This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Rmv no-longer used variables and code
[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 88#include "dquote_static.c"
a3e1f3a6 89#include "charclass_invlists.h"
04e98a4d 90
d4cce5f1 91#ifdef op
11343788 92#undef op
d4cce5f1 93#endif /* op */
11343788 94
fe14fcc3 95#ifdef MSDOS
7e4e8c89 96# if defined(BUGGY_MSC6)
fe14fcc3 97 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
7e4e8c89 98# pragma optimize("a",off)
fe14fcc3 99 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
7e4e8c89
NC
100# pragma optimize("w",on )
101# endif /* BUGGY_MSC6 */
fe14fcc3
LW
102#endif /* MSDOS */
103
a687059c
LW
104#ifndef STATIC
105#define STATIC static
106#endif
107
830247a4 108typedef struct RExC_state_t {
e2509266 109 U32 flags; /* are we folding, multilining? */
830247a4 110 char *precomp; /* uncompiled string. */
288b8c02 111 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf
YO
112 regexp *rx; /* perl core regexp structure */
113 regexp_internal *rxi; /* internal data for regexp object pprivate field */
fac92740 114 char *start; /* Start of input for compile */
830247a4
IZ
115 char *end; /* End of input for compile */
116 char *parse; /* Input-scan pointer. */
117 I32 whilem_seen; /* number of WHILEM in this expr */
fac92740 118 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 119 regnode *emit_bound; /* First regnode outside of the allocated space */
ffc61ed2 120 regnode *emit; /* Code-emit pointer; &regdummy = don't = compiling */
830247a4
IZ
121 I32 naughty; /* How bad is this pattern? */
122 I32 sawback; /* Did we see \1, ...? */
123 U32 seen;
124 I32 size; /* Code size. */
c74340f9
YO
125 I32 npar; /* Capture buffer count, (OPEN). */
126 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 127 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
128 I32 extralen;
129 I32 seen_zerolen;
130 I32 seen_evals;
40d049e4
YO
131 regnode **open_parens; /* pointers to open parens */
132 regnode **close_parens; /* pointers to close parens */
133 regnode *opend; /* END node in program */
02daf0ab
YO
134 I32 utf8; /* whether the pattern is utf8 or not */
135 I32 orig_utf8; /* whether the pattern was originally in utf8 */
136 /* XXX use this for future optimisation of case
137 * where pattern must be upgraded to utf8. */
e40e74fe
KW
138 I32 uni_semantics; /* If a d charset modifier should use unicode
139 rules, even if the pattern is not in
140 utf8 */
81714fb9 141 HV *paren_names; /* Paren names */
1f1031fe 142
40d049e4
YO
143 regnode **recurse; /* Recurse regops */
144 I32 recurse_count; /* Number of recurse regops */
b57e4118 145 I32 in_lookbehind;
4624b182 146 I32 contains_locale;
bb3f3ed2 147 I32 override_recoding;
830247a4
IZ
148#if ADD_TO_REGEXEC
149 char *starttry; /* -Dr: where regtry was called. */
150#define RExC_starttry (pRExC_state->starttry)
151#endif
3dab1dad 152#ifdef DEBUGGING
be8e71aa 153 const char *lastparse;
3dab1dad 154 I32 lastnum;
1f1031fe 155 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
156#define RExC_lastparse (pRExC_state->lastparse)
157#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 158#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 159#endif
830247a4
IZ
160} RExC_state_t;
161
e2509266 162#define RExC_flags (pRExC_state->flags)
830247a4 163#define RExC_precomp (pRExC_state->precomp)
288b8c02 164#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 165#define RExC_rx (pRExC_state->rx)
f8fc2ecf 166#define RExC_rxi (pRExC_state->rxi)
fac92740 167#define RExC_start (pRExC_state->start)
830247a4
IZ
168#define RExC_end (pRExC_state->end)
169#define RExC_parse (pRExC_state->parse)
170#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
171#ifdef RE_TRACK_PATTERN_OFFSETS
172#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
173#endif
830247a4 174#define RExC_emit (pRExC_state->emit)
fac92740 175#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 176#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
177#define RExC_naughty (pRExC_state->naughty)
178#define RExC_sawback (pRExC_state->sawback)
179#define RExC_seen (pRExC_state->seen)
180#define RExC_size (pRExC_state->size)
181#define RExC_npar (pRExC_state->npar)
e2e6a0f1 182#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
183#define RExC_extralen (pRExC_state->extralen)
184#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
185#define RExC_seen_evals (pRExC_state->seen_evals)
1aa99e6b 186#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 187#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 188#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
189#define RExC_open_parens (pRExC_state->open_parens)
190#define RExC_close_parens (pRExC_state->close_parens)
191#define RExC_opend (pRExC_state->opend)
81714fb9 192#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
193#define RExC_recurse (pRExC_state->recurse)
194#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 195#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 196#define RExC_contains_locale (pRExC_state->contains_locale)
bb3f3ed2 197#define RExC_override_recoding (pRExC_state->override_recoding)
830247a4 198
cde0cee5 199
a687059c
LW
200#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
201#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
202 ((*s) == '{' && regcurly(s)))
a687059c 203
35c8bce7
LW
204#ifdef SPSTART
205#undef SPSTART /* dratted cpp namespace... */
206#endif
a687059c
LW
207/*
208 * Flags to be passed up and down.
209 */
a687059c 210#define WORST 0 /* Worst case. */
a3b492c3 211#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee
KW
212
213/* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
d7b56a3c 214 * character, and if utf8, must be invariant. Note that this is not the same thing as REGNODE_SIMPLE */
fda99bee 215#define SIMPLE 0x02
a3b492c3
YO
216#define SPSTART 0x04 /* Starts with * or +. */
217#define TRYAGAIN 0x08 /* Weeded out a declaration. */
218#define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
a687059c 219
3dab1dad
YO
220#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
221
07be1b83
YO
222/* whether trie related optimizations are enabled */
223#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
224#define TRIE_STUDY_OPT
786e8c11 225#define FULL_TRIE_STUDY
07be1b83
YO
226#define TRIE_STCLASS
227#endif
1de06328
YO
228
229
40d049e4
YO
230
231#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
232#define PBITVAL(paren) (1 << ((paren) & 7))
233#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
234#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
235#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
236
bbd61b5f
KW
237/* If not already in utf8, do a longjmp back to the beginning */
238#define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
239#define REQUIRE_UTF8 STMT_START { \
240 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
241 } STMT_END
40d049e4 242
1de06328
YO
243/* About scan_data_t.
244
245 During optimisation we recurse through the regexp program performing
246 various inplace (keyhole style) optimisations. In addition study_chunk
247 and scan_commit populate this data structure with information about
248 what strings MUST appear in the pattern. We look for the longest
3b753521 249 string that must appear at a fixed location, and we look for the
1de06328
YO
250 longest string that may appear at a floating location. So for instance
251 in the pattern:
252
253 /FOO[xX]A.*B[xX]BAR/
254
255 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
256 strings (because they follow a .* construct). study_chunk will identify
257 both FOO and BAR as being the longest fixed and floating strings respectively.
258
259 The strings can be composites, for instance
260
261 /(f)(o)(o)/
262
263 will result in a composite fixed substring 'foo'.
264
265 For each string some basic information is maintained:
266
267 - offset or min_offset
268 This is the position the string must appear at, or not before.
269 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
270 characters must match before the string we are searching for.
271 Likewise when combined with minlenp and the length of the string it
1de06328
YO
272 tells us how many characters must appear after the string we have
273 found.
274
275 - max_offset
276 Only used for floating strings. This is the rightmost point that
3b753521 277 the string can appear at. If set to I32 max it indicates that the
1de06328
YO
278 string can occur infinitely far to the right.
279
280 - minlenp
281 A pointer to the minimum length of the pattern that the string
282 was found inside. This is important as in the case of positive
283 lookahead or positive lookbehind we can have multiple patterns
284 involved. Consider
285
286 /(?=FOO).*F/
287
288 The minimum length of the pattern overall is 3, the minimum length
289 of the lookahead part is 3, but the minimum length of the part that
290 will actually match is 1. So 'FOO's minimum length is 3, but the
291 minimum length for the F is 1. This is important as the minimum length
292 is used to determine offsets in front of and behind the string being
293 looked for. Since strings can be composites this is the length of the
486ec47a 294 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
295 the length is calculated by study_chunk, so that the minimum lengths
296 are not known until the full pattern has been compiled, thus the
297 pointer to the value.
298
299 - lookbehind
300
301 In the case of lookbehind the string being searched for can be
302 offset past the start point of the final matching string.
303 If this value was just blithely removed from the min_offset it would
304 invalidate some of the calculations for how many chars must match
305 before or after (as they are derived from min_offset and minlen and
306 the length of the string being searched for).
307 When the final pattern is compiled and the data is moved from the
308 scan_data_t structure into the regexp structure the information
309 about lookbehind is factored in, with the information that would
310 have been lost precalculated in the end_shift field for the
311 associated string.
312
313 The fields pos_min and pos_delta are used to store the minimum offset
314 and the delta to the maximum offset at the current point in the pattern.
315
316*/
2c2d71f5
JH
317
318typedef struct scan_data_t {
1de06328
YO
319 /*I32 len_min; unused */
320 /*I32 len_delta; unused */
2c2d71f5
JH
321 I32 pos_min;
322 I32 pos_delta;
323 SV *last_found;
1de06328 324 I32 last_end; /* min value, <0 unless valid. */
2c2d71f5
JH
325 I32 last_start_min;
326 I32 last_start_max;
1de06328
YO
327 SV **longest; /* Either &l_fixed, or &l_float. */
328 SV *longest_fixed; /* longest fixed string found in pattern */
329 I32 offset_fixed; /* offset where it starts */
486ec47a 330 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
331 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
332 SV *longest_float; /* longest floating string found in pattern */
333 I32 offset_float_min; /* earliest point in string it can appear */
334 I32 offset_float_max; /* latest point in string it can appear */
486ec47a 335 I32 *minlen_float; /* pointer to the minlen relevant to the string */
1de06328 336 I32 lookbehind_float; /* is the position of the string modified by LB */
2c2d71f5
JH
337 I32 flags;
338 I32 whilem_c;
cb434fcc 339 I32 *last_closep;
653099ff 340 struct regnode_charclass_class *start_class;
2c2d71f5
JH
341} scan_data_t;
342
a687059c 343/*
e50aee73 344 * Forward declarations for pregcomp()'s friends.
a687059c 345 */
a0d0e21e 346
27da23d5 347static const scan_data_t zero_scan_data =
1de06328 348 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
349
350#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
351#define SF_BEFORE_SEOL 0x0001
352#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
353#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
354#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
355
09b7f37c
CB
356#ifdef NO_UNARY_PLUS
357# define SF_FIX_SHIFT_EOL (0+2)
358# define SF_FL_SHIFT_EOL (0+4)
359#else
360# define SF_FIX_SHIFT_EOL (+2)
361# define SF_FL_SHIFT_EOL (+4)
362#endif
c277df42
IZ
363
364#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
365#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
366
367#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
368#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
369#define SF_IS_INF 0x0040
370#define SF_HAS_PAR 0x0080
371#define SF_IN_PAR 0x0100
372#define SF_HAS_EVAL 0x0200
373#define SCF_DO_SUBSTR 0x0400
653099ff
GS
374#define SCF_DO_STCLASS_AND 0x0800
375#define SCF_DO_STCLASS_OR 0x1000
376#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 377#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 378
786e8c11 379#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 380#define SCF_SEEN_ACCEPT 0x8000
07be1b83 381
43fead97 382#define UTF cBOOL(RExC_utf8)
a62b1201
KW
383#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
384#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
385#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
386#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
387#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1
KW
388#define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
389#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a62b1201 390
43fead97 391#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 392
ffc61ed2 393#define OOB_UNICODE 12345678
93733859 394#define OOB_NAMEDCLASS -1
b8c5462f 395
a0ed51b3
LW
396#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
397#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
398
8615cb43 399
b45f050a
JF
400/* length of regex to show in messages that don't mark a position within */
401#define RegexLengthToShowInErrorMessages 127
402
403/*
404 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
405 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
406 * op/pragma/warn/regcomp.
407 */
7253e4e3
RK
408#define MARKER1 "<-- HERE" /* marker as it appears in the description */
409#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 410
7253e4e3 411#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
412
413/*
414 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
415 * arg. Show regex, up to a maximum length. If it's too long, chop and add
416 * "...".
417 */
58e23c8d 418#define _FAIL(code) STMT_START { \
bfed75c6 419 const char *ellipses = ""; \
ccb2c380
MP
420 IV len = RExC_end - RExC_precomp; \
421 \
422 if (!SIZE_ONLY) \
288b8c02 423 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
424 if (len > RegexLengthToShowInErrorMessages) { \
425 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
426 len = RegexLengthToShowInErrorMessages - 10; \
427 ellipses = "..."; \
428 } \
58e23c8d 429 code; \
ccb2c380 430} STMT_END
8615cb43 431
58e23c8d
YO
432#define FAIL(msg) _FAIL( \
433 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
434 msg, (int)len, RExC_precomp, ellipses))
435
436#define FAIL2(msg,arg) _FAIL( \
437 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
438 arg, (int)len, RExC_precomp, ellipses))
439
b45f050a 440/*
b45f050a
JF
441 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
442 */
ccb2c380 443#define Simple_vFAIL(m) STMT_START { \
a28509cc 444 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
445 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
446 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
447} STMT_END
b45f050a
JF
448
449/*
450 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
451 */
ccb2c380
MP
452#define vFAIL(m) STMT_START { \
453 if (!SIZE_ONLY) \
288b8c02 454 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
455 Simple_vFAIL(m); \
456} STMT_END
b45f050a
JF
457
458/*
459 * Like Simple_vFAIL(), but accepts two arguments.
460 */
ccb2c380 461#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 462 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
463 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
464 (int)offset, RExC_precomp, RExC_precomp + offset); \
465} STMT_END
b45f050a
JF
466
467/*
468 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
469 */
ccb2c380
MP
470#define vFAIL2(m,a1) STMT_START { \
471 if (!SIZE_ONLY) \
288b8c02 472 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
473 Simple_vFAIL2(m, a1); \
474} STMT_END
b45f050a
JF
475
476
477/*
478 * Like Simple_vFAIL(), but accepts three arguments.
479 */
ccb2c380 480#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 481 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
482 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
483 (int)offset, RExC_precomp, RExC_precomp + offset); \
484} STMT_END
b45f050a
JF
485
486/*
487 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
488 */
ccb2c380
MP
489#define vFAIL3(m,a1,a2) STMT_START { \
490 if (!SIZE_ONLY) \
288b8c02 491 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
492 Simple_vFAIL3(m, a1, a2); \
493} STMT_END
b45f050a
JF
494
495/*
496 * Like Simple_vFAIL(), but accepts four arguments.
497 */
ccb2c380 498#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 499 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
500 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
501 (int)offset, RExC_precomp, RExC_precomp + offset); \
502} STMT_END
b45f050a 503
668c081a 504#define ckWARNreg(loc,m) STMT_START { \
a28509cc 505 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
506 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
507 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
508} STMT_END
509
668c081a 510#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 511 const IV offset = loc - RExC_precomp; \
d1d15184 512 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
513 m REPORT_LOCATION, \
514 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
515} STMT_END
516
2335b3d3
KW
517#define ckWARN2regdep(loc,m, a1) STMT_START { \
518 const IV offset = loc - RExC_precomp; \
519 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
520 m REPORT_LOCATION, \
521 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
522} STMT_END
523
668c081a 524#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 525 const IV offset = loc - RExC_precomp; \
668c081a 526 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
527 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
528} STMT_END
529
530#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 531 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
532 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
533 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
534} STMT_END
535
668c081a
NC
536#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
537 const IV offset = loc - RExC_precomp; \
538 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
539 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
540} STMT_END
541
ccb2c380 542#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 543 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
544 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
545 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
546} STMT_END
547
668c081a
NC
548#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
549 const IV offset = loc - RExC_precomp; \
550 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
551 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
552} STMT_END
553
ccb2c380 554#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 555 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
556 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
557 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
558} STMT_END
9d1d55b5 559
8615cb43 560
cd439c50 561/* Allow for side effects in s */
ccb2c380
MP
562#define REGC(c,s) STMT_START { \
563 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
564} STMT_END
cd439c50 565
fac92740
MJD
566/* Macros for recording node offsets. 20001227 mjd@plover.com
567 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
568 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
569 * Element 0 holds the number n.
07be1b83 570 * Position is 1 indexed.
fac92740 571 */
7122b237
YO
572#ifndef RE_TRACK_PATTERN_OFFSETS
573#define Set_Node_Offset_To_R(node,byte)
574#define Set_Node_Offset(node,byte)
575#define Set_Cur_Node_Offset
576#define Set_Node_Length_To_R(node,len)
577#define Set_Node_Length(node,len)
578#define Set_Node_Cur_Length(node)
579#define Node_Offset(n)
580#define Node_Length(n)
581#define Set_Node_Offset_Length(node,offset,len)
582#define ProgLen(ri) ri->u.proglen
583#define SetProgLen(ri,x) ri->u.proglen = x
584#else
585#define ProgLen(ri) ri->u.offsets[0]
586#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
587#define Set_Node_Offset_To_R(node,byte) STMT_START { \
588 if (! SIZE_ONLY) { \
589 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 590 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 591 if((node) < 0) { \
551405c4 592 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
593 } else { \
594 RExC_offsets[2*(node)-1] = (byte); \
595 } \
596 } \
597} STMT_END
598
599#define Set_Node_Offset(node,byte) \
600 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
601#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
602
603#define Set_Node_Length_To_R(node,len) STMT_START { \
604 if (! SIZE_ONLY) { \
605 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 606 __LINE__, (int)(node), (int)(len))); \
ccb2c380 607 if((node) < 0) { \
551405c4 608 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
609 } else { \
610 RExC_offsets[2*(node)] = (len); \
611 } \
612 } \
613} STMT_END
614
615#define Set_Node_Length(node,len) \
616 Set_Node_Length_To_R((node)-RExC_emit_start, len)
617#define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
618#define Set_Node_Cur_Length(node) \
619 Set_Node_Length(node, RExC_parse - parse_start)
fac92740
MJD
620
621/* Get offsets and lengths */
622#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
623#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
624
07be1b83
YO
625#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
626 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
627 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
628} STMT_END
7122b237 629#endif
07be1b83
YO
630
631#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
632#define EXPERIMENTAL_INPLACESCAN
f427392e 633#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 634
304ee84b
YO
635#define DEBUG_STUDYDATA(str,data,depth) \
636DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 637 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
638 "%*s" str "Pos:%"IVdf"/%"IVdf \
639 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
640 (int)(depth)*2, "", \
641 (IV)((data)->pos_min), \
642 (IV)((data)->pos_delta), \
304ee84b 643 (UV)((data)->flags), \
1de06328 644 (IV)((data)->whilem_c), \
304ee84b
YO
645 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
646 is_inf ? "INF " : "" \
1de06328
YO
647 ); \
648 if ((data)->last_found) \
649 PerlIO_printf(Perl_debug_log, \
650 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
651 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
652 SvPVX_const((data)->last_found), \
653 (IV)((data)->last_end), \
654 (IV)((data)->last_start_min), \
655 (IV)((data)->last_start_max), \
656 ((data)->longest && \
657 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
658 SvPVX_const((data)->longest_fixed), \
659 (IV)((data)->offset_fixed), \
660 ((data)->longest && \
661 (data)->longest==&((data)->longest_float)) ? "*" : "", \
662 SvPVX_const((data)->longest_float), \
663 (IV)((data)->offset_float_min), \
664 (IV)((data)->offset_float_max) \
665 ); \
666 PerlIO_printf(Perl_debug_log,"\n"); \
667});
668
acfe0abc 669static void clear_re(pTHX_ void *r);
4327152a 670
653099ff 671/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 672 Update the longest found anchored substring and the longest found
653099ff
GS
673 floating substrings if needed. */
674
4327152a 675STATIC void
304ee84b 676S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 677{
e1ec3a88
AL
678 const STRLEN l = CHR_SVLEN(data->last_found);
679 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 680 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 681
7918f24d
NC
682 PERL_ARGS_ASSERT_SCAN_COMMIT;
683
c277df42 684 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 685 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
686 if (*data->longest == data->longest_fixed) {
687 data->offset_fixed = l ? data->last_start_min : data->pos_min;
688 if (data->flags & SF_BEFORE_EOL)
b81d288d 689 data->flags
c277df42
IZ
690 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
691 else
692 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 693 data->minlen_fixed=minlenp;
1de06328 694 data->lookbehind_fixed=0;
a0ed51b3 695 }
304ee84b 696 else { /* *data->longest == data->longest_float */
c277df42 697 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
698 data->offset_float_max = (l
699 ? data->last_start_max
c277df42 700 : data->pos_min + data->pos_delta);
304ee84b 701 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 702 data->offset_float_max = I32_MAX;
c277df42 703 if (data->flags & SF_BEFORE_EOL)
b81d288d 704 data->flags
c277df42
IZ
705 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
706 else
707 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
708 data->minlen_float=minlenp;
709 data->lookbehind_float=0;
c277df42
IZ
710 }
711 }
712 SvCUR_set(data->last_found, 0);
0eda9292 713 {
a28509cc 714 SV * const sv = data->last_found;
097eb12c
AL
715 if (SvUTF8(sv) && SvMAGICAL(sv)) {
716 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
717 if (mg)
718 mg->mg_len = 0;
719 }
0eda9292 720 }
c277df42
IZ
721 data->last_end = -1;
722 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 723 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
724}
725
653099ff
GS
726/* Can match anything (initialization) */
727STATIC void
3fffb88a 728S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 729{
7918f24d
NC
730 PERL_ARGS_ASSERT_CL_ANYTHING;
731
f8bef550 732 ANYOF_BITMAP_SETALL(cl);
dd58aee1 733 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
3ad98780 734 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
3fffb88a
KW
735
736 /* If any portion of the regex is to operate under locale rules,
737 * initialization includes it. The reason this isn't done for all regexes
738 * is that the optimizer was written under the assumption that locale was
739 * all-or-nothing. Given the complexity and lack of documentation in the
740 * optimizer, and that there are inadequate test cases for locale, so many
741 * parts of it may not work properly, it is safest to avoid locale unless
742 * necessary. */
743 if (RExC_contains_locale) {
9d7a1e63 744 ANYOF_CLASS_SETALL(cl); /* /l uses class */
3fffb88a
KW
745 cl->flags |= ANYOF_LOCALE;
746 }
9d7a1e63
KW
747 else {
748 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
749 }
653099ff
GS
750}
751
752/* Can match anything (initialization) */
753STATIC int
5f66b61c 754S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
755{
756 int value;
757
7918f24d
NC
758 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
759
aaa51d5e 760 for (value = 0; value <= ANYOF_MAX; value += 2)
653099ff
GS
761 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
762 return 1;
1aa99e6b
IH
763 if (!(cl->flags & ANYOF_UNICODE_ALL))
764 return 0;
10edeb5d 765 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 766 return 0;
653099ff
GS
767 return 1;
768}
769
770/* Can match anything (initialization) */
771STATIC void
e755fd73 772S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 773{
7918f24d
NC
774 PERL_ARGS_ASSERT_CL_INIT;
775
8ecf7187 776 Zero(cl, 1, struct regnode_charclass_class);
653099ff 777 cl->type = ANYOF;
3fffb88a 778 cl_anything(pRExC_state, cl);
1411dba4 779 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
780}
781
1051e1c4
KW
782/* These two functions currently do the exact same thing */
783#define cl_init_zero S_cl_init
653099ff 784
dd58aee1
KW
785/* 'AND' a given class with another one. Can create false positives. 'cl'
786 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
787 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 788STATIC void
5f66b61c 789S_cl_and(struct regnode_charclass_class *cl,
a28509cc 790 const struct regnode_charclass_class *and_with)
653099ff 791{
7918f24d 792 PERL_ARGS_ASSERT_CL_AND;
40d049e4
YO
793
794 assert(and_with->type == ANYOF);
1e6ade67 795
c6b76537 796 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
797 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
798 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 799 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
800 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
801 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
653099ff
GS
802 int i;
803
804 if (and_with->flags & ANYOF_INVERT)
805 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
806 cl->bitmap[i] &= ~and_with->bitmap[i];
807 else
808 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
809 cl->bitmap[i] &= and_with->bitmap[i];
810 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 811
c6b76537 812 if (and_with->flags & ANYOF_INVERT) {
8951c461 813
c6b76537
KW
814 /* Here, the and'ed node is inverted. Get the AND of the flags that
815 * aren't affected by the inversion. Those that are affected are
816 * handled individually below */
817 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
818 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
819 cl->flags |= affected_flags;
820
821 /* We currently don't know how to deal with things that aren't in the
822 * bitmap, but we know that the intersection is no greater than what
823 * is already in cl, so let there be false positives that get sorted
824 * out after the synthetic start class succeeds, and the node is
825 * matched for real. */
826
827 /* The inversion of these two flags indicate that the resulting
828 * intersection doesn't have them */
829 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
830 cl->flags &= ~ANYOF_UNICODE_ALL;
831 }
c6b76537
KW
832 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
833 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 834 }
1aa99e6b 835 }
c6b76537 836 else { /* and'd node is not inverted */
3ad98780
KW
837 U8 outside_bitmap_but_not_utf8; /* Temp variable */
838
137165a6 839 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
840
841 /* Here 'and_with' doesn't match anything outside the bitmap
842 * (except possibly ANYOF_UNICODE_ALL), which means the
843 * intersection can't either, except for ANYOF_UNICODE_ALL, in
844 * which case we don't know what the intersection is, but it's no
845 * greater than what cl already has, so can just leave it alone,
846 * with possible false positives */
847 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
848 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 849 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 850 }
137165a6 851 }
c6b76537
KW
852 else if (! ANYOF_NONBITMAP(cl)) {
853
854 /* Here, 'and_with' does match something outside the bitmap, and cl
855 * doesn't have a list of things to match outside the bitmap. If
856 * cl can match all code points above 255, the intersection will
3ad98780
KW
857 * be those above-255 code points that 'and_with' matches. If cl
858 * can't match all Unicode code points, it means that it can't
859 * match anything outside the bitmap (since the 'if' that got us
860 * into this block tested for that), so we leave the bitmap empty.
861 */
c6b76537
KW
862 if (cl->flags & ANYOF_UNICODE_ALL) {
863 ARG_SET(cl, ARG(and_with));
3ad98780
KW
864
865 /* and_with's ARG may match things that don't require UTF8.
866 * And now cl's will too, in spite of this being an 'and'. See
867 * the comments below about the kludge */
868 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
869 }
870 }
871 else {
872 /* Here, both 'and_with' and cl match something outside the
873 * bitmap. Currently we do not do the intersection, so just match
874 * whatever cl had at the beginning. */
875 }
876
877
3ad98780
KW
878 /* Take the intersection of the two sets of flags. However, the
879 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
880 * kludge around the fact that this flag is not treated like the others
881 * which are initialized in cl_anything(). The way the optimizer works
882 * is that the synthetic start class (SSC) is initialized to match
883 * anything, and then the first time a real node is encountered, its
884 * values are AND'd with the SSC's with the result being the values of
885 * the real node. However, there are paths through the optimizer where
886 * the AND never gets called, so those initialized bits are set
887 * inappropriately, which is not usually a big deal, as they just cause
888 * false positives in the SSC, which will just mean a probably
889 * imperceptible slow down in execution. However this bit has a
890 * higher false positive consequence in that it can cause utf8.pm,
891 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
892 * bigger slowdown and also causes significant extra memory to be used.
893 * In order to prevent this, the code now takes a different tack. The
894 * bit isn't set unless some part of the regular expression needs it,
895 * but once set it won't get cleared. This means that these extra
896 * modules won't get loaded unless there was some path through the
897 * pattern that would have required them anyway, and so any false
898 * positives that occur by not ANDing them out when they could be
899 * aren't as severe as they would be if we treated this bit like all
900 * the others */
901 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
902 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 903 cl->flags &= and_with->flags;
3ad98780 904 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 905 }
653099ff
GS
906}
907
dd58aee1
KW
908/* 'OR' a given class with another one. Can create false positives. 'cl'
909 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
910 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 911STATIC void
3fffb88a 912S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 913{
7918f24d
NC
914 PERL_ARGS_ASSERT_CL_OR;
915
653099ff 916 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
917
918 /* Here, the or'd node is to be inverted. This means we take the
919 * complement of everything not in the bitmap, but currently we don't
920 * know what that is, so give up and match anything */
921 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 922 cl_anything(pRExC_state, cl);
c6b76537 923 }
653099ff
GS
924 /* We do not use
925 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
926 * <= (B1 | !B2) | (CL1 | !CL2)
927 * which is wasteful if CL2 is small, but we ignore CL2:
928 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
929 * XXXX Can we handle case-fold? Unclear:
930 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
931 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
932 */
c6b76537 933 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
934 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
935 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
653099ff
GS
936 int i;
937
938 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
939 cl->bitmap[i] |= ~or_with->bitmap[i];
940 } /* XXXX: logic is complicated otherwise */
941 else {
3fffb88a 942 cl_anything(pRExC_state, cl);
653099ff 943 }
c6b76537
KW
944
945 /* And, we can just take the union of the flags that aren't affected
946 * by the inversion */
947 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
948
949 /* For the remaining flags:
950 ANYOF_UNICODE_ALL and inverted means to not match anything above
951 255, which means that the union with cl should just be
952 what cl has in it, so can ignore this flag
953 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
954 is 127-255 to match them, but then invert that, so the
955 union with cl should just be what cl has in it, so can
956 ignore this flag
957 */
958 } else { /* 'or_with' is not inverted */
653099ff
GS
959 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
960 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
961 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
962 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
653099ff
GS
963 int i;
964
965 /* OR char bitmap and class bitmap separately */
966 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
967 cl->bitmap[i] |= or_with->bitmap[i];
1e6ade67 968 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
653099ff
GS
969 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
970 cl->classflags[i] |= or_with->classflags[i];
971 cl->flags |= ANYOF_CLASS;
972 }
973 }
974 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 975 cl_anything(pRExC_state, cl);
653099ff 976 }
9826f543 977
c6b76537
KW
978 if (ANYOF_NONBITMAP(or_with)) {
979
980 /* Use the added node's outside-the-bit-map match if there isn't a
981 * conflict. If there is a conflict (both nodes match something
982 * outside the bitmap, but what they match outside is not the same
983 * pointer, and hence not easily compared until XXX we extend
984 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
985 * match everything outside the bitmap. If that stuff is all above
986 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
987 if (! ANYOF_NONBITMAP(cl)) {
988 ARG_SET(cl, ARG(or_with));
989 }
990 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
991
992 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
993 cl_anything(pRExC_state, cl);
994 }
995 else {
996 cl->flags |= ANYOF_UNICODE_ALL;
997 }
c6b76537 998 }
4c34a693 999 }
0b9668ee
KW
1000
1001 /* Take the union */
1002 cl->flags |= or_with->flags;
1aa99e6b 1003 }
653099ff
GS
1004}
1005
a3621e74
YO
1006#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1007#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1008#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1009#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1010
3dab1dad
YO
1011
1012#ifdef DEBUGGING
07be1b83 1013/*
2b8b4781
NC
1014 dump_trie(trie,widecharmap,revcharmap)
1015 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1016 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1017
1018 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1019 The _interim_ variants are used for debugging the interim
1020 tables that are used to generate the final compressed
1021 representation which is what dump_trie expects.
1022
486ec47a 1023 Part of the reason for their existence is to provide a form
3dab1dad 1024 of documentation as to how the different representations function.
07be1b83
YO
1025
1026*/
3dab1dad
YO
1027
1028/*
3dab1dad
YO
1029 Dumps the final compressed table form of the trie to Perl_debug_log.
1030 Used for debugging make_trie().
1031*/
b9a59e08 1032
3dab1dad 1033STATIC void
2b8b4781
NC
1034S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1035 AV *revcharmap, U32 depth)
3dab1dad
YO
1036{
1037 U32 state;
ab3bbdeb 1038 SV *sv=sv_newmortal();
55eed653 1039 int colwidth= widecharmap ? 6 : 4;
2e64971a 1040 U16 word;
3dab1dad
YO
1041 GET_RE_DEBUG_FLAGS_DECL;
1042
7918f24d 1043 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1044
3dab1dad
YO
1045 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1046 (int)depth * 2 + 2,"",
1047 "Match","Base","Ofs" );
1048
1049 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1050 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1051 if ( tmp ) {
ab3bbdeb
YO
1052 PerlIO_printf( Perl_debug_log, "%*s",
1053 colwidth,
ddc5bc0f 1054 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1055 PL_colors[0], PL_colors[1],
1056 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1057 PERL_PV_ESCAPE_FIRSTCHAR
1058 )
1059 );
3dab1dad
YO
1060 }
1061 }
1062 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1063 (int)depth * 2 + 2,"");
1064
1065 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1066 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1067 PerlIO_printf( Perl_debug_log, "\n");
1068
1e2e3d02 1069 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1070 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1071
1072 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1073
1074 if ( trie->states[ state ].wordnum ) {
1075 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1076 } else {
1077 PerlIO_printf( Perl_debug_log, "%6s", "" );
1078 }
1079
1080 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1081
1082 if ( base ) {
1083 U32 ofs = 0;
1084
1085 while( ( base + ofs < trie->uniquecharcount ) ||
1086 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1087 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1088 ofs++;
1089
1090 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1091
1092 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1093 if ( ( base + ofs >= trie->uniquecharcount ) &&
1094 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1095 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1096 {
ab3bbdeb
YO
1097 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1098 colwidth,
3dab1dad
YO
1099 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1100 } else {
ab3bbdeb 1101 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1102 }
1103 }
1104
1105 PerlIO_printf( Perl_debug_log, "]");
1106
1107 }
1108 PerlIO_printf( Perl_debug_log, "\n" );
1109 }
2e64971a
DM
1110 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1111 for (word=1; word <= trie->wordcount; word++) {
1112 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1113 (int)word, (int)(trie->wordinfo[word].prev),
1114 (int)(trie->wordinfo[word].len));
1115 }
1116 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1117}
1118/*
3dab1dad
YO
1119 Dumps a fully constructed but uncompressed trie in list form.
1120 List tries normally only are used for construction when the number of
1121 possible chars (trie->uniquecharcount) is very high.
1122 Used for debugging make_trie().
1123*/
1124STATIC void
55eed653 1125S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1126 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1127 U32 depth)
3dab1dad
YO
1128{
1129 U32 state;
ab3bbdeb 1130 SV *sv=sv_newmortal();
55eed653 1131 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1132 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1133
1134 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1135
3dab1dad 1136 /* print out the table precompression. */
ab3bbdeb
YO
1137 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1138 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1139 "------:-----+-----------------\n" );
3dab1dad
YO
1140
1141 for( state=1 ; state < next_alloc ; state ++ ) {
1142 U16 charid;
1143
ab3bbdeb 1144 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1145 (int)depth * 2 + 2,"", (UV)state );
1146 if ( ! trie->states[ state ].wordnum ) {
1147 PerlIO_printf( Perl_debug_log, "%5s| ","");
1148 } else {
1149 PerlIO_printf( Perl_debug_log, "W%4x| ",
1150 trie->states[ state ].wordnum
1151 );
1152 }
1153 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1154 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1155 if ( tmp ) {
1156 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1157 colwidth,
ddc5bc0f 1158 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1159 PL_colors[0], PL_colors[1],
1160 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1161 PERL_PV_ESCAPE_FIRSTCHAR
1162 ) ,
1e2e3d02
YO
1163 TRIE_LIST_ITEM(state,charid).forid,
1164 (UV)TRIE_LIST_ITEM(state,charid).newstate
1165 );
1166 if (!(charid % 10))
664e119d
RGS
1167 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1168 (int)((depth * 2) + 14), "");
1e2e3d02 1169 }
ab3bbdeb
YO
1170 }
1171 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1172 }
1173}
1174
1175/*
3dab1dad
YO
1176 Dumps a fully constructed but uncompressed trie in table form.
1177 This is the normal DFA style state transition table, with a few
1178 twists to facilitate compression later.
1179 Used for debugging make_trie().
1180*/
1181STATIC void
55eed653 1182S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1183 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1184 U32 depth)
3dab1dad
YO
1185{
1186 U32 state;
1187 U16 charid;
ab3bbdeb 1188 SV *sv=sv_newmortal();
55eed653 1189 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1190 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1191
1192 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1193
1194 /*
1195 print out the table precompression so that we can do a visual check
1196 that they are identical.
1197 */
1198
1199 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1200
1201 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1202 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1203 if ( tmp ) {
ab3bbdeb
YO
1204 PerlIO_printf( Perl_debug_log, "%*s",
1205 colwidth,
ddc5bc0f 1206 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1207 PL_colors[0], PL_colors[1],
1208 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1209 PERL_PV_ESCAPE_FIRSTCHAR
1210 )
1211 );
3dab1dad
YO
1212 }
1213 }
1214
1215 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1216
1217 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1218 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1219 }
1220
1221 PerlIO_printf( Perl_debug_log, "\n" );
1222
1223 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1224
1225 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1226 (int)depth * 2 + 2,"",
1227 (UV)TRIE_NODENUM( state ) );
1228
1229 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1230 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1231 if (v)
1232 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1233 else
1234 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1235 }
1236 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1237 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1238 } else {
1239 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1240 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1241 }
1242 }
07be1b83 1243}
3dab1dad
YO
1244
1245#endif
1246
2e64971a 1247
786e8c11
YO
1248/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1249 startbranch: the first branch in the whole branch sequence
1250 first : start branch of sequence of branch-exact nodes.
1251 May be the same as startbranch
1252 last : Thing following the last branch.
1253 May be the same as tail.
1254 tail : item following the branch sequence
1255 count : words in the sequence
1256 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1257 depth : indent depth
3dab1dad 1258
786e8c11 1259Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1260
786e8c11
YO
1261A trie is an N'ary tree where the branches are determined by digital
1262decomposition of the key. IE, at the root node you look up the 1st character and
1263follow that branch repeat until you find the end of the branches. Nodes can be
1264marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1265
786e8c11 1266 /he|she|his|hers/
72f13be8 1267
786e8c11
YO
1268would convert into the following structure. Numbers represent states, letters
1269following numbers represent valid transitions on the letter from that state, if
1270the number is in square brackets it represents an accepting state, otherwise it
1271will be in parenthesis.
07be1b83 1272
786e8c11
YO
1273 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1274 | |
1275 | (2)
1276 | |
1277 (1) +-i->(6)-+-s->[7]
1278 |
1279 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1280
786e8c11
YO
1281 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1282
1283This shows that when matching against the string 'hers' we will begin at state 1
1284read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1285then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1286is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1287single traverse. We store a mapping from accepting to state to which word was
1288matched, and then when we have multiple possibilities we try to complete the
1289rest of the regex in the order in which they occured in the alternation.
1290
1291The only prior NFA like behaviour that would be changed by the TRIE support is
1292the silent ignoring of duplicate alternations which are of the form:
1293
1294 / (DUPE|DUPE) X? (?{ ... }) Y /x
1295
4b714af6 1296Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1297and without the optimisation. With the optimisations dupes will be silently
486ec47a 1298ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1299the following demonstrates:
1300
1301 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1302
1303which prints out 'word' three times, but
1304
1305 'words'=~/(word|word|word)(?{ print $1 })S/
1306
1307which doesnt print it out at all. This is due to other optimisations kicking in.
1308
1309Example of what happens on a structural level:
1310
486ec47a 1311The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1312
1313 1: CURLYM[1] {1,32767}(18)
1314 5: BRANCH(8)
1315 6: EXACT <ac>(16)
1316 8: BRANCH(11)
1317 9: EXACT <ad>(16)
1318 11: BRANCH(14)
1319 12: EXACT <ab>(16)
1320 16: SUCCEED(0)
1321 17: NOTHING(18)
1322 18: END(0)
1323
1324This would be optimizable with startbranch=5, first=5, last=16, tail=16
1325and should turn into:
1326
1327 1: CURLYM[1] {1,32767}(18)
1328 5: TRIE(16)
1329 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1330 <ac>
1331 <ad>
1332 <ab>
1333 16: SUCCEED(0)
1334 17: NOTHING(18)
1335 18: END(0)
1336
1337Cases where tail != last would be like /(?foo|bar)baz/:
1338
1339 1: BRANCH(4)
1340 2: EXACT <foo>(8)
1341 4: BRANCH(7)
1342 5: EXACT <bar>(8)
1343 7: TAIL(8)
1344 8: EXACT <baz>(10)
1345 10: END(0)
1346
1347which would be optimizable with startbranch=1, first=1, last=7, tail=8
1348and would end up looking like:
1349
1350 1: TRIE(8)
1351 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1352 <foo>
1353 <bar>
1354 7: TAIL(8)
1355 8: EXACT <baz>(10)
1356 10: END(0)
1357
1358 d = uvuni_to_utf8_flags(d, uv, 0);
1359
1360is the recommended Unicode-aware way of saying
1361
1362 *(d++) = uv;
1363*/
1364
1e2e3d02 1365#define TRIE_STORE_REVCHAR \
786e8c11 1366 STMT_START { \
73031816
NC
1367 if (UTF) { \
1368 SV *zlopp = newSV(2); \
88c9ea1e
CB
1369 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1370 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, uvc & 0xFF); \
73031816
NC
1371 SvCUR_set(zlopp, kapow - flrbbbbb); \
1372 SvPOK_on(zlopp); \
1373 SvUTF8_on(zlopp); \
1374 av_push(revcharmap, zlopp); \
1375 } else { \
6bdeddd2 1376 char ooooff = (char)uvc; \
73031816
NC
1377 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1378 } \
1379 } STMT_END
786e8c11
YO
1380
1381#define TRIE_READ_CHAR STMT_START { \
1382 wordlen++; \
1383 if ( UTF ) { \
1384 if ( folder ) { \
1385 if ( foldlen > 0 ) { \
1386 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1387 foldlen -= len; \
1388 scan += len; \
1389 len = 0; \
1390 } else { \
c81f2f9e
KW
1391 len = UTF8SKIP(uc);\
1392 uvc = to_utf8_fold( uc, foldbuf, &foldlen); \
786e8c11
YO
1393 foldlen -= UNISKIP( uvc ); \
1394 scan = foldbuf + UNISKIP( uvc ); \
1395 } \
1396 } else { \
1397 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1398 } \
1399 } else { \
1400 uvc = (U32)*uc; \
1401 len = 1; \
1402 } \
1403} STMT_END
1404
1405
1406
1407#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1408 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1409 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1410 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1411 } \
1412 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1413 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1414 TRIE_LIST_CUR( state )++; \
1415} STMT_END
07be1b83 1416
786e8c11
YO
1417#define TRIE_LIST_NEW(state) STMT_START { \
1418 Newxz( trie->states[ state ].trans.list, \
1419 4, reg_trie_trans_le ); \
1420 TRIE_LIST_CUR( state ) = 1; \
1421 TRIE_LIST_LEN( state ) = 4; \
1422} STMT_END
07be1b83 1423
786e8c11
YO
1424#define TRIE_HANDLE_WORD(state) STMT_START { \
1425 U16 dupe= trie->states[ state ].wordnum; \
1426 regnode * const noper_next = regnext( noper ); \
1427 \
786e8c11
YO
1428 DEBUG_r({ \
1429 /* store the word for dumping */ \
1430 SV* tmp; \
1431 if (OP(noper) != NOTHING) \
740cce10 1432 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1433 else \
740cce10 1434 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1435 av_push( trie_words, tmp ); \
786e8c11
YO
1436 }); \
1437 \
1438 curword++; \
2e64971a
DM
1439 trie->wordinfo[curword].prev = 0; \
1440 trie->wordinfo[curword].len = wordlen; \
1441 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1442 \
1443 if ( noper_next < tail ) { \
1444 if (!trie->jump) \
c944940b 1445 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1446 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1447 if (!jumper) \
1448 jumper = noper_next; \
1449 if (!nextbranch) \
1450 nextbranch= regnext(cur); \
1451 } \
1452 \
1453 if ( dupe ) { \
2e64971a
DM
1454 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1455 /* chain, so that when the bits of chain are later */\
1456 /* linked together, the dups appear in the chain */\
1457 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1458 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1459 } else { \
1460 /* we haven't inserted this word yet. */ \
1461 trie->states[ state ].wordnum = curword; \
1462 } \
1463} STMT_END
07be1b83 1464
3dab1dad 1465
786e8c11
YO
1466#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1467 ( ( base + charid >= ucharcount \
1468 && base + charid < ubound \
1469 && state == trie->trans[ base - ucharcount + charid ].check \
1470 && trie->trans[ base - ucharcount + charid ].next ) \
1471 ? trie->trans[ base - ucharcount + charid ].next \
1472 : ( state==1 ? special : 0 ) \
1473 )
3dab1dad 1474
786e8c11
YO
1475#define MADE_TRIE 1
1476#define MADE_JUMP_TRIE 2
1477#define MADE_EXACT_TRIE 4
3dab1dad 1478
a3621e74 1479STATIC I32
786e8c11 1480S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1481{
27da23d5 1482 dVAR;
a3621e74
YO
1483 /* first pass, loop through and scan words */
1484 reg_trie_data *trie;
55eed653 1485 HV *widecharmap = NULL;
2b8b4781 1486 AV *revcharmap = newAV();
a3621e74 1487 regnode *cur;
9f7f3913 1488 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1489 STRLEN len = 0;
1490 UV uvc = 0;
1491 U16 curword = 0;
1492 U32 next_alloc = 0;
786e8c11
YO
1493 regnode *jumper = NULL;
1494 regnode *nextbranch = NULL;
7f69552c 1495 regnode *convert = NULL;
2e64971a 1496 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1497 /* we just use folder as a flag in utf8 */
1e696034 1498 const U8 * folder = NULL;
a3621e74 1499
2b8b4781
NC
1500#ifdef DEBUGGING
1501 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1502 AV *trie_words = NULL;
1503 /* along with revcharmap, this only used during construction but both are
1504 * useful during debugging so we store them in the struct when debugging.
8e11feef 1505 */
2b8b4781
NC
1506#else
1507 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1508 STRLEN trie_charcount=0;
3dab1dad 1509#endif
2b8b4781 1510 SV *re_trie_maxbuff;
a3621e74 1511 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1512
1513 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1514#ifndef DEBUGGING
1515 PERL_UNUSED_ARG(depth);
1516#endif
a3621e74 1517
1e696034 1518 switch (flags) {
c46d03cf 1519 case EXACT: break;
2f7f8cb1 1520 case EXACTFA:
1e696034
KW
1521 case EXACTFU: folder = PL_fold_latin1; break;
1522 case EXACTF: folder = PL_fold; break;
1523 case EXACTFL: folder = PL_fold_locale; break;
c46d03cf 1524 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u", (unsigned) flags );
1e696034
KW
1525 }
1526
c944940b 1527 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1528 trie->refcount = 1;
3dab1dad 1529 trie->startstate = 1;
786e8c11 1530 trie->wordcount = word_count;
f8fc2ecf 1531 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1532 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
3dab1dad 1533 if (!(UTF && folder))
c944940b 1534 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1535 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1536 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1537
a3621e74 1538 DEBUG_r({
2b8b4781 1539 trie_words = newAV();
a3621e74 1540 });
a3621e74 1541
0111c4fd 1542 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1543 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1544 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1545 }
3dab1dad
YO
1546 DEBUG_OPTIMISE_r({
1547 PerlIO_printf( Perl_debug_log,
786e8c11 1548 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1549 (int)depth * 2 + 2, "",
1550 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1551 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1552 (int)depth);
3dab1dad 1553 });
7f69552c
YO
1554
1555 /* Find the node we are going to overwrite */
1556 if ( first == startbranch && OP( last ) != BRANCH ) {
1557 /* whole branch chain */
1558 convert = first;
1559 } else {
1560 /* branch sub-chain */
1561 convert = NEXTOPER( first );
1562 }
1563
a3621e74
YO
1564 /* -- First loop and Setup --
1565
1566 We first traverse the branches and scan each word to determine if it
1567 contains widechars, and how many unique chars there are, this is
1568 important as we have to build a table with at least as many columns as we
1569 have unique chars.
1570
1571 We use an array of integers to represent the character codes 0..255
38a44b82 1572 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1573 native representation of the character value as the key and IV's for the
1574 coded index.
1575
1576 *TODO* If we keep track of how many times each character is used we can
1577 remap the columns so that the table compression later on is more
3b753521 1578 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1579 middle and the least common are on the outside. IMO this would be better
1580 than a most to least common mapping as theres a decent chance the most
1581 common letter will share a node with the least common, meaning the node
486ec47a 1582 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1583 case is when we have the least common nodes twice.
1584
1585 */
1586
a3621e74 1587 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
c445ea15 1588 regnode * const noper = NEXTOPER( cur );
e1ec3a88 1589 const U8 *uc = (U8*)STRING( noper );
a28509cc 1590 const U8 * const e = uc + STR_LEN( noper );
a3621e74
YO
1591 STRLEN foldlen = 0;
1592 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2af232bd 1593 const U8 *scan = (U8*)NULL;
07be1b83 1594 U32 wordlen = 0; /* required init */
02daf0ab
YO
1595 STRLEN chars = 0;
1596 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1597
3dab1dad
YO
1598 if (OP(noper) == NOTHING) {
1599 trie->minlen= 0;
1600 continue;
1601 }
02daf0ab
YO
1602 if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
1603 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1604 regardless of encoding */
1605
a3621e74 1606 for ( ; uc < e ; uc += len ) {
3dab1dad 1607 TRIE_CHARCOUNT(trie)++;
a3621e74 1608 TRIE_READ_CHAR;
3dab1dad 1609 chars++;
a3621e74
YO
1610 if ( uvc < 256 ) {
1611 if ( !trie->charmap[ uvc ] ) {
1612 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1613 if ( folder )
1614 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
3dab1dad 1615 TRIE_STORE_REVCHAR;
a3621e74 1616 }
02daf0ab 1617 if ( set_bit ) {
62012aee
KW
1618 /* store the codepoint in the bitmap, and its folded
1619 * equivalent. */
02daf0ab 1620 TRIE_BITMAP_SET(trie,uvc);
0921ee73
T
1621
1622 /* store the folded codepoint */
1623 if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);
1624
1625 if ( !UTF ) {
1626 /* store first byte of utf8 representation of
acdf4139
KW
1627 variant codepoints */
1628 if (! UNI_IS_INVARIANT(uvc)) {
1629 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1630 }
1631 }
02daf0ab
YO
1632 set_bit = 0; /* We've done our bit :-) */
1633 }
a3621e74
YO
1634 } else {
1635 SV** svpp;
55eed653
NC
1636 if ( !widecharmap )
1637 widecharmap = newHV();
a3621e74 1638
55eed653 1639 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1640
1641 if ( !svpp )
e4584336 1642 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1643
1644 if ( !SvTRUE( *svpp ) ) {
1645 sv_setiv( *svpp, ++trie->uniquecharcount );
3dab1dad 1646 TRIE_STORE_REVCHAR;
a3621e74
YO
1647 }
1648 }
1649 }
3dab1dad
YO
1650 if( cur == first ) {
1651 trie->minlen=chars;
1652 trie->maxlen=chars;
1653 } else if (chars < trie->minlen) {
1654 trie->minlen=chars;
1655 } else if (chars > trie->maxlen) {
1656 trie->maxlen=chars;
1657 }
1658
a3621e74
YO
1659 } /* end first pass */
1660 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1661 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1662 (int)depth * 2 + 2,"",
55eed653 1663 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1664 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1665 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1666 );
a3621e74
YO
1667
1668 /*
1669 We now know what we are dealing with in terms of unique chars and
1670 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1671 representation using a flat table will take. If it's over a reasonable
1672 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1673 conservative but potentially much slower representation using an array
1674 of lists.
1675
1676 At the end we convert both representations into the same compressed
1677 form that will be used in regexec.c for matching with. The latter
1678 is a form that cannot be used to construct with but has memory
1679 properties similar to the list form and access properties similar
1680 to the table form making it both suitable for fast searches and
1681 small enough that its feasable to store for the duration of a program.
1682
1683 See the comment in the code where the compressed table is produced
1684 inplace from the flat tabe representation for an explanation of how
1685 the compression works.
1686
1687 */
1688
1689
2e64971a
DM
1690 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1691 prev_states[1] = 0;
1692
3dab1dad 1693 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1694 /*
1695 Second Pass -- Array Of Lists Representation
1696
1697 Each state will be represented by a list of charid:state records
1698 (reg_trie_trans_le) the first such element holds the CUR and LEN
1699 points of the allocated array. (See defines above).
1700
1701 We build the initial structure using the lists, and then convert
1702 it into the compressed table form which allows faster lookups
1703 (but cant be modified once converted).
a3621e74
YO
1704 */
1705
a3621e74
YO
1706 STRLEN transcount = 1;
1707
1e2e3d02
YO
1708 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1709 "%*sCompiling trie using list compiler\n",
1710 (int)depth * 2 + 2, ""));
686b73d4 1711
c944940b
JH
1712 trie->states = (reg_trie_state *)
1713 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1714 sizeof(reg_trie_state) );
a3621e74
YO
1715 TRIE_LIST_NEW(1);
1716 next_alloc = 2;
1717
1718 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1719
c445ea15
AL
1720 regnode * const noper = NEXTOPER( cur );
1721 U8 *uc = (U8*)STRING( noper );
1722 const U8 * const e = uc + STR_LEN( noper );
1723 U32 state = 1; /* required init */
1724 U16 charid = 0; /* sanity init */
1725 U8 *scan = (U8*)NULL; /* sanity init */
1726 STRLEN foldlen = 0; /* required init */
07be1b83 1727 U32 wordlen = 0; /* required init */
c445ea15
AL
1728 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1729
3dab1dad 1730 if (OP(noper) != NOTHING) {
786e8c11 1731 for ( ; uc < e ; uc += len ) {
c445ea15 1732
786e8c11 1733 TRIE_READ_CHAR;
c445ea15 1734
786e8c11
YO
1735 if ( uvc < 256 ) {
1736 charid = trie->charmap[ uvc ];
c445ea15 1737 } else {
55eed653 1738 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1739 if ( !svpp ) {
1740 charid = 0;
1741 } else {
1742 charid=(U16)SvIV( *svpp );
1743 }
c445ea15 1744 }
786e8c11
YO
1745 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1746 if ( charid ) {
a3621e74 1747
786e8c11
YO
1748 U16 check;
1749 U32 newstate = 0;
a3621e74 1750
786e8c11
YO
1751 charid--;
1752 if ( !trie->states[ state ].trans.list ) {
1753 TRIE_LIST_NEW( state );
c445ea15 1754 }
786e8c11
YO
1755 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1756 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1757 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1758 break;
1759 }
1760 }
1761 if ( ! newstate ) {
1762 newstate = next_alloc++;
2e64971a 1763 prev_states[newstate] = state;
786e8c11
YO
1764 TRIE_LIST_PUSH( state, charid, newstate );
1765 transcount++;
1766 }
1767 state = newstate;
1768 } else {
1769 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1770 }
a28509cc 1771 }
c445ea15 1772 }
3dab1dad 1773 TRIE_HANDLE_WORD(state);
a3621e74
YO
1774
1775 } /* end second pass */
1776
1e2e3d02
YO
1777 /* next alloc is the NEXT state to be allocated */
1778 trie->statecount = next_alloc;
c944940b
JH
1779 trie->states = (reg_trie_state *)
1780 PerlMemShared_realloc( trie->states,
1781 next_alloc
1782 * sizeof(reg_trie_state) );
a3621e74 1783
3dab1dad 1784 /* and now dump it out before we compress it */
2b8b4781
NC
1785 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1786 revcharmap, next_alloc,
1787 depth+1)
1e2e3d02 1788 );
a3621e74 1789
c944940b
JH
1790 trie->trans = (reg_trie_trans *)
1791 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1792 {
1793 U32 state;
a3621e74
YO
1794 U32 tp = 0;
1795 U32 zp = 0;
1796
1797
1798 for( state=1 ; state < next_alloc ; state ++ ) {
1799 U32 base=0;
1800
1801 /*
1802 DEBUG_TRIE_COMPILE_MORE_r(
1803 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1804 );
1805 */
1806
1807 if (trie->states[state].trans.list) {
1808 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1809 U16 maxid=minid;
a28509cc 1810 U16 idx;
a3621e74
YO
1811
1812 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1813 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1814 if ( forid < minid ) {
1815 minid=forid;
1816 } else if ( forid > maxid ) {
1817 maxid=forid;
1818 }
a3621e74
YO
1819 }
1820 if ( transcount < tp + maxid - minid + 1) {
1821 transcount *= 2;
c944940b
JH
1822 trie->trans = (reg_trie_trans *)
1823 PerlMemShared_realloc( trie->trans,
446bd890
NC
1824 transcount
1825 * sizeof(reg_trie_trans) );
a3621e74
YO
1826 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1827 }
1828 base = trie->uniquecharcount + tp - minid;
1829 if ( maxid == minid ) {
1830 U32 set = 0;
1831 for ( ; zp < tp ; zp++ ) {
1832 if ( ! trie->trans[ zp ].next ) {
1833 base = trie->uniquecharcount + zp - minid;
1834 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1835 trie->trans[ zp ].check = state;
1836 set = 1;
1837 break;
1838 }
1839 }
1840 if ( !set ) {
1841 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1842 trie->trans[ tp ].check = state;
1843 tp++;
1844 zp = tp;
1845 }
1846 } else {
1847 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1848 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1849 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1850 trie->trans[ tid ].check = state;
1851 }
1852 tp += ( maxid - minid + 1 );
1853 }
1854 Safefree(trie->states[ state ].trans.list);
1855 }
1856 /*
1857 DEBUG_TRIE_COMPILE_MORE_r(
1858 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1859 );
1860 */
1861 trie->states[ state ].trans.base=base;
1862 }
cc601c31 1863 trie->lasttrans = tp + 1;
a3621e74
YO
1864 }
1865 } else {
1866 /*
1867 Second Pass -- Flat Table Representation.
1868
1869 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1870 We know that we will need Charcount+1 trans at most to store the data
1871 (one row per char at worst case) So we preallocate both structures
1872 assuming worst case.
1873
1874 We then construct the trie using only the .next slots of the entry
1875 structs.
1876
3b753521 1877 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1878 make compression both faster and easier by keeping track of how many non
1879 zero fields are in the node.
1880
1881 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1882 transition.
1883
1884 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1885 number representing the first entry of the node, and state as a
1886 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1887 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1888 are 2 entrys per node. eg:
1889
1890 A B A B
1891 1. 2 4 1. 3 7
1892 2. 0 3 3. 0 5
1893 3. 0 0 5. 0 0
1894 4. 0 0 7. 0 0
1895
1896 The table is internally in the right hand, idx form. However as we also
1897 have to deal with the states array which is indexed by nodenum we have to
1898 use TRIE_NODENUM() to convert.
1899
1900 */
1e2e3d02
YO
1901 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1902 "%*sCompiling trie using table compiler\n",
1903 (int)depth * 2 + 2, ""));
3dab1dad 1904
c944940b
JH
1905 trie->trans = (reg_trie_trans *)
1906 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1907 * trie->uniquecharcount + 1,
1908 sizeof(reg_trie_trans) );
1909 trie->states = (reg_trie_state *)
1910 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1911 sizeof(reg_trie_state) );
a3621e74
YO
1912 next_alloc = trie->uniquecharcount + 1;
1913
3dab1dad 1914
a3621e74
YO
1915 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1916
c445ea15 1917 regnode * const noper = NEXTOPER( cur );
a28509cc
AL
1918 const U8 *uc = (U8*)STRING( noper );
1919 const U8 * const e = uc + STR_LEN( noper );
a3621e74
YO
1920
1921 U32 state = 1; /* required init */
1922
1923 U16 charid = 0; /* sanity init */
1924 U32 accept_state = 0; /* sanity init */
1925 U8 *scan = (U8*)NULL; /* sanity init */
1926
1927 STRLEN foldlen = 0; /* required init */
07be1b83 1928 U32 wordlen = 0; /* required init */
a3621e74
YO
1929 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1930
3dab1dad 1931 if ( OP(noper) != NOTHING ) {
786e8c11 1932 for ( ; uc < e ; uc += len ) {
a3621e74 1933
786e8c11 1934 TRIE_READ_CHAR;
a3621e74 1935
786e8c11
YO
1936 if ( uvc < 256 ) {
1937 charid = trie->charmap[ uvc ];
1938 } else {
55eed653 1939 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 1940 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 1941 }
786e8c11
YO
1942 if ( charid ) {
1943 charid--;
1944 if ( !trie->trans[ state + charid ].next ) {
1945 trie->trans[ state + charid ].next = next_alloc;
1946 trie->trans[ state ].check++;
2e64971a
DM
1947 prev_states[TRIE_NODENUM(next_alloc)]
1948 = TRIE_NODENUM(state);
786e8c11
YO
1949 next_alloc += trie->uniquecharcount;
1950 }
1951 state = trie->trans[ state + charid ].next;
1952 } else {
1953 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1954 }
1955 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 1956 }
a3621e74 1957 }
3dab1dad
YO
1958 accept_state = TRIE_NODENUM( state );
1959 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
1960
1961 } /* end second pass */
1962
3dab1dad 1963 /* and now dump it out before we compress it */
2b8b4781
NC
1964 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1965 revcharmap,
1966 next_alloc, depth+1));
a3621e74 1967
a3621e74
YO
1968 {
1969 /*
1970 * Inplace compress the table.*
1971
1972 For sparse data sets the table constructed by the trie algorithm will
1973 be mostly 0/FAIL transitions or to put it another way mostly empty.
1974 (Note that leaf nodes will not contain any transitions.)
1975
1976 This algorithm compresses the tables by eliminating most such
1977 transitions, at the cost of a modest bit of extra work during lookup:
1978
1979 - Each states[] entry contains a .base field which indicates the
1980 index in the state[] array wheres its transition data is stored.
1981
3b753521 1982 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
1983
1984 - If .base is nonzero then charid is added to it to find an entry in
1985 the trans array.
1986
1987 -If trans[states[state].base+charid].check!=state then the
1988 transition is taken to be a 0/Fail transition. Thus if there are fail
1989 transitions at the front of the node then the .base offset will point
1990 somewhere inside the previous nodes data (or maybe even into a node
1991 even earlier), but the .check field determines if the transition is
1992 valid.
1993
786e8c11 1994 XXX - wrong maybe?
a3621e74 1995 The following process inplace converts the table to the compressed
3b753521 1996 table: We first do not compress the root node 1,and mark all its
a3621e74 1997 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
1998 allows us to do a DFA construction from the compressed table later,
1999 and ensures that any .base pointers we calculate later are greater
2000 than 0.
a3621e74
YO
2001
2002 - We set 'pos' to indicate the first entry of the second node.
2003
2004 - We then iterate over the columns of the node, finding the first and
2005 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2006 and set the .check pointers accordingly, and advance pos
2007 appropriately and repreat for the next node. Note that when we copy
2008 the next pointers we have to convert them from the original
2009 NODEIDX form to NODENUM form as the former is not valid post
2010 compression.
2011
2012 - If a node has no transitions used we mark its base as 0 and do not
2013 advance the pos pointer.
2014
2015 - If a node only has one transition we use a second pointer into the
2016 structure to fill in allocated fail transitions from other states.
2017 This pointer is independent of the main pointer and scans forward
2018 looking for null transitions that are allocated to a state. When it
2019 finds one it writes the single transition into the "hole". If the
786e8c11 2020 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2021
2022 - Once compressed we can Renew/realloc the structures to release the
2023 excess space.
2024
2025 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2026 specifically Fig 3.47 and the associated pseudocode.
2027
2028 demq
2029 */
a3b680e6 2030 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2031 U32 state, charid;
a3621e74 2032 U32 pos = 0, zp=0;
1e2e3d02 2033 trie->statecount = laststate;
a3621e74
YO
2034
2035 for ( state = 1 ; state < laststate ; state++ ) {
2036 U8 flag = 0;
a28509cc
AL
2037 const U32 stateidx = TRIE_NODEIDX( state );
2038 const U32 o_used = trie->trans[ stateidx ].check;
2039 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2040 trie->trans[ stateidx ].check = 0;
2041
2042 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2043 if ( flag || trie->trans[ stateidx + charid ].next ) {
2044 if ( trie->trans[ stateidx + charid ].next ) {
2045 if (o_used == 1) {
2046 for ( ; zp < pos ; zp++ ) {
2047 if ( ! trie->trans[ zp ].next ) {
2048 break;
2049 }
2050 }
2051 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2052 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2053 trie->trans[ zp ].check = state;
2054 if ( ++zp > pos ) pos = zp;
2055 break;
2056 }
2057 used--;
2058 }
2059 if ( !flag ) {
2060 flag = 1;
2061 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2062 }
2063 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2064 trie->trans[ pos ].check = state;
2065 pos++;
2066 }
2067 }
2068 }
cc601c31 2069 trie->lasttrans = pos + 1;
c944940b
JH
2070 trie->states = (reg_trie_state *)
2071 PerlMemShared_realloc( trie->states, laststate
2072 * sizeof(reg_trie_state) );
a3621e74 2073 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2074 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2075 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2076 (int)depth * 2 + 2,"",
2077 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2078 (IV)next_alloc,
2079 (IV)pos,
a3621e74
YO
2080 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2081 );
2082
2083 } /* end table compress */
2084 }
1e2e3d02
YO
2085 DEBUG_TRIE_COMPILE_MORE_r(
2086 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2087 (int)depth * 2 + 2, "",
2088 (UV)trie->statecount,
2089 (UV)trie->lasttrans)
2090 );
cc601c31 2091 /* resize the trans array to remove unused space */
c944940b
JH
2092 trie->trans = (reg_trie_trans *)
2093 PerlMemShared_realloc( trie->trans, trie->lasttrans
2094 * sizeof(reg_trie_trans) );
a3621e74 2095
3b753521 2096 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2097 U8 nodetype =(U8)(flags & 0xFF);
2098 char *str=NULL;
786e8c11 2099
07be1b83 2100#ifdef DEBUGGING
e62cc96a 2101 regnode *optimize = NULL;
7122b237
YO
2102#ifdef RE_TRACK_PATTERN_OFFSETS
2103
b57a0404
JH
2104 U32 mjd_offset = 0;
2105 U32 mjd_nodelen = 0;
7122b237
YO
2106#endif /* RE_TRACK_PATTERN_OFFSETS */
2107#endif /* DEBUGGING */
a3621e74 2108 /*
3dab1dad
YO
2109 This means we convert either the first branch or the first Exact,
2110 depending on whether the thing following (in 'last') is a branch
2111 or not and whther first is the startbranch (ie is it a sub part of
2112 the alternation or is it the whole thing.)
3b753521 2113 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2114 the whole branch sequence, including the first.
a3621e74 2115 */
3dab1dad 2116 /* Find the node we are going to overwrite */
7f69552c 2117 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2118 /* branch sub-chain */
3dab1dad 2119 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2120#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2121 DEBUG_r({
2122 mjd_offset= Node_Offset((convert));
2123 mjd_nodelen= Node_Length((convert));
2124 });
7122b237 2125#endif
7f69552c 2126 /* whole branch chain */
7122b237
YO
2127 }
2128#ifdef RE_TRACK_PATTERN_OFFSETS
2129 else {
7f69552c
YO
2130 DEBUG_r({
2131 const regnode *nop = NEXTOPER( convert );
2132 mjd_offset= Node_Offset((nop));
2133 mjd_nodelen= Node_Length((nop));
2134 });
07be1b83
YO
2135 }
2136 DEBUG_OPTIMISE_r(
2137 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2138 (int)depth * 2 + 2, "",
786e8c11 2139 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2140 );
7122b237 2141#endif
3dab1dad
YO
2142 /* But first we check to see if there is a common prefix we can
2143 split out as an EXACT and put in front of the TRIE node. */
2144 trie->startstate= 1;
55eed653 2145 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2146 U32 state;
1e2e3d02 2147 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2148 U32 ofs = 0;
8e11feef
RGS
2149 I32 idx = -1;
2150 U32 count = 0;
2151 const U32 base = trie->states[ state ].trans.base;
a3621e74 2152
3dab1dad 2153 if ( trie->states[state].wordnum )
8e11feef 2154 count = 1;
a3621e74 2155
8e11feef 2156 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2157 if ( ( base + ofs >= trie->uniquecharcount ) &&
2158 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2159 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2160 {
3dab1dad 2161 if ( ++count > 1 ) {
2b8b4781 2162 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2163 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2164 if ( state == 1 ) break;
3dab1dad
YO
2165 if ( count == 2 ) {
2166 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2167 DEBUG_OPTIMISE_r(
8e11feef
RGS
2168 PerlIO_printf(Perl_debug_log,
2169 "%*sNew Start State=%"UVuf" Class: [",
2170 (int)depth * 2 + 2, "",
786e8c11 2171 (UV)state));
be8e71aa 2172 if (idx >= 0) {
2b8b4781 2173 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2174 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2175
3dab1dad 2176 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2177 if ( folder )
2178 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2179 DEBUG_OPTIMISE_r(
f1f66076 2180 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2181 );
8e11feef
RGS
2182 }
2183 }
2184 TRIE_BITMAP_SET(trie,*ch);
2185 if ( folder )
2186 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2187 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2188 }
2189 idx = ofs;
2190 }
3dab1dad
YO
2191 }
2192 if ( count == 1 ) {
2b8b4781 2193 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2194 STRLEN len;
2195 char *ch = SvPV( *tmp, len );
de734bd5
A
2196 DEBUG_OPTIMISE_r({
2197 SV *sv=sv_newmortal();
8e11feef
RGS
2198 PerlIO_printf( Perl_debug_log,
2199 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2200 (int)depth * 2 + 2, "",
de734bd5
A
2201 (UV)state, (UV)idx,
2202 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2203 PL_colors[0], PL_colors[1],
2204 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2205 PERL_PV_ESCAPE_FIRSTCHAR
2206 )
2207 );
2208 });
3dab1dad
YO
2209 if ( state==1 ) {
2210 OP( convert ) = nodetype;
2211 str=STRING(convert);
2212 STR_LEN(convert)=0;
2213 }
c490c714
YO
2214 STR_LEN(convert) += len;
2215 while (len--)
de734bd5 2216 *str++ = *ch++;
8e11feef 2217 } else {
f9049ba1 2218#ifdef DEBUGGING
8e11feef
RGS
2219 if (state>1)
2220 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2221#endif
8e11feef
RGS
2222 break;
2223 }
2224 }
2e64971a 2225 trie->prefixlen = (state-1);
3dab1dad 2226 if (str) {
8e11feef 2227 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2228 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2229 trie->startstate = state;
07be1b83
YO
2230 trie->minlen -= (state - 1);
2231 trie->maxlen -= (state - 1);
33809eae
JH
2232#ifdef DEBUGGING
2233 /* At least the UNICOS C compiler choked on this
2234 * being argument to DEBUG_r(), so let's just have
2235 * it right here. */
2236 if (
2237#ifdef PERL_EXT_RE_BUILD
2238 1
2239#else
2240 DEBUG_r_TEST
2241#endif
2242 ) {
2243 regnode *fix = convert;
2244 U32 word = trie->wordcount;
2245 mjd_nodelen++;
2246 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2247 while( ++fix < n ) {
2248 Set_Node_Offset_Length(fix, 0, 0);
2249 }
2250 while (word--) {
2251 SV ** const tmp = av_fetch( trie_words, word, 0 );
2252 if (tmp) {
2253 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2254 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2255 else
2256 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2257 }
2258 }
2259 }
2260#endif
8e11feef
RGS
2261 if (trie->maxlen) {
2262 convert = n;
2263 } else {
3dab1dad 2264 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2265 DEBUG_r(optimize= n);
3dab1dad
YO
2266 }
2267 }
2268 }
a5ca303d
YO
2269 if (!jumper)
2270 jumper = last;
3dab1dad 2271 if ( trie->maxlen ) {
8e11feef
RGS
2272 NEXT_OFF( convert ) = (U16)(tail - convert);
2273 ARG_SET( convert, data_slot );
786e8c11
YO
2274 /* Store the offset to the first unabsorbed branch in
2275 jump[0], which is otherwise unused by the jump logic.
2276 We use this when dumping a trie and during optimisation. */
2277 if (trie->jump)
7f69552c 2278 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2279
6c48061a
YO
2280 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2281 * and there is a bitmap
2282 * and the first "jump target" node we found leaves enough room
2283 * then convert the TRIE node into a TRIEC node, with the bitmap
2284 * embedded inline in the opcode - this is hypothetically faster.
2285 */
2286 if ( !trie->states[trie->startstate].wordnum
2287 && trie->bitmap
2288 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2289 {
2290 OP( convert ) = TRIEC;
2291 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2292 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2293 trie->bitmap= NULL;
2294 } else
2295 OP( convert ) = TRIE;
a3621e74 2296
3dab1dad
YO
2297 /* store the type in the flags */
2298 convert->flags = nodetype;
a5ca303d
YO
2299 DEBUG_r({
2300 optimize = convert
2301 + NODE_STEP_REGNODE
2302 + regarglen[ OP( convert ) ];
2303 });
2304 /* XXX We really should free up the resource in trie now,
2305 as we won't use them - (which resources?) dmq */
3dab1dad 2306 }
a3621e74 2307 /* needed for dumping*/
e62cc96a 2308 DEBUG_r(if (optimize) {
07be1b83 2309 regnode *opt = convert;
bcdf7404 2310
e62cc96a 2311 while ( ++opt < optimize) {
07be1b83
YO
2312 Set_Node_Offset_Length(opt,0,0);
2313 }
786e8c11
YO
2314 /*
2315 Try to clean up some of the debris left after the
2316 optimisation.
a3621e74 2317 */
786e8c11 2318 while( optimize < jumper ) {
07be1b83 2319 mjd_nodelen += Node_Length((optimize));
a3621e74 2320 OP( optimize ) = OPTIMIZED;
07be1b83 2321 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2322 optimize++;
2323 }
07be1b83 2324 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2325 });
2326 } /* end node insert */
2e64971a
DM
2327
2328 /* Finish populating the prev field of the wordinfo array. Walk back
2329 * from each accept state until we find another accept state, and if
2330 * so, point the first word's .prev field at the second word. If the
2331 * second already has a .prev field set, stop now. This will be the
2332 * case either if we've already processed that word's accept state,
3b753521
FN
2333 * or that state had multiple words, and the overspill words were
2334 * already linked up earlier.
2e64971a
DM
2335 */
2336 {
2337 U16 word;
2338 U32 state;
2339 U16 prev;
2340
2341 for (word=1; word <= trie->wordcount; word++) {
2342 prev = 0;
2343 if (trie->wordinfo[word].prev)
2344 continue;
2345 state = trie->wordinfo[word].accept;
2346 while (state) {
2347 state = prev_states[state];
2348 if (!state)
2349 break;
2350 prev = trie->states[state].wordnum;
2351 if (prev)
2352 break;
2353 }
2354 trie->wordinfo[word].prev = prev;
2355 }
2356 Safefree(prev_states);
2357 }
2358
2359
2360 /* and now dump out the compressed format */
2361 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2362
55eed653 2363 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2364#ifdef DEBUGGING
2365 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2366 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2367#else
2368 SvREFCNT_dec(revcharmap);
07be1b83 2369#endif
786e8c11
YO
2370 return trie->jump
2371 ? MADE_JUMP_TRIE
2372 : trie->startstate>1
2373 ? MADE_EXACT_TRIE
2374 : MADE_TRIE;
2375}
2376
2377STATIC void
2378S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2379{
3b753521 2380/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2381
2382 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2383 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2384 ISBN 0-201-10088-6
2385
2386 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2387 suffix of the current state's 'word' that is also a proper prefix of another word in our
2388 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2389 the DFA not to have to restart after its tried and failed a word at a given point, it
2390 simply continues as though it had been matching the other word in the first place.
2391 Consider
2392 'abcdgu'=~/abcdefg|cdgu/
2393 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2394 fail, which would bring us to the state representing 'd' in the second word where we would
2395 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2396 */
2397 /* add a fail transition */
3251b653
NC
2398 const U32 trie_offset = ARG(source);
2399 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2400 U32 *q;
2401 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2402 const U32 numstates = trie->statecount;
786e8c11
YO
2403 const U32 ubound = trie->lasttrans + ucharcount;
2404 U32 q_read = 0;
2405 U32 q_write = 0;
2406 U32 charid;
2407 U32 base = trie->states[ 1 ].trans.base;
2408 U32 *fail;
2409 reg_ac_data *aho;
2410 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2411 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2412
2413 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2414#ifndef DEBUGGING
2415 PERL_UNUSED_ARG(depth);
2416#endif
2417
2418
2419 ARG_SET( stclass, data_slot );
c944940b 2420 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2421 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2422 aho->trie=trie_offset;
446bd890
NC
2423 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2424 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2425 Newxz( q, numstates, U32);
c944940b 2426 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2427 aho->refcount = 1;
2428 fail = aho->fail;
2429 /* initialize fail[0..1] to be 1 so that we always have
2430 a valid final fail state */
2431 fail[ 0 ] = fail[ 1 ] = 1;
2432
2433 for ( charid = 0; charid < ucharcount ; charid++ ) {
2434 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2435 if ( newstate ) {
2436 q[ q_write ] = newstate;
2437 /* set to point at the root */
2438 fail[ q[ q_write++ ] ]=1;
2439 }
2440 }
2441 while ( q_read < q_write) {
2442 const U32 cur = q[ q_read++ % numstates ];
2443 base = trie->states[ cur ].trans.base;
2444
2445 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2446 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2447 if (ch_state) {
2448 U32 fail_state = cur;
2449 U32 fail_base;
2450 do {
2451 fail_state = fail[ fail_state ];
2452 fail_base = aho->states[ fail_state ].trans.base;
2453 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2454
2455 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2456 fail[ ch_state ] = fail_state;
2457 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2458 {
2459 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2460 }
2461 q[ q_write++ % numstates] = ch_state;
2462 }
2463 }
2464 }
2465 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2466 when we fail in state 1, this allows us to use the
2467 charclass scan to find a valid start char. This is based on the principle
2468 that theres a good chance the string being searched contains lots of stuff
2469 that cant be a start char.
2470 */
2471 fail[ 0 ] = fail[ 1 ] = 0;
2472 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2473 PerlIO_printf(Perl_debug_log,
2474 "%*sStclass Failtable (%"UVuf" states): 0",
2475 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2476 );
786e8c11
YO
2477 for( q_read=1; q_read<numstates; q_read++ ) {
2478 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2479 }
2480 PerlIO_printf(Perl_debug_log, "\n");
2481 });
2482 Safefree(q);
2483 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2484}
2485
786e8c11 2486
a3621e74 2487/*
5d1c421c
JH
2488 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2489 * These need to be revisited when a newer toolchain becomes available.
2490 */
2491#if defined(__sparc64__) && defined(__GNUC__)
2492# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2493# undef SPARC64_GCC_WORKAROUND
2494# define SPARC64_GCC_WORKAROUND 1
2495# endif
2496#endif
2497
07be1b83 2498#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2499 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2500 SV * const mysv=sv_newmortal(); \
2501 regnode *Next = regnext(scan); \
2502 regprop(RExC_rx, mysv, scan); \
7f69552c 2503 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2504 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2505 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2506 }});
07be1b83 2507
1de06328 2508
bb914485
KW
2509/* The below joins as many adjacent EXACTish nodes as possible into a single
2510 * one, and looks for problematic sequences of characters whose folds vs.
2511 * non-folds have sufficiently different lengths, that the optimizer would be
2512 * fooled into rejecting legitimate matches of them, and the trie construction
2513 * code can't cope with them. The joining is only done if:
2514 * 1) there is room in the current conglomerated node to entirely contain the
2515 * next one.
2516 * 2) they are the exact same node type
2517 *
2518 * The adjacent nodes actually may be separated by NOTHING kind nodes, and
2519 * these get optimized out
2520 *
9d071ca8
KW
2521 * If there are problematic code sequences, *min_subtract is set to the delta
2522 * that the minimum size of the node can be less than its actual size. And,
2523 * the node type of the result is changed to reflect that it contains these
bb914485
KW
2524 * sequences.
2525 *
a0c4c608
KW
2526 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2527 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2528 *
bb914485
KW
2529 * This is as good a place as any to discuss the design of handling these
2530 * problematic sequences. It's been wrong in Perl for a very long time. There
2531 * are three code points in Unicode whose folded lengths differ so much from
2532 * the un-folded lengths that it causes problems for the optimizer and trie
2533 * construction. Why only these are problematic, and not others where lengths
2534 * also differ is something I (khw) do not understand. New versions of Unicode
2535 * might add more such code points. Hopefully the logic in fold_grind.t that
287722f3 2536 * figures out what to test (in part by verifying that each size-combination
bb914485 2537 * gets tested) will catch any that do come along, so they can be added to the
287722f3
KW
2538 * special handling below. The chances of new ones are actually rather small,
2539 * as most, if not all, of the world's scripts that have casefolding have
2540 * already been encoded by Unicode. Also, a number of Unicode's decisions were
2541 * made to allow compatibility with pre-existing standards, and almost all of
2542 * those have already been dealt with. These would otherwise be the most
2543 * likely candidates for generating further tricky sequences. In other words,
2544 * Unicode by itself is unlikely to add new ones unless it is for compatibility
a0c4c608 2545 * with pre-existing standards, and there aren't many of those left.
bb914485
KW
2546 *
2547 * The previous designs for dealing with these involved assigning a special
2548 * node for them. This approach doesn't work, as evidenced by this example:
a0c4c608 2549 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
bb914485
KW
2550 * Both these fold to "sss", but if the pattern is parsed to create a node of
2551 * that would match just the \xDF, it won't be able to handle the case where a
2552 * successful match would have to cross the node's boundary. The new approach
2553 * that hopefully generally solves the problem generates an EXACTFU_SS node
2554 * that is "sss".
2555 *
2556 * There are a number of components to the approach (a lot of work for just
2557 * three code points!):
2558 * 1) This routine examines each EXACTFish node that could contain the
9d071ca8
KW
2559 * problematic sequences. It returns in *min_subtract how much to
2560 * subtract from the the actual length of the string to get a real minimum
2561 * for one that could match it. This number is usually 0 except for the
2562 * problematic sequences. This delta is used by the caller to adjust the
2563 * min length of the match, and the delta between min and max, so that the
2564 * optimizer doesn't reject these possibilities based on size constraints.
bb914485
KW
2565 * 2) These sequences are not currently correctly handled by the trie code
2566 * either, so it changes the joined node type to ops that are not handled
2567 * by trie's, those new ops being EXACTFU_SS and EXACTFU_NO_TRIE.
2568 * 3) This is sufficient for the two Greek sequences (described below), but
2569 * the one involving the Sharp s (\xDF) needs more. The node type
2570 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2571 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2572 * case where there is a possible fold length change. That means that a
2573 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2574 * itself with length changes, and so can be processed faster. regexec.c
2575 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2576 * is pre-folded by regcomp.c. This saves effort in regex matching.
2577 * However, probably mostly for historical reasons, the pre-folding isn't
a0c4c608
KW
2578 * done for non-UTF8 patterns (and it can't be for EXACTF and EXACTFL
2579 * nodes, as what they fold to isn't known until runtime.) The fold
2580 * possibilities for the non-UTF8 patterns are quite simple, except for
2581 * the sharp s. All the ones that don't involve a UTF-8 target string
2582 * are members of a fold-pair, and arrays are set up for all of them
2583 * that quickly find the other member of the pair. It might actually
2584 * be faster to pre-fold these, but it isn't currently done, except for
2585 * the sharp s. Code elsewhere in this file makes sure that it gets
2586 * folded to 'ss', even if the pattern isn't UTF-8. This avoids the
2587 * issues described in the next item.
bb914485
KW
2588 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2589 * 'ss' or not is not knowable at compile time. It will match iff the
2590 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2591 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2592 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2593 * described in item 3). An assumption that the optimizer part of
2594 * regexec.c (probably unwittingly) makes is that a character in the
2595 * pattern corresponds to at most a single character in the target string.
2596 * (And I do mean character, and not byte here, unlike other parts of the
2597 * documentation that have never been updated to account for multibyte
2598 * Unicode.) This assumption is wrong only in this case, as all other
2599 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2600 * virtue of having this file pre-fold UTF-8 patterns. I'm
2601 * reluctant to try to change this assumption, so instead the code punts.
9d071ca8
KW
2602 * This routine examines EXACTF nodes for the sharp s, and returns a
2603 * boolean indicating whether or not the node is an EXACTF node that
2604 * contains a sharp s. When it is true, the caller sets a flag that later
2605 * causes the optimizer in this file to not set values for the floating
2606 * and fixed string lengths, and thus avoids the optimizer code in
2607 * regexec.c that makes the invalid assumption. Thus, there is no
2608 * optimization based on string lengths for EXACTF nodes that contain the
2609 * sharp s. This only happens for /id rules (which means the pattern
2610 * isn't in UTF-8).
bb914485 2611 */
1de06328 2612
9d071ca8 2613#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2614 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2615 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2616
be8e71aa 2617STATIC U32
9d071ca8 2618S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, UV *min_subtract, bool *has_exactf_sharp_s, U32 flags,regnode *val, U32 depth) {
07be1b83
YO
2619 /* Merge several consecutive EXACTish nodes into one. */
2620 regnode *n = regnext(scan);
2621 U32 stringok = 1;
2622 regnode *next = scan + NODE_SZ_STR(scan);
2623 U32 merged = 0;
2624 U32 stopnow = 0;
2625#ifdef DEBUGGING
2626 regnode *stop = scan;
72f13be8 2627 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2628#else
d47053eb
RGS
2629 PERL_UNUSED_ARG(depth);
2630#endif
7918f24d
NC
2631
2632 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2633#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2634 PERL_UNUSED_ARG(flags);
2635 PERL_UNUSED_ARG(val);
07be1b83 2636#endif
07be1b83 2637 DEBUG_PEEP("join",scan,depth);
bb914485 2638
3f410cf6
KW
2639 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2640 * EXACT ones that are mergeable to the current one. */
2641 while (n
2642 && (PL_regkind[OP(n)] == NOTHING
2643 || (stringok && OP(n) == OP(scan)))
07be1b83 2644 && NEXT_OFF(n)
3f410cf6
KW
2645 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2646 {
07be1b83
YO
2647
2648 if (OP(n) == TAIL || n > next)
2649 stringok = 0;
2650 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2651 DEBUG_PEEP("skip:",n,depth);
2652 NEXT_OFF(scan) += NEXT_OFF(n);
2653 next = n + NODE_STEP_REGNODE;
2654#ifdef DEBUGGING
2655 if (stringok)
2656 stop = n;
2657#endif
2658 n = regnext(n);
2659 }
2660 else if (stringok) {
786e8c11 2661 const unsigned int oldl = STR_LEN(scan);
07be1b83 2662 regnode * const nnext = regnext(n);
b2230d39
KW
2663
2664 if (oldl + STR_LEN(n) > U8_MAX)
2665 break;
07be1b83
YO
2666
2667 DEBUG_PEEP("merg",n,depth);
07be1b83 2668 merged++;
b2230d39 2669
07be1b83
YO
2670 NEXT_OFF(scan) += NEXT_OFF(n);
2671 STR_LEN(scan) += STR_LEN(n);
2672 next = n + NODE_SZ_STR(n);
2673 /* Now we can overwrite *n : */
2674 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2675#ifdef DEBUGGING
2676 stop = next - 1;
2677#endif
2678 n = nnext;
2679 if (stopnow) break;
2680 }
2681
d47053eb
RGS
2682#ifdef EXPERIMENTAL_INPLACESCAN
2683 if (flags && !NEXT_OFF(n)) {
2684 DEBUG_PEEP("atch", val, depth);
2685 if (reg_off_by_arg[OP(n)]) {
2686 ARG_SET(n, val - n);
2687 }
2688 else {
2689 NEXT_OFF(n) = val - n;
2690 }
2691 stopnow = 1;
2692 }
07be1b83
YO
2693#endif
2694 }
2c2b7f86 2695
9d071ca8 2696 *min_subtract = 0;
f758bddf 2697 *has_exactf_sharp_s = FALSE;
f646642f 2698
3f410cf6
KW
2699 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2700 * can now analyze for sequences of problematic code points. (Prior to
2701 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2702 * hence missed). The sequences only happen in folding, hence for any
2703 * non-EXACT EXACTish node */
86d6fcad 2704 if (OP(scan) != EXACT) {
f758bddf
KW
2705 U8 *s;
2706 U8 * s0 = (U8*) STRING(scan);
2707 U8 * const s_end = s0 + STR_LEN(scan);
2708
2709 /* The below is perhaps overboard, but this allows us to save a test
2710 * each time through the loop at the expense of a mask. This is
2711 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2712 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2713 * This uses an exclusive 'or' to find that bit and then inverts it to
2714 * form a mask, with just a single 0, in the bit position where 'S' and
2715 * 's' differ. */
2716 const U8 S_or_s_mask = ~ ('S' ^ 's');
2717 const U8 s_masked = 's' & S_or_s_mask;
2718
2719 /* One pass is made over the node's string looking for all the
2720 * possibilities. to avoid some tests in the loop, there are two main
2721 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2722 * non-UTF-8 */
2723 if (UTF) {
86d6fcad 2724
f758bddf
KW
2725 /* There are two problematic Greek code points in Unicode
2726 * casefolding
86d6fcad
KW
2727 *
2728 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2729 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2730 *
2731 * which casefold to
2732 *
2733 * Unicode UTF-8
2734 *
2735 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2736 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2737 *
2738 * This means that in case-insensitive matching (or "loose
2739 * matching", as Unicode calls it), an EXACTF of length six (the
2740 * UTF-8 encoded byte length of the above casefolded versions) can
2741 * match a target string of length two (the byte length of UTF-8
2742 * encoded U+0390 or U+03B0). This would rather mess up the
2743 * minimum length computation. (there are other code points that
2744 * also fold to these two sequences, but the delta is smaller)
2745 *
f758bddf
KW
2746 * If these sequences are found, the minimum length is decreased by
2747 * four (six minus two).
86d6fcad 2748 *
f758bddf
KW
2749 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2750 * LETTER SHARP S. We decrease the min length by 1 for each
2751 * occurrence of 'ss' found */
3f410cf6 2752
e294cc5d 2753#ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
f758bddf
KW
2754# define U390_first_byte 0xb4
2755 const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
2756# define U3B0_first_byte 0xb5
2757 const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
e294cc5d 2758#else
f758bddf
KW
2759# define U390_first_byte 0xce
2760 const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
2761# define U3B0_first_byte 0xcf
2762 const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
e294cc5d 2763#endif
f758bddf
KW
2764 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2765 yields a net of 0 */
2766 /* Examine the string for one of the problematic sequences */
2767 for (s = s0;
2768 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2769 * sequence we are looking for is 2 */
2770 s += UTF8SKIP(s))
86d6fcad 2771 {
bb914485 2772
f758bddf
KW
2773 /* Look for the first byte in each problematic sequence */
2774 switch (*s) {
2775 /* We don't have to worry about other things that fold to
2776 * 's' (such as the long s, U+017F), as all above-latin1
2777 * code points have been pre-folded */
2778 case 's':
2779 case 'S':
2780
a0c4c608
KW
2781 /* Current character is an 's' or 'S'. If next one is
2782 * as well, we have the dreaded sequence */
f758bddf
KW
2783 if (((*(s+1) & S_or_s_mask) == s_masked)
2784 /* These two node types don't have special handling
2785 * for 'ss' */
2786 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2787 {
9d071ca8 2788 *min_subtract += 1;
f758bddf
KW
2789 OP(scan) = EXACTFU_SS;
2790 s++; /* No need to look at this character again */
2791 }
2792 break;
2793
2794 case U390_first_byte:
2795 if (s_end - s >= len
2796
2797 /* The 1's are because are skipping comparing the
2798 * first byte */
2799 && memEQ(s + 1, U390_tail, len - 1))
2800 {
2801 goto greek_sequence;
2802 }
2803 break;
2804
2805 case U3B0_first_byte:
2806 if (! (s_end - s >= len
2807 && memEQ(s + 1, U3B0_tail, len - 1)))
2808 {
2809 break;
2810 }
2811 greek_sequence:
9d071ca8 2812 *min_subtract += 4;
f758bddf
KW
2813
2814 /* This can't currently be handled by trie's, so change
2815 * the node type to indicate this. If EXACTFA and
2816 * EXACTFL were ever to be handled by trie's, this
2817 * would have to be changed. If this node has already
2818 * been changed to EXACTFU_SS in this loop, leave it as
2819 * is. (I (khw) think it doesn't matter in regexec.c
2820 * for UTF patterns, but no need to change it */
2821 if (OP(scan) == EXACTFU) {
2822 OP(scan) = EXACTFU_NO_TRIE;
2823 }
2824 s += 6; /* We already know what this sequence is. Skip
2825 the rest of it */
2826 break;
bb914485
KW
2827 }
2828 }
2829 }
f758bddf 2830 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
bb914485 2831
f758bddf
KW
2832 /* Here, the pattern is not UTF-8. We need to look only for the
2833 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2834 * in the final position. Otherwise we can stop looking 1 byte
2835 * earlier because have to find both the first and second 's' */
2836 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2837
2838 for (s = s0; s < upper; s++) {
2839 switch (*s) {
2840 case 'S':
2841 case 's':
2842 if (s_end - s > 1
2843 && ((*(s+1) & S_or_s_mask) == s_masked))
2844 {
9d071ca8 2845 *min_subtract += 1;
f758bddf
KW
2846
2847 /* EXACTF nodes need to know that the minimum
2848 * length changed so that a sharp s in the string
2849 * can match this ss in the pattern, but they
2850 * remain EXACTF nodes, as they are not trie'able,
2851 * so don't have to invent a new node type to
2852 * exclude them from the trie code */
2853 if (OP(scan) != EXACTF) {
2854 OP(scan) = EXACTFU_SS;
2855 }
2856 s++;
2857 }
2858 break;
2859 case LATIN_SMALL_LETTER_SHARP_S:
2860 if (OP(scan) == EXACTF) {
2861 *has_exactf_sharp_s = TRUE;
2862 }
2863 break;
86d6fcad
KW
2864 }
2865 }
2866 }
07be1b83 2867 }
3f410cf6 2868
07be1b83 2869#ifdef DEBUGGING
bb789b09
DM
2870 /* Allow dumping but overwriting the collection of skipped
2871 * ops and/or strings with fake optimized ops */
07be1b83
YO
2872 n = scan + NODE_SZ_STR(scan);
2873 while (n <= stop) {
bb789b09
DM
2874 OP(n) = OPTIMIZED;
2875 FLAGS(n) = 0;
2876 NEXT_OFF(n) = 0;
07be1b83
YO
2877 n++;
2878 }
2879#endif
2880 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2881 return stopnow;
2882}
2883
486ec47a 2884/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2885 Finds fixed substrings. */
2886
a0288114 2887/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2888 to the position after last scanned or to NULL. */
2889
40d049e4
YO
2890#define INIT_AND_WITHP \
2891 assert(!and_withp); \
2892 Newx(and_withp,1,struct regnode_charclass_class); \
2893 SAVEFREEPV(and_withp)
07be1b83 2894
b515a41d 2895/* this is a chain of data about sub patterns we are processing that
486ec47a 2896 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2897 we can simulate recursion without losing state. */
2898struct scan_frame;
2899typedef struct scan_frame {
2900 regnode *last; /* last node to process in this frame */
2901 regnode *next; /* next node to process when last is reached */
2902 struct scan_frame *prev; /*previous frame*/
2903 I32 stop; /* what stopparen do we use */
2904} scan_frame;
2905
304ee84b
YO
2906
2907#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2908
e1d1eefb
YO
2909#define CASE_SYNST_FNC(nAmE) \
2910case nAmE: \
2911 if (flags & SCF_DO_STCLASS_AND) { \
2912 for (value = 0; value < 256; value++) \
2913 if (!is_ ## nAmE ## _cp(value)) \
2914 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2915 } \
2916 else { \
2917 for (value = 0; value < 256; value++) \
2918 if (is_ ## nAmE ## _cp(value)) \
2919 ANYOF_BITMAP_SET(data->start_class, value); \
2920 } \
2921 break; \
2922case N ## nAmE: \
2923 if (flags & SCF_DO_STCLASS_AND) { \
2924 for (value = 0; value < 256; value++) \
2925 if (is_ ## nAmE ## _cp(value)) \
2926 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2927 } \
2928 else { \
2929 for (value = 0; value < 256; value++) \
2930 if (!is_ ## nAmE ## _cp(value)) \
2931 ANYOF_BITMAP_SET(data->start_class, value); \
2932 } \
2933 break
2934
2935
2936
76e3520e 2937STATIC I32
40d049e4 2938S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 2939 I32 *minlenp, I32 *deltap,
40d049e4
YO
2940 regnode *last,
2941 scan_data_t *data,
2942 I32 stopparen,
2943 U8* recursed,
2944 struct regnode_charclass_class *and_withp,
2945 U32 flags, U32 depth)
c277df42
IZ
2946 /* scanp: Start here (read-write). */
2947 /* deltap: Write maxlen-minlen here. */
2948 /* last: Stop before this one. */
40d049e4
YO
2949 /* data: string data about the pattern */
2950 /* stopparen: treat close N as END */
2951 /* recursed: which subroutines have we recursed into */
2952 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 2953{
97aff369 2954 dVAR;
c277df42
IZ
2955 I32 min = 0, pars = 0, code;
2956 regnode *scan = *scanp, *next;
2957 I32 delta = 0;
2958 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 2959 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
2960 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2961 scan_data_t data_fake;
a3621e74 2962 SV *re_trie_maxbuff = NULL;
786e8c11 2963 regnode *first_non_open = scan;
e2e6a0f1 2964 I32 stopmin = I32_MAX;
8aa23a47 2965 scan_frame *frame = NULL;
a3621e74 2966 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 2967
7918f24d
NC
2968 PERL_ARGS_ASSERT_STUDY_CHUNK;
2969
13a24bad 2970#ifdef DEBUGGING
40d049e4 2971 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 2972#endif
40d049e4 2973
786e8c11 2974 if ( depth == 0 ) {
40d049e4 2975 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
2976 first_non_open=regnext(first_non_open);
2977 }
2978
b81d288d 2979
8aa23a47
YO
2980 fake_study_recurse:
2981 while ( scan && OP(scan) != END && scan < last ){
9d071ca8
KW
2982 UV min_subtract = 0; /* How much to subtract from the minimum node
2983 length to get a real minimum (because the
2984 folded version may be shorter) */
f758bddf 2985 bool has_exactf_sharp_s = FALSE;
8aa23a47 2986 /* Peephole optimizer: */
304ee84b 2987 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 2988 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
2989
2990 /* Its not clear to khw or hv why this is done here, and not in the
2991 * clauses that deal with EXACT nodes. khw's guess is that it's
2992 * because of a previous design */
9d071ca8 2993 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
2994
2995 /* Follow the next-chain of the current node and optimize
2996 away all the NOTHINGs from it. */
2997 if (OP(scan) != CURLYX) {
2998 const int max = (reg_off_by_arg[OP(scan)]
2999 ? I32_MAX
3000 /* I32 may be smaller than U16 on CRAYs! */
3001 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3002 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3003 int noff;
3004 regnode *n = scan;
686b73d4 3005
8aa23a47
YO
3006 /* Skip NOTHING and LONGJMP. */
3007 while ((n = regnext(n))
3008 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3009 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3010 && off + noff < max)
3011 off += noff;
3012 if (reg_off_by_arg[OP(scan)])
3013 ARG(scan) = off;
3014 else
3015 NEXT_OFF(scan) = off;
3016 }
a3621e74 3017
c277df42 3018
8aa23a47
YO
3019
3020 /* The principal pseudo-switch. Cannot be a switch, since we
3021 look into several different things. */
3022 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3023 || OP(scan) == IFTHEN) {
3024 next = regnext(scan);
3025 code = OP(scan);
3026 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3027
8aa23a47
YO
3028 if (OP(next) == code || code == IFTHEN) {
3029 /* NOTE - There is similar code to this block below for handling
3030 TRIE nodes on a re-study. If you change stuff here check there
3031 too. */
3032 I32 max1 = 0, min1 = I32_MAX, num = 0;
3033 struct regnode_charclass_class accum;
3034 regnode * const startbranch=scan;
686b73d4 3035
8aa23a47 3036 if (flags & SCF_DO_SUBSTR)
304ee84b 3037 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3038 if (flags & SCF_DO_STCLASS)
e755fd73 3039 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3040
3041 while (OP(scan) == code) {
3042 I32 deltanext, minnext, f = 0, fake;
3043 struct regnode_charclass_class this_class;
3044
3045 num++;
3046 data_fake.flags = 0;
3047 if (data) {
3048 data_fake.whilem_c = data->whilem_c;
3049 data_fake.last_closep = data->last_closep;
3050 }
3051 else
3052 data_fake.last_closep = &fake;
58e23c8d
YO
3053
3054 data_fake.pos_delta = delta;
8aa23a47
YO
3055 next = regnext(scan);
3056 scan = NEXTOPER(scan);
3057 if (code != BRANCH)
c277df42 3058 scan = NEXTOPER(scan);
8aa23a47 3059 if (flags & SCF_DO_STCLASS) {
e755fd73 3060 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3061 data_fake.start_class = &this_class;
3062 f = SCF_DO_STCLASS_AND;
58e23c8d 3063 }
8aa23a47
YO
3064 if (flags & SCF_WHILEM_VISITED_POS)
3065 f |= SCF_WHILEM_VISITED_POS;
3066
3067 /* we suppose the run is continuous, last=next...*/
3068 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3069 next, &data_fake,
3070 stopparen, recursed, NULL, f,depth+1);
3071 if (min1 > minnext)
3072 min1 = minnext;
3073 if (max1 < minnext + deltanext)
3074 max1 = minnext + deltanext;
3075 if (deltanext == I32_MAX)
3076 is_inf = is_inf_internal = 1;
3077 scan = next;
3078 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3079 pars++;
3080 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3081 if ( stopmin > minnext)
3082 stopmin = min + min1;
3083 flags &= ~SCF_DO_SUBSTR;
3084 if (data)
3085 data->flags |= SCF_SEEN_ACCEPT;
3086 }
3087 if (data) {
3088 if (data_fake.flags & SF_HAS_EVAL)
3089 data->flags |= SF_HAS_EVAL;
3090 data->whilem_c = data_fake.whilem_c;
3dab1dad 3091 }
8aa23a47 3092 if (flags & SCF_DO_STCLASS)
3fffb88a 3093 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3094 }
3095 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3096 min1 = 0;
3097 if (flags & SCF_DO_SUBSTR) {
3098 data->pos_min += min1;
3099 data->pos_delta += max1 - min1;
3100 if (max1 != min1 || is_inf)
3101 data->longest = &(data->longest_float);
3102 }
3103 min += min1;
3104 delta += max1 - min1;
3105 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3106 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3107 if (min1) {
3108 cl_and(data->start_class, and_withp);
3109 flags &= ~SCF_DO_STCLASS;
653099ff 3110 }
8aa23a47
YO
3111 }
3112 else if (flags & SCF_DO_STCLASS_AND) {
3113 if (min1) {
3114 cl_and(data->start_class, &accum);
3115 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3116 }
8aa23a47
YO
3117 else {
3118 /* Switch to OR mode: cache the old value of
3119 * data->start_class */
3120 INIT_AND_WITHP;
3121 StructCopy(data->start_class, and_withp,
3122 struct regnode_charclass_class);
3123 flags &= ~SCF_DO_STCLASS_AND;
3124 StructCopy(&accum, data->start_class,
3125 struct regnode_charclass_class);
3126 flags |= SCF_DO_STCLASS_OR;
3127 data->start_class->flags |= ANYOF_EOS;
de0c8cb8 3128 }
8aa23a47 3129 }
a3621e74 3130
8aa23a47
YO
3131 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3132 /* demq.
a3621e74 3133
8aa23a47
YO
3134 Assuming this was/is a branch we are dealing with: 'scan' now
3135 points at the item that follows the branch sequence, whatever
3136 it is. We now start at the beginning of the sequence and look
3137 for subsequences of
a3621e74 3138
8aa23a47
YO
3139 BRANCH->EXACT=>x1
3140 BRANCH->EXACT=>x2
3141 tail
a3621e74 3142
8aa23a47 3143 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3144
486ec47a 3145 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3146 element into a trie and then add the subsequent branch exact
3147 strings to the trie.
a3621e74 3148
8aa23a47 3149 We have two cases
a3621e74 3150
3b753521 3151 1. patterns where the whole set of branches can be converted.
a3621e74 3152
8aa23a47 3153 2. patterns where only a subset can be converted.
a3621e74 3154
8aa23a47
YO
3155 In case 1 we can replace the whole set with a single regop
3156 for the trie. In case 2 we need to keep the start and end
3b753521 3157 branches so
a3621e74 3158
8aa23a47
YO
3159 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3160 becomes BRANCH TRIE; BRANCH X;
786e8c11 3161
8aa23a47
YO
3162 There is an additional case, that being where there is a
3163 common prefix, which gets split out into an EXACT like node
3164 preceding the TRIE node.
a3621e74 3165
8aa23a47
YO
3166 If x(1..n)==tail then we can do a simple trie, if not we make
3167 a "jump" trie, such that when we match the appropriate word
486ec47a 3168 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3169 a nested if into a case structure of sorts.
b515a41d 3170
8aa23a47 3171 */
686b73d4 3172
8aa23a47
YO
3173 int made=0;
3174 if (!re_trie_maxbuff) {
3175 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3176 if (!SvIOK(re_trie_maxbuff))
3177 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3178 }
3179 if ( SvIV(re_trie_maxbuff)>=0 ) {
3180 regnode *cur;
3181 regnode *first = (regnode *)NULL;
3182 regnode *last = (regnode *)NULL;
3183 regnode *tail = scan;
3184 U8 optype = 0;
3185 U32 count=0;
a3621e74
YO
3186
3187#ifdef DEBUGGING
8aa23a47 3188 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3189#endif
8aa23a47
YO
3190 /* var tail is used because there may be a TAIL
3191 regop in the way. Ie, the exacts will point to the
3192 thing following the TAIL, but the last branch will
3193 point at the TAIL. So we advance tail. If we
3194 have nested (?:) we may have to move through several
3195 tails.
3196 */
3197
3198 while ( OP( tail ) == TAIL ) {
3199 /* this is the TAIL generated by (?:) */
3200 tail = regnext( tail );
3201 }
a3621e74 3202
8aa23a47
YO
3203
3204 DEBUG_OPTIMISE_r({
3205 regprop(RExC_rx, mysv, tail );
3206 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3207 (int)depth * 2 + 2, "",
3208 "Looking for TRIE'able sequences. Tail node is: ",
3209 SvPV_nolen_const( mysv )
3210 );
3211 });
3212
3213 /*
3214
3215 step through the branches, cur represents each
3216 branch, noper is the first thing to be matched
3217 as part of that branch and noper_next is the
3218 regnext() of that node. if noper is an EXACT
3219 and noper_next is the same as scan (our current
3220 position in the regex) then the EXACT branch is
3221 a possible optimization target. Once we have
486ec47a 3222 two or more consecutive such branches we can
8aa23a47
YO
3223 create a trie of the EXACT's contents and stich
3224 it in place. If the sequence represents all of
3225 the branches we eliminate the whole thing and
3226 replace it with a single TRIE. If it is a
3227 subsequence then we need to stitch it in. This
3228 means the first branch has to remain, and needs
3229 to be repointed at the item on the branch chain
3230 following the last branch optimized. This could
3231 be either a BRANCH, in which case the
3232 subsequence is internal, or it could be the
3233 item following the branch sequence in which
3234 case the subsequence is at the end.
3235
3236 */
3237
3238 /* dont use tail as the end marker for this traverse */
3239 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3240 regnode * const noper = NEXTOPER( cur );
b515a41d 3241#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3242 regnode * const noper_next = regnext( noper );
b515a41d
YO
3243#endif
3244
8aa23a47
YO
3245 DEBUG_OPTIMISE_r({
3246 regprop(RExC_rx, mysv, cur);
3247 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3248 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3249
3250 regprop(RExC_rx, mysv, noper);
3251 PerlIO_printf( Perl_debug_log, " -> %s",
3252 SvPV_nolen_const(mysv));
3253
3254 if ( noper_next ) {
3255 regprop(RExC_rx, mysv, noper_next );
3256 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3257 SvPV_nolen_const(mysv));
3258 }
3259 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
3260 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
3261 });
3262 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
3263 : PL_regkind[ OP( noper ) ] == EXACT )
3264 || OP(noper) == NOTHING )
786e8c11 3265#ifdef NOJUMPTRIE
8aa23a47 3266 && noper_next == tail
786e8c11 3267#endif
8aa23a47
YO
3268 && count < U16_MAX)
3269 {
3270 count++;
3271 if ( !first || optype == NOTHING ) {
3272 if (!first) first = cur;
3273 optype = OP( noper );
3274 } else {
3275 last = cur;
3276 }
3277 } else {
a0a388a1 3278/*
fbebf34e
KW
3279 Currently the trie logic handles case insensitive matching properly only
3280 when the pattern is UTF-8 and the node is EXACTFU (thus forcing unicode
3281 semantics).
0abd0d78
YO
3282
3283 If/when this is fixed the following define can be swapped
3284 in below to fully enable trie logic.
3285
a0a388a1 3286#define TRIE_TYPE_IS_SAFE 1
0abd0d78 3287
a0c4c608
KW
3288Note that join_exact() assumes that the other types of EXACTFish nodes are not
3289used in tries, so that would have to be updated if this changed
3290
0abd0d78 3291*/
fbebf34e 3292#define TRIE_TYPE_IS_SAFE ((UTF && optype == EXACTFU) || optype==EXACT)
0abd0d78 3293
a0a388a1 3294 if ( last && TRIE_TYPE_IS_SAFE ) {
8aa23a47
YO
3295 make_trie( pRExC_state,
3296 startbranch, first, cur, tail, count,
3297 optype, depth+1 );
3298 }
3299 if ( PL_regkind[ OP( noper ) ] == EXACT
786e8c11 3300#ifdef NOJUMPTRIE
8aa23a47 3301 && noper_next == tail
786e8c11 3302#endif
8aa23a47
YO
3303 ){
3304 count = 1;
3305 first = cur;
3306 optype = OP( noper );
3307 } else {
3308 count = 0;
3309 first = NULL;
3310 optype = 0;
3311 }
3312 last = NULL;
3313 }
3314 }
3315 DEBUG_OPTIMISE_r({
3316 regprop(RExC_rx, mysv, cur);
3317 PerlIO_printf( Perl_debug_log,
3318 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3319 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3320
3321 });
a0a388a1
YO
3322
3323 if ( last && TRIE_TYPE_IS_SAFE ) {
8aa23a47 3324 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
686b73d4 3325#ifdef TRIE_STUDY_OPT
8aa23a47
YO
3326 if ( ((made == MADE_EXACT_TRIE &&
3327 startbranch == first)
3328 || ( first_non_open == first )) &&
3329 depth==0 ) {
3330 flags |= SCF_TRIE_RESTUDY;
3331 if ( startbranch == first
3332 && scan == tail )
3333 {
3334 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3335 }
3336 }
3dab1dad 3337#endif
8aa23a47
YO
3338 }
3339 }
3340
3341 } /* do trie */
3342
653099ff 3343 }
8aa23a47
YO
3344 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3345 scan = NEXTOPER(NEXTOPER(scan));
3346 } else /* single branch is optimized. */
3347 scan = NEXTOPER(scan);
3348 continue;
3349 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3350 scan_frame *newframe = NULL;
3351 I32 paren;
3352 regnode *start;
3353 regnode *end;
3354
3355 if (OP(scan) != SUSPEND) {
3356 /* set the pointer */
3357 if (OP(scan) == GOSUB) {
3358 paren = ARG(scan);
3359 RExC_recurse[ARG2L(scan)] = scan;
3360 start = RExC_open_parens[paren-1];
3361 end = RExC_close_parens[paren-1];
3362 } else {
3363 paren = 0;
f8fc2ecf 3364 start = RExC_rxi->program + 1;
8aa23a47
YO
3365 end = RExC_opend;
3366 }
3367 if (!recursed) {
3368 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3369 SAVEFREEPV(recursed);
3370 }
3371 if (!PAREN_TEST(recursed,paren+1)) {
3372 PAREN_SET(recursed,paren+1);
3373 Newx(newframe,1,scan_frame);
3374 } else {
3375 if (flags & SCF_DO_SUBSTR) {
304ee84b 3376 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3377 data->longest = &(data->longest_float);
3378 }
3379 is_inf = is_inf_internal = 1;
3380 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3381 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3382 flags &= ~SCF_DO_STCLASS;
3383 }
3384 } else {
3385 Newx(newframe,1,scan_frame);
3386 paren = stopparen;
3387 start = scan+2;
3388 end = regnext(scan);
3389 }
3390 if (newframe) {
3391 assert(start);
3392 assert(end);
3393 SAVEFREEPV(newframe);
3394 newframe->next = regnext(scan);
3395 newframe->last = last;
3396 newframe->stop = stopparen;
3397 newframe->prev = frame;
3398
3399 frame = newframe;
3400 scan = start;
3401 stopparen = paren;
3402 last = end;
3403
3404 continue;
3405 }
3406 }
3407 else if (OP(scan) == EXACT) {
3408 I32 l = STR_LEN(scan);
3409 UV uc;
3410 if (UTF) {
3411 const U8 * const s = (U8*)STRING(scan);
3412 l = utf8_length(s, s + l);
3413 uc = utf8_to_uvchr(s, NULL);
3414 } else {
3415 uc = *((U8*)STRING(scan));
3416 }
3417 min += l;
3418 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3419 /* The code below prefers earlier match for fixed
3420 offset, later match for variable offset. */
3421 if (data->last_end == -1) { /* Update the start info. */
3422 data->last_start_min = data->pos_min;
3423 data->last_start_max = is_inf
3424 ? I32_MAX : data->pos_min + data->pos_delta;
b515a41d 3425 }
8aa23a47
YO
3426 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3427 if (UTF)
3428 SvUTF8_on(data->last_found);
3429 {
3430 SV * const sv = data->last_found;
3431 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3432 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3433 if (mg && mg->mg_len >= 0)
3434 mg->mg_len += utf8_length((U8*)STRING(scan),
3435 (U8*)STRING(scan)+STR_LEN(scan));
b515a41d 3436 }
8aa23a47
YO
3437 data->last_end = data->pos_min + l;
3438 data->pos_min += l; /* As in the first entry. */
3439 data->flags &= ~SF_BEFORE_EOL;
3440 }
3441 if (flags & SCF_DO_STCLASS_AND) {
3442 /* Check whether it is compatible with what we know already! */
3443 int compat = 1;
3444
54251c2e 3445
486ec47a 3446 /* If compatible, we or it in below. It is compatible if is
54251c2e
KW
3447 * in the bitmp and either 1) its bit or its fold is set, or 2)
3448 * it's for a locale. Even if there isn't unicode semantics
3449 * here, at runtime there may be because of matching against a
3450 * utf8 string, so accept a possible false positive for
3451 * latin1-range folds */
8aa23a47
YO
3452 if (uc >= 0x100 ||
3453 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3454 && !ANYOF_BITMAP_TEST(data->start_class, uc)
39065660 3455 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
54251c2e 3456 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
8aa23a47 3457 )
d18bf9dc 3458 {
8aa23a47 3459 compat = 0;
d18bf9dc 3460 }
8aa23a47
YO
3461 ANYOF_CLASS_ZERO(data->start_class);
3462 ANYOF_BITMAP_ZERO(data->start_class);
3463 if (compat)
3464 ANYOF_BITMAP_SET(data->start_class, uc);
d18bf9dc
KW
3465 else if (uc >= 0x100) {
3466 int i;
3467
3468 /* Some Unicode code points fold to the Latin1 range; as
3469 * XXX temporary code, instead of figuring out if this is
3470 * one, just assume it is and set all the start class bits
3471 * that could be some such above 255 code point's fold
3472 * which will generate fals positives. As the code
3473 * elsewhere that does compute the fold settles down, it
3474 * can be extracted out and re-used here */
3475 for (i = 0; i < 256; i++){
3476 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3477 ANYOF_BITMAP_SET(data->start_class, i);
3478 }
3479 }
3480 }
8aa23a47
YO
3481 data->start_class->flags &= ~ANYOF_EOS;
3482 if (uc < 0x100)
3483 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3484 }
3485 else if (flags & SCF_DO_STCLASS_OR) {
3486 /* false positive possible if the class is case-folded */
3487 if (uc < 0x100)
3488 ANYOF_BITMAP_SET(data->start_class, uc);
3489 else
3490 data->start_class->flags |= ANYOF_UNICODE_ALL;
3491 data->start_class->flags &= ~ANYOF_EOS;
3492 cl_and(data->start_class, and_withp);
3493 }
3494 flags &= ~SCF_DO_STCLASS;
3495 }
3496 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3497 I32 l = STR_LEN(scan);
3498 UV uc = *((U8*)STRING(scan));
3499
3500 /* Search for fixed substrings supports EXACT only. */
3501 if (flags & SCF_DO_SUBSTR) {
3502 assert(data);
304ee84b 3503 SCAN_COMMIT(pRExC_state, data, minlenp);
8aa23a47
YO
3504 }
3505 if (UTF) {
3506 const U8 * const s = (U8 *)STRING(scan);
3507 l = utf8_length(s, s + l);
3508 uc = utf8_to_uvchr(s, NULL);
3509 }
f758bddf
KW
3510 else if (has_exactf_sharp_s) {
3511 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
bb914485 3512 }
9d071ca8 3513 min += l - min_subtract;
f646642f
KW
3514 if (min < 0) {
3515 min = 0;
3516 }
9d071ca8 3517 delta += min_subtract;
f646642f 3518 if (flags & SCF_DO_SUBSTR) {
9d071ca8 3519 data->pos_min += l - min_subtract;
f646642f
KW
3520 if (data->pos_min < 0) {
3521 data->pos_min = 0;
3522 }
9d071ca8
KW
3523 data->pos_delta += min_subtract;
3524 if (min_subtract) {
d2197104
KW
3525 data->longest = &(data->longest_float);
3526 }
f646642f 3527 }
8aa23a47
YO
3528 if (flags & SCF_DO_STCLASS_AND) {
3529 /* Check whether it is compatible with what we know already! */
3530 int compat = 1;
8aa23a47 3531 if (uc >= 0x100 ||
54251c2e
KW
3532 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3533 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3534 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3535 {
8aa23a47 3536 compat = 0;
54251c2e 3537 }
8aa23a47
YO
3538 ANYOF_CLASS_ZERO(data->start_class);
3539 ANYOF_BITMAP_ZERO(data->start_class);
3540 if (compat) {
3541 ANYOF_BITMAP_SET(data->start_class, uc);
653099ff 3542 data->start_class->flags &= ~ANYOF_EOS;
39065660 3543 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
970c8436 3544 if (OP(scan) == EXACTFL) {
af302e7f
KW
3545 /* XXX This set is probably no longer necessary, and
3546 * probably wrong as LOCALE now is on in the initial
3547 * state */
8aa23a47 3548 data->start_class->flags |= ANYOF_LOCALE;
970c8436
KW
3549 }
3550 else {
3551
54251c2e
KW
3552 /* Also set the other member of the fold pair. In case
3553 * that unicode semantics is called for at runtime, use
3554 * the full latin1 fold. (Can't do this for locale,
a0c4c608 3555 * because not known until runtime) */
54251c2e 3556 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
e22b340a 3557
a0c4c608
KW
3558 /* All other (EXACTFL handled above) folds except under
3559 * /iaa that include s, S, and sharp_s also may include
3560 * the others */
e22b340a
KW
3561 if (OP(scan) != EXACTFA) {
3562 if (uc == 's' || uc == 'S') {
3563 ANYOF_BITMAP_SET(data->start_class,
3564 LATIN_SMALL_LETTER_SHARP_S);
3565 }
3566 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3567 ANYOF_BITMAP_SET(data->start_class, 's');
3568 ANYOF_BITMAP_SET(data->start_class, 'S');
3569 }
3570 }
970c8436 3571 }
653099ff 3572 }
d18bf9dc
KW
3573 else if (uc >= 0x100) {
3574 int i;
3575 for (i = 0; i < 256; i++){
3576 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3577 ANYOF_BITMAP_SET(data->start_class, i);
3578 }
3579 }
3580 }
8aa23a47
YO
3581 }
3582 else if (flags & SCF_DO_STCLASS_OR) {
39065660 3583 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
8aa23a47
YO
3584 /* false positive possible if the class is case-folded.
3585 Assume that the locale settings are the same... */
970c8436 3586 if (uc < 0x100) {
1aa99e6b 3587 ANYOF_BITMAP_SET(data->start_class, uc);
970c8436
KW
3588 if (OP(scan) != EXACTFL) {
3589
3590 /* And set the other member of the fold pair, but
3591 * can't do that in locale because not known until
3592 * run-time */
3593 ANYOF_BITMAP_SET(data->start_class,
54251c2e 3594 PL_fold_latin1[uc]);
e22b340a
KW
3595
3596 /* All folds except under /iaa that include s, S,
3597 * and sharp_s also may include the others */
3598 if (OP(scan) != EXACTFA) {
3599 if (uc == 's' || uc == 'S') {
3600 ANYOF_BITMAP_SET(data->start_class,
3601 LATIN_SMALL_LETTER_SHARP_S);
3602 }
3603 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3604 ANYOF_BITMAP_SET(data->start_class, 's');
3605 ANYOF_BITMAP_SET(data->start_class, 'S');
3606 }
3607 }
970c8436
KW
3608 }
3609 }
653099ff
GS
3610 data->start_class->flags &= ~ANYOF_EOS;
3611 }
8aa23a47 3612 cl_and(data->start_class, and_withp);
653099ff 3613 }
8aa23a47
YO
3614 flags &= ~SCF_DO_STCLASS;
3615 }
e52fc539 3616 else if (REGNODE_VARIES(OP(scan))) {
8aa23a47
YO
3617 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3618 I32 f = flags, pos_before = 0;
3619 regnode * const oscan = scan;
3620 struct regnode_charclass_class this_class;
3621 struct regnode_charclass_class *oclass = NULL;
3622 I32 next_is_eval = 0;
3623
3624 switch (PL_regkind[OP(scan)]) {
3625 case WHILEM: /* End of (?:...)* . */
3626 scan = NEXTOPER(scan);
3627 goto finish;
3628 case PLUS:
3629 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3630 next = NEXTOPER(scan);
3631 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3632 mincount = 1;
3633 maxcount = REG_INFTY;
3634 next = regnext(scan);<