab1dad 1550 });
7f69552c
YO
1551
1552 /* Find the node we are going to overwrite */
1553 if ( first == startbranch && OP( last ) != BRANCH ) {
1554 /* whole branch chain */
1555 convert = first;
1556 } else {
1557 /* branch sub-chain */
1558 convert = NEXTOPER( first );
1559 }
1560
a3621e74
YO
1561 /* -- First loop and Setup --
1562
1563 We first traverse the branches and scan each word to determine if it
1564 contains widechars, and how many unique chars there are, this is
1565 important as we have to build a table with at least as many columns as we
1566 have unique chars.
1567
1568 We use an array of integers to represent the character codes 0..255
38a44b82 1569 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1570 native representation of the character value as the key and IV's for the
1571 coded index.
1572
1573 *TODO* If we keep track of how many times each character is used we can
1574 remap the columns so that the table compression later on is more
3b753521 1575 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1576 middle and the least common are on the outside. IMO this would be better
1577 than a most to least common mapping as theres a decent chance the most
1578 common letter will share a node with the least common, meaning the node
486ec47a 1579 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1580 case is when we have the least common nodes twice.
1581
1582 */
1583
a3621e74 1584 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
c445ea15 1585 regnode * const noper = NEXTOPER( cur );
e1ec3a88 1586 const U8 *uc = (U8*)STRING( noper );
a28509cc 1587 const U8 * const e = uc + STR_LEN( noper );
a3621e74
YO
1588 STRLEN foldlen = 0;
1589 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2af232bd 1590 const U8 *scan = (U8*)NULL;
07be1b83 1591 U32 wordlen = 0; /* required init */
02daf0ab
YO
1592 STRLEN chars = 0;
1593 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1594
3dab1dad
YO
1595 if (OP(noper) == NOTHING) {
1596 trie->minlen= 0;
1597 continue;
1598 }
02daf0ab
YO
1599 if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
1600 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1601 regardless of encoding */
1602
a3621e74 1603 for ( ; uc < e ; uc += len ) {
3dab1dad 1604 TRIE_CHARCOUNT(trie)++;
a3621e74 1605 TRIE_READ_CHAR;
3dab1dad 1606 chars++;
a3621e74
YO
1607 if ( uvc < 256 ) {
1608 if ( !trie->charmap[ uvc ] ) {
1609 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1610 if ( folder )
1611 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
3dab1dad 1612 TRIE_STORE_REVCHAR;
a3621e74 1613 }
02daf0ab 1614 if ( set_bit ) {
62012aee
KW
1615 /* store the codepoint in the bitmap, and its folded
1616 * equivalent. */
02daf0ab 1617 TRIE_BITMAP_SET(trie,uvc);
0921ee73
T
1618
1619 /* store the folded codepoint */
1620 if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);
1621
1622 if ( !UTF ) {
1623 /* store first byte of utf8 representation of
acdf4139
KW
1624 variant codepoints */
1625 if (! UNI_IS_INVARIANT(uvc)) {
1626 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1627 }
1628 }
02daf0ab
YO
1629 set_bit = 0; /* We've done our bit :-) */
1630 }
a3621e74
YO
1631 } else {
1632 SV** svpp;
55eed653
NC
1633 if ( !widecharmap )
1634 widecharmap = newHV();
a3621e74 1635
55eed653 1636 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1637
1638 if ( !svpp )
e4584336 1639 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1640
1641 if ( !SvTRUE( *svpp ) ) {
1642 sv_setiv( *svpp, ++trie->uniquecharcount );
3dab1dad 1643 TRIE_STORE_REVCHAR;
a3621e74
YO
1644 }
1645 }
1646 }
3dab1dad
YO
1647 if( cur == first ) {
1648 trie->minlen=chars;
1649 trie->maxlen=chars;
1650 } else if (chars < trie->minlen) {
1651 trie->minlen=chars;
1652 } else if (chars > trie->maxlen) {
1653 trie->maxlen=chars;
1654 }
1655
a3621e74
YO
1656 } /* end first pass */
1657 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1658 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1659 (int)depth * 2 + 2,"",
55eed653 1660 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1661 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1662 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1663 );
a3621e74
YO
1664
1665 /*
1666 We now know what we are dealing with in terms of unique chars and
1667 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1668 representation using a flat table will take. If it's over a reasonable
1669 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1670 conservative but potentially much slower representation using an array
1671 of lists.
1672
1673 At the end we convert both representations into the same compressed
1674 form that will be used in regexec.c for matching with. The latter
1675 is a form that cannot be used to construct with but has memory
1676 properties similar to the list form and access properties similar
1677 to the table form making it both suitable for fast searches and
1678 small enough that its feasable to store for the duration of a program.
1679
1680 See the comment in the code where the compressed table is produced
1681 inplace from the flat tabe representation for an explanation of how
1682 the compression works.
1683
1684 */
1685
1686
2e64971a
DM
1687 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1688 prev_states[1] = 0;
1689
3dab1dad 1690 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1691 /*
1692 Second Pass -- Array Of Lists Representation
1693
1694 Each state will be represented by a list of charid:state records
1695 (reg_trie_trans_le) the first such element holds the CUR and LEN
1696 points of the allocated array. (See defines above).
1697
1698 We build the initial structure using the lists, and then convert
1699 it into the compressed table form which allows faster lookups
1700 (but cant be modified once converted).
a3621e74
YO
1701 */
1702
a3621e74
YO
1703 STRLEN transcount = 1;
1704
1e2e3d02
YO
1705 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1706 "%*sCompiling trie using list compiler\n",
1707 (int)depth * 2 + 2, ""));
446bd890 1708
c944940b
JH
1709 trie->states = (reg_trie_state *)
1710 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1711 sizeof(reg_trie_state) );
a3621e74
YO
1712 TRIE_LIST_NEW(1);
1713 next_alloc = 2;
1714
1715 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1716
c445ea15
AL
1717 regnode * const noper = NEXTOPER( cur );
1718 U8 *uc = (U8*)STRING( noper );
1719 const U8 * const e = uc + STR_LEN( noper );
1720 U32 state = 1; /* required init */
1721 U16 charid = 0; /* sanity init */
1722 U8 *scan = (U8*)NULL; /* sanity init */
1723 STRLEN foldlen = 0; /* required init */
07be1b83 1724 U32 wordlen = 0; /* required init */
c445ea15
AL
1725 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1726
3dab1dad 1727 if (OP(noper) != NOTHING) {
786e8c11 1728 for ( ; uc < e ; uc += len ) {
c445ea15 1729
786e8c11 1730 TRIE_READ_CHAR;
c445ea15 1731
786e8c11
YO
1732 if ( uvc < 256 ) {
1733 charid = trie->charmap[ uvc ];
c445ea15 1734 } else {
55eed653 1735 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1736 if ( !svpp ) {
1737 charid = 0;
1738 } else {
1739 charid=(U16)SvIV( *svpp );
1740 }
c445ea15 1741 }
786e8c11
YO
1742 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1743 if ( charid ) {
a3621e74 1744
786e8c11
YO
1745 U16 check;
1746 U32 newstate = 0;
a3621e74 1747
786e8c11
YO
1748 charid--;
1749 if ( !trie->states[ state ].trans.list ) {
1750 TRIE_LIST_NEW( state );
c445ea15 1751 }
786e8c11
YO
1752 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1753 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1754 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1755 break;
1756 }
1757 }
1758 if ( ! newstate ) {
1759 newstate = next_alloc++;
2e64971a 1760 prev_states[newstate] = state;
786e8c11
YO
1761 TRIE_LIST_PUSH( state, charid, newstate );
1762 transcount++;
1763 }
1764 state = newstate;
1765 } else {
1766 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1767 }
a28509cc 1768 }
c445ea15 1769 }
3dab1dad 1770 TRIE_HANDLE_WORD(state);
a3621e74
YO
1771
1772 } /* end second pass */
1773
1e2e3d02
YO
1774 /* next alloc is the NEXT state to be allocated */
1775 trie->statecount = next_alloc;
c944940b
JH
1776 trie->states = (reg_trie_state *)
1777 PerlMemShared_realloc( trie->states,
1778 next_alloc
1779 * sizeof(reg_trie_state) );
a3621e74 1780
3dab1dad 1781 /* and now dump it out before we compress it */
2b8b4781
NC
1782 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1783 revcharmap, next_alloc,
1784 depth+1)
1e2e3d02 1785 );
a3621e74 1786
c944940b
JH
1787 trie->trans = (reg_trie_trans *)
1788 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1789 {
1790 U32 state;
a3621e74
YO
1791 U32 tp = 0;
1792 U32 zp = 0;
1793
1794
1795 for( state=1 ; state < next_alloc ; state ++ ) {
1796 U32 base=0;
1797
1798 /*
1799 DEBUG_TRIE_COMPILE_MORE_r(
1800 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1801 );
1802 */
1803
1804 if (trie->states[state].trans.list) {
1805 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1806 U16 maxid=minid;
a28509cc 1807 U16 idx;
a3621e74
YO
1808
1809 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1810 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1811 if ( forid < minid ) {
1812 minid=forid;
1813 } else if ( forid > maxid ) {
1814 maxid=forid;
1815 }
a3621e74
YO
1816 }
1817 if ( transcount < tp + maxid - minid + 1) {
1818 transcount *= 2;
c944940b
JH
1819 trie->trans = (reg_trie_trans *)
1820 PerlMemShared_realloc( trie->trans,
446bd890
NC
1821 transcount
1822 * sizeof(reg_trie_trans) );
a3621e74
YO
1823 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1824 }
1825 base = trie->uniquecharcount + tp - minid;
1826 if ( maxid == minid ) {
1827 U32 set = 0;
1828 for ( ; zp < tp ; zp++ ) {
1829 if ( ! trie->trans[ zp ].next ) {
1830 base = trie->uniquecharcount + zp - minid;
1831 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1832 trie->trans[ zp ].check = state;
1833 set = 1;
1834 break;
1835 }
1836 }
1837 if ( !set ) {
1838 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1839 trie->trans[ tp ].check = state;
1840 tp++;
1841 zp = tp;
1842 }
1843 } else {
1844 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1845 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1846 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1847 trie->trans[ tid ].check = state;
1848 }
1849 tp += ( maxid - minid + 1 );
1850 }
1851 Safefree(trie->states[ state ].trans.list);
1852 }
1853 /*
1854 DEBUG_TRIE_COMPILE_MORE_r(
1855 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1856 );
1857 */
1858 trie->states[ state ].trans.base=base;
1859 }
cc601c31 1860 trie->lasttrans = tp + 1;
a3621e74
YO
1861 }
1862 } else {
1863 /*
1864 Second Pass -- Flat Table Representation.
1865
1866 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1867 We know that we will need Charcount+1 trans at most to store the data
1868 (one row per char at worst case) So we preallocate both structures
1869 assuming worst case.
1870
1871 We then construct the trie using only the .next slots of the entry
1872 structs.
1873
3b753521 1874 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1875 make compression both faster and easier by keeping track of how many non
1876 zero fields are in the node.
1877
1878 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1879 transition.
1880
1881 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1882 number representing the first entry of the node, and state as a
1883 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1884 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1885 are 2 entrys per node. eg:
1886
1887 A B A B
1888 1. 2 4 1. 3 7
1889 2. 0 3 3. 0 5
1890 3. 0 0 5. 0 0
1891 4. 0 0 7. 0 0
1892
1893 The table is internally in the right hand, idx form. However as we also
1894 have to deal with the states array which is indexed by nodenum we have to
1895 use TRIE_NODENUM() to convert.
1896
1897 */
1e2e3d02
YO
1898 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1899 "%*sCompiling trie using table compiler\n",
1900 (int)depth * 2 + 2, ""));
3dab1dad 1901
c944940b
JH
1902 trie->trans = (reg_trie_trans *)
1903 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1904 * trie->uniquecharcount + 1,
1905 sizeof(reg_trie_trans) );
1906 trie->states = (reg_trie_state *)
1907 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1908 sizeof(reg_trie_state) );
a3621e74
YO
1909 next_alloc = trie->uniquecharcount + 1;
1910
3dab1dad 1911
a3621e74
YO
1912 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1913
c445ea15 1914 regnode * const noper = NEXTOPER( cur );
a28509cc
AL
1915 const U8 *uc = (U8*)STRING( noper );
1916 const U8 * const e = uc + STR_LEN( noper );
a3621e74
YO
1917
1918 U32 state = 1; /* required init */
1919
1920 U16 charid = 0; /* sanity init */
1921 U32 accept_state = 0; /* sanity init */
1922 U8 *scan = (U8*)NULL; /* sanity init */
1923
1924 STRLEN foldlen = 0; /* required init */
07be1b83 1925 U32 wordlen = 0; /* required init */
a3621e74
YO
1926 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1927
3dab1dad 1928 if ( OP(noper) != NOTHING ) {
786e8c11 1929 for ( ; uc < e ; uc += len ) {
a3621e74 1930
786e8c11 1931 TRIE_READ_CHAR;
a3621e74 1932
786e8c11
YO
1933 if ( uvc < 256 ) {
1934 charid = trie->charmap[ uvc ];
1935 } else {
55eed653 1936 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 1937 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 1938 }
786e8c11
YO
1939 if ( charid ) {
1940 charid--;
1941 if ( !trie->trans[ state + charid ].next ) {
1942 trie->trans[ state + charid ].next = next_alloc;
1943 trie->trans[ state ].check++;
2e64971a
DM
1944 prev_states[TRIE_NODENUM(next_alloc)]
1945 = TRIE_NODENUM(state);
786e8c11
YO
1946 next_alloc += trie->uniquecharcount;
1947 }
1948 state = trie->trans[ state + charid ].next;
1949 } else {
1950 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1951 }
1952 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 1953 }
a3621e74 1954 }
3dab1dad
YO
1955 accept_state = TRIE_NODENUM( state );
1956 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
1957
1958 } /* end second pass */
1959
3dab1dad 1960 /* and now dump it out before we compress it */
2b8b4781
NC
1961 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1962 revcharmap,
1963 next_alloc, depth+1));
a3621e74 1964
a3621e74
YO
1965 {
1966 /*
1967 * Inplace compress the table.*
1968
1969 For sparse data sets the table constructed by the trie algorithm will
1970 be mostly 0/FAIL transitions or to put it another way mostly empty.
1971 (Note that leaf nodes will not contain any transitions.)
1972
1973 This algorithm compresses the tables by eliminating most such
1974 transitions, at the cost of a modest bit of extra work during lookup:
1975
1976 - Each states[] entry contains a .base field which indicates the
1977 index in the state[] array wheres its transition data is stored.
1978
3b753521 1979 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
1980
1981 - If .base is nonzero then charid is added to it to find an entry in
1982 the trans array.
1983
1984 -If trans[states[state].base+charid].check!=state then the
1985 transition is taken to be a 0/Fail transition. Thus if there are fail
1986 transitions at the front of the node then the .base offset will point
1987 somewhere inside the previous nodes data (or maybe even into a node
1988 even earlier), but the .check field determines if the transition is
1989 valid.
1990
786e8c11 1991 XXX - wrong maybe?
a3621e74 1992 The following process inplace converts the table to the compressed
3b753521 1993 table: We first do not compress the root node 1,and mark all its
a3621e74 1994 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
1995 allows us to do a DFA construction from the compressed table later,
1996 and ensures that any .base pointers we calculate later are greater
1997 than 0.
a3621e74
YO
1998
1999 - We set 'pos' to indicate the first entry of the second node.
2000
2001 - We then iterate over the columns of the node, finding the first and
2002 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2003 and set the .check pointers accordingly, and advance pos
2004 appropriately and repreat for the next node. Note that when we copy
2005 the next pointers we have to convert them from the original
2006 NODEIDX form to NODENUM form as the former is not valid post
2007 compression.
2008
2009 - If a node has no transitions used we mark its base as 0 and do not
2010 advance the pos pointer.
2011
2012 - If a node only has one transition we use a second pointer into the
2013 structure to fill in allocated fail transitions from other states.
2014 This pointer is independent of the main pointer and scans forward
2015 looking for null transitions that are allocated to a state. When it
2016 finds one it writes the single transition into the "hole". If the
786e8c11 2017 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2018
2019 - Once compressed we can Renew/realloc the structures to release the
2020 excess space.
2021
2022 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2023 specifically Fig 3.47 and the associated pseudocode.
2024
2025 demq
2026 */
a3b680e6 2027 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2028 U32 state, charid;
a3621e74 2029 U32 pos = 0, zp=0;
1e2e3d02 2030 trie->statecount = laststate;
a3621e74
YO
2031
2032 for ( state = 1 ; state < laststate ; state++ ) {
2033 U8 flag = 0;
a28509cc
AL
2034 const U32 stateidx = TRIE_NODEIDX( state );
2035 const U32 o_used = trie->trans[ stateidx ].check;
2036 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2037 trie->trans[ stateidx ].check = 0;
2038
2039 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2040 if ( flag || trie->trans[ stateidx + charid ].next ) {
2041 if ( trie->trans[ stateidx + charid ].next ) {
2042 if (o_used == 1) {
2043 for ( ; zp < pos ; zp++ ) {
2044 if ( ! trie->trans[ zp ].next ) {
2045 break;
2046 }
2047 }
2048 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2049 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2050 trie->trans[ zp ].check = state;
2051 if ( ++zp > pos ) pos = zp;
2052 break;
2053 }
2054 used--;
2055 }
2056 if ( !flag ) {
2057 flag = 1;
2058 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2059 }
2060 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2061 trie->trans[ pos ].check = state;
2062 pos++;
2063 }
2064 }
2065 }
cc601c31 2066 trie->lasttrans = pos + 1;
c944940b
JH
2067 trie->states = (reg_trie_state *)
2068 PerlMemShared_realloc( trie->states, laststate
2069 * sizeof(reg_trie_state) );
a3621e74 2070 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2071 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2072 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2073 (int)depth * 2 + 2,"",
2074 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2075 (IV)next_alloc,
2076 (IV)pos,
a3621e74
YO
2077 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2078 );
2079
2080 } /* end table compress */
2081 }
1e2e3d02
YO
2082 DEBUG_TRIE_COMPILE_MORE_r(
2083 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2084 (int)depth * 2 + 2, "",
2085 (UV)trie->statecount,
2086 (UV)trie->lasttrans)
2087 );
cc601c31 2088 /* resize the trans array to remove unused space */
c944940b
JH
2089 trie->trans = (reg_trie_trans *)
2090 PerlMemShared_realloc( trie->trans, trie->lasttrans
2091 * sizeof(reg_trie_trans) );
a3621e74 2092
3b753521 2093 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2094 U8 nodetype =(U8)(flags & 0xFF);
2095 char *str=NULL;
786e8c11 2096
07be1b83 2097#ifdef DEBUGGING
e62cc96a 2098 regnode *optimize = NULL;
7122b237
YO
2099#ifdef RE_TRACK_PATTERN_OFFSETS
2100
b57a0404
JH
2101 U32 mjd_offset = 0;
2102 U32 mjd_nodelen = 0;
7122b237
YO
2103#endif /* RE_TRACK_PATTERN_OFFSETS */
2104#endif /* DEBUGGING */
a3621e74 2105 /*
3dab1dad
YO
2106 This means we convert either the first branch or the first Exact,
2107 depending on whether the thing following (in 'last') is a branch
2108 or not and whther first is the startbranch (ie is it a sub part of
2109 the alternation or is it the whole thing.)
3b753521 2110 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2111 the whole branch sequence, including the first.
a3621e74 2112 */
3dab1dad 2113 /* Find the node we are going to overwrite */
7f69552c 2114 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2115 /* branch sub-chain */
3dab1dad 2116 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2117#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2118 DEBUG_r({
2119 mjd_offset= Node_Offset((convert));
2120 mjd_nodelen= Node_Length((convert));
2121 });
7122b237 2122#endif
7f69552c 2123 /* whole branch chain */
7122b237
YO
2124 }
2125#ifdef RE_TRACK_PATTERN_OFFSETS
2126 else {
7f69552c
YO
2127 DEBUG_r({
2128 const regnode *nop = NEXTOPER( convert );
2129 mjd_offset= Node_Offset((nop));
2130 mjd_nodelen= Node_Length((nop));
2131 });
07be1b83
YO
2132 }
2133 DEBUG_OPTIMISE_r(
2134 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2135 (int)depth * 2 + 2, "",
786e8c11 2136 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2137 );
7122b237 2138#endif
3dab1dad
YO
2139 /* But first we check to see if there is a common prefix we can
2140 split out as an EXACT and put in front of the TRIE node. */
2141 trie->startstate= 1;
55eed653 2142 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2143 U32 state;
1e2e3d02 2144 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2145 U32 ofs = 0;
8e11feef
RGS
2146 I32 idx = -1;
2147 U32 count = 0;
2148 const U32 base = trie->states[ state ].trans.base;
a3621e74 2149
3dab1dad 2150 if ( trie->states[state].wordnum )
8e11feef 2151 count = 1;
a3621e74 2152
8e11feef 2153 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2154 if ( ( base + ofs >= trie->uniquecharcount ) &&
2155 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2156 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2157 {
3dab1dad 2158 if ( ++count > 1 ) {
2b8b4781 2159 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2160 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2161 if ( state == 1 ) break;
3dab1dad
YO
2162 if ( count == 2 ) {
2163 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2164 DEBUG_OPTIMISE_r(
8e11feef
RGS
2165 PerlIO_printf(Perl_debug_log,
2166 "%*sNew Start State=%"UVuf" Class: [",
2167 (int)depth * 2 + 2, "",
786e8c11 2168 (UV)state));
be8e71aa 2169 if (idx >= 0) {
2b8b4781 2170 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2171 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2172
3dab1dad 2173 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2174 if ( folder )
2175 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2176 DEBUG_OPTIMISE_r(
f1f66076 2177 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2178 );
8e11feef
RGS
2179 }
2180 }
2181 TRIE_BITMAP_SET(trie,*ch);
2182 if ( folder )
2183 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2184 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2185 }
2186 idx = ofs;
2187 }
3dab1dad
YO
2188 }
2189 if ( count == 1 ) {
2b8b4781 2190 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2191 STRLEN len;
2192 char *ch = SvPV( *tmp, len );
de734bd5
A
2193 DEBUG_OPTIMISE_r({
2194 SV *sv=sv_newmortal();
8e11feef
RGS
2195 PerlIO_printf( Perl_debug_log,
2196 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2197 (int)depth * 2 + 2, "",
de734bd5
A
2198 (UV)state, (UV)idx,
2199 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2200 PL_colors[0], PL_colors[1],
2201 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2202 PERL_PV_ESCAPE_FIRSTCHAR
2203 )
2204 );
2205 });
3dab1dad
YO
2206 if ( state==1 ) {
2207 OP( convert ) = nodetype;
2208 str=STRING(convert);
2209 STR_LEN(convert)=0;
2210 }
c490c714
YO
2211 STR_LEN(convert) += len;
2212 while (len--)
de734bd5 2213 *str++ = *ch++;
8e11feef 2214 } else {
f9049ba1 2215#ifdef DEBUGGING
8e11feef
RGS
2216 if (state>1)
2217 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2218#endif
8e11feef
RGS
2219 break;
2220 }
2221 }
2e64971a 2222 trie->prefixlen = (state-1);
3dab1dad 2223 if (str) {
8e11feef 2224 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2225 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2226 trie->startstate = state;
07be1b83
YO
2227 trie->minlen -= (state - 1);
2228 trie->maxlen -= (state - 1);
33809eae
JH
2229#ifdef DEBUGGING
2230 /* At least the UNICOS C compiler choked on this
2231 * being argument to DEBUG_r(), so let's just have
2232 * it right here. */
2233 if (
2234#ifdef PERL_EXT_RE_BUILD
2235 1
2236#else
2237 DEBUG_r_TEST
2238#endif
2239 ) {
2240 regnode *fix = convert;
2241 U32 word = trie->wordcount;
2242 mjd_nodelen++;
2243 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2244 while( ++fix < n ) {
2245 Set_Node_Offset_Length(fix, 0, 0);
2246 }
2247 while (word--) {
2248 SV ** const tmp = av_fetch( trie_words, word, 0 );
2249 if (tmp) {
2250 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2251 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2252 else
2253 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2254 }
2255 }
2256 }
2257#endif
8e11feef
RGS
2258 if (trie->maxlen) {
2259 convert = n;
2260 } else {
3dab1dad 2261 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2262 DEBUG_r(optimize= n);
3dab1dad
YO
2263 }
2264 }
2265 }
a5ca303d
YO
2266 if (!jumper)
2267 jumper = last;
3dab1dad 2268 if ( trie->maxlen ) {
8e11feef
RGS
2269 NEXT_OFF( convert ) = (U16)(tail - convert);
2270 ARG_SET( convert, data_slot );
786e8c11
YO
2271 /* Store the offset to the first unabsorbed branch in
2272 jump[0], which is otherwise unused by the jump logic.
2273 We use this when dumping a trie and during optimisation. */
2274 if (trie->jump)
7f69552c 2275 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2276
6c48061a
YO
2277 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2278 * and there is a bitmap
2279 * and the first "jump target" node we found leaves enough room
2280 * then convert the TRIE node into a TRIEC node, with the bitmap
2281 * embedded inline in the opcode - this is hypothetically faster.
2282 */
2283 if ( !trie->states[trie->startstate].wordnum
2284 && trie->bitmap
2285 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2286 {
2287 OP( convert ) = TRIEC;
2288 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2289 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2290 trie->bitmap= NULL;
2291 } else
2292 OP( convert ) = TRIE;
a3621e74 2293
3dab1dad
YO
2294 /* store the type in the flags */
2295 convert->flags = nodetype;
a5ca303d
YO
2296 DEBUG_r({
2297 optimize = convert
2298 + NODE_STEP_REGNODE
2299 + regarglen[ OP( convert ) ];
2300 });
2301 /* XXX We really should free up the resource in trie now,
2302 as we won't use them - (which resources?) dmq */
3dab1dad 2303 }
a3621e74 2304 /* needed for dumping*/
e62cc96a 2305 DEBUG_r(if (optimize) {
07be1b83 2306 regnode *opt = convert;
bcdf7404 2307
e62cc96a 2308 while ( ++opt < optimize) {
07be1b83
YO
2309 Set_Node_Offset_Length(opt,0,0);
2310 }
786e8c11
YO
2311 /*
2312 Try to clean up some of the debris left after the
2313 optimisation.
a3621e74 2314 */
786e8c11 2315 while( optimize < jumper ) {
07be1b83 2316 mjd_nodelen += Node_Length((optimize));
a3621e74 2317 OP( optimize ) = OPTIMIZED;
07be1b83 2318 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2319 optimize++;
2320 }
07be1b83 2321 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2322 });
2323 } /* end node insert */
2e64971a
DM
2324
2325 /* Finish populating the prev field of the wordinfo array. Walk back
2326 * from each accept state until we find another accept state, and if
2327 * so, point the first word's .prev field at the second word. If the
2328 * second already has a .prev field set, stop now. This will be the
2329 * case either if we've already processed that word's accept state,
3b753521
FN
2330 * or that state had multiple words, and the overspill words were
2331 * already linked up earlier.
2e64971a
DM
2332 */
2333 {
2334 U16 word;
2335 U32 state;
2336 U16 prev;
2337
2338 for (word=1; word <= trie->wordcount; word++) {
2339 prev = 0;
2340 if (trie->wordinfo[word].prev)
2341 continue;
2342 state = trie->wordinfo[word].accept;
2343 while (state) {
2344 state = prev_states[state];
2345 if (!state)
2346 break;
2347 prev = trie->states[state].wordnum;
2348 if (prev)
2349 break;
2350 }
2351 trie->wordinfo[word].prev = prev;
2352 }
2353 Safefree(prev_states);
2354 }
2355
2356
2357 /* and now dump out the compressed format */
2358 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2359
55eed653 2360 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2361#ifdef DEBUGGING
2362 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2363 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2364#else
2365 SvREFCNT_dec(revcharmap);
07be1b83 2366#endif
786e8c11
YO
2367 return trie->jump
2368 ? MADE_JUMP_TRIE
2369 : trie->startstate>1
2370 ? MADE_EXACT_TRIE
2371 : MADE_TRIE;
2372}
2373
2374STATIC void
2375S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2376{
3b753521 2377/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2378
2379 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2380 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2381 ISBN 0-201-10088-6
2382
2383 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2384 suffix of the current state's 'word' that is also a proper prefix of another word in our
2385 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2386 the DFA not to have to restart after its tried and failed a word at a given point, it
2387 simply continues as though it had been matching the other word in the first place.
2388 Consider
2389 'abcdgu'=~/abcdefg|cdgu/
2390 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2391 fail, which would bring us to the state representing 'd' in the second word where we would
2392 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2393 */
2394 /* add a fail transition */
3251b653
NC
2395 const U32 trie_offset = ARG(source);
2396 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2397 U32 *q;
2398 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2399 const U32 numstates = trie->statecount;
786e8c11
YO
2400 const U32 ubound = trie->lasttrans + ucharcount;
2401 U32 q_read = 0;
2402 U32 q_write = 0;
2403 U32 charid;
2404 U32 base = trie->states[ 1 ].trans.base;
2405 U32 *fail;
2406 reg_ac_data *aho;
2407 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2408 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2409
2410 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2411#ifndef DEBUGGING
2412 PERL_UNUSED_ARG(depth);
2413#endif
2414
2415
2416 ARG_SET( stclass, data_slot );
c944940b 2417 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2418 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2419 aho->trie=trie_offset;
446bd890
NC
2420 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2421 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2422 Newxz( q, numstates, U32);
c944940b 2423 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2424 aho->refcount = 1;
2425 fail = aho->fail;
2426 /* initialize fail[0..1] to be 1 so that we always have
2427 a valid final fail state */
2428 fail[ 0 ] = fail[ 1 ] = 1;
2429
2430 for ( charid = 0; charid < ucharcount ; charid++ ) {
2431 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2432 if ( newstate ) {
2433 q[ q_write ] = newstate;
2434 /* set to point at the root */
2435 fail[ q[ q_write++ ] ]=1;
2436 }
2437 }
2438 while ( q_read < q_write) {
2439 const U32 cur = q[ q_read++ % numstates ];
2440 base = trie->states[ cur ].trans.base;
2441
2442 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2443 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2444 if (ch_state) {
2445 U32 fail_state = cur;
2446 U32 fail_base;
2447 do {
2448 fail_state = fail[ fail_state ];
2449 fail_base = aho->states[ fail_state ].trans.base;
2450 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2451
2452 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2453 fail[ ch_state ] = fail_state;
2454 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2455 {
2456 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2457 }
2458 q[ q_write++ % numstates] = ch_state;
2459 }
2460 }
2461 }
2462 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2463 when we fail in state 1, this allows us to use the
2464 charclass scan to find a valid start char. This is based on the principle
2465 that theres a good chance the string being searched contains lots of stuff
2466 that cant be a start char.
2467 */
2468 fail[ 0 ] = fail[ 1 ] = 0;
2469 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2470 PerlIO_printf(Perl_debug_log,
2471 "%*sStclass Failtable (%"UVuf" states): 0",
2472 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2473 );
786e8c11
YO
2474 for( q_read=1; q_read<numstates; q_read++ ) {
2475 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2476 }
2477 PerlIO_printf(Perl_debug_log, "\n");
2478 });
2479 Safefree(q);
2480 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2481}
2482
786e8c11 2483
a3621e74 2484/*
5d1c421c
JH
2485 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2486 * These need to be revisited when a newer toolchain becomes available.
2487 */
2488#if defined(__sparc64__) && defined(__GNUC__)
2489# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2490# undef SPARC64_GCC_WORKAROUND
2491# define SPARC64_GCC_WORKAROUND 1
2492# endif
2493#endif
2494
07be1b83 2495#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2496 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2497 SV * const mysv=sv_newmortal(); \
2498 regnode *Next = regnext(scan); \
2499 regprop(RExC_rx, mysv, scan); \
7f69552c 2500 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2501 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2502 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2503 }});
07be1b83 2504
1de06328
YO
2505
2506
2507
2508
07be1b83
YO
2509#define JOIN_EXACT(scan,min,flags) \
2510 if (PL_regkind[OP(scan)] == EXACT) \
2511 join_exact(pRExC_state,(scan),(min),(flags),NULL,depth+1)
2512
be8e71aa 2513STATIC U32
07be1b83
YO
2514S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, I32 *min, U32 flags,regnode *val, U32 depth) {
2515 /* Merge several consecutive EXACTish nodes into one. */
2516 regnode *n = regnext(scan);
2517 U32 stringok = 1;
2518 regnode *next = scan + NODE_SZ_STR(scan);
2519 U32 merged = 0;
2520 U32 stopnow = 0;
2521#ifdef DEBUGGING
2522 regnode *stop = scan;
72f13be8 2523 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2524#else
d47053eb
RGS
2525 PERL_UNUSED_ARG(depth);
2526#endif
7918f24d
NC
2527
2528 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2529#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2530 PERL_UNUSED_ARG(flags);
2531 PERL_UNUSED_ARG(val);
07be1b83 2532#endif
07be1b83
YO
2533 DEBUG_PEEP("join",scan,depth);
2534
2535 /* Skip NOTHING, merge EXACT*. */
2536 while (n &&
2537 ( PL_regkind[OP(n)] == NOTHING ||
2538 (stringok && (OP(n) == OP(scan))))
2539 && NEXT_OFF(n)
2540 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX) {
2541
2542 if (OP(n) == TAIL || n > next)
2543 stringok = 0;
2544 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2545 DEBUG_PEEP("skip:",n,depth);
2546 NEXT_OFF(scan) += NEXT_OFF(n);
2547 next = n + NODE_STEP_REGNODE;
2548#ifdef DEBUGGING
2549 if (stringok)
2550 stop = n;
2551#endif
2552 n = regnext(n);
2553 }
2554 else if (stringok) {
786e8c11 2555 const unsigned int oldl = STR_LEN(scan);
07be1b83
YO
2556 regnode * const nnext = regnext(n);
2557
2558 DEBUG_PEEP("merg",n,depth);
2559
2560 merged++;
2561 if (oldl + STR_LEN(n) > U8_MAX)
2562 break;
2563 NEXT_OFF(scan) += NEXT_OFF(n);
2564 STR_LEN(scan) += STR_LEN(n);
2565 next = n + NODE_SZ_STR(n);
2566 /* Now we can overwrite *n : */
2567 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2568#ifdef DEBUGGING
2569 stop = next - 1;
2570#endif
2571 n = nnext;
2572 if (stopnow) break;
2573 }
2574
d47053eb
RGS
2575#ifdef EXPERIMENTAL_INPLACESCAN
2576 if (flags && !NEXT_OFF(n)) {
2577 DEBUG_PEEP("atch", val, depth);
2578 if (reg_off_by_arg[OP(n)]) {
2579 ARG_SET(n, val - n);
2580 }
2581 else {
2582 NEXT_OFF(n) = val - n;
2583 }
2584 stopnow = 1;
2585 }
07be1b83
YO
2586#endif
2587 }
ced7f090
KW
2588#define GREEK_SMALL_LETTER_IOTA_WITH_DIALYTIKA_AND_TONOS 0x0390
2589#define IOTA_D_T GREEK_SMALL_LETTER_IOTA_WITH_DIALYTIKA_AND_TONOS
2590#define GREEK_SMALL_LETTER_UPSILON_WITH_DIALYTIKA_AND_TONOS 0x03B0
2591#define UPSILON_D_T GREEK_SMALL_LETTER_UPSILON_WITH_DIALYTIKA_AND_TONOS
2c2b7f86
KW
2592
2593 if (UTF
2f7f8cb1 2594 && ( OP(scan) == EXACTF || OP(scan) == EXACTFU || OP(scan) == EXACTFA)
2c2b7f86
KW
2595 && ( STR_LEN(scan) >= 6 ) )
2596 {
07be1b83
YO
2597 /*
2598 Two problematic code points in Unicode casefolding of EXACT nodes:
2599
2600 U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2601 U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2602
2603 which casefold to
2604
2605 Unicode UTF-8
2606
2607 U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2608 U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2609
2610 This means that in case-insensitive matching (or "loose matching",
2611 as Unicode calls it), an EXACTF of length six (the UTF-8 encoded byte
2612 length of the above casefolded versions) can match a target string
2613 of length two (the byte length of UTF-8 encoded U+0390 or U+03B0).
2614 This would rather mess up the minimum length computation.
2615
2616 What we'll do is to look for the tail four bytes, and then peek
2617 at the preceding two bytes to see whether we need to decrease
2618 the minimum length by four (six minus two).
2619
2620 Thanks to the design of UTF-8, there cannot be false matches:
2621 A sequence of valid UTF-8 bytes cannot be a subsequence of
2622 another valid sequence of UTF-8 bytes.
2623
2624 */
2625 char * const s0 = STRING(scan), *s, *t;
2626 char * const s1 = s0 + STR_LEN(scan) - 1;
2627 char * const s2 = s1 - 4;
e294cc5d
JH
2628#ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
2629 const char t0[] = "\xaf\x49\xaf\x42";
2630#else
07be1b83 2631 const char t0[] = "\xcc\x88\xcc\x81";
e294cc5d 2632#endif
07be1b83
YO
2633 const char * const t1 = t0 + 3;
2634
2635 for (s = s0 + 2;
2636 s < s2 && (t = ninstr(s, s1, t0, t1));
2637 s = t + 4) {
e294cc5d
JH
2638#ifdef EBCDIC
2639 if (((U8)t[-1] == 0x68 && (U8)t[-2] == 0xB4) ||
2640 ((U8)t[-1] == 0x46 && (U8)t[-2] == 0xB5))
2641#else
07be1b83
YO
2642 if (((U8)t[-1] == 0xB9 && (U8)t[-2] == 0xCE) ||
2643 ((U8)t[-1] == 0x85 && (U8)t[-2] == 0xCF))
e294cc5d 2644#endif
07be1b83
YO
2645 *min -= 4;
2646 }
2647 }
2648
2649#ifdef DEBUGGING
bb789b09
DM
2650 /* Allow dumping but overwriting the collection of skipped
2651 * ops and/or strings with fake optimized ops */
07be1b83
YO
2652 n = scan + NODE_SZ_STR(scan);
2653 while (n <= stop) {
bb789b09
DM
2654 OP(n) = OPTIMIZED;
2655 FLAGS(n) = 0;
2656 NEXT_OFF(n) = 0;
07be1b83
YO
2657 n++;
2658 }
2659#endif
2660 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2661 return stopnow;
2662}
2663
486ec47a 2664/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2665 Finds fixed substrings. */
2666
a0288114 2667/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2668 to the position after last scanned or to NULL. */
2669
40d049e4
YO
2670#define INIT_AND_WITHP \
2671 assert(!and_withp); \
2672 Newx(and_withp,1,struct regnode_charclass_class); \
2673 SAVEFREEPV(and_withp)
07be1b83 2674
b515a41d 2675/* this is a chain of data about sub patterns we are processing that
486ec47a 2676 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2677 we can simulate recursion without losing state. */
2678struct scan_frame;
2679typedef struct scan_frame {
2680 regnode *last; /* last node to process in this frame */
2681 regnode *next; /* next node to process when last is reached */
2682 struct scan_frame *prev; /*previous frame*/
2683 I32 stop; /* what stopparen do we use */
2684} scan_frame;
2685
304ee84b
YO
2686
2687#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2688
e1d1eefb
YO
2689#define CASE_SYNST_FNC(nAmE) \
2690case nAmE: \
2691 if (flags & SCF_DO_STCLASS_AND) { \
2692 for (value = 0; value < 256; value++) \
2693 if (!is_ ## nAmE ## _cp(value)) \
2694 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2695 } \
2696 else { \
2697 for (value = 0; value < 256; value++) \
2698 if (is_ ## nAmE ## _cp(value)) \
2699 ANYOF_BITMAP_SET(data->start_class, value); \
2700 } \
2701 break; \
2702case N ## nAmE: \
2703 if (flags & SCF_DO_STCLASS_AND) { \
2704 for (value = 0; value < 256; value++) \
2705 if (is_ ## nAmE ## _cp(value)) \
2706 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2707 } \
2708 else { \
2709 for (value = 0; value < 256; value++) \
2710 if (!is_ ## nAmE ## _cp(value)) \
2711 ANYOF_BITMAP_SET(data->start_class, value); \
2712 } \
2713 break
2714
2715
2716
76e3520e 2717STATIC I32
40d049e4 2718S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 2719 I32 *minlenp, I32 *deltap,
40d049e4
YO
2720 regnode *last,
2721 scan_data_t *data,
2722 I32 stopparen,
2723 U8* recursed,
2724 struct regnode_charclass_class *and_withp,
2725 U32 flags, U32 depth)
c277df42
IZ
2726 /* scanp: Start here (read-write). */
2727 /* deltap: Write maxlen-minlen here. */
2728 /* last: Stop before this one. */
40d049e4
YO
2729 /* data: string data about the pattern */
2730 /* stopparen: treat close N as END */
2731 /* recursed: which subroutines have we recursed into */
2732 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 2733{
97aff369 2734 dVAR;
c277df42
IZ
2735 I32 min = 0, pars = 0, code;
2736 regnode *scan = *scanp, *next;
2737 I32 delta = 0;
2738 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 2739 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
2740 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2741 scan_data_t data_fake;
a3621e74 2742 SV *re_trie_maxbuff = NULL;
786e8c11 2743 regnode *first_non_open = scan;
e2e6a0f1 2744 I32 stopmin = I32_MAX;
8aa23a47 2745 scan_frame *frame = NULL;
a3621e74 2746 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 2747
7918f24d
NC
2748 PERL_ARGS_ASSERT_STUDY_CHUNK;
2749
13a24bad 2750#ifdef DEBUGGING
40d049e4 2751 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 2752#endif
40d049e4 2753
786e8c11 2754 if ( depth == 0 ) {
40d049e4 2755 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
2756 first_non_open=regnext(first_non_open);
2757 }
2758
b81d288d 2759
8aa23a47
YO
2760 fake_study_recurse:
2761 while ( scan && OP(scan) != END && scan < last ){
2762 /* Peephole optimizer: */
304ee84b 2763 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47
YO
2764 DEBUG_PEEP("Peep",scan,depth);
2765 JOIN_EXACT(scan,&min,0);
2766
2767 /* Follow the next-chain of the current node and optimize
2768 away all the NOTHINGs from it. */
2769 if (OP(scan) != CURLYX) {
2770 const int max = (reg_off_by_arg[OP(scan)]
2771 ? I32_MAX
2772 /* I32 may be smaller than U16 on CRAYs! */
2773 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
2774 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
2775 int noff;
2776 regnode *n = scan;
2777
2778 /* Skip NOTHING and LONGJMP. */
2779 while ((n = regnext(n))
2780 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
2781 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
2782 && off + noff < max)
2783 off += noff;
2784 if (reg_off_by_arg[OP(scan)])
2785 ARG(scan) = off;
2786 else
2787 NEXT_OFF(scan) = off;
2788 }
a3621e74 2789
c277df42 2790
8aa23a47
YO
2791
2792 /* The principal pseudo-switch. Cannot be a switch, since we
2793 look into several different things. */
2794 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
2795 || OP(scan) == IFTHEN) {
2796 next = regnext(scan);
2797 code = OP(scan);
2798 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
2799
2800 if (OP(next) == code || code == IFTHEN) {
2801 /* NOTE - There is similar code to this block below for handling
2802 TRIE nodes on a re-study. If you change stuff here check there
2803 too. */
2804 I32 max1 = 0, min1 = I32_MAX, num = 0;
2805 struct regnode_charclass_class accum;
2806 regnode * const startbranch=scan;
2807
2808 if (flags & SCF_DO_SUBSTR)
304ee84b 2809 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 2810 if (flags & SCF_DO_STCLASS)
e755fd73 2811 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
2812
2813 while (OP(scan) == code) {
2814 I32 deltanext, minnext, f = 0, fake;
2815 struct regnode_charclass_class this_class;
2816
2817 num++;
2818 data_fake.flags = 0;
2819 if (data) {
2820 data_fake.whilem_c = data->whilem_c;
2821 data_fake.last_closep = data->last_closep;
2822 }
2823 else
2824 data_fake.last_closep = &fake;
58e23c8d
YO
2825
2826 data_fake.pos_delta = delta;
8aa23a47
YO
2827 next = regnext(scan);
2828 scan = NEXTOPER(scan);
2829 if (code != BRANCH)
c277df42 2830 scan = NEXTOPER(scan);
8aa23a47 2831 if (flags & SCF_DO_STCLASS) {
e755fd73 2832 cl_init(pRExC_state, &this_class);
8aa23a47
YO
2833 data_fake.start_class = &this_class;
2834 f = SCF_DO_STCLASS_AND;
58e23c8d 2835 }
8aa23a47
YO
2836 if (flags & SCF_WHILEM_VISITED_POS)
2837 f |= SCF_WHILEM_VISITED_POS;
2838
2839 /* we suppose the run is continuous, last=next...*/
2840 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
2841 next, &data_fake,
2842 stopparen, recursed, NULL, f,depth+1);
2843 if (min1 > minnext)
2844 min1 = minnext;
2845 if (max1 < minnext + deltanext)
2846 max1 = minnext + deltanext;
2847 if (deltanext == I32_MAX)
2848 is_inf = is_inf_internal = 1;
2849 scan = next;
2850 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
2851 pars++;
2852 if (data_fake.flags & SCF_SEEN_ACCEPT) {
2853 if ( stopmin > minnext)
2854 stopmin = min + min1;
2855 flags &= ~SCF_DO_SUBSTR;
2856 if (data)
2857 data->flags |= SCF_SEEN_ACCEPT;
2858 }
2859 if (data) {
2860 if (data_fake.flags & SF_HAS_EVAL)
2861 data->flags |= SF_HAS_EVAL;
2862 data->whilem_c = data_fake.whilem_c;
3dab1dad 2863 }
8aa23a47 2864 if (flags & SCF_DO_STCLASS)
3fffb88a 2865 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
2866 }
2867 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
2868 min1 = 0;
2869 if (flags & SCF_DO_SUBSTR) {
2870 data->pos_min += min1;
2871 data->pos_delta += max1 - min1;
2872 if (max1 != min1 || is_inf)
2873 data->longest = &(data->longest_float);
2874 }
2875 min += min1;
2876 delta += max1 - min1;
2877 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 2878 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
2879 if (min1) {
2880 cl_and(data->start_class, and_withp);
2881 flags &= ~SCF_DO_STCLASS;
653099ff 2882 }
8aa23a47
YO
2883 }
2884 else if (flags & SCF_DO_STCLASS_AND) {
2885 if (min1) {
2886 cl_and(data->start_class, &accum);
2887 flags &= ~SCF_DO_STCLASS;
de0c8cb8 2888 }
8aa23a47
YO
2889 else {
2890 /* Switch to OR mode: cache the old value of
2891 * data->start_class */
2892 INIT_AND_WITHP;
2893 StructCopy(data->start_class, and_withp,
2894 struct regnode_charclass_class);
2895 flags &= ~SCF_DO_STCLASS_AND;
2896 StructCopy(&accum, data->start_class,
2897 struct regnode_charclass_class);
2898 flags |= SCF_DO_STCLASS_OR;
2899 data->start_class->flags |= ANYOF_EOS;
de0c8cb8 2900 }
8aa23a47 2901 }
a3621e74 2902
8aa23a47
YO
2903 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
2904 /* demq.
a3621e74 2905
8aa23a47
YO
2906 Assuming this was/is a branch we are dealing with: 'scan' now
2907 points at the item that follows the branch sequence, whatever
2908 it is. We now start at the beginning of the sequence and look
2909 for subsequences of
a3621e74 2910
8aa23a47
YO
2911 BRANCH->EXACT=>x1
2912 BRANCH->EXACT=>x2
2913 tail
a3621e74 2914
8aa23a47 2915 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 2916
486ec47a 2917 If we can find such a subsequence we need to turn the first
8aa23a47
YO
2918 element into a trie and then add the subsequent branch exact
2919 strings to the trie.
a3621e74 2920
8aa23a47 2921 We have two cases
a3621e74 2922
3b753521 2923 1. patterns where the whole set of branches can be converted.
a3621e74 2924
8aa23a47 2925 2. patterns where only a subset can be converted.
a3621e74 2926
8aa23a47
YO
2927 In case 1 we can replace the whole set with a single regop
2928 for the trie. In case 2 we need to keep the start and end
3b753521 2929 branches so
a3621e74 2930
8aa23a47
YO
2931 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
2932 becomes BRANCH TRIE; BRANCH X;
786e8c11 2933
8aa23a47
YO
2934 There is an additional case, that being where there is a
2935 common prefix, which gets split out into an EXACT like node
2936 preceding the TRIE node.
a3621e74 2937
8aa23a47
YO
2938 If x(1..n)==tail then we can do a simple trie, if not we make
2939 a "jump" trie, such that when we match the appropriate word
486ec47a 2940 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 2941 a nested if into a case structure of sorts.
b515a41d 2942
8aa23a47
YO
2943 */
2944
2945 int made=0;
2946 if (!re_trie_maxbuff) {
2947 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2948 if (!SvIOK(re_trie_maxbuff))
2949 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2950 }
2951 if ( SvIV(re_trie_maxbuff)>=0 ) {
2952 regnode *cur;
2953 regnode *first = (regnode *)NULL;
2954 regnode *last = (regnode *)NULL;
2955 regnode *tail = scan;
2956 U8 optype = 0;
2957 U32 count=0;
a3621e74
YO
2958
2959#ifdef DEBUGGING
8aa23a47 2960 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 2961#endif
8aa23a47
YO
2962 /* var tail is used because there may be a TAIL
2963 regop in the way. Ie, the exacts will point to the
2964 thing following the TAIL, but the last branch will
2965 point at the TAIL. So we advance tail. If we
2966 have nested (?:) we may have to move through several
2967 tails.
2968 */
2969
2970 while ( OP( tail ) == TAIL ) {
2971 /* this is the TAIL generated by (?:) */
2972 tail = regnext( tail );
2973 }
a3621e74 2974
8aa23a47
YO
2975
2976 DEBUG_OPTIMISE_r({
2977 regprop(RExC_rx, mysv, tail );
2978 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
2979 (int)depth * 2 + 2, "",
2980 "Looking for TRIE'able sequences. Tail node is: ",
2981 SvPV_nolen_const( mysv )
2982 );
2983 });
2984
2985 /*
2986
2987 step through the branches, cur represents each
2988 branch, noper is the first thing to be matched
2989 as part of that branch and noper_next is the
2990 regnext() of that node. if noper is an EXACT
2991 and noper_next is the same as scan (our current
2992 position in the regex) then the EXACT branch is
2993 a possible optimization target. Once we have
486ec47a 2994 two or more consecutive such branches we can
8aa23a47
YO
2995 create a trie of the EXACT's contents and stich
2996 it in place. If the sequence represents all of
2997 the branches we eliminate the whole thing and
2998 replace it with a single TRIE. If it is a
2999 subsequence then we need to stitch it in. This
3000 means the first branch has to remain, and needs
3001 to be repointed at the item on the branch chain
3002 following the last branch optimized. This could
3003 be either a BRANCH, in which case the
3004 subsequence is internal, or it could be the
3005 item following the branch sequence in which
3006 case the subsequence is at the end.
3007
3008 */
3009
3010 /* dont use tail as the end marker for this traverse */
3011 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3012 regnode * const noper = NEXTOPER( cur );
b515a41d 3013#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3014 regnode * const noper_next = regnext( noper );
b515a41d
YO
3015#endif
3016
8aa23a47
YO
3017 DEBUG_OPTIMISE_r({
3018 regprop(RExC_rx, mysv, cur);
3019 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3020 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3021
3022 regprop(RExC_rx, mysv, noper);
3023 PerlIO_printf( Perl_debug_log, " -> %s",
3024 SvPV_nolen_const(mysv));
3025
3026 if ( noper_next ) {
3027 regprop(RExC_rx, mysv, noper_next );
3028 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3029 SvPV_nolen_const(mysv));
3030 }
3031 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
3032 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
3033 });
3034 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
3035 : PL_regkind[ OP( noper ) ] == EXACT )
3036 || OP(noper) == NOTHING )
786e8c11 3037#ifdef NOJUMPTRIE
8aa23a47 3038 && noper_next == tail
786e8c11 3039#endif
8aa23a47
YO
3040 && count < U16_MAX)
3041 {
3042 count++;
3043 if ( !first || optype == NOTHING ) {
3044 if (!first) first = cur;
3045 optype = OP( noper );
3046 } else {
3047 last = cur;
3048 }
3049 } else {
a0a388a1 3050/*
fbebf34e
KW
3051 Currently the trie logic handles case insensitive matching properly only
3052 when the pattern is UTF-8 and the node is EXACTFU (thus forcing unicode
3053 semantics).
0abd0d78
YO
3054
3055 If/when this is fixed the following define can be swapped
3056 in below to fully enable trie logic.
3057
a0a388a1 3058#define TRIE_TYPE_IS_SAFE 1
0abd0d78
YO
3059
3060*/
fbebf34e 3061#define TRIE_TYPE_IS_SAFE ((UTF && optype == EXACTFU) || optype==EXACT)
0abd0d78 3062
a0a388a1 3063 if ( last && TRIE_TYPE_IS_SAFE ) {
8aa23a47
YO
3064 make_trie( pRExC_state,
3065 startbranch, first, cur, tail, count,
3066 optype, depth+1 );
3067 }
3068 if ( PL_regkind[ OP( noper ) ] == EXACT
786e8c11 3069#ifdef NOJUMPTRIE
8aa23a47 3070 && noper_next == tail
786e8c11 3071#endif
8aa23a47
YO
3072 ){
3073 count = 1;
3074 first = cur;
3075 optype = OP( noper );
3076 } else {
3077 count = 0;
3078 first = NULL;
3079 optype = 0;
3080 }
3081 last = NULL;
3082 }
3083 }
3084 DEBUG_OPTIMISE_r({
3085 regprop(RExC_rx, mysv, cur);
3086 PerlIO_printf( Perl_debug_log,
3087 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3088 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3089
3090 });
a0a388a1
YO
3091
3092 if ( last && TRIE_TYPE_IS_SAFE ) {
8aa23a47 3093 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
3dab1dad 3094#ifdef TRIE_STUDY_OPT
8aa23a47
YO
3095 if ( ((made == MADE_EXACT_TRIE &&
3096 startbranch == first)
3097 || ( first_non_open == first )) &&
3098 depth==0 ) {
3099 flags |= SCF_TRIE_RESTUDY;
3100 if ( startbranch == first
3101 && scan == tail )
3102 {
3103 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3104 }
3105 }
3dab1dad 3106#endif
8aa23a47
YO
3107 }
3108 }
3109
3110 } /* do trie */
3111
653099ff 3112 }
8aa23a47
YO
3113 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3114 scan = NEXTOPER(NEXTOPER(scan));
3115 } else /* single branch is optimized. */
3116 scan = NEXTOPER(scan);
3117 continue;
3118 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3119 scan_frame *newframe = NULL;
3120 I32 paren;
3121 regnode *start;
3122 regnode *end;
3123
3124 if (OP(scan) != SUSPEND) {
3125 /* set the pointer */
3126 if (OP(scan) == GOSUB) {
3127 paren = ARG(scan);
3128 RExC_recurse[ARG2L(scan)] = scan;
3129 start = RExC_open_parens[paren-1];
3130 end = RExC_close_parens[paren-1];
3131 } else {
3132 paren = 0;
f8fc2ecf 3133 start = RExC_rxi->program + 1;
8aa23a47
YO
3134 end = RExC_opend;
3135 }
3136 if (!recursed) {
3137 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3138 SAVEFREEPV(recursed);
3139 }
3140 if (!PAREN_TEST(recursed,paren+1)) {
3141 PAREN_SET(recursed,paren+1);
3142 Newx(newframe,1,scan_frame);
3143 } else {
3144 if (flags & SCF_DO_SUBSTR) {
304ee84b 3145 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3146 data->longest = &(data->longest_float);
3147 }
3148 is_inf = is_inf_internal = 1;
3149 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3150 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3151 flags &= ~SCF_DO_STCLASS;
3152 }
3153 } else {
3154 Newx(newframe,1,scan_frame);
3155 paren = stopparen;
3156 start = scan+2;
3157 end = regnext(scan);
3158 }
3159 if (newframe) {
3160 assert(start);
3161 assert(end);
3162 SAVEFREEPV(newframe);
3163 newframe->next = regnext(scan);
3164 newframe->last = last;
3165 newframe->stop = stopparen;
3166 newframe->prev = frame;
3167
3168 frame = newframe;
3169 scan = start;
3170 stopparen = paren;
3171 last = end;
3172
3173 continue;
3174 }
3175 }
3176 else if (OP(scan) == EXACT) {
3177 I32 l = STR_LEN(scan);
3178 UV uc;
3179 if (UTF) {
3180 const U8 * const s = (U8*)STRING(scan);
3181 l = utf8_length(s, s + l);
3182 uc = utf8_to_uvchr(s, NULL);
3183 } else {
3184 uc = *((U8*)STRING(scan));
3185 }
3186 min += l;
3187 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3188 /* The code below prefers earlier match for fixed
3189 offset, later match for variable offset. */
3190 if (data->last_end == -1) { /* Update the start info. */
3191 data->last_start_min = data->pos_min;
3192 data->last_start_max = is_inf
3193 ? I32_MAX : data->pos_min + data->pos_delta;
b515a41d 3194 }
8aa23a47
YO
3195 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3196 if (UTF)
3197 SvUTF8_on(data->last_found);
3198 {
3199 SV * const sv = data->last_found;
3200 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3201 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3202 if (mg && mg->mg_len >= 0)
3203 mg->mg_len += utf8_length((U8*)STRING(scan),
3204 (U8*)STRING(scan)+STR_LEN(scan));
b515a41d 3205 }
8aa23a47
YO
3206 data->last_end = data->pos_min + l;
3207 data->pos_min += l; /* As in the first entry. */
3208 data->flags &= ~SF_BEFORE_EOL;
3209 }
3210 if (flags & SCF_DO_STCLASS_AND) {
3211 /* Check whether it is compatible with what we know already! */
3212 int compat = 1;
3213
54251c2e 3214
486ec47a 3215 /* If compatible, we or it in below. It is compatible if is
54251c2e
KW
3216 * in the bitmp and either 1) its bit or its fold is set, or 2)
3217 * it's for a locale. Even if there isn't unicode semantics
3218 * here, at runtime there may be because of matching against a
3219 * utf8 string, so accept a possible false positive for
3220 * latin1-range folds */
8aa23a47
YO
3221 if (uc >= 0x100 ||
3222 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3223 && !ANYOF_BITMAP_TEST(data->start_class, uc)
39065660 3224 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
54251c2e 3225 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
8aa23a47 3226 )
d18bf9dc 3227 {
8aa23a47 3228 compat = 0;
d18bf9dc 3229 }
8aa23a47
YO
3230 ANYOF_CLASS_ZERO(data->start_class);
3231 ANYOF_BITMAP_ZERO(data->start_class);
3232 if (compat)
3233 ANYOF_BITMAP_SET(data->start_class, uc);
d18bf9dc
KW
3234 else if (uc >= 0x100) {
3235 int i;
3236
3237 /* Some Unicode code points fold to the Latin1 range; as
3238 * XXX temporary code, instead of figuring out if this is
3239 * one, just assume it is and set all the start class bits
3240 * that could be some such above 255 code point's fold
3241 * which will generate fals positives. As the code
3242 * elsewhere that does compute the fold settles down, it
3243 * can be extracted out and re-used here */
3244 for (i = 0; i < 256; i++){
3245 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3246 ANYOF_BITMAP_SET(data->start_class, i);
3247 }
3248 }
3249 }
8aa23a47
YO
3250 data->start_class->flags &= ~ANYOF_EOS;
3251 if (uc < 0x100)
3252 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3253 }
3254 else if (flags & SCF_DO_STCLASS_OR) {
3255 /* false positive possible if the class is case-folded */
3256 if (uc < 0x100)
3257 ANYOF_BITMAP_SET(data->start_class, uc);
3258 else
3259 data->start_class->flags |= ANYOF_UNICODE_ALL;
3260 data->start_class->flags &= ~ANYOF_EOS;
3261 cl_and(data->start_class, and_withp);
3262 }
3263 flags &= ~SCF_DO_STCLASS;
3264 }
3265 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3266 I32 l = STR_LEN(scan);
3267 UV uc = *((U8*)STRING(scan));
3268
3269 /* Search for fixed substrings supports EXACT only. */
3270 if (flags & SCF_DO_SUBSTR) {
3271 assert(data);
304ee84b 3272 SCAN_COMMIT(pRExC_state, data, minlenp);
8aa23a47
YO
3273 }
3274 if (UTF) {
3275 const U8 * const s = (U8 *)STRING(scan);
3276 l = utf8_length(s, s + l);
3277 uc = utf8_to_uvchr(s, NULL);
3278 }
3279 min += l;
3280 if (flags & SCF_DO_SUBSTR)
3281 data->pos_min += l;
3282 if (flags & SCF_DO_STCLASS_AND) {
3283 /* Check whether it is compatible with what we know already! */
3284 int compat = 1;
8aa23a47 3285 if (uc >= 0x100 ||
54251c2e
KW
3286 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3287 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3288 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3289 {
8aa23a47 3290 compat = 0;
54251c2e 3291 }
8aa23a47
YO
3292 ANYOF_CLASS_ZERO(data->start_class);
3293 ANYOF_BITMAP_ZERO(data->start_class);
3294 if (compat) {
3295 ANYOF_BITMAP_SET(data->start_class, uc);
653099ff 3296 data->start_class->flags &= ~ANYOF_EOS;
39065660 3297 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
970c8436 3298 if (OP(scan) == EXACTFL) {
af302e7f
KW
3299 /* XXX This set is probably no longer necessary, and
3300 * probably wrong as LOCALE now is on in the initial
3301 * state */
8aa23a47 3302 data->start_class->flags |= ANYOF_LOCALE;
970c8436
KW
3303 }
3304 else {
3305
54251c2e
KW
3306 /* Also set the other member of the fold pair. In case
3307 * that unicode semantics is called for at runtime, use
3308 * the full latin1 fold. (Can't do this for locale,
3309 * because not known until runtime */
3310 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
970c8436 3311 }
653099ff 3312 }
d18bf9dc
KW
3313 else if (uc >= 0x100) {
3314 int i;
3315 for (i = 0; i < 256; i++){
3316 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3317 ANYOF_BITMAP_SET(data->start_class, i);
3318 }
3319 }
3320 }
8aa23a47
YO
3321 }
3322 else if (flags & SCF_DO_STCLASS_OR) {
39065660 3323 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
8aa23a47
YO
3324 /* false positive possible if the class is case-folded.
3325 Assume that the locale settings are the same... */
970c8436 3326 if (uc < 0x100) {
1aa99e6b 3327 ANYOF_BITMAP_SET(data->start_class, uc);
970c8436
KW
3328 if (OP(scan) != EXACTFL) {
3329
3330 /* And set the other member of the fold pair, but
3331 * can't do that in locale because not known until
3332 * run-time */
3333 ANYOF_BITMAP_SET(data->start_class,
54251c2e 3334 PL_fold_latin1[uc]);
970c8436
KW
3335 }
3336 }
653099ff
GS
3337 data->start_class->flags &= ~ANYOF_EOS;
3338 }
8aa23a47 3339 cl_and(data->start_class, and_withp);
653099ff 3340 }
8aa23a47
YO
3341 flags &= ~SCF_DO_STCLASS;
3342 }
e52fc539 3343 else if (REGNODE_VARIES(OP(scan))) {
8aa23a47
YO
3344 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3345 I32 f = flags, pos_before = 0;
3346 regnode * const oscan = scan;
3347 struct regnode_charclass_class this_class;
3348 struct regnode_charclass_class *oclass = NULL;
3349 I32 next_is_eval = 0;
3350
3351 switch (PL_regkind[OP(scan)]) {
3352 case WHILEM: /* End of (?:...)* . */
3353 scan = NEXTOPER(scan);
3354 goto finish;
3355 case PLUS:
3356 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3357 next = NEXTOPER(scan);
3358 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3359 mincount = 1;
3360 maxcount = REG_INFTY;
3361 next = regnext(scan);
3362 scan = NEXTOPER(scan);
3363 goto do_curly;
3364 }
3365 }
3366 if (flags & SCF_DO_SUBSTR)
3367 data->pos_min++;
3368 min++;
3369 /* Fall through. */
3370 case STAR:
3371 if (flags & SCF_DO_STCLASS) {
3372 mincount = 0;
3373 maxcount = REG_INFTY;
3374 next = regnext(scan);
3375 scan = NEXTOPER(scan);
3376 goto do_curly;
3377 }
3378 is_inf = is_inf_internal = 1;
3379 scan = regnext(scan);
c277df42 3380 if (flags & SCF_DO_SUBSTR) {
304ee84b 3381 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
8aa23a47 3382 data->longest = &(data->longest_float);
c277df42 3383 }
8aa23a47
YO
3384 goto optimize_curly_tail;
3385 case CURLY:
3386 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3387 && (scan->flags == stopparen))
3388 {
3389 mincount = 1;
3390 maxcount = 1;
3391 } else {
3392 mincount = ARG1(scan);
3393 maxcount = ARG2(scan);
653099ff 3394 }
8aa23a47
YO
3395 next = regnext(scan);
3396 if (OP(scan) == CURLYX) {
3397 I32 lp = (data ? *(data->last_closep) : 0);
3398 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
653099ff 3399 }
8aa23a47
YO
3400 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3401 next_is_eval = (OP(scan) == EVAL);
3402 do_curly:
3403 if (flags & SCF_DO_SUBSTR) {
304ee84b 3404 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
8aa23a47 3405 pos_before = data->pos_min;
b45f050a 3406 }
8aa23a47
YO
3407 if (data) {
3408 fl = data->flags;
3409 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3410 if (is_inf)
3411 data->flags |= SF_IS_INF;
3412 }
3413 if (flags & SCF_DO_STCLASS) {
e755fd73 3414 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3415 oclass = data->start_class;
3416 data->start_class = &this_class;
3417 f |= SCF_DO_STCLASS_AND;
3418 f &= ~SCF_DO_STCLASS_OR;
3419 }
779bcb7d
NC
3420 /* Exclude from super-linear cache processing any {n,m}
3421 regops for which the combination of input pos and regex
3422 pos is not enough information to determine if a match
3423 will be possible.
3424
3425 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3426 regex pos at the \s*, the prospects for a match depend not
3427 only on the input position but also on how many (bar\s*)
3428 repeats into the {4,8} we are. */
3429 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
8aa23a47 3430 f &= ~SCF_WHILEM_VISITED_POS;
b45f050a 3431
8aa23a47
YO
3432 /* This will finish on WHILEM, setting scan, or on NULL: */
3433 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3434 last, data, stopparen, recursed, NULL,
3435 (mincount == 0
3436 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
b515a41d 3437
8aa23a47
YO
3438 if (flags & SCF_DO_STCLASS)
3439 data->start_class = oclass;
3440 if (mincount == 0 || minnext == 0) {
3441 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3442 cl_or(pRExC_state, data->start_class, &this_class);
8aa23a47
YO
3443 }
3444 else if (flags & SCF_DO_STCLASS_AND) {
3445 /* Switch to OR mode: cache the old value of
3446 * data->start_class */
3447 INIT_AND_WITHP;
3448 StructCopy(data->start_class, and_withp,
3449 struct regnode_charclass_class);
3450 flags &= ~SCF_DO_STCLASS_AND;
3451 StructCopy(&this_class, data->start_class,
3452 struct regnode_charclass_class);
3453 flags |= SCF_DO_STCLASS_OR;
3454 data->start_class->flags |= ANYOF_EOS;
3455 }
3456 } else { /* Non-zero len */
3457 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3458 cl_or(pRExC_state, data->start_class, &this_class);
8aa23a47
YO
3459 cl_and(data->start_class, and_withp);
3460 }
3461 else if (flags & SCF_DO_STCLASS_AND)
3462 cl_and(data->start_class, &this_class);
3463 flags &= ~SCF_DO_STCLASS;
3464 }
3465 if (!scan) /* It was not CURLYX, but CURLY. */
3466 scan = next;
3467 if ( /* ? quantifier ok, except for (?{ ... }) */
3468 (next_is_eval || !(mincount == 0 && maxcount == 1))
3469 && (minnext == 0) && (deltanext == 0)
3470 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
668c081a 3471 && maxcount <= REG_INFTY/3) /* Complement check for big count */
8aa23a47 3472 {
668c081a
NC
3473 ckWARNreg(RExC_parse,
3474 "Quantifier unexpected on zero-length expression");
8aa23a47
YO
3475 }
3476
3477 min += minnext * mincount;
3478 is_inf_internal |= ((maxcount == REG_INFTY
3479 && (minnext + deltanext) > 0)
3480 || deltanext == I32_MAX);
3481 is_inf |= is_inf_internal;
3482 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3483
3484 /* Try powerful optimization CURLYX => CURLYN. */
3485 if ( OP(oscan) == CURLYX && data
3486 && data->flags & SF_IN_PAR
3487 && !(data->flags & SF_HAS_EVAL)
3488 && !deltanext && minnext == 1 ) {
3489 /* Try to optimize to CURLYN. */
3490 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3491 regnode * const nxt1 = nxt;
497b47a8 3492#ifdef DEBUGGING
8aa23a47 3493 regnode *nxt2;
497b47a8 3494#endif
c277df42 3495
8aa23a47
YO
3496 /* Skip open. */
3497 nxt = regnext(nxt);
e52fc539 3498 if (!REGNODE_SIMPLE(OP(nxt))
8aa23a47
YO
3499 && !(PL_regkind[OP(nxt)] == EXACT
3500 && STR_LEN(nxt) == 1))
3501 goto nogo;
497b47a8 3502#ifdef DEBUGGING
8aa23a47 3503 nxt2 = nxt;
497b47a8 3504#endif
8aa23a47
YO
3505 nxt = regnext(nxt);
3506 if (OP(nxt) != CLOSE)
3507 goto nogo;
3508 if (RExC_open_parens) {
3509 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3510 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3511 }
3512 /* Now we know that nxt2 is the only contents: */
3513 oscan->flags = (U8)ARG(nxt);
3514 OP(oscan) = CURLYN;
3515 OP(nxt1) = NOTHING; /* was OPEN. */
40d049e4 3516
c277df42 3517#ifdef DEBUGGING
8aa23a47 3518 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
fda99bee
KW
3519 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3520 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
8aa23a47
YO
3521 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3522 OP(nxt + 1) = OPTIMIZED; /* was count. */
fda99bee 3523 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
b81d288d 3524#endif
8aa23a47
YO
3525 }
3526 nogo:
3527
3528 /* Try optimization CURLYX => CURLYM. */
3529 if ( OP(oscan) == CURLYX && data
3530 && !(data->flags & SF_HAS_PAR)
3531 && !(data->flags & SF_HAS_EVAL)
3532 && !deltanext /* atom is fixed width */
3533 && minnext != 0 /* CURLYM can't handle zero width */
3534 ) {
3535 /* XXXX How to optimize if data == 0? */
3536 /* Optimize to a simpler form. */
3537 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3538 regnode *nxt2;
3539
3540 OP(oscan) = CURLYM;
3541 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3542 && (OP(nxt2) != WHILEM))
3543 nxt = nxt2;
3544 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3545 /* Need to optimize away parenths. */
b3c0965f 3546 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
8aa23a47
YO
3547 /* Set the parenth number. */
3548 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3549
8aa23a47
YO
3550 oscan->flags = (U8)ARG(nxt);
3551 if (RExC_open_parens) {
3552 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3553 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
40d049e4 3554 }
8aa23a47
YO
3555 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3556 OP(nxt) = OPTIMIZED; /* was CLOSE. */
40d049e4 3557
c277df42 3558#ifdef DEBUGGING
8aa23a47
YO
3559 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3560 OP(nxt + 1) = OPTIMIZED; /* was count. */
486ec47a
PAR
3561 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
3562 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
b81d288d 3563#endif
c277df42 3564#if 0
8aa23a47
YO
3565 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3566 regnode *nnxt = regnext(nxt1);
8aa23a47
YO
3567 if (nnxt == nxt) {
3568 if (reg_off_by_arg[OP(nxt1)])
3569 ARG_SET(nxt1, nxt2 - nxt1);
3570 else if (nxt2 - nxt1 < U16_MAX)
3571 NEXT_OFF(nxt1) = nxt2 - nxt1;
3572 else
3573 OP(nxt) = NOTHING; /* Cannot beautify */
c277df42 3574 }
8aa23a47 3575 nxt1 = nnxt;
c277df42 3576 }
5d1c421c 3577#endif
8aa23a47
YO
3578 /* Optimize again: */
3579 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
3580 NULL, stopparen, recursed, NULL, 0,depth+1);
3581 }
3582 else
3583 oscan->flags = 0;
3584 }
3585 else if ((OP(oscan) == CURLYX)
3586 && (flags & SCF_WHILEM_VISITED_POS)
3587 /* See the comment on a similar expression above.
3b753521 3588 However, this time it's not a subexpression
8aa23a47
YO
3589 we care about, but the expression itself. */
3590 && (maxcount == REG_INFTY)
3591 && data && ++data->whilem_c < 16) {
3592 /* This stays as CURLYX, we can put the count/of pair. */
3593 /* Find WHILEM (as in regexec.c) */
3594 regnode *nxt = oscan + NEXT_OFF(oscan);
3595
3596 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
3597 nxt += ARG(nxt);
3598 PREVOPER(nxt)->flags = (U8)(data->whilem_c
3599 | (RExC_whilem_seen << 4)); /* On WHILEM */
3600 }
3601 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
3602 pars++;
3603 if (flags & SCF_DO_SUBSTR) {
3604 SV *last_str = NULL;
3605 int counted = mincount != 0;
a0ed51b3 3606
8aa23a47
YO
3607 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
3608#if defined(SPARC64_GCC_WORKAROUND)
3609 I32 b = 0;
3610 STRLEN l = 0;
3611 const char *s = NULL;
3612 I32 old = 0;
b515a41d 3613
8aa23a47
YO
3614 if (pos_before >= data->last_start_min)
3615 b = pos_before;
3616 else
3617 b = data->last_start_min;
b515a41d 3618
8aa23a47
YO
3619 l = 0;
3620 s = SvPV_const(data->last_found, l);
3621 old = b - data->last_start_min;
3622
3623#else
3624 I32 b = pos_before >= data->last_start_min
3625 ? pos_before : data->last_start_min;
3626 STRLEN l;
3627 const char * const s = SvPV_const(data->last_found, l);
3628 I32 old = b - data->last_start_min;
3629#endif
3630
3631 if (UTF)
3632 old = utf8_hop((U8*)s, old) - (U8*)s;
8aa23a47
YO
3633 l -= old;
3634 /* Get the added string: */
740cce10 3635 last_str = newSVpvn_utf8(s + old, l, UTF);
8aa23a47
YO
3636 if (deltanext == 0 && pos_before == b) {
3637 /* What was added is a constant string */
3638 if (mincount > 1) {
3639 SvGROW(last_str, (mincount * l) + 1);
3640 repeatcpy(SvPVX(last_str) + l,
3641 SvPVX_const(last_str), l, mincount - 1);
3642 SvCUR_set(last_str, SvCUR(last_str) * mincount);
3643 /* Add additional parts. */
3644 SvCUR_set(data->last_found,
3645 SvCUR(data->last_found) - l);
3646 sv_catsv(data->last_found, last_str);
3647 {
3648 SV * sv = data->last_found;
3649 MAGIC *mg =
3650 SvUTF8(sv) && SvMAGICAL(sv) ?
3651 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3652 if (mg && mg->mg_len >= 0)
bd94e887 3653 mg->mg_len += CHR_SVLEN(last_str) - l;
b515a41d 3654 }
8aa23a47 3655 data->last_end += l * (mincount - 1);
b515a41d 3656 }
8aa23a47
YO
3657 } else {
3658 /* start offset must point into the last copy */
3659 data->last_start_min += minnext * (mincount - 1);
3660 data->last_start_max += is_inf ? I32_MAX
3661 : (maxcount - 1) * (minnext + data->pos_delta);
3662 }
c277df42 3663 }
8aa23a47
YO
3664 /* It is counted once already... */
3665 data->pos_min += minnext * (mincount - counted);