This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Add comment
[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"
370b8f2f
TC
89#ifndef PERL_IN_XSUB_RE
90# include "charclass_invlists.h"
91#endif
04e98a4d 92
d4cce5f1 93#ifdef op
11343788 94#undef op
d4cce5f1 95#endif /* op */
11343788 96
fe14fcc3 97#ifdef MSDOS
7e4e8c89 98# if defined(BUGGY_MSC6)
fe14fcc3 99 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
7e4e8c89 100# pragma optimize("a",off)
fe14fcc3 101 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
7e4e8c89
NC
102# pragma optimize("w",on )
103# endif /* BUGGY_MSC6 */
fe14fcc3
LW
104#endif /* MSDOS */
105
a687059c
LW
106#ifndef STATIC
107#define STATIC static
108#endif
109
830247a4 110typedef struct RExC_state_t {
e2509266 111 U32 flags; /* are we folding, multilining? */
830247a4 112 char *precomp; /* uncompiled string. */
288b8c02 113 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf
YO
114 regexp *rx; /* perl core regexp structure */
115 regexp_internal *rxi; /* internal data for regexp object pprivate field */
fac92740 116 char *start; /* Start of input for compile */
830247a4
IZ
117 char *end; /* End of input for compile */
118 char *parse; /* Input-scan pointer. */
119 I32 whilem_seen; /* number of WHILEM in this expr */
fac92740 120 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 121 regnode *emit_bound; /* First regnode outside of the allocated space */
ffc61ed2 122 regnode *emit; /* Code-emit pointer; &regdummy = don't = compiling */
830247a4
IZ
123 I32 naughty; /* How bad is this pattern? */
124 I32 sawback; /* Did we see \1, ...? */
125 U32 seen;
126 I32 size; /* Code size. */
c74340f9
YO
127 I32 npar; /* Capture buffer count, (OPEN). */
128 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 129 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
130 I32 extralen;
131 I32 seen_zerolen;
132 I32 seen_evals;
40d049e4
YO
133 regnode **open_parens; /* pointers to open parens */
134 regnode **close_parens; /* pointers to close parens */
135 regnode *opend; /* END node in program */
02daf0ab
YO
136 I32 utf8; /* whether the pattern is utf8 or not */
137 I32 orig_utf8; /* whether the pattern was originally in utf8 */
138 /* XXX use this for future optimisation of case
139 * where pattern must be upgraded to utf8. */
e40e74fe
KW
140 I32 uni_semantics; /* If a d charset modifier should use unicode
141 rules, even if the pattern is not in
142 utf8 */
81714fb9 143 HV *paren_names; /* Paren names */
1f1031fe 144
40d049e4
YO
145 regnode **recurse; /* Recurse regops */
146 I32 recurse_count; /* Number of recurse regops */
b57e4118 147 I32 in_lookbehind;
4624b182 148 I32 contains_locale;
bb3f3ed2 149 I32 override_recoding;
830247a4
IZ
150#if ADD_TO_REGEXEC
151 char *starttry; /* -Dr: where regtry was called. */
152#define RExC_starttry (pRExC_state->starttry)
153#endif
3dab1dad 154#ifdef DEBUGGING
be8e71aa 155 const char *lastparse;
3dab1dad 156 I32 lastnum;
1f1031fe 157 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
158#define RExC_lastparse (pRExC_state->lastparse)
159#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 160#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 161#endif
830247a4
IZ
162} RExC_state_t;
163
e2509266 164#define RExC_flags (pRExC_state->flags)
830247a4 165#define RExC_precomp (pRExC_state->precomp)
288b8c02 166#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 167#define RExC_rx (pRExC_state->rx)
f8fc2ecf 168#define RExC_rxi (pRExC_state->rxi)
fac92740 169#define RExC_start (pRExC_state->start)
830247a4
IZ
170#define RExC_end (pRExC_state->end)
171#define RExC_parse (pRExC_state->parse)
172#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
173#ifdef RE_TRACK_PATTERN_OFFSETS
174#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
175#endif
830247a4 176#define RExC_emit (pRExC_state->emit)
fac92740 177#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 178#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
179#define RExC_naughty (pRExC_state->naughty)
180#define RExC_sawback (pRExC_state->sawback)
181#define RExC_seen (pRExC_state->seen)
182#define RExC_size (pRExC_state->size)
183#define RExC_npar (pRExC_state->npar)
e2e6a0f1 184#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
185#define RExC_extralen (pRExC_state->extralen)
186#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
187#define RExC_seen_evals (pRExC_state->seen_evals)
1aa99e6b 188#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 189#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 190#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
191#define RExC_open_parens (pRExC_state->open_parens)
192#define RExC_close_parens (pRExC_state->close_parens)
193#define RExC_opend (pRExC_state->opend)
81714fb9 194#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
195#define RExC_recurse (pRExC_state->recurse)
196#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 197#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 198#define RExC_contains_locale (pRExC_state->contains_locale)
bb3f3ed2 199#define RExC_override_recoding (pRExC_state->override_recoding)
830247a4 200
cde0cee5 201
a687059c
LW
202#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
203#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
204 ((*s) == '{' && regcurly(s)))
a687059c 205
35c8bce7
LW
206#ifdef SPSTART
207#undef SPSTART /* dratted cpp namespace... */
208#endif
a687059c
LW
209/*
210 * Flags to be passed up and down.
211 */
a687059c 212#define WORST 0 /* Worst case. */
a3b492c3 213#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee
KW
214
215/* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
d7b56a3c 216 * character, and if utf8, must be invariant. Note that this is not the same thing as REGNODE_SIMPLE */
fda99bee 217#define SIMPLE 0x02
a3b492c3
YO
218#define SPSTART 0x04 /* Starts with * or +. */
219#define TRYAGAIN 0x08 /* Weeded out a declaration. */
220#define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
a687059c 221
3dab1dad
YO
222#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
223
07be1b83
YO
224/* whether trie related optimizations are enabled */
225#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
226#define TRIE_STUDY_OPT
786e8c11 227#define FULL_TRIE_STUDY
07be1b83
YO
228#define TRIE_STCLASS
229#endif
1de06328
YO
230
231
40d049e4
YO
232
233#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
234#define PBITVAL(paren) (1 << ((paren) & 7))
235#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
236#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
237#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
238
bbd61b5f
KW
239/* If not already in utf8, do a longjmp back to the beginning */
240#define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
241#define REQUIRE_UTF8 STMT_START { \
242 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
243 } STMT_END
40d049e4 244
1de06328
YO
245/* About scan_data_t.
246
247 During optimisation we recurse through the regexp program performing
248 various inplace (keyhole style) optimisations. In addition study_chunk
249 and scan_commit populate this data structure with information about
250 what strings MUST appear in the pattern. We look for the longest
3b753521 251 string that must appear at a fixed location, and we look for the
1de06328
YO
252 longest string that may appear at a floating location. So for instance
253 in the pattern:
254
255 /FOO[xX]A.*B[xX]BAR/
256
257 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
258 strings (because they follow a .* construct). study_chunk will identify
259 both FOO and BAR as being the longest fixed and floating strings respectively.
260
261 The strings can be composites, for instance
262
263 /(f)(o)(o)/
264
265 will result in a composite fixed substring 'foo'.
266
267 For each string some basic information is maintained:
268
269 - offset or min_offset
270 This is the position the string must appear at, or not before.
271 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
272 characters must match before the string we are searching for.
273 Likewise when combined with minlenp and the length of the string it
1de06328
YO
274 tells us how many characters must appear after the string we have
275 found.
276
277 - max_offset
278 Only used for floating strings. This is the rightmost point that
3b753521 279 the string can appear at. If set to I32 max it indicates that the
1de06328
YO
280 string can occur infinitely far to the right.
281
282 - minlenp
283 A pointer to the minimum length of the pattern that the string
284 was found inside. This is important as in the case of positive
285 lookahead or positive lookbehind we can have multiple patterns
286 involved. Consider
287
288 /(?=FOO).*F/
289
290 The minimum length of the pattern overall is 3, the minimum length
291 of the lookahead part is 3, but the minimum length of the part that
292 will actually match is 1. So 'FOO's minimum length is 3, but the
293 minimum length for the F is 1. This is important as the minimum length
294 is used to determine offsets in front of and behind the string being
295 looked for. Since strings can be composites this is the length of the
486ec47a 296 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
297 the length is calculated by study_chunk, so that the minimum lengths
298 are not known until the full pattern has been compiled, thus the
299 pointer to the value.
300
301 - lookbehind
302
303 In the case of lookbehind the string being searched for can be
304 offset past the start point of the final matching string.
305 If this value was just blithely removed from the min_offset it would
306 invalidate some of the calculations for how many chars must match
307 before or after (as they are derived from min_offset and minlen and
308 the length of the string being searched for).
309 When the final pattern is compiled and the data is moved from the
310 scan_data_t structure into the regexp structure the information
311 about lookbehind is factored in, with the information that would
312 have been lost precalculated in the end_shift field for the
313 associated string.
314
315 The fields pos_min and pos_delta are used to store the minimum offset
316 and the delta to the maximum offset at the current point in the pattern.
317
318*/
2c2d71f5
JH
319
320typedef struct scan_data_t {
1de06328
YO
321 /*I32 len_min; unused */
322 /*I32 len_delta; unused */
2c2d71f5
JH
323 I32 pos_min;
324 I32 pos_delta;
325 SV *last_found;
1de06328 326 I32 last_end; /* min value, <0 unless valid. */
2c2d71f5
JH
327 I32 last_start_min;
328 I32 last_start_max;
1de06328
YO
329 SV **longest; /* Either &l_fixed, or &l_float. */
330 SV *longest_fixed; /* longest fixed string found in pattern */
331 I32 offset_fixed; /* offset where it starts */
486ec47a 332 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
333 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
334 SV *longest_float; /* longest floating string found in pattern */
335 I32 offset_float_min; /* earliest point in string it can appear */
336 I32 offset_float_max; /* latest point in string it can appear */
486ec47a 337 I32 *minlen_float; /* pointer to the minlen relevant to the string */
1de06328 338 I32 lookbehind_float; /* is the position of the string modified by LB */
2c2d71f5
JH
339 I32 flags;
340 I32 whilem_c;
cb434fcc 341 I32 *last_closep;
653099ff 342 struct regnode_charclass_class *start_class;
2c2d71f5
JH
343} scan_data_t;
344
a687059c 345/*
e50aee73 346 * Forward declarations for pregcomp()'s friends.
a687059c 347 */
a0d0e21e 348
27da23d5 349static const scan_data_t zero_scan_data =
1de06328 350 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
351
352#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
353#define SF_BEFORE_SEOL 0x0001
354#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
355#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
356#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
357
09b7f37c
CB
358#ifdef NO_UNARY_PLUS
359# define SF_FIX_SHIFT_EOL (0+2)
360# define SF_FL_SHIFT_EOL (0+4)
361#else
362# define SF_FIX_SHIFT_EOL (+2)
363# define SF_FL_SHIFT_EOL (+4)
364#endif
c277df42
IZ
365
366#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
367#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
368
369#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
370#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
371#define SF_IS_INF 0x0040
372#define SF_HAS_PAR 0x0080
373#define SF_IN_PAR 0x0100
374#define SF_HAS_EVAL 0x0200
375#define SCF_DO_SUBSTR 0x0400
653099ff
GS
376#define SCF_DO_STCLASS_AND 0x0800
377#define SCF_DO_STCLASS_OR 0x1000
378#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 379#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 380
786e8c11 381#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 382#define SCF_SEEN_ACCEPT 0x8000
07be1b83 383
43fead97 384#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
385
386/* The enums for all these are ordered so things work out correctly */
a62b1201 387#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
cfaf538b 388#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
00b27cfc 389#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
390#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
391#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1
KW
392#define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
393#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a62b1201 394
43fead97 395#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 396
ffc61ed2 397#define OOB_UNICODE 12345678
93733859 398#define OOB_NAMEDCLASS -1
b8c5462f 399
a0ed51b3
LW
400#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
401#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
402
8615cb43 403
b45f050a
JF
404/* length of regex to show in messages that don't mark a position within */
405#define RegexLengthToShowInErrorMessages 127
406
407/*
408 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
409 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
410 * op/pragma/warn/regcomp.
411 */
7253e4e3
RK
412#define MARKER1 "<-- HERE" /* marker as it appears in the description */
413#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 414
7253e4e3 415#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
416
417/*
418 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
419 * arg. Show regex, up to a maximum length. If it's too long, chop and add
420 * "...".
421 */
58e23c8d 422#define _FAIL(code) STMT_START { \
bfed75c6 423 const char *ellipses = ""; \
ccb2c380
MP
424 IV len = RExC_end - RExC_precomp; \
425 \
426 if (!SIZE_ONLY) \
288b8c02 427 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
428 if (len > RegexLengthToShowInErrorMessages) { \
429 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
430 len = RegexLengthToShowInErrorMessages - 10; \
431 ellipses = "..."; \
432 } \
58e23c8d 433 code; \
ccb2c380 434} STMT_END
8615cb43 435
58e23c8d
YO
436#define FAIL(msg) _FAIL( \
437 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
438 msg, (int)len, RExC_precomp, ellipses))
439
440#define FAIL2(msg,arg) _FAIL( \
441 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
442 arg, (int)len, RExC_precomp, ellipses))
443
b45f050a 444/*
b45f050a
JF
445 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
446 */
ccb2c380 447#define Simple_vFAIL(m) STMT_START { \
a28509cc 448 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
449 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
450 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
451} STMT_END
b45f050a
JF
452
453/*
454 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
455 */
ccb2c380
MP
456#define vFAIL(m) STMT_START { \
457 if (!SIZE_ONLY) \
288b8c02 458 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
459 Simple_vFAIL(m); \
460} STMT_END
b45f050a
JF
461
462/*
463 * Like Simple_vFAIL(), but accepts two arguments.
464 */
ccb2c380 465#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 466 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
467 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
468 (int)offset, RExC_precomp, RExC_precomp + offset); \
469} STMT_END
b45f050a
JF
470
471/*
472 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
473 */
ccb2c380
MP
474#define vFAIL2(m,a1) STMT_START { \
475 if (!SIZE_ONLY) \
288b8c02 476 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
477 Simple_vFAIL2(m, a1); \
478} STMT_END
b45f050a
JF
479
480
481/*
482 * Like Simple_vFAIL(), but accepts three arguments.
483 */
ccb2c380 484#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 485 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
486 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
487 (int)offset, RExC_precomp, RExC_precomp + offset); \
488} STMT_END
b45f050a
JF
489
490/*
491 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
492 */
ccb2c380
MP
493#define vFAIL3(m,a1,a2) STMT_START { \
494 if (!SIZE_ONLY) \
288b8c02 495 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
496 Simple_vFAIL3(m, a1, a2); \
497} STMT_END
b45f050a
JF
498
499/*
500 * Like Simple_vFAIL(), but accepts four arguments.
501 */
ccb2c380 502#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 503 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
504 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
505 (int)offset, RExC_precomp, RExC_precomp + offset); \
506} STMT_END
b45f050a 507
668c081a 508#define ckWARNreg(loc,m) STMT_START { \
a28509cc 509 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
510 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
511 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
512} STMT_END
513
668c081a 514#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 515 const IV offset = loc - RExC_precomp; \
d1d15184 516 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
517 m REPORT_LOCATION, \
518 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
519} STMT_END
520
2335b3d3
KW
521#define ckWARN2regdep(loc,m, a1) STMT_START { \
522 const IV offset = loc - RExC_precomp; \
523 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
524 m REPORT_LOCATION, \
525 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
526} STMT_END
527
668c081a 528#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 529 const IV offset = loc - RExC_precomp; \
668c081a 530 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
531 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
532} STMT_END
533
534#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 535 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
536 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
537 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
538} STMT_END
539
668c081a
NC
540#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
541 const IV offset = loc - RExC_precomp; \
542 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
543 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
544} STMT_END
545
ccb2c380 546#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 547 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
548 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
549 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
550} STMT_END
551
668c081a
NC
552#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
553 const IV offset = loc - RExC_precomp; \
554 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
555 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
556} STMT_END
557
ccb2c380 558#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 559 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
560 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
561 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
562} STMT_END
9d1d55b5 563
8615cb43 564
cd439c50 565/* Allow for side effects in s */
ccb2c380
MP
566#define REGC(c,s) STMT_START { \
567 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
568} STMT_END
cd439c50 569
fac92740
MJD
570/* Macros for recording node offsets. 20001227 mjd@plover.com
571 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
572 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
573 * Element 0 holds the number n.
07be1b83 574 * Position is 1 indexed.
fac92740 575 */
7122b237
YO
576#ifndef RE_TRACK_PATTERN_OFFSETS
577#define Set_Node_Offset_To_R(node,byte)
578#define Set_Node_Offset(node,byte)
579#define Set_Cur_Node_Offset
580#define Set_Node_Length_To_R(node,len)
581#define Set_Node_Length(node,len)
582#define Set_Node_Cur_Length(node)
583#define Node_Offset(n)
584#define Node_Length(n)
585#define Set_Node_Offset_Length(node,offset,len)
586#define ProgLen(ri) ri->u.proglen
587#define SetProgLen(ri,x) ri->u.proglen = x
588#else
589#define ProgLen(ri) ri->u.offsets[0]
590#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
591#define Set_Node_Offset_To_R(node,byte) STMT_START { \
592 if (! SIZE_ONLY) { \
593 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 594 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 595 if((node) < 0) { \
551405c4 596 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
597 } else { \
598 RExC_offsets[2*(node)-1] = (byte); \
599 } \
600 } \
601} STMT_END
602
603#define Set_Node_Offset(node,byte) \
604 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
605#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
606
607#define Set_Node_Length_To_R(node,len) STMT_START { \
608 if (! SIZE_ONLY) { \
609 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 610 __LINE__, (int)(node), (int)(len))); \
ccb2c380 611 if((node) < 0) { \
551405c4 612 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
613 } else { \
614 RExC_offsets[2*(node)] = (len); \
615 } \
616 } \
617} STMT_END
618
619#define Set_Node_Length(node,len) \
620 Set_Node_Length_To_R((node)-RExC_emit_start, len)
621#define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
622#define Set_Node_Cur_Length(node) \
623 Set_Node_Length(node, RExC_parse - parse_start)
fac92740
MJD
624
625/* Get offsets and lengths */
626#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
627#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
628
07be1b83
YO
629#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
630 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
631 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
632} STMT_END
7122b237 633#endif
07be1b83
YO
634
635#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
636#define EXPERIMENTAL_INPLACESCAN
f427392e 637#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 638
304ee84b
YO
639#define DEBUG_STUDYDATA(str,data,depth) \
640DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 641 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
642 "%*s" str "Pos:%"IVdf"/%"IVdf \
643 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
644 (int)(depth)*2, "", \
645 (IV)((data)->pos_min), \
646 (IV)((data)->pos_delta), \
304ee84b 647 (UV)((data)->flags), \
1de06328 648 (IV)((data)->whilem_c), \
304ee84b
YO
649 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
650 is_inf ? "INF " : "" \
1de06328
YO
651 ); \
652 if ((data)->last_found) \
653 PerlIO_printf(Perl_debug_log, \
654 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
655 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
656 SvPVX_const((data)->last_found), \
657 (IV)((data)->last_end), \
658 (IV)((data)->last_start_min), \
659 (IV)((data)->last_start_max), \
660 ((data)->longest && \
661 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
662 SvPVX_const((data)->longest_fixed), \
663 (IV)((data)->offset_fixed), \
664 ((data)->longest && \
665 (data)->longest==&((data)->longest_float)) ? "*" : "", \
666 SvPVX_const((data)->longest_float), \
667 (IV)((data)->offset_float_min), \
668 (IV)((data)->offset_float_max) \
669 ); \
670 PerlIO_printf(Perl_debug_log,"\n"); \
671});
672
acfe0abc 673static void clear_re(pTHX_ void *r);
4327152a 674
653099ff 675/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 676 Update the longest found anchored substring and the longest found
653099ff
GS
677 floating substrings if needed. */
678
4327152a 679STATIC void
304ee84b 680S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 681{
e1ec3a88
AL
682 const STRLEN l = CHR_SVLEN(data->last_found);
683 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 684 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 685
7918f24d
NC
686 PERL_ARGS_ASSERT_SCAN_COMMIT;
687
c277df42 688 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 689 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
690 if (*data->longest == data->longest_fixed) {
691 data->offset_fixed = l ? data->last_start_min : data->pos_min;
692 if (data->flags & SF_BEFORE_EOL)
b81d288d 693 data->flags
c277df42
IZ
694 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
695 else
696 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 697 data->minlen_fixed=minlenp;
1de06328 698 data->lookbehind_fixed=0;
a0ed51b3 699 }
304ee84b 700 else { /* *data->longest == data->longest_float */
c277df42 701 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
702 data->offset_float_max = (l
703 ? data->last_start_max
c277df42 704 : data->pos_min + data->pos_delta);
304ee84b 705 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 706 data->offset_float_max = I32_MAX;
c277df42 707 if (data->flags & SF_BEFORE_EOL)
b81d288d 708 data->flags
c277df42
IZ
709 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
710 else
711 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
712 data->minlen_float=minlenp;
713 data->lookbehind_float=0;
c277df42
IZ
714 }
715 }
716 SvCUR_set(data->last_found, 0);
0eda9292 717 {
a28509cc 718 SV * const sv = data->last_found;
097eb12c
AL
719 if (SvUTF8(sv) && SvMAGICAL(sv)) {
720 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
721 if (mg)
722 mg->mg_len = 0;
723 }
0eda9292 724 }
c277df42
IZ
725 data->last_end = -1;
726 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 727 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
728}
729
653099ff
GS
730/* Can match anything (initialization) */
731STATIC void
3fffb88a 732S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 733{
7918f24d
NC
734 PERL_ARGS_ASSERT_CL_ANYTHING;
735
f8bef550 736 ANYOF_BITMAP_SETALL(cl);
dd58aee1 737 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
3ad98780 738 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
3fffb88a
KW
739
740 /* If any portion of the regex is to operate under locale rules,
741 * initialization includes it. The reason this isn't done for all regexes
742 * is that the optimizer was written under the assumption that locale was
743 * all-or-nothing. Given the complexity and lack of documentation in the
744 * optimizer, and that there are inadequate test cases for locale, so many
745 * parts of it may not work properly, it is safest to avoid locale unless
746 * necessary. */
747 if (RExC_contains_locale) {
9d7a1e63 748 ANYOF_CLASS_SETALL(cl); /* /l uses class */
3fffb88a
KW
749 cl->flags |= ANYOF_LOCALE;
750 }
9d7a1e63
KW
751 else {
752 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
753 }
653099ff
GS
754}
755
756/* Can match anything (initialization) */
757STATIC int
5f66b61c 758S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
759{
760 int value;
761
7918f24d
NC
762 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
763
aaa51d5e 764 for (value = 0; value <= ANYOF_MAX; value += 2)
653099ff
GS
765 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
766 return 1;
1aa99e6b
IH
767 if (!(cl->flags & ANYOF_UNICODE_ALL))
768 return 0;
10edeb5d 769 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 770 return 0;
653099ff
GS
771 return 1;
772}
773
774/* Can match anything (initialization) */
775STATIC void
e755fd73 776S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 777{
7918f24d
NC
778 PERL_ARGS_ASSERT_CL_INIT;
779
8ecf7187 780 Zero(cl, 1, struct regnode_charclass_class);
653099ff 781 cl->type = ANYOF;
3fffb88a 782 cl_anything(pRExC_state, cl);
1411dba4 783 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
784}
785
1051e1c4
KW
786/* These two functions currently do the exact same thing */
787#define cl_init_zero S_cl_init
653099ff 788
dd58aee1
KW
789/* 'AND' a given class with another one. Can create false positives. 'cl'
790 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
791 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 792STATIC void
5f66b61c 793S_cl_and(struct regnode_charclass_class *cl,
a28509cc 794 const struct regnode_charclass_class *and_with)
653099ff 795{
7918f24d 796 PERL_ARGS_ASSERT_CL_AND;
40d049e4
YO
797
798 assert(and_with->type == ANYOF);
1e6ade67 799
c6b76537 800 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
801 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
802 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 803 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
804 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
805 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
653099ff
GS
806 int i;
807
808 if (and_with->flags & ANYOF_INVERT)
809 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
810 cl->bitmap[i] &= ~and_with->bitmap[i];
811 else
812 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
813 cl->bitmap[i] &= and_with->bitmap[i];
814 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 815
c6b76537 816 if (and_with->flags & ANYOF_INVERT) {
8951c461 817
c6b76537
KW
818 /* Here, the and'ed node is inverted. Get the AND of the flags that
819 * aren't affected by the inversion. Those that are affected are
820 * handled individually below */
821 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
822 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
823 cl->flags |= affected_flags;
824
825 /* We currently don't know how to deal with things that aren't in the
826 * bitmap, but we know that the intersection is no greater than what
827 * is already in cl, so let there be false positives that get sorted
828 * out after the synthetic start class succeeds, and the node is
829 * matched for real. */
830
831 /* The inversion of these two flags indicate that the resulting
832 * intersection doesn't have them */
833 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
834 cl->flags &= ~ANYOF_UNICODE_ALL;
835 }
c6b76537
KW
836 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
837 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 838 }
1aa99e6b 839 }
c6b76537 840 else { /* and'd node is not inverted */
3ad98780
KW
841 U8 outside_bitmap_but_not_utf8; /* Temp variable */
842
137165a6 843 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
844
845 /* Here 'and_with' doesn't match anything outside the bitmap
846 * (except possibly ANYOF_UNICODE_ALL), which means the
847 * intersection can't either, except for ANYOF_UNICODE_ALL, in
848 * which case we don't know what the intersection is, but it's no
849 * greater than what cl already has, so can just leave it alone,
850 * with possible false positives */
851 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
852 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 853 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 854 }
137165a6 855 }
c6b76537
KW
856 else if (! ANYOF_NONBITMAP(cl)) {
857
858 /* Here, 'and_with' does match something outside the bitmap, and cl
859 * doesn't have a list of things to match outside the bitmap. If
860 * cl can match all code points above 255, the intersection will
3ad98780
KW
861 * be those above-255 code points that 'and_with' matches. If cl
862 * can't match all Unicode code points, it means that it can't
863 * match anything outside the bitmap (since the 'if' that got us
864 * into this block tested for that), so we leave the bitmap empty.
865 */
c6b76537
KW
866 if (cl->flags & ANYOF_UNICODE_ALL) {
867 ARG_SET(cl, ARG(and_with));
3ad98780
KW
868
869 /* and_with's ARG may match things that don't require UTF8.
870 * And now cl's will too, in spite of this being an 'and'. See
871 * the comments below about the kludge */
872 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
873 }
874 }
875 else {
876 /* Here, both 'and_with' and cl match something outside the
877 * bitmap. Currently we do not do the intersection, so just match
878 * whatever cl had at the beginning. */
879 }
880
881
3ad98780
KW
882 /* Take the intersection of the two sets of flags. However, the
883 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
884 * kludge around the fact that this flag is not treated like the others
885 * which are initialized in cl_anything(). The way the optimizer works
886 * is that the synthetic start class (SSC) is initialized to match
887 * anything, and then the first time a real node is encountered, its
888 * values are AND'd with the SSC's with the result being the values of
889 * the real node. However, there are paths through the optimizer where
890 * the AND never gets called, so those initialized bits are set
891 * inappropriately, which is not usually a big deal, as they just cause
892 * false positives in the SSC, which will just mean a probably
893 * imperceptible slow down in execution. However this bit has a
894 * higher false positive consequence in that it can cause utf8.pm,
895 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
896 * bigger slowdown and also causes significant extra memory to be used.
897 * In order to prevent this, the code now takes a different tack. The
898 * bit isn't set unless some part of the regular expression needs it,
899 * but once set it won't get cleared. This means that these extra
900 * modules won't get loaded unless there was some path through the
901 * pattern that would have required them anyway, and so any false
902 * positives that occur by not ANDing them out when they could be
903 * aren't as severe as they would be if we treated this bit like all
904 * the others */
905 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
906 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 907 cl->flags &= and_with->flags;
3ad98780 908 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 909 }
653099ff
GS
910}
911
dd58aee1
KW
912/* 'OR' a given class with another one. Can create false positives. 'cl'
913 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
914 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 915STATIC void
3fffb88a 916S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 917{
7918f24d
NC
918 PERL_ARGS_ASSERT_CL_OR;
919
653099ff 920 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
921
922 /* Here, the or'd node is to be inverted. This means we take the
923 * complement of everything not in the bitmap, but currently we don't
924 * know what that is, so give up and match anything */
925 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 926 cl_anything(pRExC_state, cl);
c6b76537 927 }
653099ff
GS
928 /* We do not use
929 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
930 * <= (B1 | !B2) | (CL1 | !CL2)
931 * which is wasteful if CL2 is small, but we ignore CL2:
932 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
933 * XXXX Can we handle case-fold? Unclear:
934 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
935 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
936 */
c6b76537 937 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
938 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
939 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
653099ff
GS
940 int i;
941
942 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
943 cl->bitmap[i] |= ~or_with->bitmap[i];
944 } /* XXXX: logic is complicated otherwise */
945 else {
3fffb88a 946 cl_anything(pRExC_state, cl);
653099ff 947 }
c6b76537
KW
948
949 /* And, we can just take the union of the flags that aren't affected
950 * by the inversion */
951 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
952
953 /* For the remaining flags:
954 ANYOF_UNICODE_ALL and inverted means to not match anything above
955 255, which means that the union with cl should just be
956 what cl has in it, so can ignore this flag
957 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
958 is 127-255 to match them, but then invert that, so the
959 union with cl should just be what cl has in it, so can
960 ignore this flag
961 */
962 } else { /* 'or_with' is not inverted */
653099ff
GS
963 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
964 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
965 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
966 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
653099ff
GS
967 int i;
968
969 /* OR char bitmap and class bitmap separately */
970 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
971 cl->bitmap[i] |= or_with->bitmap[i];
1e6ade67 972 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
653099ff
GS
973 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
974 cl->classflags[i] |= or_with->classflags[i];
975 cl->flags |= ANYOF_CLASS;
976 }
977 }
978 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 979 cl_anything(pRExC_state, cl);
653099ff 980 }
9826f543 981
c6b76537
KW
982 if (ANYOF_NONBITMAP(or_with)) {
983
984 /* Use the added node's outside-the-bit-map match if there isn't a
985 * conflict. If there is a conflict (both nodes match something
986 * outside the bitmap, but what they match outside is not the same
987 * pointer, and hence not easily compared until XXX we extend
988 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
989 * match everything outside the bitmap. If that stuff is all above
990 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
991 if (! ANYOF_NONBITMAP(cl)) {
992 ARG_SET(cl, ARG(or_with));
993 }
994 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
995
996 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
997 cl_anything(pRExC_state, cl);
998 }
999 else {
1000 cl->flags |= ANYOF_UNICODE_ALL;
1001 }
c6b76537 1002 }
4c34a693 1003 }
0b9668ee
KW
1004
1005 /* Take the union */
1006 cl->flags |= or_with->flags;
1aa99e6b 1007 }
653099ff
GS
1008}
1009
a3621e74
YO
1010#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1011#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1012#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1013#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1014
3dab1dad
YO
1015
1016#ifdef DEBUGGING
07be1b83 1017/*
2b8b4781
NC
1018 dump_trie(trie,widecharmap,revcharmap)
1019 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1020 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1021
1022 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1023 The _interim_ variants are used for debugging the interim
1024 tables that are used to generate the final compressed
1025 representation which is what dump_trie expects.
1026
486ec47a 1027 Part of the reason for their existence is to provide a form
3dab1dad 1028 of documentation as to how the different representations function.
07be1b83
YO
1029
1030*/
3dab1dad
YO
1031
1032/*
3dab1dad
YO
1033 Dumps the final compressed table form of the trie to Perl_debug_log.
1034 Used for debugging make_trie().
1035*/
b9a59e08 1036
3dab1dad 1037STATIC void
2b8b4781
NC
1038S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1039 AV *revcharmap, U32 depth)
3dab1dad
YO
1040{
1041 U32 state;
ab3bbdeb 1042 SV *sv=sv_newmortal();
55eed653 1043 int colwidth= widecharmap ? 6 : 4;
2e64971a 1044 U16 word;
3dab1dad
YO
1045 GET_RE_DEBUG_FLAGS_DECL;
1046
7918f24d 1047 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1048
3dab1dad
YO
1049 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1050 (int)depth * 2 + 2,"",
1051 "Match","Base","Ofs" );
1052
1053 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1054 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1055 if ( tmp ) {
ab3bbdeb
YO
1056 PerlIO_printf( Perl_debug_log, "%*s",
1057 colwidth,
ddc5bc0f 1058 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1059 PL_colors[0], PL_colors[1],
1060 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1061 PERL_PV_ESCAPE_FIRSTCHAR
1062 )
1063 );
3dab1dad
YO
1064 }
1065 }
1066 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1067 (int)depth * 2 + 2,"");
1068
1069 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1070 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1071 PerlIO_printf( Perl_debug_log, "\n");
1072
1e2e3d02 1073 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1074 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1075
1076 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1077
1078 if ( trie->states[ state ].wordnum ) {
1079 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1080 } else {
1081 PerlIO_printf( Perl_debug_log, "%6s", "" );
1082 }
1083
1084 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1085
1086 if ( base ) {
1087 U32 ofs = 0;
1088
1089 while( ( base + ofs < trie->uniquecharcount ) ||
1090 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1091 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1092 ofs++;
1093
1094 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1095
1096 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1097 if ( ( base + ofs >= trie->uniquecharcount ) &&
1098 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1099 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1100 {
ab3bbdeb
YO
1101 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1102 colwidth,
3dab1dad
YO
1103 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1104 } else {
ab3bbdeb 1105 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1106 }
1107 }
1108
1109 PerlIO_printf( Perl_debug_log, "]");
1110
1111 }
1112 PerlIO_printf( Perl_debug_log, "\n" );
1113 }
2e64971a
DM
1114 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1115 for (word=1; word <= trie->wordcount; word++) {
1116 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1117 (int)word, (int)(trie->wordinfo[word].prev),
1118 (int)(trie->wordinfo[word].len));
1119 }
1120 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1121}
1122/*
3dab1dad
YO
1123 Dumps a fully constructed but uncompressed trie in list form.
1124 List tries normally only are used for construction when the number of
1125 possible chars (trie->uniquecharcount) is very high.
1126 Used for debugging make_trie().
1127*/
1128STATIC void
55eed653 1129S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1130 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1131 U32 depth)
3dab1dad
YO
1132{
1133 U32 state;
ab3bbdeb 1134 SV *sv=sv_newmortal();
55eed653 1135 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1136 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1137
1138 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1139
3dab1dad 1140 /* print out the table precompression. */
ab3bbdeb
YO
1141 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1142 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1143 "------:-----+-----------------\n" );
3dab1dad
YO
1144
1145 for( state=1 ; state < next_alloc ; state ++ ) {
1146 U16 charid;
1147
ab3bbdeb 1148 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1149 (int)depth * 2 + 2,"", (UV)state );
1150 if ( ! trie->states[ state ].wordnum ) {
1151 PerlIO_printf( Perl_debug_log, "%5s| ","");
1152 } else {
1153 PerlIO_printf( Perl_debug_log, "W%4x| ",
1154 trie->states[ state ].wordnum
1155 );
1156 }
1157 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1158 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1159 if ( tmp ) {
1160 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1161 colwidth,
ddc5bc0f 1162 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1163 PL_colors[0], PL_colors[1],
1164 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1165 PERL_PV_ESCAPE_FIRSTCHAR
1166 ) ,
1e2e3d02
YO
1167 TRIE_LIST_ITEM(state,charid).forid,
1168 (UV)TRIE_LIST_ITEM(state,charid).newstate
1169 );
1170 if (!(charid % 10))
664e119d
RGS
1171 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1172 (int)((depth * 2) + 14), "");
1e2e3d02 1173 }
ab3bbdeb
YO
1174 }
1175 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1176 }
1177}
1178
1179/*
3dab1dad
YO
1180 Dumps a fully constructed but uncompressed trie in table form.
1181 This is the normal DFA style state transition table, with a few
1182 twists to facilitate compression later.
1183 Used for debugging make_trie().
1184*/
1185STATIC void
55eed653 1186S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1187 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1188 U32 depth)
3dab1dad
YO
1189{
1190 U32 state;
1191 U16 charid;
ab3bbdeb 1192 SV *sv=sv_newmortal();
55eed653 1193 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1194 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1195
1196 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1197
1198 /*
1199 print out the table precompression so that we can do a visual check
1200 that they are identical.
1201 */
1202
1203 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1204
1205 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1206 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1207 if ( tmp ) {
ab3bbdeb
YO
1208 PerlIO_printf( Perl_debug_log, "%*s",
1209 colwidth,
ddc5bc0f 1210 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1211 PL_colors[0], PL_colors[1],
1212 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1213 PERL_PV_ESCAPE_FIRSTCHAR
1214 )
1215 );
3dab1dad
YO
1216 }
1217 }
1218
1219 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1220
1221 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1222 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1223 }
1224
1225 PerlIO_printf( Perl_debug_log, "\n" );
1226
1227 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1228
1229 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1230 (int)depth * 2 + 2,"",
1231 (UV)TRIE_NODENUM( state ) );
1232
1233 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1234 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1235 if (v)
1236 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1237 else
1238 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1239 }
1240 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1241 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1242 } else {
1243 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1244 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1245 }
1246 }
07be1b83 1247}
3dab1dad
YO
1248
1249#endif
1250
2e64971a 1251
786e8c11
YO
1252/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1253 startbranch: the first branch in the whole branch sequence
1254 first : start branch of sequence of branch-exact nodes.
1255 May be the same as startbranch
1256 last : Thing following the last branch.
1257 May be the same as tail.
1258 tail : item following the branch sequence
1259 count : words in the sequence
1260 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1261 depth : indent depth
3dab1dad 1262
786e8c11 1263Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1264
786e8c11
YO
1265A trie is an N'ary tree where the branches are determined by digital
1266decomposition of the key. IE, at the root node you look up the 1st character and
1267follow that branch repeat until you find the end of the branches. Nodes can be
1268marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1269
786e8c11 1270 /he|she|his|hers/
72f13be8 1271
786e8c11
YO
1272would convert into the following structure. Numbers represent states, letters
1273following numbers represent valid transitions on the letter from that state, if
1274the number is in square brackets it represents an accepting state, otherwise it
1275will be in parenthesis.
07be1b83 1276
786e8c11
YO
1277 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1278 | |
1279 | (2)
1280 | |
1281 (1) +-i->(6)-+-s->[7]
1282 |
1283 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1284
786e8c11
YO
1285 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1286
1287This shows that when matching against the string 'hers' we will begin at state 1
1288read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1289then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1290is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1291single traverse. We store a mapping from accepting to state to which word was
1292matched, and then when we have multiple possibilities we try to complete the
1293rest of the regex in the order in which they occured in the alternation.
1294
1295The only prior NFA like behaviour that would be changed by the TRIE support is
1296the silent ignoring of duplicate alternations which are of the form:
1297
1298 / (DUPE|DUPE) X? (?{ ... }) Y /x
1299
4b714af6 1300Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1301and without the optimisation. With the optimisations dupes will be silently
486ec47a 1302ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1303the following demonstrates:
1304
1305 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1306
1307which prints out 'word' three times, but
1308
1309 'words'=~/(word|word|word)(?{ print $1 })S/
1310
1311which doesnt print it out at all. This is due to other optimisations kicking in.
1312
1313Example of what happens on a structural level:
1314
486ec47a 1315The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1316
1317 1: CURLYM[1] {1,32767}(18)
1318 5: BRANCH(8)
1319 6: EXACT <ac>(16)
1320 8: BRANCH(11)
1321 9: EXACT <ad>(16)
1322 11: BRANCH(14)
1323 12: EXACT <ab>(16)
1324 16: SUCCEED(0)
1325 17: NOTHING(18)
1326 18: END(0)
1327
1328This would be optimizable with startbranch=5, first=5, last=16, tail=16
1329and should turn into:
1330
1331 1: CURLYM[1] {1,32767}(18)
1332 5: TRIE(16)
1333 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1334 <ac>
1335 <ad>
1336 <ab>
1337 16: SUCCEED(0)
1338 17: NOTHING(18)
1339 18: END(0)
1340
1341Cases where tail != last would be like /(?foo|bar)baz/:
1342
1343 1: BRANCH(4)
1344 2: EXACT <foo>(8)
1345 4: BRANCH(7)
1346 5: EXACT <bar>(8)
1347 7: TAIL(8)
1348 8: EXACT <baz>(10)
1349 10: END(0)
1350
1351which would be optimizable with startbranch=1, first=1, last=7, tail=8
1352and would end up looking like:
1353
1354 1: TRIE(8)
1355 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1356 <foo>
1357 <bar>
1358 7: TAIL(8)
1359 8: EXACT <baz>(10)
1360 10: END(0)
1361
1362 d = uvuni_to_utf8_flags(d, uv, 0);
1363
1364is the recommended Unicode-aware way of saying
1365
1366 *(d++) = uv;
1367*/
1368
1e2e3d02 1369#define TRIE_STORE_REVCHAR \
786e8c11 1370 STMT_START { \
73031816
NC
1371 if (UTF) { \
1372 SV *zlopp = newSV(2); \
88c9ea1e
CB
1373 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1374 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, uvc & 0xFF); \
73031816
NC
1375 SvCUR_set(zlopp, kapow - flrbbbbb); \
1376 SvPOK_on(zlopp); \
1377 SvUTF8_on(zlopp); \
1378 av_push(revcharmap, zlopp); \
1379 } else { \
6bdeddd2 1380 char ooooff = (char)uvc; \
73031816
NC
1381 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1382 } \
1383 } STMT_END
786e8c11
YO
1384
1385#define TRIE_READ_CHAR STMT_START { \
1386 wordlen++; \
1387 if ( UTF ) { \
1388 if ( folder ) { \
1389 if ( foldlen > 0 ) { \
1390 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1391 foldlen -= len; \
1392 scan += len; \
1393 len = 0; \
1394 } else { \
c81f2f9e
KW
1395 len = UTF8SKIP(uc);\
1396 uvc = to_utf8_fold( uc, foldbuf, &foldlen); \
786e8c11
YO
1397 foldlen -= UNISKIP( uvc ); \
1398 scan = foldbuf + UNISKIP( uvc ); \
1399 } \
1400 } else { \
1401 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1402 } \
1403 } else { \
1404 uvc = (U32)*uc; \
1405 len = 1; \
1406 } \
1407} STMT_END
1408
1409
1410
1411#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1412 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1413 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1414 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1415 } \
1416 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1417 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1418 TRIE_LIST_CUR( state )++; \
1419} STMT_END
07be1b83 1420
786e8c11
YO
1421#define TRIE_LIST_NEW(state) STMT_START { \
1422 Newxz( trie->states[ state ].trans.list, \
1423 4, reg_trie_trans_le ); \
1424 TRIE_LIST_CUR( state ) = 1; \
1425 TRIE_LIST_LEN( state ) = 4; \
1426} STMT_END
07be1b83 1427
786e8c11
YO
1428#define TRIE_HANDLE_WORD(state) STMT_START { \
1429 U16 dupe= trie->states[ state ].wordnum; \
1430 regnode * const noper_next = regnext( noper ); \
1431 \
786e8c11
YO
1432 DEBUG_r({ \
1433 /* store the word for dumping */ \
1434 SV* tmp; \
1435 if (OP(noper) != NOTHING) \
740cce10 1436 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1437 else \
740cce10 1438 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1439 av_push( trie_words, tmp ); \
786e8c11
YO
1440 }); \
1441 \
1442 curword++; \
2e64971a
DM
1443 trie->wordinfo[curword].prev = 0; \
1444 trie->wordinfo[curword].len = wordlen; \
1445 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1446 \
1447 if ( noper_next < tail ) { \
1448 if (!trie->jump) \
c944940b 1449 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1450 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1451 if (!jumper) \
1452 jumper = noper_next; \
1453 if (!nextbranch) \
1454 nextbranch= regnext(cur); \
1455 } \
1456 \
1457 if ( dupe ) { \
2e64971a
DM
1458 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1459 /* chain, so that when the bits of chain are later */\
1460 /* linked together, the dups appear in the chain */\
1461 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1462 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1463 } else { \
1464 /* we haven't inserted this word yet. */ \
1465 trie->states[ state ].wordnum = curword; \
1466 } \
1467} STMT_END
07be1b83 1468
3dab1dad 1469
786e8c11
YO
1470#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1471 ( ( base + charid >= ucharcount \
1472 && base + charid < ubound \
1473 && state == trie->trans[ base - ucharcount + charid ].check \
1474 && trie->trans[ base - ucharcount + charid ].next ) \
1475 ? trie->trans[ base - ucharcount + charid ].next \
1476 : ( state==1 ? special : 0 ) \
1477 )
3dab1dad 1478
786e8c11
YO
1479#define MADE_TRIE 1
1480#define MADE_JUMP_TRIE 2
1481#define MADE_EXACT_TRIE 4
3dab1dad 1482
a3621e74 1483STATIC I32
786e8c11 1484S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1485{
27da23d5 1486 dVAR;
a3621e74
YO
1487 /* first pass, loop through and scan words */
1488 reg_trie_data *trie;
55eed653 1489 HV *widecharmap = NULL;
2b8b4781 1490 AV *revcharmap = newAV();
a3621e74 1491 regnode *cur;
9f7f3913 1492 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1493 STRLEN len = 0;
1494 UV uvc = 0;
1495 U16 curword = 0;
1496 U32 next_alloc = 0;
786e8c11
YO
1497 regnode *jumper = NULL;
1498 regnode *nextbranch = NULL;
7f69552c 1499 regnode *convert = NULL;
2e64971a 1500 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1501 /* we just use folder as a flag in utf8 */
1e696034 1502 const U8 * folder = NULL;
a3621e74 1503
2b8b4781
NC
1504#ifdef DEBUGGING
1505 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1506 AV *trie_words = NULL;
1507 /* along with revcharmap, this only used during construction but both are
1508 * useful during debugging so we store them in the struct when debugging.
8e11feef 1509 */
2b8b4781
NC
1510#else
1511 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1512 STRLEN trie_charcount=0;
3dab1dad 1513#endif
2b8b4781 1514 SV *re_trie_maxbuff;
a3621e74 1515 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1516
1517 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1518#ifndef DEBUGGING
1519 PERL_UNUSED_ARG(depth);
1520#endif
a3621e74 1521
1e696034 1522 switch (flags) {
c46d03cf 1523 case EXACT: break;
2f7f8cb1 1524 case EXACTFA:
1e696034
KW
1525 case EXACTFU: folder = PL_fold_latin1; break;
1526 case EXACTF: folder = PL_fold; break;
1527 case EXACTFL: folder = PL_fold_locale; break;
c46d03cf 1528 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u", (unsigned) flags );
1e696034
KW
1529 }
1530
c944940b 1531 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1532 trie->refcount = 1;
3dab1dad 1533 trie->startstate = 1;
786e8c11 1534 trie->wordcount = word_count;
f8fc2ecf 1535 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1536 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
3dab1dad 1537 if (!(UTF && folder))
c944940b 1538 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1539 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1540 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1541
a3621e74 1542 DEBUG_r({
2b8b4781 1543 trie_words = newAV();
a3621e74 1544 });
a3621e74 1545
0111c4fd 1546 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1547 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1548 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1549 }
3dab1dad
YO
1550 DEBUG_OPTIMISE_r({
1551 PerlIO_printf( Perl_debug_log,
786e8c11 1552 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1553 (int)depth * 2 + 2, "",
1554 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1555 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1556 (int)depth);
3dab1dad 1557 });
7f69552c
YO
1558
1559 /* Find the node we are going to overwrite */
1560 if ( first == startbranch && OP( last ) != BRANCH ) {
1561 /* whole branch chain */
1562 convert = first;
1563 } else {
1564 /* branch sub-chain */
1565 convert = NEXTOPER( first );
1566 }
1567
a3621e74
YO
1568 /* -- First loop and Setup --
1569
1570 We first traverse the branches and scan each word to determine if it
1571 contains widechars, and how many unique chars there are, this is
1572 important as we have to build a table with at least as many columns as we
1573 have unique chars.
1574
1575 We use an array of integers to represent the character codes 0..255
38a44b82 1576 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1577 native representation of the character value as the key and IV's for the
1578 coded index.
1579
1580 *TODO* If we keep track of how many times each character is used we can
1581 remap the columns so that the table compression later on is more
3b753521 1582 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1583 middle and the least common are on the outside. IMO this would be better
1584 than a most to least common mapping as theres a decent chance the most
1585 common letter will share a node with the least common, meaning the node
486ec47a 1586 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1587 case is when we have the least common nodes twice.
1588
1589 */
1590
a3621e74 1591 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
c445ea15 1592 regnode * const noper = NEXTOPER( cur );
e1ec3a88 1593 const U8 *uc = (U8*)STRING( noper );
a28509cc 1594 const U8 * const e = uc + STR_LEN( noper );
a3621e74
YO
1595 STRLEN foldlen = 0;
1596 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2af232bd 1597 const U8 *scan = (U8*)NULL;
07be1b83 1598 U32 wordlen = 0; /* required init */
02daf0ab
YO
1599 STRLEN chars = 0;
1600 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1601
3dab1dad
YO
1602 if (OP(noper) == NOTHING) {
1603 trie->minlen= 0;
1604 continue;
1605 }
02daf0ab
YO
1606 if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
1607 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1608 regardless of encoding */
1609
a3621e74 1610 for ( ; uc < e ; uc += len ) {
3dab1dad 1611 TRIE_CHARCOUNT(trie)++;
a3621e74 1612 TRIE_READ_CHAR;
3dab1dad 1613 chars++;
a3621e74
YO
1614 if ( uvc < 256 ) {
1615 if ( !trie->charmap[ uvc ] ) {
1616 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1617 if ( folder )
1618 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
3dab1dad 1619 TRIE_STORE_REVCHAR;
a3621e74 1620 }
02daf0ab 1621 if ( set_bit ) {
62012aee
KW
1622 /* store the codepoint in the bitmap, and its folded
1623 * equivalent. */
02daf0ab 1624 TRIE_BITMAP_SET(trie,uvc);
0921ee73
T
1625
1626 /* store the folded codepoint */
1627 if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);
1628
1629 if ( !UTF ) {
1630 /* store first byte of utf8 representation of
acdf4139
KW
1631 variant codepoints */
1632 if (! UNI_IS_INVARIANT(uvc)) {
1633 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1634 }
1635 }
02daf0ab
YO
1636 set_bit = 0; /* We've done our bit :-) */
1637 }
a3621e74
YO
1638 } else {
1639 SV** svpp;
55eed653
NC
1640 if ( !widecharmap )
1641 widecharmap = newHV();
a3621e74 1642
55eed653 1643 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1644
1645 if ( !svpp )
e4584336 1646 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1647
1648 if ( !SvTRUE( *svpp ) ) {
1649 sv_setiv( *svpp, ++trie->uniquecharcount );
3dab1dad 1650 TRIE_STORE_REVCHAR;
a3621e74
YO
1651 }
1652 }
1653 }
3dab1dad
YO
1654 if( cur == first ) {
1655 trie->minlen=chars;
1656 trie->maxlen=chars;
1657 } else if (chars < trie->minlen) {
1658 trie->minlen=chars;
1659 } else if (chars > trie->maxlen) {
1660 trie->maxlen=chars;
1661 }
1662
a3621e74
YO
1663 } /* end first pass */
1664 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1665 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1666 (int)depth * 2 + 2,"",
55eed653 1667 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1668 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1669 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1670 );
a3621e74
YO
1671
1672 /*
1673 We now know what we are dealing with in terms of unique chars and
1674 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1675 representation using a flat table will take. If it's over a reasonable
1676 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1677 conservative but potentially much slower representation using an array
1678 of lists.
1679
1680 At the end we convert both representations into the same compressed
1681 form that will be used in regexec.c for matching with. The latter
1682 is a form that cannot be used to construct with but has memory
1683 properties similar to the list form and access properties similar
1684 to the table form making it both suitable for fast searches and
1685 small enough that its feasable to store for the duration of a program.
1686
1687 See the comment in the code where the compressed table is produced
1688 inplace from the flat tabe representation for an explanation of how
1689 the compression works.
1690
1691 */
1692
1693
2e64971a
DM
1694 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1695 prev_states[1] = 0;
1696
3dab1dad 1697 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1698 /*
1699 Second Pass -- Array Of Lists Representation
1700
1701 Each state will be represented by a list of charid:state records
1702 (reg_trie_trans_le) the first such element holds the CUR and LEN
1703 points of the allocated array. (See defines above).
1704
1705 We build the initial structure using the lists, and then convert
1706 it into the compressed table form which allows faster lookups
1707 (but cant be modified once converted).
a3621e74
YO
1708 */
1709
a3621e74
YO
1710 STRLEN transcount = 1;
1711
1e2e3d02
YO
1712 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1713 "%*sCompiling trie using list compiler\n",
1714 (int)depth * 2 + 2, ""));
686b73d4 1715
c944940b
JH
1716 trie->states = (reg_trie_state *)
1717 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1718 sizeof(reg_trie_state) );
a3621e74
YO
1719 TRIE_LIST_NEW(1);
1720 next_alloc = 2;
1721
1722 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1723
c445ea15
AL
1724 regnode * const noper = NEXTOPER( cur );
1725 U8 *uc = (U8*)STRING( noper );
1726 const U8 * const e = uc + STR_LEN( noper );
1727 U32 state = 1; /* required init */
1728 U16 charid = 0; /* sanity init */
1729 U8 *scan = (U8*)NULL; /* sanity init */
1730 STRLEN foldlen = 0; /* required init */
07be1b83 1731 U32 wordlen = 0; /* required init */
c445ea15
AL
1732 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1733
3dab1dad 1734 if (OP(noper) != NOTHING) {
786e8c11 1735 for ( ; uc < e ; uc += len ) {
c445ea15 1736
786e8c11 1737 TRIE_READ_CHAR;
c445ea15 1738
786e8c11
YO
1739 if ( uvc < 256 ) {
1740 charid = trie->charmap[ uvc ];
c445ea15 1741 } else {
55eed653 1742 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1743 if ( !svpp ) {
1744 charid = 0;
1745 } else {
1746 charid=(U16)SvIV( *svpp );
1747 }
c445ea15 1748 }
786e8c11
YO
1749 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1750 if ( charid ) {
a3621e74 1751
786e8c11
YO
1752 U16 check;
1753 U32 newstate = 0;
a3621e74 1754
786e8c11
YO
1755 charid--;
1756 if ( !trie->states[ state ].trans.list ) {
1757 TRIE_LIST_NEW( state );
c445ea15 1758 }
786e8c11
YO
1759 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1760 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1761 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1762 break;
1763 }
1764 }
1765 if ( ! newstate ) {
1766 newstate = next_alloc++;
2e64971a 1767 prev_states[newstate] = state;
786e8c11
YO
1768 TRIE_LIST_PUSH( state, charid, newstate );
1769 transcount++;
1770 }
1771 state = newstate;
1772 } else {
1773 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1774 }
a28509cc 1775 }
c445ea15 1776 }
3dab1dad 1777 TRIE_HANDLE_WORD(state);
a3621e74
YO
1778
1779 } /* end second pass */
1780
1e2e3d02
YO
1781 /* next alloc is the NEXT state to be allocated */
1782 trie->statecount = next_alloc;
c944940b
JH
1783 trie->states = (reg_trie_state *)
1784 PerlMemShared_realloc( trie->states,
1785 next_alloc
1786 * sizeof(reg_trie_state) );
a3621e74 1787
3dab1dad 1788 /* and now dump it out before we compress it */
2b8b4781
NC
1789 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1790 revcharmap, next_alloc,
1791 depth+1)
1e2e3d02 1792 );
a3621e74 1793
c944940b
JH
1794 trie->trans = (reg_trie_trans *)
1795 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1796 {
1797 U32 state;
a3621e74
YO
1798 U32 tp = 0;
1799 U32 zp = 0;
1800
1801
1802 for( state=1 ; state < next_alloc ; state ++ ) {
1803 U32 base=0;
1804
1805 /*
1806 DEBUG_TRIE_COMPILE_MORE_r(
1807 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1808 );
1809 */
1810
1811 if (trie->states[state].trans.list) {
1812 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1813 U16 maxid=minid;
a28509cc 1814 U16 idx;
a3621e74
YO
1815
1816 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1817 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1818 if ( forid < minid ) {
1819 minid=forid;
1820 } else if ( forid > maxid ) {
1821 maxid=forid;
1822 }
a3621e74
YO
1823 }
1824 if ( transcount < tp + maxid - minid + 1) {
1825 transcount *= 2;
c944940b
JH
1826 trie->trans = (reg_trie_trans *)
1827 PerlMemShared_realloc( trie->trans,
446bd890
NC
1828 transcount
1829 * sizeof(reg_trie_trans) );
a3621e74
YO
1830 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1831 }
1832 base = trie->uniquecharcount + tp - minid;
1833 if ( maxid == minid ) {
1834 U32 set = 0;
1835 for ( ; zp < tp ; zp++ ) {
1836 if ( ! trie->trans[ zp ].next ) {
1837 base = trie->uniquecharcount + zp - minid;
1838 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1839 trie->trans[ zp ].check = state;
1840 set = 1;
1841 break;
1842 }
1843 }
1844 if ( !set ) {
1845 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1846 trie->trans[ tp ].check = state;
1847 tp++;
1848 zp = tp;
1849 }
1850 } else {
1851 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1852 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1853 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1854 trie->trans[ tid ].check = state;
1855 }
1856 tp += ( maxid - minid + 1 );
1857 }
1858 Safefree(trie->states[ state ].trans.list);
1859 }
1860 /*
1861 DEBUG_TRIE_COMPILE_MORE_r(
1862 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1863 );
1864 */
1865 trie->states[ state ].trans.base=base;
1866 }
cc601c31 1867 trie->lasttrans = tp + 1;
a3621e74
YO
1868 }
1869 } else {
1870 /*
1871 Second Pass -- Flat Table Representation.
1872
1873 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1874 We know that we will need Charcount+1 trans at most to store the data
1875 (one row per char at worst case) So we preallocate both structures
1876 assuming worst case.
1877
1878 We then construct the trie using only the .next slots of the entry
1879 structs.
1880
3b753521 1881 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1882 make compression both faster and easier by keeping track of how many non
1883 zero fields are in the node.
1884
1885 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1886 transition.
1887
1888 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1889 number representing the first entry of the node, and state as a
1890 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1891 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1892 are 2 entrys per node. eg:
1893
1894 A B A B
1895 1. 2 4 1. 3 7
1896 2. 0 3 3. 0 5
1897 3. 0 0 5. 0 0
1898 4. 0 0 7. 0 0
1899
1900 The table is internally in the right hand, idx form. However as we also
1901 have to deal with the states array which is indexed by nodenum we have to
1902 use TRIE_NODENUM() to convert.
1903
1904 */
1e2e3d02
YO
1905 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1906 "%*sCompiling trie using table compiler\n",
1907 (int)depth * 2 + 2, ""));
3dab1dad 1908
c944940b
JH
1909 trie->trans = (reg_trie_trans *)
1910 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1911 * trie->uniquecharcount + 1,
1912 sizeof(reg_trie_trans) );
1913 trie->states = (reg_trie_state *)
1914 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1915 sizeof(reg_trie_state) );
a3621e74
YO
1916 next_alloc = trie->uniquecharcount + 1;
1917
3dab1dad 1918
a3621e74
YO
1919 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1920
c445ea15 1921 regnode * const noper = NEXTOPER( cur );
a28509cc
AL
1922 const U8 *uc = (U8*)STRING( noper );
1923 const U8 * const e = uc + STR_LEN( noper );
a3621e74
YO
1924
1925 U32 state = 1; /* required init */
1926
1927 U16 charid = 0; /* sanity init */
1928 U32 accept_state = 0; /* sanity init */
1929 U8 *scan = (U8*)NULL; /* sanity init */
1930
1931 STRLEN foldlen = 0; /* required init */
07be1b83 1932 U32 wordlen = 0; /* required init */
a3621e74
YO
1933 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1934
3dab1dad 1935 if ( OP(noper) != NOTHING ) {
786e8c11 1936 for ( ; uc < e ; uc += len ) {
a3621e74 1937
786e8c11 1938 TRIE_READ_CHAR;
a3621e74 1939
786e8c11
YO
1940 if ( uvc < 256 ) {
1941 charid = trie->charmap[ uvc ];
1942 } else {
55eed653 1943 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 1944 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 1945 }
786e8c11
YO
1946 if ( charid ) {
1947 charid--;
1948 if ( !trie->trans[ state + charid ].next ) {
1949 trie->trans[ state + charid ].next = next_alloc;
1950 trie->trans[ state ].check++;
2e64971a
DM
1951 prev_states[TRIE_NODENUM(next_alloc)]
1952 = TRIE_NODENUM(state);
786e8c11
YO
1953 next_alloc += trie->uniquecharcount;
1954 }
1955 state = trie->trans[ state + charid ].next;
1956 } else {
1957 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1958 }
1959 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 1960 }
a3621e74 1961 }
3dab1dad
YO
1962 accept_state = TRIE_NODENUM( state );
1963 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
1964
1965 } /* end second pass */
1966
3dab1dad 1967 /* and now dump it out before we compress it */
2b8b4781
NC
1968 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1969 revcharmap,
1970 next_alloc, depth+1));
a3621e74 1971
a3621e74
YO
1972 {
1973 /*
1974 * Inplace compress the table.*
1975
1976 For sparse data sets the table constructed by the trie algorithm will
1977 be mostly 0/FAIL transitions or to put it another way mostly empty.
1978 (Note that leaf nodes will not contain any transitions.)
1979
1980 This algorithm compresses the tables by eliminating most such
1981 transitions, at the cost of a modest bit of extra work during lookup:
1982
1983 - Each states[] entry contains a .base field which indicates the
1984 index in the state[] array wheres its transition data is stored.
1985
3b753521 1986 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
1987
1988 - If .base is nonzero then charid is added to it to find an entry in
1989 the trans array.
1990
1991 -If trans[states[state].base+charid].check!=state then the
1992 transition is taken to be a 0/Fail transition. Thus if there are fail
1993 transitions at the front of the node then the .base offset will point
1994 somewhere inside the previous nodes data (or maybe even into a node
1995 even earlier), but the .check field determines if the transition is
1996 valid.
1997
786e8c11 1998 XXX - wrong maybe?
a3621e74 1999 The following process inplace converts the table to the compressed
3b753521 2000 table: We first do not compress the root node 1,and mark all its
a3621e74 2001 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2002 allows us to do a DFA construction from the compressed table later,
2003 and ensures that any .base pointers we calculate later are greater
2004 than 0.
a3621e74
YO
2005
2006 - We set 'pos' to indicate the first entry of the second node.
2007
2008 - We then iterate over the columns of the node, finding the first and
2009 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2010 and set the .check pointers accordingly, and advance pos
2011 appropriately and repreat for the next node. Note that when we copy
2012 the next pointers we have to convert them from the original
2013 NODEIDX form to NODENUM form as the former is not valid post
2014 compression.
2015
2016 - If a node has no transitions used we mark its base as 0 and do not
2017 advance the pos pointer.
2018
2019 - If a node only has one transition we use a second pointer into the
2020 structure to fill in allocated fail transitions from other states.
2021 This pointer is independent of the main pointer and scans forward
2022 looking for null transitions that are allocated to a state. When it
2023 finds one it writes the single transition into the "hole". If the
786e8c11 2024 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2025
2026 - Once compressed we can Renew/realloc the structures to release the
2027 excess space.
2028
2029 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2030 specifically Fig 3.47 and the associated pseudocode.
2031
2032 demq
2033 */
a3b680e6 2034 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2035 U32 state, charid;
a3621e74 2036 U32 pos = 0, zp=0;
1e2e3d02 2037 trie->statecount = laststate;
a3621e74
YO
2038
2039 for ( state = 1 ; state < laststate ; state++ ) {
2040 U8 flag = 0;
a28509cc
AL
2041 const U32 stateidx = TRIE_NODEIDX( state );
2042 const U32 o_used = trie->trans[ stateidx ].check;
2043 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2044 trie->trans[ stateidx ].check = 0;
2045
2046 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2047 if ( flag || trie->trans[ stateidx + charid ].next ) {
2048 if ( trie->trans[ stateidx + charid ].next ) {
2049 if (o_used == 1) {
2050 for ( ; zp < pos ; zp++ ) {
2051 if ( ! trie->trans[ zp ].next ) {
2052 break;
2053 }
2054 }
2055 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2056 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2057 trie->trans[ zp ].check = state;
2058 if ( ++zp > pos ) pos = zp;
2059 break;
2060 }
2061 used--;
2062 }
2063 if ( !flag ) {
2064 flag = 1;
2065 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2066 }
2067 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2068 trie->trans[ pos ].check = state;
2069 pos++;
2070 }
2071 }
2072 }
cc601c31 2073 trie->lasttrans = pos + 1;
c944940b
JH
2074 trie->states = (reg_trie_state *)
2075 PerlMemShared_realloc( trie->states, laststate
2076 * sizeof(reg_trie_state) );
a3621e74 2077 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2078 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2079 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2080 (int)depth * 2 + 2,"",
2081 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2082 (IV)next_alloc,
2083 (IV)pos,
a3621e74
YO
2084 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2085 );
2086
2087 } /* end table compress */
2088 }
1e2e3d02
YO
2089 DEBUG_TRIE_COMPILE_MORE_r(
2090 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2091 (int)depth * 2 + 2, "",
2092 (UV)trie->statecount,
2093 (UV)trie->lasttrans)
2094 );
cc601c31 2095 /* resize the trans array to remove unused space */
c944940b
JH
2096 trie->trans = (reg_trie_trans *)
2097 PerlMemShared_realloc( trie->trans, trie->lasttrans
2098 * sizeof(reg_trie_trans) );
a3621e74 2099
3b753521 2100 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2101 U8 nodetype =(U8)(flags & 0xFF);
2102 char *str=NULL;
786e8c11 2103
07be1b83 2104#ifdef DEBUGGING
e62cc96a 2105 regnode *optimize = NULL;
7122b237
YO
2106#ifdef RE_TRACK_PATTERN_OFFSETS
2107
b57a0404
JH
2108 U32 mjd_offset = 0;
2109 U32 mjd_nodelen = 0;
7122b237
YO
2110#endif /* RE_TRACK_PATTERN_OFFSETS */
2111#endif /* DEBUGGING */
a3621e74 2112 /*
3dab1dad
YO
2113 This means we convert either the first branch or the first Exact,
2114 depending on whether the thing following (in 'last') is a branch
2115 or not and whther first is the startbranch (ie is it a sub part of
2116 the alternation or is it the whole thing.)
3b753521 2117 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2118 the whole branch sequence, including the first.
a3621e74 2119 */
3dab1dad 2120 /* Find the node we are going to overwrite */
7f69552c 2121 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2122 /* branch sub-chain */
3dab1dad 2123 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2124#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2125 DEBUG_r({
2126 mjd_offset= Node_Offset((convert));
2127 mjd_nodelen= Node_Length((convert));
2128 });
7122b237 2129#endif
7f69552c 2130 /* whole branch chain */
7122b237
YO
2131 }
2132#ifdef RE_TRACK_PATTERN_OFFSETS
2133 else {
7f69552c
YO
2134 DEBUG_r({
2135 const regnode *nop = NEXTOPER( convert );
2136 mjd_offset= Node_Offset((nop));
2137 mjd_nodelen= Node_Length((nop));
2138 });
07be1b83
YO
2139 }
2140 DEBUG_OPTIMISE_r(
2141 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2142 (int)depth * 2 + 2, "",
786e8c11 2143 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2144 );
7122b237 2145#endif
3dab1dad
YO
2146 /* But first we check to see if there is a common prefix we can
2147 split out as an EXACT and put in front of the TRIE node. */
2148 trie->startstate= 1;
55eed653 2149 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2150 U32 state;
1e2e3d02 2151 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2152 U32 ofs = 0;
8e11feef
RGS
2153 I32 idx = -1;
2154 U32 count = 0;
2155 const U32 base = trie->states[ state ].trans.base;
a3621e74 2156
3dab1dad 2157 if ( trie->states[state].wordnum )
8e11feef 2158 count = 1;
a3621e74 2159
8e11feef 2160 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2161 if ( ( base + ofs >= trie->uniquecharcount ) &&
2162 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2163 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2164 {
3dab1dad 2165 if ( ++count > 1 ) {
2b8b4781 2166 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2167 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2168 if ( state == 1 ) break;
3dab1dad
YO
2169 if ( count == 2 ) {
2170 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2171 DEBUG_OPTIMISE_r(
8e11feef
RGS
2172 PerlIO_printf(Perl_debug_log,
2173 "%*sNew Start State=%"UVuf" Class: [",
2174 (int)depth * 2 + 2, "",
786e8c11 2175 (UV)state));
be8e71aa 2176 if (idx >= 0) {
2b8b4781 2177 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2178 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2179
3dab1dad 2180 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2181 if ( folder )
2182 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2183 DEBUG_OPTIMISE_r(
f1f66076 2184 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2185 );
8e11feef
RGS
2186 }
2187 }
2188 TRIE_BITMAP_SET(trie,*ch);
2189 if ( folder )
2190 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2191 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2192 }
2193 idx = ofs;
2194 }
3dab1dad
YO
2195 }
2196 if ( count == 1 ) {
2b8b4781 2197 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2198 STRLEN len;
2199 char *ch = SvPV( *tmp, len );
de734bd5
A
2200 DEBUG_OPTIMISE_r({
2201 SV *sv=sv_newmortal();
8e11feef
RGS
2202 PerlIO_printf( Perl_debug_log,
2203 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2204 (int)depth * 2 + 2, "",
de734bd5
A
2205 (UV)state, (UV)idx,
2206 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2207 PL_colors[0], PL_colors[1],
2208 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2209 PERL_PV_ESCAPE_FIRSTCHAR
2210 )
2211 );
2212 });
3dab1dad
YO
2213 if ( state==1 ) {
2214 OP( convert ) = nodetype;
2215 str=STRING(convert);
2216 STR_LEN(convert)=0;
2217 }
c490c714
YO
2218 STR_LEN(convert) += len;
2219 while (len--)
de734bd5 2220 *str++ = *ch++;
8e11feef 2221 } else {
f9049ba1 2222#ifdef DEBUGGING
8e11feef
RGS
2223 if (state>1)
2224 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2225#endif
8e11feef
RGS
2226 break;
2227 }
2228 }
2e64971a 2229 trie->prefixlen = (state-1);
3dab1dad 2230 if (str) {
8e11feef 2231 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2232 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2233 trie->startstate = state;
07be1b83
YO
2234 trie->minlen -= (state - 1);
2235 trie->maxlen -= (state - 1);
33809eae
JH
2236#ifdef DEBUGGING
2237 /* At least the UNICOS C compiler choked on this
2238 * being argument to DEBUG_r(), so let's just have
2239 * it right here. */
2240 if (
2241#ifdef PERL_EXT_RE_BUILD
2242 1
2243#else
2244 DEBUG_r_TEST
2245#endif
2246 ) {
2247 regnode *fix = convert;
2248 U32 word = trie->wordcount;
2249 mjd_nodelen++;
2250 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2251 while( ++fix < n ) {
2252 Set_Node_Offset_Length(fix, 0, 0);
2253 }
2254 while (word--) {
2255 SV ** const tmp = av_fetch( trie_words, word, 0 );
2256 if (tmp) {
2257 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2258 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2259 else
2260 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2261 }
2262 }
2263 }
2264#endif
8e11feef
RGS
2265 if (trie->maxlen) {
2266 convert = n;
2267 } else {
3dab1dad 2268 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2269 DEBUG_r(optimize= n);
3dab1dad
YO
2270 }
2271 }
2272 }
a5ca303d
YO
2273 if (!jumper)
2274 jumper = last;
3dab1dad 2275 if ( trie->maxlen ) {
8e11feef
RGS
2276 NEXT_OFF( convert ) = (U16)(tail - convert);
2277 ARG_SET( convert, data_slot );
786e8c11
YO
2278 /* Store the offset to the first unabsorbed branch in
2279 jump[0], which is otherwise unused by the jump logic.
2280 We use this when dumping a trie and during optimisation. */
2281 if (trie->jump)
7f69552c 2282 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2283
6c48061a
YO
2284 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2285 * and there is a bitmap
2286 * and the first "jump target" node we found leaves enough room
2287 * then convert the TRIE node into a TRIEC node, with the bitmap
2288 * embedded inline in the opcode - this is hypothetically faster.
2289 */
2290 if ( !trie->states[trie->startstate].wordnum
2291 && trie->bitmap
2292 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2293 {
2294 OP( convert ) = TRIEC;
2295 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2296 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2297 trie->bitmap= NULL;
2298 } else
2299 OP( convert ) = TRIE;
a3621e74 2300
3dab1dad
YO
2301 /* store the type in the flags */
2302 convert->flags = nodetype;
a5ca303d
YO
2303 DEBUG_r({
2304 optimize = convert
2305 + NODE_STEP_REGNODE
2306 + regarglen[ OP( convert ) ];
2307 });
2308 /* XXX We really should free up the resource in trie now,
2309 as we won't use them - (which resources?) dmq */
3dab1dad 2310 }
a3621e74 2311 /* needed for dumping*/
e62cc96a 2312 DEBUG_r(if (optimize) {
07be1b83 2313 regnode *opt = convert;
bcdf7404 2314
e62cc96a 2315 while ( ++opt < optimize) {
07be1b83
YO
2316 Set_Node_Offset_Length(opt,0,0);
2317 }
786e8c11
YO
2318 /*
2319 Try to clean up some of the debris left after the
2320 optimisation.
a3621e74 2321 */
786e8c11 2322 while( optimize < jumper ) {
07be1b83 2323 mjd_nodelen += Node_Length((optimize));
a3621e74 2324 OP( optimize ) = OPTIMIZED;
07be1b83 2325 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2326 optimize++;
2327 }
07be1b83 2328 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2329 });
2330 } /* end node insert */
2e64971a
DM
2331
2332 /* Finish populating the prev field of the wordinfo array. Walk back
2333 * from each accept state until we find another accept state, and if
2334 * so, point the first word's .prev field at the second word. If the
2335 * second already has a .prev field set, stop now. This will be the
2336 * case either if we've already processed that word's accept state,
3b753521
FN
2337 * or that state had multiple words, and the overspill words were
2338 * already linked up earlier.
2e64971a
DM
2339 */
2340 {
2341 U16 word;
2342 U32 state;
2343 U16 prev;
2344
2345 for (word=1; word <= trie->wordcount; word++) {
2346 prev = 0;
2347 if (trie->wordinfo[word].prev)
2348 continue;
2349 state = trie->wordinfo[word].accept;
2350 while (state) {
2351 state = prev_states[state];
2352 if (!state)
2353 break;
2354 prev = trie->states[state].wordnum;
2355 if (prev)
2356 break;
2357 }
2358 trie->wordinfo[word].prev = prev;
2359 }
2360 Safefree(prev_states);
2361 }
2362
2363
2364 /* and now dump out the compressed format */
2365 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2366
55eed653 2367 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2368#ifdef DEBUGGING
2369 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2370 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2371#else
2372 SvREFCNT_dec(revcharmap);
07be1b83 2373#endif
786e8c11
YO
2374 return trie->jump
2375 ? MADE_JUMP_TRIE
2376 : trie->startstate>1
2377 ? MADE_EXACT_TRIE
2378 : MADE_TRIE;
2379}
2380
2381STATIC void
2382S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2383{
3b753521 2384/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2385
2386 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2387 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2388 ISBN 0-201-10088-6
2389
2390 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2391 suffix of the current state's 'word' that is also a proper prefix of another word in our
2392 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2393 the DFA not to have to restart after its tried and failed a word at a given point, it
2394 simply continues as though it had been matching the other word in the first place.
2395 Consider
2396 'abcdgu'=~/abcdefg|cdgu/
2397 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2398 fail, which would bring us to the state representing 'd' in the second word where we would
2399 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2400 */
2401 /* add a fail transition */
3251b653
NC
2402 const U32 trie_offset = ARG(source);
2403 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2404 U32 *q;
2405 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2406 const U32 numstates = trie->statecount;
786e8c11
YO
2407 const U32 ubound = trie->lasttrans + ucharcount;
2408 U32 q_read = 0;
2409 U32 q_write = 0;
2410 U32 charid;
2411 U32 base = trie->states[ 1 ].trans.base;
2412 U32 *fail;
2413 reg_ac_data *aho;
2414 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2415 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2416
2417 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2418#ifndef DEBUGGING
2419 PERL_UNUSED_ARG(depth);
2420#endif
2421
2422
2423 ARG_SET( stclass, data_slot );
c944940b 2424 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2425 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2426 aho->trie=trie_offset;
446bd890
NC
2427 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2428 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2429 Newxz( q, numstates, U32);
c944940b 2430 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2431 aho->refcount = 1;
2432 fail = aho->fail;
2433 /* initialize fail[0..1] to be 1 so that we always have
2434 a valid final fail state */
2435 fail[ 0 ] = fail[ 1 ] = 1;
2436
2437 for ( charid = 0; charid < ucharcount ; charid++ ) {
2438 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2439 if ( newstate ) {
2440 q[ q_write ] = newstate;
2441 /* set to point at the root */
2442 fail[ q[ q_write++ ] ]=1;
2443 }
2444 }
2445 while ( q_read < q_write) {
2446 const U32 cur = q[ q_read++ % numstates ];
2447 base = trie->states[ cur ].trans.base;
2448
2449 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2450 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2451 if (ch_state) {
2452 U32 fail_state = cur;
2453 U32 fail_base;
2454 do {
2455 fail_state = fail[ fail_state ];
2456 fail_base = aho->states[ fail_state ].trans.base;
2457 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2458
2459 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2460 fail[ ch_state ] = fail_state;
2461 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2462 {
2463 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2464 }
2465 q[ q_write++ % numstates] = ch_state;
2466 }
2467 }
2468 }
2469 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2470 when we fail in state 1, this allows us to use the
2471 charclass scan to find a valid start char. This is based on the principle
2472 that theres a good chance the string being searched contains lots of stuff
2473 that cant be a start char.
2474 */
2475 fail[ 0 ] = fail[ 1 ] = 0;
2476 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2477 PerlIO_printf(Perl_debug_log,
2478 "%*sStclass Failtable (%"UVuf" states): 0",
2479 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2480 );
786e8c11
YO
2481 for( q_read=1; q_read<numstates; q_read++ ) {
2482 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2483 }
2484 PerlIO_printf(Perl_debug_log, "\n");
2485 });
2486 Safefree(q);
2487 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2488}
2489
786e8c11 2490
a3621e74 2491/*
5d1c421c
JH
2492 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2493 * These need to be revisited when a newer toolchain becomes available.
2494 */
2495#if defined(__sparc64__) && defined(__GNUC__)
2496# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2497# undef SPARC64_GCC_WORKAROUND
2498# define SPARC64_GCC_WORKAROUND 1
2499# endif
2500#endif
2501
07be1b83 2502#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2503 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2504 SV * const mysv=sv_newmortal(); \
2505 regnode *Next = regnext(scan); \
2506 regprop(RExC_rx, mysv, scan); \
7f69552c 2507 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2508 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2509 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2510 }});
07be1b83 2511
1de06328 2512
bb914485
KW
2513/* The below joins as many adjacent EXACTish nodes as possible into a single
2514 * one, and looks for problematic sequences of characters whose folds vs.
2515 * non-folds have sufficiently different lengths, that the optimizer would be
2516 * fooled into rejecting legitimate matches of them, and the trie construction
2517 * code can't cope with them. The joining is only done if:
2518 * 1) there is room in the current conglomerated node to entirely contain the
2519 * next one.
2520 * 2) they are the exact same node type
2521 *
2522 * The adjacent nodes actually may be separated by NOTHING kind nodes, and
2523 * these get optimized out
2524 *
9d071ca8
KW
2525 * If there are problematic code sequences, *min_subtract is set to the delta
2526 * that the minimum size of the node can be less than its actual size. And,
2527 * the node type of the result is changed to reflect that it contains these
bb914485
KW
2528 * sequences.
2529 *
a0c4c608
KW
2530 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2531 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2532 *
bb914485
KW
2533 * This is as good a place as any to discuss the design of handling these
2534 * problematic sequences. It's been wrong in Perl for a very long time. There
2535 * are three code points in Unicode whose folded lengths differ so much from
2536 * the un-folded lengths that it causes problems for the optimizer and trie
2537 * construction. Why only these are problematic, and not others where lengths
2538 * also differ is something I (khw) do not understand. New versions of Unicode
2539 * might add more such code points. Hopefully the logic in fold_grind.t that
287722f3 2540 * figures out what to test (in part by verifying that each size-combination
bb914485 2541 * gets tested) will catch any that do come along, so they can be added to the
287722f3
KW
2542 * special handling below. The chances of new ones are actually rather small,
2543 * as most, if not all, of the world's scripts that have casefolding have
2544 * already been encoded by Unicode. Also, a number of Unicode's decisions were
2545 * made to allow compatibility with pre-existing standards, and almost all of
2546 * those have already been dealt with. These would otherwise be the most
2547 * likely candidates for generating further tricky sequences. In other words,
2548 * Unicode by itself is unlikely to add new ones unless it is for compatibility
a0c4c608 2549 * with pre-existing standards, and there aren't many of those left.
bb914485
KW
2550 *
2551 * The previous designs for dealing with these involved assigning a special
2552 * node for them. This approach doesn't work, as evidenced by this example:
a0c4c608 2553 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
bb914485
KW
2554 * Both these fold to "sss", but if the pattern is parsed to create a node of
2555 * that would match just the \xDF, it won't be able to handle the case where a
2556 * successful match would have to cross the node's boundary. The new approach
2557 * that hopefully generally solves the problem generates an EXACTFU_SS node
2558 * that is "sss".
2559 *
2560 * There are a number of components to the approach (a lot of work for just
2561 * three code points!):
2562 * 1) This routine examines each EXACTFish node that could contain the
9d071ca8
KW
2563 * problematic sequences. It returns in *min_subtract how much to
2564 * subtract from the the actual length of the string to get a real minimum
2565 * for one that could match it. This number is usually 0 except for the
2566 * problematic sequences. This delta is used by the caller to adjust the
2567 * min length of the match, and the delta between min and max, so that the
2568 * optimizer doesn't reject these possibilities based on size constraints.
bb914485
KW
2569 * 2) These sequences are not currently correctly handled by the trie code
2570 * either, so it changes the joined node type to ops that are not handled
2571 * by trie's, those new ops being EXACTFU_SS and EXACTFU_NO_TRIE.
2572 * 3) This is sufficient for the two Greek sequences (described below), but
2573 * the one involving the Sharp s (\xDF) needs more. The node type
2574 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2575 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2576 * case where there is a possible fold length change. That means that a
2577 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2578 * itself with length changes, and so can be processed faster. regexec.c
2579 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2580 * is pre-folded by regcomp.c. This saves effort in regex matching.
2581 * However, probably mostly for historical reasons, the pre-folding isn't
a0c4c608
KW
2582 * done for non-UTF8 patterns (and it can't be for EXACTF and EXACTFL
2583 * nodes, as what they fold to isn't known until runtime.) The fold
2584 * possibilities for the non-UTF8 patterns are quite simple, except for
2585 * the sharp s. All the ones that don't involve a UTF-8 target string
2586 * are members of a fold-pair, and arrays are set up for all of them
2587 * that quickly find the other member of the pair. It might actually
2588 * be faster to pre-fold these, but it isn't currently done, except for
2589 * the sharp s. Code elsewhere in this file makes sure that it gets
2590 * folded to 'ss', even if the pattern isn't UTF-8. This avoids the
2591 * issues described in the next item.
bb914485
KW
2592 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2593 * 'ss' or not is not knowable at compile time. It will match iff the
2594 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2595 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2596 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2597 * described in item 3). An assumption that the optimizer part of
2598 * regexec.c (probably unwittingly) makes is that a character in the
2599 * pattern corresponds to at most a single character in the target string.
2600 * (And I do mean character, and not byte here, unlike other parts of the
2601 * documentation that have never been updated to account for multibyte
2602 * Unicode.) This assumption is wrong only in this case, as all other
2603 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2604 * virtue of having this file pre-fold UTF-8 patterns. I'm
2605 * reluctant to try to change this assumption, so instead the code punts.
9d071ca8
KW
2606 * This routine examines EXACTF nodes for the sharp s, and returns a
2607 * boolean indicating whether or not the node is an EXACTF node that
2608 * contains a sharp s. When it is true, the caller sets a flag that later
2609 * causes the optimizer in this file to not set values for the floating
2610 * and fixed string lengths, and thus avoids the optimizer code in
2611 * regexec.c that makes the invalid assumption. Thus, there is no
2612 * optimization based on string lengths for EXACTF nodes that contain the
2613 * sharp s. This only happens for /id rules (which means the pattern
2614 * isn't in UTF-8).
bb914485 2615 */
1de06328 2616
9d071ca8 2617#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2618 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2619 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2620
be8e71aa 2621STATIC U32
9d071ca8 2622S_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
2623 /* Merge several consecutive EXACTish nodes into one. */
2624 regnode *n = regnext(scan);
2625 U32 stringok = 1;
2626 regnode *next = scan + NODE_SZ_STR(scan);
2627 U32 merged = 0;
2628 U32 stopnow = 0;
2629#ifdef DEBUGGING
2630 regnode *stop = scan;
72f13be8 2631 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2632#else
d47053eb
RGS
2633 PERL_UNUSED_ARG(depth);
2634#endif
7918f24d
NC
2635
2636 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2637#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2638 PERL_UNUSED_ARG(flags);
2639 PERL_UNUSED_ARG(val);
07be1b83 2640#endif
07be1b83 2641 DEBUG_PEEP("join",scan,depth);
bb914485 2642
3f410cf6
KW
2643 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2644 * EXACT ones that are mergeable to the current one. */
2645 while (n
2646 && (PL_regkind[OP(n)] == NOTHING
2647 || (stringok && OP(n) == OP(scan)))
07be1b83 2648 && NEXT_OFF(n)
3f410cf6
KW
2649 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2650 {
07be1b83
YO
2651
2652 if (OP(n) == TAIL || n > next)
2653 stringok = 0;
2654 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2655 DEBUG_PEEP("skip:",n,depth);
2656 NEXT_OFF(scan) += NEXT_OFF(n);
2657 next = n + NODE_STEP_REGNODE;
2658#ifdef DEBUGGING
2659 if (stringok)
2660 stop = n;
2661#endif
2662 n = regnext(n);
2663 }
2664 else if (stringok) {
786e8c11 2665 const unsigned int oldl = STR_LEN(scan);
07be1b83 2666 regnode * const nnext = regnext(n);
b2230d39
KW
2667
2668 if (oldl + STR_LEN(n) > U8_MAX)
2669 break;
07be1b83
YO
2670
2671 DEBUG_PEEP("merg",n,depth);
07be1b83 2672 merged++;
b2230d39 2673
07be1b83
YO
2674 NEXT_OFF(scan) += NEXT_OFF(n);
2675 STR_LEN(scan) += STR_LEN(n);
2676 next = n + NODE_SZ_STR(n);
2677 /* Now we can overwrite *n : */
2678 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2679#ifdef DEBUGGING
2680 stop = next - 1;
2681#endif
2682 n = nnext;
2683 if (stopnow) break;
2684 }
2685
d47053eb
RGS
2686#ifdef EXPERIMENTAL_INPLACESCAN
2687 if (flags && !NEXT_OFF(n)) {
2688 DEBUG_PEEP("atch", val, depth);
2689 if (reg_off_by_arg[OP(n)]) {
2690 ARG_SET(n, val - n);
2691 }
2692 else {
2693 NEXT_OFF(n) = val - n;
2694 }
2695 stopnow = 1;
2696 }
07be1b83
YO
2697#endif
2698 }
2c2b7f86 2699
9d071ca8 2700 *min_subtract = 0;
f758bddf 2701 *has_exactf_sharp_s = FALSE;
f646642f 2702
3f410cf6
KW
2703 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2704 * can now analyze for sequences of problematic code points. (Prior to
2705 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2706 * hence missed). The sequences only happen in folding, hence for any
2707 * non-EXACT EXACTish node */
86d6fcad 2708 if (OP(scan) != EXACT) {
f758bddf
KW
2709 U8 *s;
2710 U8 * s0 = (U8*) STRING(scan);
2711 U8 * const s_end = s0 + STR_LEN(scan);
2712
2713 /* The below is perhaps overboard, but this allows us to save a test
2714 * each time through the loop at the expense of a mask. This is
2715 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2716 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2717 * This uses an exclusive 'or' to find that bit and then inverts it to
2718 * form a mask, with just a single 0, in the bit position where 'S' and
2719 * 's' differ. */
dbeb8947 2720 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
f758bddf
KW
2721 const U8 s_masked = 's' & S_or_s_mask;
2722
2723 /* One pass is made over the node's string looking for all the
2724 * possibilities. to avoid some tests in the loop, there are two main
2725 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2726 * non-UTF-8 */
2727 if (UTF) {
86d6fcad 2728
f758bddf
KW
2729 /* There are two problematic Greek code points in Unicode
2730 * casefolding
86d6fcad
KW
2731 *
2732 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2733 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2734 *
2735 * which casefold to
2736 *
2737 * Unicode UTF-8
2738 *
2739 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2740 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2741 *
2742 * This means that in case-insensitive matching (or "loose
2743 * matching", as Unicode calls it), an EXACTF of length six (the
2744 * UTF-8 encoded byte length of the above casefolded versions) can
2745 * match a target string of length two (the byte length of UTF-8
2746 * encoded U+0390 or U+03B0). This would rather mess up the
2747 * minimum length computation. (there are other code points that
2748 * also fold to these two sequences, but the delta is smaller)
2749 *
f758bddf
KW
2750 * If these sequences are found, the minimum length is decreased by
2751 * four (six minus two).
86d6fcad 2752 *
f758bddf
KW
2753 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2754 * LETTER SHARP S. We decrease the min length by 1 for each
2755 * occurrence of 'ss' found */
3f410cf6 2756
e294cc5d 2757#ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
f758bddf
KW
2758# define U390_first_byte 0xb4
2759 const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
2760# define U3B0_first_byte 0xb5
2761 const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
e294cc5d 2762#else
f758bddf
KW
2763# define U390_first_byte 0xce
2764 const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
2765# define U3B0_first_byte 0xcf
2766 const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
e294cc5d 2767#endif
f758bddf
KW
2768 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2769 yields a net of 0 */
2770 /* Examine the string for one of the problematic sequences */
2771 for (s = s0;
2772 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2773 * sequence we are looking for is 2 */
2774 s += UTF8SKIP(s))
86d6fcad 2775 {
bb914485 2776
f758bddf
KW
2777 /* Look for the first byte in each problematic sequence */
2778 switch (*s) {
2779 /* We don't have to worry about other things that fold to
2780 * 's' (such as the long s, U+017F), as all above-latin1
2781 * code points have been pre-folded */
2782 case 's':
2783 case 'S':
2784
a0c4c608
KW
2785 /* Current character is an 's' or 'S'. If next one is
2786 * as well, we have the dreaded sequence */
f758bddf
KW
2787 if (((*(s+1) & S_or_s_mask) == s_masked)
2788 /* These two node types don't have special handling
2789 * for 'ss' */
2790 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2791 {
9d071ca8 2792 *min_subtract += 1;
f758bddf
KW
2793 OP(scan) = EXACTFU_SS;
2794 s++; /* No need to look at this character again */
2795 }
2796 break;
2797
2798 case U390_first_byte:
2799 if (s_end - s >= len
2800
2801 /* The 1's are because are skipping comparing the
2802 * first byte */
2803 && memEQ(s + 1, U390_tail, len - 1))
2804 {
2805 goto greek_sequence;
2806 }
2807 break;
2808
2809 case U3B0_first_byte:
2810 if (! (s_end - s >= len
2811 && memEQ(s + 1, U3B0_tail, len - 1)))
2812 {
2813 break;
2814 }
2815 greek_sequence:
9d071ca8 2816 *min_subtract += 4;
f758bddf
KW
2817
2818 /* This can't currently be handled by trie's, so change
2819 * the node type to indicate this. If EXACTFA and
2820 * EXACTFL were ever to be handled by trie's, this
2821 * would have to be changed. If this node has already
2822 * been changed to EXACTFU_SS in this loop, leave it as
2823 * is. (I (khw) think it doesn't matter in regexec.c
2824 * for UTF patterns, but no need to change it */
2825 if (OP(scan) == EXACTFU) {
2826 OP(scan) = EXACTFU_NO_TRIE;
2827 }
2828 s += 6; /* We already know what this sequence is. Skip
2829 the rest of it */
2830 break;
bb914485
KW
2831 }
2832 }
2833 }
f758bddf 2834 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
bb914485 2835
f758bddf
KW
2836 /* Here, the pattern is not UTF-8. We need to look only for the
2837 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2838 * in the final position. Otherwise we can stop looking 1 byte
2839 * earlier because have to find both the first and second 's' */
2840 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2841
2842 for (s = s0; s < upper; s++) {
2843 switch (*s) {
2844 case 'S':
2845 case 's':
2846 if (s_end - s > 1
2847 && ((*(s+1) & S_or_s_mask) == s_masked))
2848 {
9d071ca8 2849 *min_subtract += 1;
f758bddf
KW
2850
2851 /* EXACTF nodes need to know that the minimum
2852 * length changed so that a sharp s in the string
2853 * can match this ss in the pattern, but they
2854 * remain EXACTF nodes, as they are not trie'able,
2855 * so don't have to invent a new node type to
2856 * exclude them from the trie code */
2857 if (OP(scan) != EXACTF) {
2858 OP(scan) = EXACTFU_SS;
2859 }
2860 s++;
2861 }
2862 break;
2863 case LATIN_SMALL_LETTER_SHARP_S:
2864 if (OP(scan) == EXACTF) {
2865 *has_exactf_sharp_s = TRUE;
2866 }
2867 break;
86d6fcad
KW
2868 }
2869 }
2870 }
07be1b83 2871 }
3f410cf6 2872
07be1b83 2873#ifdef DEBUGGING
bb789b09
DM
2874 /* Allow dumping but overwriting the collection of skipped
2875 * ops and/or strings with fake optimized ops */
07be1b83
YO
2876 n = scan + NODE_SZ_STR(scan);
2877 while (n <= stop) {
bb789b09
DM
2878 OP(n) = OPTIMIZED;
2879 FLAGS(n) = 0;
2880 NEXT_OFF(n) = 0;
07be1b83
YO
2881 n++;
2882 }
2883#endif
2884 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2885 return stopnow;
2886}
2887
486ec47a 2888/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2889 Finds fixed substrings. */
2890
a0288114 2891/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2892 to the position after last scanned or to NULL. */
2893
40d049e4
YO
2894#define INIT_AND_WITHP \
2895 assert(!and_withp); \
2896 Newx(and_withp,1,struct regnode_charclass_class); \
2897 SAVEFREEPV(and_withp)
07be1b83 2898
b515a41d 2899/* this is a chain of data about sub patterns we are processing that
486ec47a 2900 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2901 we can simulate recursion without losing state. */
2902struct scan_frame;
2903typedef struct scan_frame {
2904 regnode *last; /* last node to process in this frame */
2905 regnode *next; /* next node to process when last is reached */
2906 struct scan_frame *prev; /*previous frame*/
2907 I32 stop; /* what stopparen do we use */
2908} scan_frame;
2909
304ee84b
YO
2910
2911#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2912
e1d1eefb
YO
2913#define CASE_SYNST_FNC(nAmE) \
2914case nAmE: \
2915 if (flags & SCF_DO_STCLASS_AND) { \
2916 for (value = 0; value < 256; value++) \
2917 if (!is_ ## nAmE ## _cp(value)) \
2918 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2919 } \
2920 else { \
2921 for (value = 0; value < 256; value++) \
2922 if (is_ ## nAmE ## _cp(value)) \
2923 ANYOF_BITMAP_SET(data->start_class, value); \
2924 } \
2925 break; \
2926case N ## nAmE: \
2927 if (flags & SCF_DO_STCLASS_AND) { \
2928 for (value = 0; value < 256; value++) \
2929 if (is_ ## nAmE ## _cp(value)) \
2930 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2931 } \
2932 else { \
2933 for (value = 0; value < 256; value++) \
2934 if (!is_ ## nAmE ## _cp(value)) \
2935 ANYOF_BITMAP_SET(data->start_class, value); \
2936 } \
2937 break
2938
2939
2940
76e3520e 2941STATIC I32
40d049e4 2942S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 2943 I32 *minlenp, I32 *deltap,
40d049e4
YO
2944 regnode *last,
2945 scan_data_t *data,
2946 I32 stopparen,
2947 U8* recursed,
2948 struct regnode_charclass_class *and_withp,
2949 U32 flags, U32 depth)
c277df42
IZ
2950 /* scanp: Start here (read-write). */
2951 /* deltap: Write maxlen-minlen here. */
2952 /* last: Stop before this one. */
40d049e4
YO
2953 /* data: string data about the pattern */
2954 /* stopparen: treat close N as END */
2955 /* recursed: which subroutines have we recursed into */
2956 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 2957{
97aff369 2958 dVAR;
c277df42
IZ
2959 I32 min = 0, pars = 0, code;
2960 regnode *scan = *scanp, *next;
2961 I32 delta = 0;
2962 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 2963 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
2964 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2965 scan_data_t data_fake;
a3621e74 2966 SV *re_trie_maxbuff = NULL;
786e8c11 2967 regnode *first_non_open = scan;
e2e6a0f1 2968 I32 stopmin = I32_MAX;
8aa23a47 2969 scan_frame *frame = NULL;
a3621e74 2970 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 2971
7918f24d
NC
2972 PERL_ARGS_ASSERT_STUDY_CHUNK;
2973
13a24bad 2974#ifdef DEBUGGING
40d049e4 2975 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 2976#endif
40d049e4 2977
786e8c11 2978 if ( depth == 0 ) {
40d049e4 2979 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
2980 first_non_open=regnext(first_non_open);
2981 }
2982
b81d288d 2983
8aa23a47
YO
2984 fake_study_recurse:
2985 while ( scan && OP(scan) != END && scan < last ){
9d071ca8
KW
2986 UV min_subtract = 0; /* How much to subtract from the minimum node
2987 length to get a real minimum (because the
2988 folded version may be shorter) */
f758bddf 2989 bool has_exactf_sharp_s = FALSE;
8aa23a47 2990 /* Peephole optimizer: */
304ee84b 2991 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 2992 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
2993
2994 /* Its not clear to khw or hv why this is done here, and not in the
2995 * clauses that deal with EXACT nodes. khw's guess is that it's
2996 * because of a previous design */
9d071ca8 2997 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
2998
2999 /* Follow the next-chain of the current node and optimize
3000 away all the NOTHINGs from it. */
3001 if (OP(scan) != CURLYX) {
3002 const int max = (reg_off_by_arg[OP(scan)]
3003 ? I32_MAX
3004 /* I32 may be smaller than U16 on CRAYs! */
3005 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3006 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3007 int noff;
3008 regnode *n = scan;
686b73d4 3009
8aa23a47
YO
3010 /* Skip NOTHING and LONGJMP. */
3011 while ((n = regnext(n))
3012 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3013 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3014 && off + noff < max)
3015 off += noff;
3016 if (reg_off_by_arg[OP(scan)])
3017 ARG(scan) = off;
3018 else
3019 NEXT_OFF(scan) = off;
3020 }
a3621e74 3021
c277df42 3022
8aa23a47
YO
3023
3024 /* The principal pseudo-switch. Cannot be a switch, since we
3025 look into several different things. */
3026 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3027 || OP(scan) == IFTHEN) {
3028 next = regnext(scan);
3029 code = OP(scan);
3030 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3031
8aa23a47
YO
3032 if (OP(next) == code || code == IFTHEN) {
3033 /* NOTE - There is similar code to this block below for handling
3034 TRIE nodes on a re-study. If you change stuff here check there
3035 too. */
3036 I32 max1 = 0, min1 = I32_MAX, num = 0;
3037 struct regnode_charclass_class accum;
3038 regnode * const startbranch=scan;
686b73d4 3039
8aa23a47 3040 if (flags & SCF_DO_SUBSTR)
304ee84b 3041 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3042 if (flags & SCF_DO_STCLASS)
e755fd73 3043 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3044
3045 while (OP(scan) == code) {
3046 I32 deltanext, minnext, f = 0, fake;
3047 struct regnode_charclass_class this_class;
3048
3049 num++;
3050 data_fake.flags = 0;
3051 if (data) {
3052 data_fake.whilem_c = data->whilem_c;
3053 data_fake.last_closep = data->last_closep;
3054 }
3055 else
3056 data_fake.last_closep = &fake;
58e23c8d
YO
3057
3058 data_fake.pos_delta = delta;
8aa23a47
YO
3059 next = regnext(scan);
3060 scan = NEXTOPER(scan);
3061 if (code != BRANCH)
c277df42 3062 scan = NEXTOPER(scan);
8aa23a47 3063 if (flags & SCF_DO_STCLASS) {
e755fd73 3064 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3065 data_fake.start_class = &this_class;
3066 f = SCF_DO_STCLASS_AND;
58e23c8d 3067 }
8aa23a47
YO
3068 if (flags & SCF_WHILEM_VISITED_POS)
3069 f |= SCF_WHILEM_VISITED_POS;
3070
3071 /* we suppose the run is continuous, last=next...*/
3072 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3073 next, &data_fake,
3074 stopparen, recursed, NULL, f,depth+1);
3075 if (min1 > minnext)
3076 min1 = minnext;
3077 if (max1 < minnext + deltanext)
3078 max1 = minnext + deltanext;
3079 if (deltanext == I32_MAX)
3080 is_inf = is_inf_internal = 1;
3081 scan = next;
3082 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3083 pars++;
3084 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3085 if ( stopmin > minnext)
3086 stopmin = min + min1;
3087 flags &= ~SCF_DO_SUBSTR;
3088 if (data)
3089 data->flags |= SCF_SEEN_ACCEPT;
3090 }
3091 if (data) {
3092 if (data_fake.flags & SF_HAS_EVAL)
3093 data->flags |= SF_HAS_EVAL;
3094 data->whilem_c = data_fake.whilem_c;
3dab1dad 3095 }
8aa23a47 3096 if (flags & SCF_DO_STCLASS)
3fffb88a 3097 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3098 }
3099 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3100 min1 = 0;
3101 if (flags & SCF_DO_SUBSTR) {
3102 data->pos_min += min1;
3103 data->pos_delta += max1 - min1;
3104 if (max1 != min1 || is_inf)
3105 data->longest = &(data->longest_float);
3106 }
3107 min += min1;
3108 delta += max1 - min1;
3109 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3110 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3111 if (min1) {
3112 cl_and(data->start_class, and_withp);
3113 flags &= ~SCF_DO_STCLASS;
653099ff 3114 }
8aa23a47
YO
3115 }
3116 else if (flags & SCF_DO_STCLASS_AND) {
3117 if (min1) {
3118 cl_and(data->start_class, &accum);
3119 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3120 }
8aa23a47
YO
3121 else {
3122 /* Switch to OR mode: cache the old value of
3123 * data->start_class */
3124 INIT_AND_WITHP;
3125 StructCopy(data->start_class, and_withp,
3126 struct regnode_charclass_class);
3127 flags &= ~SCF_DO_STCLASS_AND;
3128 StructCopy(&accum, data->start_class,
3129 struct regnode_charclass_class);
3130 flags |= SCF_DO_STCLASS_OR;
3131 data->start_class->flags |= ANYOF_EOS;
de0c8cb8 3132 }
8aa23a47 3133 }
a3621e74 3134
8aa23a47
YO
3135 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3136 /* demq.
a3621e74 3137
8aa23a47
YO
3138 Assuming this was/is a branch we are dealing with: 'scan' now
3139 points at the item that follows the branch sequence, whatever
3140 it is. We now start at the beginning of the sequence and look
3141 for subsequences of
a3621e74 3142
8aa23a47
YO
3143 BRANCH->EXACT=>x1
3144 BRANCH->EXACT=>x2
3145 tail
a3621e74 3146
8aa23a47 3147 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3148
486ec47a 3149 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3150 element into a trie and then add the subsequent branch exact
3151 strings to the trie.
a3621e74 3152
8aa23a47 3153 We have two cases
a3621e74 3154
3b753521 3155 1. patterns where the whole set of branches can be converted.
a3621e74 3156
8aa23a47 3157 2. patterns where only a subset can be converted.
a3621e74 3158
8aa23a47
YO
3159 In case 1 we can replace the whole set with a single regop
3160 for the trie. In case 2 we need to keep the start and end
3b753521 3161 branches so
a3621e74 3162
8aa23a47
YO
3163 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3164 becomes BRANCH TRIE; BRANCH X;
786e8c11 3165
8aa23a47
YO
3166 There is an additional case, that being where there is a
3167 common prefix, which gets split out into an EXACT like node
3168 preceding the TRIE node.
a3621e74 3169
8aa23a47
YO
3170 If x(1..n)==tail then we can do a simple trie, if not we make
3171 a "jump" trie, such that when we match the appropriate word
486ec47a 3172 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3173 a nested if into a case structure of sorts.
b515a41d 3174
8aa23a47 3175 */
686b73d4 3176
8aa23a47
YO
3177 int made=0;
3178 if (!re_trie_maxbuff) {
3179 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3180 if (!SvIOK(re_trie_maxbuff))
3181 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3182 }
3183 if ( SvIV(re_trie_maxbuff)>=0 ) {
3184 regnode *cur;
3185 regnode *first = (regnode *)NULL;
3186 regnode *last = (regnode *)NULL;
3187 regnode *tail = scan;
3188 U8 optype = 0;
3189 U32 count=0;
a3621e74
YO
3190
3191#ifdef DEBUGGING
8aa23a47 3192 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3193#endif
8aa23a47
YO
3194 /* var tail is used because there may be a TAIL
3195 regop in the way. Ie, the exacts will point to the
3196 thing following the TAIL, but the last branch will
3197 point at the TAIL. So we advance tail. If we
3198 have nested (?:) we may have to move through several
3199 tails.
3200 */
3201
3202 while ( OP( tail ) == TAIL ) {
3203 /* this is the TAIL generated by (?:) */
3204 tail = regnext( tail );
3205 }
a3621e74 3206
8aa23a47
YO
3207
3208 DEBUG_OPTIMISE_r({
3209 regprop(RExC_rx, mysv, tail );
3210 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3211 (int)depth * 2 + 2, "",
3212 "Looking for TRIE'able sequences. Tail node is: ",
3213 SvPV_nolen_const( mysv )
3214 );
3215 });
3216
3217 /*
3218
3219 step through the branches, cur represents each
3220 branch, noper is the first thing to be matched
3221 as part of that branch and noper_next is the
3222 regnext() of that node. if noper is an EXACT
3223 and noper_next is the same as scan (our current
3224 position in the regex) then the EXACT branch is
3225 a possible optimization target. Once we have
486ec47a 3226 two or more consecutive such branches we can
8aa23a47
YO
3227 create a trie of the EXACT's contents and stich
3228 it in place. If the sequence represents all of
3229 the branches we eliminate the whole thing and
3230 replace it with a single TRIE. If it is a
3231 subsequence then we need to stitch it in. This
3232 means the first branch has to remain, and needs
3233 to be repointed at the item on the branch chain
3234 following the last branch optimized. This could
3235 be either a BRANCH, in which case the
3236 subsequence is internal, or it could be the
3237 item following the branch sequence in which
3238 case the subsequence is at the end.
3239
3240 */
3241
3242 /* dont use tail as the end marker for this traverse */
3243 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3244 regnode * const noper = NEXTOPER( cur );
b515a41d 3245#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3246 regnode * const noper_next = regnext( noper );
b515a41d
YO
3247#endif
3248
8aa23a47
YO
3249 DEBUG_OPTIMISE_r({
3250 regprop(RExC_rx, mysv, cur);
3251 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3252 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3253
3254 regprop(RExC_rx, mysv, noper);
3255 PerlIO_printf( Perl_debug_log, " -> %s",
3256 SvPV_nolen_const(mysv));
3257
3258 if ( noper_next ) {
3259 regprop(RExC_rx, mysv, noper_next );
3260 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3261 SvPV_nolen_const(mysv));
3262 }
3263 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
3264 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
3265 });
3266 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
3267 : PL_regkind[ OP( noper ) ] == EXACT )
3268 || OP(noper) == NOTHING )
786e8c11 3269#ifdef NOJUMPTRIE
8aa23a47 3270 && noper_next == tail
786e8c11 3271#endif
8aa23a47
YO
3272 && count < U16_MAX)
3273 {
3274 count++;
3275 if ( !first || optype == NOTHING ) {
3276 if (!first) first = cur;
3277 optype = OP( noper );
3278 } else {
3279 last = cur;
3280 }
3281 } else {
a0a388a1 3282/*
fbebf34e
KW
3283 Currently the trie logic handles case insensitive matching properly only
3284 when the pattern is UTF-8 and the node is EXACTFU (thus forcing unicode
3285 semantics).
0abd0d78
YO
3286
3287 If/when this is fixed the following define can be swapped
3288 in below to fully enable trie logic.
3289
a0a388a1 3290#define TRIE_TYPE_IS_SAFE 1
0abd0d78 3291
a0c4c608
KW
3292Note that join_exact() assumes that the other types of EXACTFish nodes are not
3293used in tries, so that would have to be updated if this changed
3294
0abd0d78 3295*/
fbebf34e 3296#define TRIE_TYPE_IS_SAFE ((UTF && optype == EXACTFU) || optype==EXACT)
0abd0d78 3297
a0a388a1 3298 if ( last && TRIE_TYPE_IS_SAFE ) {
8aa23a47
YO
3299 make_trie( pRExC_state,
3300 startbranch, first, cur, tail, count,
3301 optype, depth+1 );
3302 }
3303 if ( PL_regkind[ OP( noper ) ] == EXACT
786e8c11 3304#ifdef NOJUMPTRIE
8aa23a47 3305 && noper_next == tail
786e8c11 3306#endif
8aa23a47
YO
3307 ){
3308 count = 1;
3309 first = cur;
3310 optype = OP( noper );
3311 } else {
3312 count = 0;
3313 first = NULL;
3314 optype = 0;
3315 }
3316 last = NULL;
3317 }
3318 }
3319 DEBUG_OPTIMISE_r({
3320 regprop(RExC_rx, mysv, cur);
3321 PerlIO_printf( Perl_debug_log,
3322 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3323 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3324
3325 });
a0a388a1
YO
3326
3327 if ( last && TRIE_TYPE_IS_SAFE ) {
8aa23a47 3328 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
686b73d4 3329#ifdef TRIE_STUDY_OPT
8aa23a47
YO
3330 if ( ((made == MADE_EXACT_TRIE &&
3331 startbranch == first)
3332 || ( first_non_open == first )) &&
3333 depth==0 ) {
3334 flags |= SCF_TRIE_RESTUDY;
3335 if ( startbranch == first
3336 && scan == tail )
3337 {
3338 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3339 }
3340 }
3dab1dad 3341#endif
8aa23a47
YO
3342 }
3343 }
3344
3345 } /* do trie */
3346
653099ff 3347 }
8aa23a47
YO
3348 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3349 scan = NEXTOPER(NEXTOPER(scan));
3350 } else /* single branch is optimized. */
3351 scan = NEXTOPER(scan);
3352 continue;
3353 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3354 scan_frame *newframe = NULL;
3355 I32 paren;
3356 regnode *start;
3357 regnode *end;
3358
3359 if (OP(scan) != SUSPEND) {
3360 /* set the pointer */
3361 if (OP(scan) == GOSUB) {
3362 paren = ARG(scan);
3363 RExC_recurse[ARG2L(scan)] = scan;
3364 start = RExC_open_parens[paren-1];
3365 end = RExC_close_parens[paren-1];
3366 } else {
3367 paren = 0;
f8fc2ecf 3368 start = RExC_rxi->program + 1;
8aa23a47
YO
3369 end = RExC_opend;
3370 }
3371 if (!recursed) {
3372 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3373 SAVEFREEPV(recursed);
3374 }
3375 if (!PAREN_TEST(recursed,paren+1)) {
3376 PAREN_SET(recursed,paren+1);
3377 Newx(newframe,1,scan_frame);
3378 } else {
3379 if (flags & SCF_DO_SUBSTR) {
304ee84b 3380 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3381 data->longest = &(data->longest_float);
3382 }
3383 is_inf = is_inf_internal = 1;
3384 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3385 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3386 flags &= ~SCF_DO_STCLASS;
3387 }
3388 } else {
3389 Newx(newframe,1,scan_frame);
3390 paren = stopparen;
3391 start = scan+2;
3392 end = regnext(scan);
3393 }
3394 if (newframe) {
3395 assert(start);
3396 assert(end);
3397 SAVEFREEPV(newframe);
3398 newframe->next = regnext(scan);
3399 newframe->last = last;
3400 newframe->stop = stopparen;
3401 newframe->prev = frame;
3402
3403 frame = newframe;
3404 scan = start;
3405 stopparen = paren;
3406 last = end;
3407
3408 continue;
3409 }
3410 }
3411 else if (OP(scan) == EXACT) {
3412 I32 l = STR_LEN(scan);
3413 UV uc;
3414 if (UTF) {
3415 const U8 * const s = (U8*)STRING(scan);
3416 l = utf8_length(s, s + l);
3417 uc = utf8_to_uvchr(s, NULL);
3418 } else {
3419 uc = *((U8*)STRING(scan));
3420 }
3421 min += l;
3422 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3423 /* The code below prefers earlier match for fixed
3424 offset, later match for variable offset. */
3425 if (data->last_end == -1) { /* Update the start info. */
3426 data->last_start_min = data->pos_min;
3427 data->last_start_max = is_inf
3428 ? I32_MAX : data->pos_min + data->pos_delta;
b515a41d 3429 }
8aa23a47
YO
3430 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3431 if (UTF)
3432 SvUTF8_on(data->last_found);
3433 {
3434 SV * const sv = data->last_found;
3435 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3436 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3437 if (mg && mg->mg_len >= 0)
3438 mg->mg_len += utf8_length((U8*)STRING(scan),
3439 (U8*)STRING(scan)+STR_LEN(scan));
b515a41d 3440 }
8aa23a47
YO
3441 data->last_end = data->pos_min + l;
3442 data->pos_min += l; /* As in the first entry. */
3443 data->flags &= ~SF_BEFORE_EOL;
3444 }
3445 if (flags & SCF_DO_STCLASS_AND) {
3446 /* Check whether it is compatible with what we know already! */
3447 int compat = 1;
3448
54251c2e 3449
486ec47a 3450 /* If compatible, we or it in below. It is compatible if is
54251c2e
KW
3451 * in the bitmp and either 1) its bit or its fold is set, or 2)
3452 * it's for a locale. Even if there isn't unicode semantics
3453 * here, at runtime there may be because of matching against a
3454 * utf8 string, so accept a possible false positive for
3455 * latin1-range folds */
8aa23a47
YO
3456 if (uc >= 0x100 ||
3457 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3458 && !ANYOF_BITMAP_TEST(data->start_class, uc)
39065660 3459 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
54251c2e 3460 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
8aa23a47 3461 )
d18bf9dc 3462 {
8aa23a47 3463 compat = 0;
d18bf9dc 3464 }
8aa23a47
YO
3465 ANYOF_CLASS_ZERO(data->start_class);
3466 ANYOF_BITMAP_ZERO(data->start_class);
3467 if (compat)
3468 ANYOF_BITMAP_SET(data->start_class, uc);
d18bf9dc
KW
3469 else if (uc >= 0x100) {
3470 int i;
3471
3472 /* Some Unicode code points fold to the Latin1 range; as
3473 * XXX temporary code, instead of figuring out if this is
3474 * one, just assume it is and set all the start class bits
3475 * that could be some such above 255 code point's fold
3476 * which will generate fals positives. As the code
3477 * elsewhere that does compute the fold settles down, it
3478 * can be extracted out and re-used here */
3479 for (i = 0; i < 256; i++){
3480 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3481 ANYOF_BITMAP_SET(data->start_class, i);
3482 }
3483 }
3484 }
8aa23a47
YO
3485 data->start_class->flags &= ~ANYOF_EOS;
3486 if (uc < 0x100)
3487 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3488 }
3489 else if (flags & SCF_DO_STCLASS_OR) {
3490 /* false positive possible if the class is case-folded */
3491 if (uc < 0x100)
3492 ANYOF_BITMAP_SET(data->start_class, uc);
3493 else
3494 data->start_class->flags |= ANYOF_UNICODE_ALL;
3495 data->start_class->flags &= ~ANYOF_EOS;
3496 cl_and(data->start_class, and_withp);
3497 }
3498 flags &= ~SCF_DO_STCLASS;
3499 }
3500 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3501 I32 l = STR_LEN(scan);
3502 UV uc = *((U8*)STRING(scan));
3503
3504 /* Search for fixed substrings supports EXACT only. */
3505 if (flags & SCF_DO_SUBSTR) {
3506 assert(data);
304ee84b 3507 SCAN_COMMIT(pRExC_state, data, minlenp);
8aa23a47
YO
3508 }
3509 if (UTF) {
3510 const U8 * const s = (U8 *)STRING(scan);
3511 l = utf8_length(s, s + l);
3512 uc = utf8_to_uvchr(s, NULL);
3513 }
f758bddf
KW
3514 else if (has_exactf_sharp_s) {
3515 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
bb914485 3516 }
9d071ca8 3517 min += l - min_subtract;
f646642f
KW
3518 if (min < 0) {
3519 min = 0;
3520 }
9d071ca8 3521 delta += min_subtract;
f646642f 3522 if (flags & SCF_DO_SUBSTR) {
9d071ca8 3523 data->pos_min += l - min_subtract;
f646642f
KW
3524 if (data->pos_min < 0) {
3525 data->pos_min = 0;
3526 }
9d071ca8
KW
3527 data->pos_delta += min_subtract;
3528 if (min_subtract) {
d2197104
KW
3529 data->longest = &(data->longest_float);
3530 }
f646642f 3531 }
8aa23a47
YO
3532 if (flags & SCF_DO_STCLASS_AND) {
3533 /* Check whether it is compatible with what we know already! */
3534 int compat = 1;
8aa23a47 3535 if (uc >= 0x100 ||
54251c2e
KW
3536 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3537 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3538 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3539 {
8aa23a47 3540 compat = 0;
54251c2e 3541 }
8aa23a47
YO
3542 ANYOF_CLASS_ZERO(data->start_class);
3543 ANYOF_BITMAP_ZERO(data->start_class);
3544 if (compat) {
3545 ANYOF_BITMAP_SET(data->start_class, uc);
653099ff 3546 data->start_class->flags &= ~ANYOF_EOS;
39065660 3547 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
970c8436 3548 if (OP(scan) == EXACTFL) {
af302e7f
KW
3549 /* XXX This set is probably no longer necessary, and
3550 * probably wrong as LOCALE now is on in the initial
3551 * state */
8aa23a47 3552 data->start_class->flags |= ANYOF_LOCALE;
970c8436
KW
3553 }
3554 else {
3555
54251c2e
KW
3556 /* Also set the other member of the fold pair. In case
3557 * that unicode semantics is called for at runtime, use
3558 * the full latin1 fold. (Can't do this for locale,
a0c4c608 3559 * because not known until runtime) */
54251c2e 3560 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
e22b340a 3561
a0c4c608
KW
3562 /* All other (EXACTFL handled above) folds except under
3563 * /iaa that include s, S, and sharp_s also may include
3564 * the others */
e22b340a
KW
3565 if (OP(scan) != EXACTFA) {
3566 if (uc == 's' || uc == 'S') {
3567 ANYOF_BITMAP_SET(data->start_class,
3568 LATIN_SMALL_LETTER_SHARP_S);
3569 }
3570 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3571 ANYOF_BITMAP_SET(data->start_class, 's');
3572 ANYOF_BITMAP_SET(data->start_class, 'S');
3573 }
3574 }
970c8436 3575 }
653099ff 3576 }
d18bf9dc
KW
3577 else if (uc >= 0x100) {
3578 int i;
3579 for (i = 0; i < 256; i++){
3580 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3581 ANYOF_BITMAP_SET(data->start_class, i);
3582 }
3583 }
3584 }
8aa23a47
YO
3585 }
3586 else if (flags & SCF_DO_STCLASS_OR) {
39065660 3587 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
8aa23a47
YO
3588 /* false positive possible if the class is case-folded.
3589 Assume that the locale settings are the same... */
970c8436 3590 if (uc < 0x100) {
1aa99e6b 3591 ANYOF_BITMAP_SET(data->start_class, uc);
970c8436
KW
3592 if (OP(scan) != EXACTFL) {
3593
3594 /* And set the other member of the fold pair, but
3595 * can't do that in locale because not known until
3596 * run-time */
3597 ANYOF_BITMAP_SET(data->start_class,
54251c2e 3598 PL_fold_latin1[uc]);
e22b340a
KW
3599
3600 /* All folds except under /iaa that include s, S,
3601 * and sharp_s also may include the others */
3602 if (OP(scan) != EXACTFA) {
3603 if (uc == 's' || uc == 'S') {
3604 ANYOF_BITMAP_SET(data->start_class,
3605 LATIN_SMALL_LETTER_SHARP_S);
3606 }
3607 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3608 ANYOF_BITMAP_SET(data->start_class, 's');
3609 ANYOF_BITMAP_SET(data->start_class, 'S');
3610 }
3611 }
970c8436
KW
3612 }
3613 }
653099ff
GS
3614 data->start_class->flags &= ~ANYOF_EOS;
3615 }
8aa23a47 3616 cl_and(data->start_class, and_withp);
653099ff 3617 }
8aa23a47
YO
3618 flags &= ~SCF_DO_STCLASS;
3619 }
e52fc539 3620 else if (REGNODE_VARIES(OP(scan))) {
8aa23a47
YO
3621 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3622 I32 f = flags, pos_before = 0;
3623 regnode * const oscan = scan;
3624 struct regnode_charclass_class this_class;
3625 struct regnode_charclass_class *oclass = NULL;
3626 I32 next_is_eval = 0;
3627
3628 switch (PL_regkind[OP(scan)]) {
3629 case WHILEM: /* End of (?:...)* . */