This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c:regprop: [bracketize] \w..., add \v
[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"
40b6423c 84extern const struct regexp_engine my_reg_engine;
54df2634
NC
85#else
86# include "regcomp.h"
87#endif
a687059c 88
04e98a4d 89#include "dquote_static.c"
26faadbd 90#include "charclass_invlists.h"
81e983c1 91#include "inline_invlist.c"
1b0f46bf 92#include "unicode_constants.h"
04e98a4d 93
963df1c9
KW
94#ifdef HAS_ISBLANK
95# define hasISBLANK 1
96#else
97# define hasISBLANK 0
98#endif
99
94dc5c2d 100#define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
26faadbd 101#define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
2c61f163 102#define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94dc5c2d 103
d4cce5f1 104#ifdef op
11343788 105#undef op
d4cce5f1 106#endif /* op */
11343788 107
fe14fcc3 108#ifdef MSDOS
7e4e8c89 109# if defined(BUGGY_MSC6)
fe14fcc3 110 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
7e4e8c89 111# pragma optimize("a",off)
fe14fcc3 112 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
7e4e8c89
NC
113# pragma optimize("w",on )
114# endif /* BUGGY_MSC6 */
fe14fcc3
LW
115#endif /* MSDOS */
116
a687059c
LW
117#ifndef STATIC
118#define STATIC static
119#endif
120
b1603ef8 121
830247a4 122typedef struct RExC_state_t {
514a91f1
DM
123 U32 flags; /* RXf_* are we folding, multilining? */
124 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 125 char *precomp; /* uncompiled string. */
288b8c02 126 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf
YO
127 regexp *rx; /* perl core regexp structure */
128 regexp_internal *rxi; /* internal data for regexp object pprivate field */
fac92740 129 char *start; /* Start of input for compile */
830247a4
IZ
130 char *end; /* End of input for compile */
131 char *parse; /* Input-scan pointer. */
132 I32 whilem_seen; /* number of WHILEM in this expr */
fac92740 133 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 134 regnode *emit_bound; /* First regnode outside of the allocated space */
ffc61ed2 135 regnode *emit; /* Code-emit pointer; &regdummy = don't = compiling */
830247a4
IZ
136 I32 naughty; /* How bad is this pattern? */
137 I32 sawback; /* Did we see \1, ...? */
138 U32 seen;
139 I32 size; /* Code size. */
c74340f9
YO
140 I32 npar; /* Capture buffer count, (OPEN). */
141 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 142 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
143 I32 extralen;
144 I32 seen_zerolen;
40d049e4
YO
145 regnode **open_parens; /* pointers to open parens */
146 regnode **close_parens; /* pointers to close parens */
147 regnode *opend; /* END node in program */
02daf0ab
YO
148 I32 utf8; /* whether the pattern is utf8 or not */
149 I32 orig_utf8; /* whether the pattern was originally in utf8 */
150 /* XXX use this for future optimisation of case
151 * where pattern must be upgraded to utf8. */
e40e74fe
KW
152 I32 uni_semantics; /* If a d charset modifier should use unicode
153 rules, even if the pattern is not in
154 utf8 */
81714fb9 155 HV *paren_names; /* Paren names */
1f1031fe 156
40d049e4
YO
157 regnode **recurse; /* Recurse regops */
158 I32 recurse_count; /* Number of recurse regops */
b57e4118 159 I32 in_lookbehind;
4624b182 160 I32 contains_locale;
bb3f3ed2 161 I32 override_recoding;
9d53c457 162 I32 in_multi_char_class;
3d2bd50a 163 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 164 within pattern */
b1603ef8
DM
165 int num_code_blocks; /* size of code_blocks[] */
166 int code_index; /* next code_blocks[] slot */
830247a4
IZ
167#if ADD_TO_REGEXEC
168 char *starttry; /* -Dr: where regtry was called. */
169#define RExC_starttry (pRExC_state->starttry)
170#endif
d24ca0c5 171 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 172#ifdef DEBUGGING
be8e71aa 173 const char *lastparse;
3dab1dad 174 I32 lastnum;
1f1031fe 175 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
176#define RExC_lastparse (pRExC_state->lastparse)
177#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 178#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 179#endif
830247a4
IZ
180} RExC_state_t;
181
e2509266 182#define RExC_flags (pRExC_state->flags)
514a91f1 183#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 184#define RExC_precomp (pRExC_state->precomp)
288b8c02 185#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 186#define RExC_rx (pRExC_state->rx)
f8fc2ecf 187#define RExC_rxi (pRExC_state->rxi)
fac92740 188#define RExC_start (pRExC_state->start)
830247a4
IZ
189#define RExC_end (pRExC_state->end)
190#define RExC_parse (pRExC_state->parse)
191#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
192#ifdef RE_TRACK_PATTERN_OFFSETS
193#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
194#endif
830247a4 195#define RExC_emit (pRExC_state->emit)
fac92740 196#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 197#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
198#define RExC_naughty (pRExC_state->naughty)
199#define RExC_sawback (pRExC_state->sawback)
200#define RExC_seen (pRExC_state->seen)
201#define RExC_size (pRExC_state->size)
202#define RExC_npar (pRExC_state->npar)
e2e6a0f1 203#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
204#define RExC_extralen (pRExC_state->extralen)
205#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 206#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 207#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 208#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
209#define RExC_open_parens (pRExC_state->open_parens)
210#define RExC_close_parens (pRExC_state->close_parens)
211#define RExC_opend (pRExC_state->opend)
81714fb9 212#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
213#define RExC_recurse (pRExC_state->recurse)
214#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 215#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 216#define RExC_contains_locale (pRExC_state->contains_locale)
9d53c457
KW
217#define RExC_override_recoding (pRExC_state->override_recoding)
218#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
830247a4 219
cde0cee5 220
a687059c
LW
221#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
222#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
223 ((*s) == '{' && regcurly(s)))
a687059c 224
35c8bce7
LW
225#ifdef SPSTART
226#undef SPSTART /* dratted cpp namespace... */
227#endif
a687059c
LW
228/*
229 * Flags to be passed up and down.
230 */
a687059c 231#define WORST 0 /* Worst case. */
a3b492c3 232#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 233
e64f369d 234/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
235 * character. (There needs to be a case: in the switch statement in regexec.c
236 * for any node marked SIMPLE.) Note that this is not the same thing as
237 * REGNODE_SIMPLE */
fda99bee 238#define SIMPLE 0x02
e64f369d 239#define SPSTART 0x04 /* Starts with * or + */
a3b492c3
YO
240#define TRYAGAIN 0x08 /* Weeded out a declaration. */
241#define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
a687059c 242
3dab1dad
YO
243#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
244
07be1b83
YO
245/* whether trie related optimizations are enabled */
246#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
247#define TRIE_STUDY_OPT
786e8c11 248#define FULL_TRIE_STUDY
07be1b83
YO
249#define TRIE_STCLASS
250#endif
1de06328
YO
251
252
40d049e4
YO
253
254#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
255#define PBITVAL(paren) (1 << ((paren) & 7))
256#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
257#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
258#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
259
bbd61b5f
KW
260/* If not already in utf8, do a longjmp back to the beginning */
261#define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
262#define REQUIRE_UTF8 STMT_START { \
263 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
264 } STMT_END
40d049e4 265
1de06328
YO
266/* About scan_data_t.
267
268 During optimisation we recurse through the regexp program performing
269 various inplace (keyhole style) optimisations. In addition study_chunk
270 and scan_commit populate this data structure with information about
271 what strings MUST appear in the pattern. We look for the longest
3b753521 272 string that must appear at a fixed location, and we look for the
1de06328
YO
273 longest string that may appear at a floating location. So for instance
274 in the pattern:
275
276 /FOO[xX]A.*B[xX]BAR/
277
278 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
279 strings (because they follow a .* construct). study_chunk will identify
280 both FOO and BAR as being the longest fixed and floating strings respectively.
281
282 The strings can be composites, for instance
283
284 /(f)(o)(o)/
285
286 will result in a composite fixed substring 'foo'.
287
288 For each string some basic information is maintained:
289
290 - offset or min_offset
291 This is the position the string must appear at, or not before.
292 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
293 characters must match before the string we are searching for.
294 Likewise when combined with minlenp and the length of the string it
1de06328
YO
295 tells us how many characters must appear after the string we have
296 found.
297
298 - max_offset
299 Only used for floating strings. This is the rightmost point that
3b753521 300 the string can appear at. If set to I32 max it indicates that the
1de06328
YO
301 string can occur infinitely far to the right.
302
303 - minlenp
2d608413
KW
304 A pointer to the minimum number of characters of the pattern that the
305 string was found inside. This is important as in the case of positive
1de06328
YO
306 lookahead or positive lookbehind we can have multiple patterns
307 involved. Consider
308
309 /(?=FOO).*F/
310
311 The minimum length of the pattern overall is 3, the minimum length
312 of the lookahead part is 3, but the minimum length of the part that
313 will actually match is 1. So 'FOO's minimum length is 3, but the
314 minimum length for the F is 1. This is important as the minimum length
315 is used to determine offsets in front of and behind the string being
316 looked for. Since strings can be composites this is the length of the
486ec47a 317 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
318 the length is calculated by study_chunk, so that the minimum lengths
319 are not known until the full pattern has been compiled, thus the
320 pointer to the value.
321
322 - lookbehind
323
324 In the case of lookbehind the string being searched for can be
325 offset past the start point of the final matching string.
326 If this value was just blithely removed from the min_offset it would
327 invalidate some of the calculations for how many chars must match
328 before or after (as they are derived from min_offset and minlen and
329 the length of the string being searched for).
330 When the final pattern is compiled and the data is moved from the
331 scan_data_t structure into the regexp structure the information
332 about lookbehind is factored in, with the information that would
333 have been lost precalculated in the end_shift field for the
334 associated string.
335
336 The fields pos_min and pos_delta are used to store the minimum offset
337 and the delta to the maximum offset at the current point in the pattern.
338
339*/
2c2d71f5
JH
340
341typedef struct scan_data_t {
1de06328
YO
342 /*I32 len_min; unused */
343 /*I32 len_delta; unused */
2c2d71f5
JH
344 I32 pos_min;
345 I32 pos_delta;
346 SV *last_found;
1de06328 347 I32 last_end; /* min value, <0 unless valid. */
2c2d71f5
JH
348 I32 last_start_min;
349 I32 last_start_max;
1de06328
YO
350 SV **longest; /* Either &l_fixed, or &l_float. */
351 SV *longest_fixed; /* longest fixed string found in pattern */
352 I32 offset_fixed; /* offset where it starts */
486ec47a 353 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
354 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
355 SV *longest_float; /* longest floating string found in pattern */
356 I32 offset_float_min; /* earliest point in string it can appear */
357 I32 offset_float_max; /* latest point in string it can appear */
486ec47a 358 I32 *minlen_float; /* pointer to the minlen relevant to the string */
1de06328 359 I32 lookbehind_float; /* is the position of the string modified by LB */
2c2d71f5
JH
360 I32 flags;
361 I32 whilem_c;
cb434fcc 362 I32 *last_closep;
653099ff 363 struct regnode_charclass_class *start_class;
2c2d71f5
JH
364} scan_data_t;
365
a687059c 366/*
e50aee73 367 * Forward declarations for pregcomp()'s friends.
a687059c 368 */
a0d0e21e 369
27da23d5 370static const scan_data_t zero_scan_data =
1de06328 371 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
372
373#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
374#define SF_BEFORE_SEOL 0x0001
375#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
376#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
377#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
378
09b7f37c
CB
379#ifdef NO_UNARY_PLUS
380# define SF_FIX_SHIFT_EOL (0+2)
381# define SF_FL_SHIFT_EOL (0+4)
382#else
383# define SF_FIX_SHIFT_EOL (+2)
384# define SF_FL_SHIFT_EOL (+4)
385#endif
c277df42
IZ
386
387#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
388#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
389
390#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
391#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
392#define SF_IS_INF 0x0040
393#define SF_HAS_PAR 0x0080
394#define SF_IN_PAR 0x0100
395#define SF_HAS_EVAL 0x0200
396#define SCF_DO_SUBSTR 0x0400
653099ff
GS
397#define SCF_DO_STCLASS_AND 0x0800
398#define SCF_DO_STCLASS_OR 0x1000
399#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 400#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 401
786e8c11 402#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 403#define SCF_SEEN_ACCEPT 0x8000
07be1b83 404
43fead97 405#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
406
407/* The enums for all these are ordered so things work out correctly */
a62b1201 408#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
cfaf538b 409#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
00b27cfc 410#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
411#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
412#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1 413#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a725e29c 414#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 415
43fead97 416#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 417
93733859 418#define OOB_NAMEDCLASS -1
b8c5462f 419
8e661ac5
KW
420/* There is no code point that is out-of-bounds, so this is problematic. But
421 * its only current use is to initialize a variable that is always set before
422 * looked at. */
423#define OOB_UNICODE 0xDEADBEEF
424
a0ed51b3
LW
425#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
426#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
427
8615cb43 428
b45f050a
JF
429/* length of regex to show in messages that don't mark a position within */
430#define RegexLengthToShowInErrorMessages 127
431
432/*
433 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
434 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
435 * op/pragma/warn/regcomp.
436 */
7253e4e3
RK
437#define MARKER1 "<-- HERE" /* marker as it appears in the description */
438#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 439
7253e4e3 440#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
441
442/*
443 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
444 * arg. Show regex, up to a maximum length. If it's too long, chop and add
445 * "...".
446 */
58e23c8d 447#define _FAIL(code) STMT_START { \
bfed75c6 448 const char *ellipses = ""; \
ccb2c380
MP
449 IV len = RExC_end - RExC_precomp; \
450 \
451 if (!SIZE_ONLY) \
a5e7bc51 452 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
453 if (len > RegexLengthToShowInErrorMessages) { \
454 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
455 len = RegexLengthToShowInErrorMessages - 10; \
456 ellipses = "..."; \
457 } \
58e23c8d 458 code; \
ccb2c380 459} STMT_END
8615cb43 460
58e23c8d
YO
461#define FAIL(msg) _FAIL( \
462 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
463 msg, (int)len, RExC_precomp, ellipses))
464
465#define FAIL2(msg,arg) _FAIL( \
466 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
467 arg, (int)len, RExC_precomp, ellipses))
468
b45f050a 469/*
b45f050a
JF
470 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
471 */
ccb2c380 472#define Simple_vFAIL(m) STMT_START { \
a28509cc 473 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
474 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
475 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
476} STMT_END
b45f050a
JF
477
478/*
479 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
480 */
ccb2c380
MP
481#define vFAIL(m) STMT_START { \
482 if (!SIZE_ONLY) \
a5e7bc51 483 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
484 Simple_vFAIL(m); \
485} STMT_END
b45f050a
JF
486
487/*
488 * Like Simple_vFAIL(), but accepts two arguments.
489 */
ccb2c380 490#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 491 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
492 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
493 (int)offset, RExC_precomp, RExC_precomp + offset); \
494} STMT_END
b45f050a
JF
495
496/*
497 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
498 */
ccb2c380
MP
499#define vFAIL2(m,a1) STMT_START { \
500 if (!SIZE_ONLY) \
a5e7bc51 501 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
502 Simple_vFAIL2(m, a1); \
503} STMT_END
b45f050a
JF
504
505
506/*
507 * Like Simple_vFAIL(), but accepts three arguments.
508 */
ccb2c380 509#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 510 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
511 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
512 (int)offset, RExC_precomp, RExC_precomp + offset); \
513} STMT_END
b45f050a
JF
514
515/*
516 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
517 */
ccb2c380
MP
518#define vFAIL3(m,a1,a2) STMT_START { \
519 if (!SIZE_ONLY) \
a5e7bc51 520 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
521 Simple_vFAIL3(m, a1, a2); \
522} STMT_END
b45f050a
JF
523
524/*
525 * Like Simple_vFAIL(), but accepts four arguments.
526 */
ccb2c380 527#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 528 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
529 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
530 (int)offset, RExC_precomp, RExC_precomp + offset); \
531} STMT_END
b45f050a 532
668c081a 533#define ckWARNreg(loc,m) STMT_START { \
a28509cc 534 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
535 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
536 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
537} STMT_END
538
668c081a 539#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 540 const IV offset = loc - RExC_precomp; \
d1d15184 541 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
542 m REPORT_LOCATION, \
543 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
544} STMT_END
545
2335b3d3
KW
546#define ckWARN2regdep(loc,m, a1) STMT_START { \
547 const IV offset = loc - RExC_precomp; \
548 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
549 m REPORT_LOCATION, \
550 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
551} STMT_END
552
668c081a 553#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 554 const IV offset = loc - RExC_precomp; \
668c081a 555 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
556 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
557} STMT_END
558
559#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 560 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
561 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
562 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
563} STMT_END
564
668c081a
NC
565#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
566 const IV offset = loc - RExC_precomp; \
567 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
568 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
569} STMT_END
570
ccb2c380 571#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 572 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
573 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
574 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
575} STMT_END
576
668c081a
NC
577#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
578 const IV offset = loc - RExC_precomp; \
579 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
580 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
581} STMT_END
582
ccb2c380 583#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 584 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
585 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
586 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
587} STMT_END
9d1d55b5 588
8615cb43 589
cd439c50 590/* Allow for side effects in s */
ccb2c380
MP
591#define REGC(c,s) STMT_START { \
592 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
593} STMT_END
cd439c50 594
fac92740
MJD
595/* Macros for recording node offsets. 20001227 mjd@plover.com
596 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
597 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
598 * Element 0 holds the number n.
07be1b83 599 * Position is 1 indexed.
fac92740 600 */
7122b237
YO
601#ifndef RE_TRACK_PATTERN_OFFSETS
602#define Set_Node_Offset_To_R(node,byte)
603#define Set_Node_Offset(node,byte)
604#define Set_Cur_Node_Offset
605#define Set_Node_Length_To_R(node,len)
606#define Set_Node_Length(node,len)
607#define Set_Node_Cur_Length(node)
608#define Node_Offset(n)
609#define Node_Length(n)
610#define Set_Node_Offset_Length(node,offset,len)
611#define ProgLen(ri) ri->u.proglen
612#define SetProgLen(ri,x) ri->u.proglen = x
613#else
614#define ProgLen(ri) ri->u.offsets[0]
615#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
616#define Set_Node_Offset_To_R(node,byte) STMT_START { \
617 if (! SIZE_ONLY) { \
618 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 619 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 620 if((node) < 0) { \
551405c4 621 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
622 } else { \
623 RExC_offsets[2*(node)-1] = (byte); \
624 } \
625 } \
626} STMT_END
627
628#define Set_Node_Offset(node,byte) \
629 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
630#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
631
632#define Set_Node_Length_To_R(node,len) STMT_START { \
633 if (! SIZE_ONLY) { \
634 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 635 __LINE__, (int)(node), (int)(len))); \
ccb2c380 636 if((node) < 0) { \
551405c4 637 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
638 } else { \
639 RExC_offsets[2*(node)] = (len); \
640 } \
641 } \
642} STMT_END
643
644#define Set_Node_Length(node,len) \
645 Set_Node_Length_To_R((node)-RExC_emit_start, len)
646#define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
647#define Set_Node_Cur_Length(node) \
648 Set_Node_Length(node, RExC_parse - parse_start)
fac92740
MJD
649
650/* Get offsets and lengths */
651#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
652#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
653
07be1b83
YO
654#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
655 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
656 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
657} STMT_END
7122b237 658#endif
07be1b83
YO
659
660#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
661#define EXPERIMENTAL_INPLACESCAN
f427392e 662#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 663
304ee84b
YO
664#define DEBUG_STUDYDATA(str,data,depth) \
665DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 666 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
667 "%*s" str "Pos:%"IVdf"/%"IVdf \
668 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
669 (int)(depth)*2, "", \
670 (IV)((data)->pos_min), \
671 (IV)((data)->pos_delta), \
304ee84b 672 (UV)((data)->flags), \
1de06328 673 (IV)((data)->whilem_c), \
304ee84b
YO
674 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
675 is_inf ? "INF " : "" \
1de06328
YO
676 ); \
677 if ((data)->last_found) \
678 PerlIO_printf(Perl_debug_log, \
679 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
680 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
681 SvPVX_const((data)->last_found), \
682 (IV)((data)->last_end), \
683 (IV)((data)->last_start_min), \
684 (IV)((data)->last_start_max), \
685 ((data)->longest && \
686 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
687 SvPVX_const((data)->longest_fixed), \
688 (IV)((data)->offset_fixed), \
689 ((data)->longest && \
690 (data)->longest==&((data)->longest_float)) ? "*" : "", \
691 SvPVX_const((data)->longest_float), \
692 (IV)((data)->offset_float_min), \
693 (IV)((data)->offset_float_max) \
694 ); \
695 PerlIO_printf(Perl_debug_log,"\n"); \
696});
697
653099ff 698/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 699 Update the longest found anchored substring and the longest found
653099ff
GS
700 floating substrings if needed. */
701
4327152a 702STATIC void
304ee84b 703S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 704{
e1ec3a88
AL
705 const STRLEN l = CHR_SVLEN(data->last_found);
706 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 707 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 708
7918f24d
NC
709 PERL_ARGS_ASSERT_SCAN_COMMIT;
710
c277df42 711 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 712 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
713 if (*data->longest == data->longest_fixed) {
714 data->offset_fixed = l ? data->last_start_min : data->pos_min;
715 if (data->flags & SF_BEFORE_EOL)
b81d288d 716 data->flags
c277df42
IZ
717 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
718 else
719 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 720 data->minlen_fixed=minlenp;
1de06328 721 data->lookbehind_fixed=0;
a0ed51b3 722 }
304ee84b 723 else { /* *data->longest == data->longest_float */
c277df42 724 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
725 data->offset_float_max = (l
726 ? data->last_start_max
c277df42 727 : data->pos_min + data->pos_delta);
304ee84b 728 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 729 data->offset_float_max = I32_MAX;
c277df42 730 if (data->flags & SF_BEFORE_EOL)
b81d288d 731 data->flags
c277df42
IZ
732 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
733 else
734 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
735 data->minlen_float=minlenp;
736 data->lookbehind_float=0;
c277df42
IZ
737 }
738 }
739 SvCUR_set(data->last_found, 0);
0eda9292 740 {
a28509cc 741 SV * const sv = data->last_found;
097eb12c
AL
742 if (SvUTF8(sv) && SvMAGICAL(sv)) {
743 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
744 if (mg)
745 mg->mg_len = 0;
746 }
0eda9292 747 }
c277df42
IZ
748 data->last_end = -1;
749 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 750 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
751}
752
653099ff
GS
753/* Can match anything (initialization) */
754STATIC void
3fffb88a 755S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 756{
7918f24d
NC
757 PERL_ARGS_ASSERT_CL_ANYTHING;
758
f8bef550 759 ANYOF_BITMAP_SETALL(cl);
dd58aee1 760 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
ae6bdd37 761 |ANYOF_NON_UTF8_LATIN1_ALL;
3fffb88a
KW
762
763 /* If any portion of the regex is to operate under locale rules,
764 * initialization includes it. The reason this isn't done for all regexes
765 * is that the optimizer was written under the assumption that locale was
766 * all-or-nothing. Given the complexity and lack of documentation in the
767 * optimizer, and that there are inadequate test cases for locale, so many
768 * parts of it may not work properly, it is safest to avoid locale unless
769 * necessary. */
770 if (RExC_contains_locale) {
9d7a1e63 771 ANYOF_CLASS_SETALL(cl); /* /l uses class */
ae6bdd37 772 cl->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3fffb88a 773 }
9d7a1e63
KW
774 else {
775 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
776 }
653099ff
GS
777}
778
779/* Can match anything (initialization) */
780STATIC int
5f66b61c 781S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
782{
783 int value;
784
7918f24d
NC
785 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
786
aaa51d5e 787 for (value = 0; value <= ANYOF_MAX; value += 2)
653099ff
GS
788 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
789 return 1;
1aa99e6b
IH
790 if (!(cl->flags & ANYOF_UNICODE_ALL))
791 return 0;
10edeb5d 792 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 793 return 0;
653099ff
GS
794 return 1;
795}
796
797/* Can match anything (initialization) */
798STATIC void
e755fd73 799S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 800{
7918f24d
NC
801 PERL_ARGS_ASSERT_CL_INIT;
802
8ecf7187 803 Zero(cl, 1, struct regnode_charclass_class);
653099ff 804 cl->type = ANYOF;
3fffb88a 805 cl_anything(pRExC_state, cl);
1411dba4 806 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
807}
808
1051e1c4
KW
809/* These two functions currently do the exact same thing */
810#define cl_init_zero S_cl_init
653099ff 811
dd58aee1
KW
812/* 'AND' a given class with another one. Can create false positives. 'cl'
813 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
814 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 815STATIC void
5f66b61c 816S_cl_and(struct regnode_charclass_class *cl,
a28509cc 817 const struct regnode_charclass_class *and_with)
653099ff 818{
7918f24d 819 PERL_ARGS_ASSERT_CL_AND;
40d049e4
YO
820
821 assert(and_with->type == ANYOF);
1e6ade67 822
c6b76537 823 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
824 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
825 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 826 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
827 && !(and_with->flags & ANYOF_LOC_FOLD)
828 && !(cl->flags & ANYOF_LOC_FOLD)) {
653099ff
GS
829 int i;
830
831 if (and_with->flags & ANYOF_INVERT)
832 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
833 cl->bitmap[i] &= ~and_with->bitmap[i];
834 else
835 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
836 cl->bitmap[i] &= and_with->bitmap[i];
837 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 838
c6b76537 839 if (and_with->flags & ANYOF_INVERT) {
8951c461 840
c6b76537
KW
841 /* Here, the and'ed node is inverted. Get the AND of the flags that
842 * aren't affected by the inversion. Those that are affected are
843 * handled individually below */
844 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
845 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
846 cl->flags |= affected_flags;
847
848 /* We currently don't know how to deal with things that aren't in the
849 * bitmap, but we know that the intersection is no greater than what
850 * is already in cl, so let there be false positives that get sorted
851 * out after the synthetic start class succeeds, and the node is
852 * matched for real. */
853
854 /* The inversion of these two flags indicate that the resulting
855 * intersection doesn't have them */
856 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
857 cl->flags &= ~ANYOF_UNICODE_ALL;
858 }
c6b76537
KW
859 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
860 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 861 }
1aa99e6b 862 }
c6b76537 863 else { /* and'd node is not inverted */
3ad98780
KW
864 U8 outside_bitmap_but_not_utf8; /* Temp variable */
865
137165a6 866 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
867
868 /* Here 'and_with' doesn't match anything outside the bitmap
869 * (except possibly ANYOF_UNICODE_ALL), which means the
870 * intersection can't either, except for ANYOF_UNICODE_ALL, in
871 * which case we don't know what the intersection is, but it's no
872 * greater than what cl already has, so can just leave it alone,
873 * with possible false positives */
874 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
875 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 876 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 877 }
137165a6 878 }
c6b76537
KW
879 else if (! ANYOF_NONBITMAP(cl)) {
880
881 /* Here, 'and_with' does match something outside the bitmap, and cl
882 * doesn't have a list of things to match outside the bitmap. If
883 * cl can match all code points above 255, the intersection will
3ad98780
KW
884 * be those above-255 code points that 'and_with' matches. If cl
885 * can't match all Unicode code points, it means that it can't
886 * match anything outside the bitmap (since the 'if' that got us
887 * into this block tested for that), so we leave the bitmap empty.
888 */
c6b76537
KW
889 if (cl->flags & ANYOF_UNICODE_ALL) {
890 ARG_SET(cl, ARG(and_with));
3ad98780
KW
891
892 /* and_with's ARG may match things that don't require UTF8.
893 * And now cl's will too, in spite of this being an 'and'. See
894 * the comments below about the kludge */
895 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
896 }
897 }
898 else {
899 /* Here, both 'and_with' and cl match something outside the
900 * bitmap. Currently we do not do the intersection, so just match
901 * whatever cl had at the beginning. */
902 }
903
904
3ad98780
KW
905 /* Take the intersection of the two sets of flags. However, the
906 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
907 * kludge around the fact that this flag is not treated like the others
908 * which are initialized in cl_anything(). The way the optimizer works
909 * is that the synthetic start class (SSC) is initialized to match
910 * anything, and then the first time a real node is encountered, its
911 * values are AND'd with the SSC's with the result being the values of
912 * the real node. However, there are paths through the optimizer where
913 * the AND never gets called, so those initialized bits are set
914 * inappropriately, which is not usually a big deal, as they just cause
915 * false positives in the SSC, which will just mean a probably
916 * imperceptible slow down in execution. However this bit has a
917 * higher false positive consequence in that it can cause utf8.pm,
918 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
919 * bigger slowdown and also causes significant extra memory to be used.
920 * In order to prevent this, the code now takes a different tack. The
921 * bit isn't set unless some part of the regular expression needs it,
922 * but once set it won't get cleared. This means that these extra
923 * modules won't get loaded unless there was some path through the
924 * pattern that would have required them anyway, and so any false
925 * positives that occur by not ANDing them out when they could be
926 * aren't as severe as they would be if we treated this bit like all
927 * the others */
928 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
929 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 930 cl->flags &= and_with->flags;
3ad98780 931 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 932 }
653099ff
GS
933}
934
dd58aee1
KW
935/* 'OR' a given class with another one. Can create false positives. 'cl'
936 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
937 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 938STATIC void
3fffb88a 939S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 940{
7918f24d
NC
941 PERL_ARGS_ASSERT_CL_OR;
942
653099ff 943 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
944
945 /* Here, the or'd node is to be inverted. This means we take the
946 * complement of everything not in the bitmap, but currently we don't
947 * know what that is, so give up and match anything */
948 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 949 cl_anything(pRExC_state, cl);
c6b76537 950 }
653099ff
GS
951 /* We do not use
952 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
953 * <= (B1 | !B2) | (CL1 | !CL2)
954 * which is wasteful if CL2 is small, but we ignore CL2:
955 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
956 * XXXX Can we handle case-fold? Unclear:
957 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
958 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
959 */
c6b76537 960 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
961 && !(or_with->flags & ANYOF_LOC_FOLD)
962 && !(cl->flags & ANYOF_LOC_FOLD) ) {
653099ff
GS
963 int i;
964
965 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
966 cl->bitmap[i] |= ~or_with->bitmap[i];
967 } /* XXXX: logic is complicated otherwise */
968 else {
3fffb88a 969 cl_anything(pRExC_state, cl);
653099ff 970 }
c6b76537
KW
971
972 /* And, we can just take the union of the flags that aren't affected
973 * by the inversion */
974 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
975
976 /* For the remaining flags:
977 ANYOF_UNICODE_ALL and inverted means to not match anything above
978 255, which means that the union with cl should just be
979 what cl has in it, so can ignore this flag
980 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
981 is 127-255 to match them, but then invert that, so the
982 union with cl should just be what cl has in it, so can
983 ignore this flag
984 */
985 } else { /* 'or_with' is not inverted */
653099ff
GS
986 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
987 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
988 && (!(or_with->flags & ANYOF_LOC_FOLD)
989 || (cl->flags & ANYOF_LOC_FOLD)) ) {
653099ff
GS
990 int i;
991
992 /* OR char bitmap and class bitmap separately */
993 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
994 cl->bitmap[i] |= or_with->bitmap[i];
1e6ade67 995 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
653099ff
GS
996 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
997 cl->classflags[i] |= or_with->classflags[i];
998 cl->flags |= ANYOF_CLASS;
999 }
1000 }
1001 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 1002 cl_anything(pRExC_state, cl);
653099ff 1003 }
9826f543 1004
c6b76537
KW
1005 if (ANYOF_NONBITMAP(or_with)) {
1006
1007 /* Use the added node's outside-the-bit-map match if there isn't a
1008 * conflict. If there is a conflict (both nodes match something
1009 * outside the bitmap, but what they match outside is not the same
1010 * pointer, and hence not easily compared until XXX we extend
1011 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
1012 * match everything outside the bitmap. If that stuff is all above
1013 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1014 if (! ANYOF_NONBITMAP(cl)) {
1015 ARG_SET(cl, ARG(or_with));
1016 }
1017 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1018
1019 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1020 cl_anything(pRExC_state, cl);
1021 }
1022 else {
1023 cl->flags |= ANYOF_UNICODE_ALL;
1024 }
c6b76537 1025 }
4c34a693 1026 }
0b9668ee
KW
1027
1028 /* Take the union */
1029 cl->flags |= or_with->flags;
1aa99e6b 1030 }
653099ff
GS
1031}
1032
a3621e74
YO
1033#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1034#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1035#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1036#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1037
3dab1dad
YO
1038
1039#ifdef DEBUGGING
07be1b83 1040/*
2b8b4781
NC
1041 dump_trie(trie,widecharmap,revcharmap)
1042 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1043 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1044
1045 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1046 The _interim_ variants are used for debugging the interim
1047 tables that are used to generate the final compressed
1048 representation which is what dump_trie expects.
1049
486ec47a 1050 Part of the reason for their existence is to provide a form
3dab1dad 1051 of documentation as to how the different representations function.
07be1b83
YO
1052
1053*/
3dab1dad
YO
1054
1055/*
3dab1dad
YO
1056 Dumps the final compressed table form of the trie to Perl_debug_log.
1057 Used for debugging make_trie().
1058*/
b9a59e08 1059
3dab1dad 1060STATIC void
2b8b4781
NC
1061S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1062 AV *revcharmap, U32 depth)
3dab1dad
YO
1063{
1064 U32 state;
ab3bbdeb 1065 SV *sv=sv_newmortal();
55eed653 1066 int colwidth= widecharmap ? 6 : 4;
2e64971a 1067 U16 word;
3dab1dad
YO
1068 GET_RE_DEBUG_FLAGS_DECL;
1069
7918f24d 1070 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1071
3dab1dad
YO
1072 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1073 (int)depth * 2 + 2,"",
1074 "Match","Base","Ofs" );
1075
1076 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1077 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1078 if ( tmp ) {
ab3bbdeb
YO
1079 PerlIO_printf( Perl_debug_log, "%*s",
1080 colwidth,
ddc5bc0f 1081 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1082 PL_colors[0], PL_colors[1],
1083 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1084 PERL_PV_ESCAPE_FIRSTCHAR
1085 )
1086 );
3dab1dad
YO
1087 }
1088 }
1089 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1090 (int)depth * 2 + 2,"");
1091
1092 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1093 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1094 PerlIO_printf( Perl_debug_log, "\n");
1095
1e2e3d02 1096 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1097 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1098
1099 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1100
1101 if ( trie->states[ state ].wordnum ) {
1102 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1103 } else {
1104 PerlIO_printf( Perl_debug_log, "%6s", "" );
1105 }
1106
1107 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1108
1109 if ( base ) {
1110 U32 ofs = 0;
1111
1112 while( ( base + ofs < trie->uniquecharcount ) ||
1113 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1114 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1115 ofs++;
1116
1117 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1118
1119 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1120 if ( ( base + ofs >= trie->uniquecharcount ) &&
1121 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1122 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1123 {
ab3bbdeb
YO
1124 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1125 colwidth,
3dab1dad
YO
1126 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1127 } else {
ab3bbdeb 1128 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1129 }
1130 }
1131
1132 PerlIO_printf( Perl_debug_log, "]");
1133
1134 }
1135 PerlIO_printf( Perl_debug_log, "\n" );
1136 }
2e64971a
DM
1137 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1138 for (word=1; word <= trie->wordcount; word++) {
1139 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1140 (int)word, (int)(trie->wordinfo[word].prev),
1141 (int)(trie->wordinfo[word].len));
1142 }
1143 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1144}
1145/*
3dab1dad
YO
1146 Dumps a fully constructed but uncompressed trie in list form.
1147 List tries normally only are used for construction when the number of
1148 possible chars (trie->uniquecharcount) is very high.
1149 Used for debugging make_trie().
1150*/
1151STATIC void
55eed653 1152S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1153 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1154 U32 depth)
3dab1dad
YO
1155{
1156 U32 state;
ab3bbdeb 1157 SV *sv=sv_newmortal();
55eed653 1158 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1159 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1160
1161 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1162
3dab1dad 1163 /* print out the table precompression. */
ab3bbdeb
YO
1164 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1165 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1166 "------:-----+-----------------\n" );
3dab1dad
YO
1167
1168 for( state=1 ; state < next_alloc ; state ++ ) {
1169 U16 charid;
1170
ab3bbdeb 1171 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1172 (int)depth * 2 + 2,"", (UV)state );
1173 if ( ! trie->states[ state ].wordnum ) {
1174 PerlIO_printf( Perl_debug_log, "%5s| ","");
1175 } else {
1176 PerlIO_printf( Perl_debug_log, "W%4x| ",
1177 trie->states[ state ].wordnum
1178 );
1179 }
1180 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1181 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1182 if ( tmp ) {
1183 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1184 colwidth,
ddc5bc0f 1185 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1186 PL_colors[0], PL_colors[1],
1187 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1188 PERL_PV_ESCAPE_FIRSTCHAR
1189 ) ,
1e2e3d02
YO
1190 TRIE_LIST_ITEM(state,charid).forid,
1191 (UV)TRIE_LIST_ITEM(state,charid).newstate
1192 );
1193 if (!(charid % 10))
664e119d
RGS
1194 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1195 (int)((depth * 2) + 14), "");
1e2e3d02 1196 }
ab3bbdeb
YO
1197 }
1198 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1199 }
1200}
1201
1202/*
3dab1dad
YO
1203 Dumps a fully constructed but uncompressed trie in table form.
1204 This is the normal DFA style state transition table, with a few
1205 twists to facilitate compression later.
1206 Used for debugging make_trie().
1207*/
1208STATIC void
55eed653 1209S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1210 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1211 U32 depth)
3dab1dad
YO
1212{
1213 U32 state;
1214 U16 charid;
ab3bbdeb 1215 SV *sv=sv_newmortal();
55eed653 1216 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1217 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1218
1219 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1220
1221 /*
1222 print out the table precompression so that we can do a visual check
1223 that they are identical.
1224 */
1225
1226 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1227
1228 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1229 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1230 if ( tmp ) {
ab3bbdeb
YO
1231 PerlIO_printf( Perl_debug_log, "%*s",
1232 colwidth,
ddc5bc0f 1233 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1234 PL_colors[0], PL_colors[1],
1235 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1236 PERL_PV_ESCAPE_FIRSTCHAR
1237 )
1238 );
3dab1dad
YO
1239 }
1240 }
1241
1242 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1243
1244 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1245 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1246 }
1247
1248 PerlIO_printf( Perl_debug_log, "\n" );
1249
1250 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1251
1252 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1253 (int)depth * 2 + 2,"",
1254 (UV)TRIE_NODENUM( state ) );
1255
1256 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1257 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1258 if (v)
1259 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1260 else
1261 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1262 }
1263 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1264 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1265 } else {
1266 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1267 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1268 }
1269 }
07be1b83 1270}
3dab1dad
YO
1271
1272#endif
1273
2e64971a 1274
786e8c11
YO
1275/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1276 startbranch: the first branch in the whole branch sequence
1277 first : start branch of sequence of branch-exact nodes.
1278 May be the same as startbranch
1279 last : Thing following the last branch.
1280 May be the same as tail.
1281 tail : item following the branch sequence
1282 count : words in the sequence
1283 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1284 depth : indent depth
3dab1dad 1285
786e8c11 1286Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1287
786e8c11
YO
1288A trie is an N'ary tree where the branches are determined by digital
1289decomposition of the key. IE, at the root node you look up the 1st character and
1290follow that branch repeat until you find the end of the branches. Nodes can be
1291marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1292
786e8c11 1293 /he|she|his|hers/
72f13be8 1294
786e8c11
YO
1295would convert into the following structure. Numbers represent states, letters
1296following numbers represent valid transitions on the letter from that state, if
1297the number is in square brackets it represents an accepting state, otherwise it
1298will be in parenthesis.
07be1b83 1299
786e8c11
YO
1300 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1301 | |
1302 | (2)
1303 | |
1304 (1) +-i->(6)-+-s->[7]
1305 |
1306 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1307
786e8c11
YO
1308 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1309
1310This shows that when matching against the string 'hers' we will begin at state 1
1311read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1312then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1313is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1314single traverse. We store a mapping from accepting to state to which word was
1315matched, and then when we have multiple possibilities we try to complete the
1316rest of the regex in the order in which they occured in the alternation.
1317
1318The only prior NFA like behaviour that would be changed by the TRIE support is
1319the silent ignoring of duplicate alternations which are of the form:
1320
1321 / (DUPE|DUPE) X? (?{ ... }) Y /x
1322
4b714af6 1323Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1324and without the optimisation. With the optimisations dupes will be silently
486ec47a 1325ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1326the following demonstrates:
1327
1328 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1329
1330which prints out 'word' three times, but
1331
1332 'words'=~/(word|word|word)(?{ print $1 })S/
1333
1334which doesnt print it out at all. This is due to other optimisations kicking in.
1335
1336Example of what happens on a structural level:
1337
486ec47a 1338The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1339
1340 1: CURLYM[1] {1,32767}(18)
1341 5: BRANCH(8)
1342 6: EXACT <ac>(16)
1343 8: BRANCH(11)
1344 9: EXACT <ad>(16)
1345 11: BRANCH(14)
1346 12: EXACT <ab>(16)
1347 16: SUCCEED(0)
1348 17: NOTHING(18)
1349 18: END(0)
1350
1351This would be optimizable with startbranch=5, first=5, last=16, tail=16
1352and should turn into:
1353
1354 1: CURLYM[1] {1,32767}(18)
1355 5: TRIE(16)
1356 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1357 <ac>
1358 <ad>
1359 <ab>
1360 16: SUCCEED(0)
1361 17: NOTHING(18)
1362 18: END(0)
1363
1364Cases where tail != last would be like /(?foo|bar)baz/:
1365
1366 1: BRANCH(4)
1367 2: EXACT <foo>(8)
1368 4: BRANCH(7)
1369 5: EXACT <bar>(8)
1370 7: TAIL(8)
1371 8: EXACT <baz>(10)
1372 10: END(0)
1373
1374which would be optimizable with startbranch=1, first=1, last=7, tail=8
1375and would end up looking like:
1376
1377 1: TRIE(8)
1378 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1379 <foo>
1380 <bar>
1381 7: TAIL(8)
1382 8: EXACT <baz>(10)
1383 10: END(0)
1384
1385 d = uvuni_to_utf8_flags(d, uv, 0);
1386
1387is the recommended Unicode-aware way of saying
1388
1389 *(d++) = uv;
1390*/
1391
fab2782b 1392#define TRIE_STORE_REVCHAR(val) \
786e8c11 1393 STMT_START { \
73031816 1394 if (UTF) { \
fab2782b 1395 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1396 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
fab2782b 1397 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
73031816
NC
1398 SvCUR_set(zlopp, kapow - flrbbbbb); \
1399 SvPOK_on(zlopp); \
1400 SvUTF8_on(zlopp); \
1401 av_push(revcharmap, zlopp); \
1402 } else { \
fab2782b 1403 char ooooff = (char)val; \
73031816
NC
1404 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1405 } \
1406 } STMT_END
786e8c11 1407
fab2782b
YO
1408#define TRIE_READ_CHAR STMT_START { \
1409 wordlen++; \
1410 if ( UTF ) { \
1411 /* if it is UTF then it is either already folded, or does not need folding */ \
1412 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1413 } \
1414 else if (folder == PL_fold_latin1) { \
1415 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1416 if ( foldlen > 0 ) { \
1417 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1418 foldlen -= len; \
1419 scan += len; \
1420 len = 0; \
1421 } else { \
1422 len = 1; \
1423 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1424 skiplen = UNISKIP(uvc); \
1425 foldlen -= skiplen; \
1426 scan = foldbuf + skiplen; \
1427 } \
1428 } else { \
1429 /* raw data, will be folded later if needed */ \
1430 uvc = (U32)*uc; \
1431 len = 1; \
1432 } \
786e8c11
YO
1433} STMT_END
1434
1435
1436
1437#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1438 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1439 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1440 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1441 } \
1442 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1443 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1444 TRIE_LIST_CUR( state )++; \
1445} STMT_END
07be1b83 1446
786e8c11
YO
1447#define TRIE_LIST_NEW(state) STMT_START { \
1448 Newxz( trie->states[ state ].trans.list, \
1449 4, reg_trie_trans_le ); \
1450 TRIE_LIST_CUR( state ) = 1; \
1451 TRIE_LIST_LEN( state ) = 4; \
1452} STMT_END
07be1b83 1453
786e8c11
YO
1454#define TRIE_HANDLE_WORD(state) STMT_START { \
1455 U16 dupe= trie->states[ state ].wordnum; \
1456 regnode * const noper_next = regnext( noper ); \
1457 \
786e8c11
YO
1458 DEBUG_r({ \
1459 /* store the word for dumping */ \
1460 SV* tmp; \
1461 if (OP(noper) != NOTHING) \
740cce10 1462 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1463 else \
740cce10 1464 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1465 av_push( trie_words, tmp ); \
786e8c11
YO
1466 }); \
1467 \
1468 curword++; \
2e64971a
DM
1469 trie->wordinfo[curword].prev = 0; \
1470 trie->wordinfo[curword].len = wordlen; \
1471 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1472 \
1473 if ( noper_next < tail ) { \
1474 if (!trie->jump) \
c944940b 1475 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1476 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1477 if (!jumper) \
1478 jumper = noper_next; \
1479 if (!nextbranch) \
1480 nextbranch= regnext(cur); \
1481 } \
1482 \
1483 if ( dupe ) { \
2e64971a
DM
1484 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1485 /* chain, so that when the bits of chain are later */\
1486 /* linked together, the dups appear in the chain */\
1487 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1488 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1489 } else { \
1490 /* we haven't inserted this word yet. */ \
1491 trie->states[ state ].wordnum = curword; \
1492 } \
1493} STMT_END
07be1b83 1494
3dab1dad 1495
786e8c11
YO
1496#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1497 ( ( base + charid >= ucharcount \
1498 && base + charid < ubound \
1499 && state == trie->trans[ base - ucharcount + charid ].check \
1500 && trie->trans[ base - ucharcount + charid ].next ) \
1501 ? trie->trans[ base - ucharcount + charid ].next \
1502 : ( state==1 ? special : 0 ) \
1503 )
3dab1dad 1504
786e8c11
YO
1505#define MADE_TRIE 1
1506#define MADE_JUMP_TRIE 2
1507#define MADE_EXACT_TRIE 4
3dab1dad 1508
a3621e74 1509STATIC I32
786e8c11 1510S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1511{
27da23d5 1512 dVAR;
a3621e74
YO
1513 /* first pass, loop through and scan words */
1514 reg_trie_data *trie;
55eed653 1515 HV *widecharmap = NULL;
2b8b4781 1516 AV *revcharmap = newAV();
a3621e74 1517 regnode *cur;
9f7f3913 1518 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1519 STRLEN len = 0;
1520 UV uvc = 0;
1521 U16 curword = 0;
1522 U32 next_alloc = 0;
786e8c11
YO
1523 regnode *jumper = NULL;
1524 regnode *nextbranch = NULL;
7f69552c 1525 regnode *convert = NULL;
2e64971a 1526 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1527 /* we just use folder as a flag in utf8 */
1e696034 1528 const U8 * folder = NULL;
a3621e74 1529
2b8b4781
NC
1530#ifdef DEBUGGING
1531 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1532 AV *trie_words = NULL;
1533 /* along with revcharmap, this only used during construction but both are
1534 * useful during debugging so we store them in the struct when debugging.
8e11feef 1535 */
2b8b4781
NC
1536#else
1537 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1538 STRLEN trie_charcount=0;
3dab1dad 1539#endif
2b8b4781 1540 SV *re_trie_maxbuff;
a3621e74 1541 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1542
1543 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1544#ifndef DEBUGGING
1545 PERL_UNUSED_ARG(depth);
1546#endif
a3621e74 1547
1e696034 1548 switch (flags) {
c46d03cf 1549 case EXACT: break;
2f7f8cb1 1550 case EXACTFA:
fab2782b
YO
1551 case EXACTFU_SS:
1552 case EXACTFU_TRICKYFOLD:
1e696034
KW
1553 case EXACTFU: folder = PL_fold_latin1; break;
1554 case EXACTF: folder = PL_fold; break;
1555 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1556 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1557 }
1558
c944940b 1559 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1560 trie->refcount = 1;
3dab1dad 1561 trie->startstate = 1;
786e8c11 1562 trie->wordcount = word_count;
f8fc2ecf 1563 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1564 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1565 if (flags == EXACT)
c944940b 1566 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1567 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1568 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1569
a3621e74 1570 DEBUG_r({
2b8b4781 1571 trie_words = newAV();
a3621e74 1572 });
a3621e74 1573
0111c4fd 1574 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1575 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1576 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1577 }
df826430 1578 DEBUG_TRIE_COMPILE_r({
3dab1dad 1579 PerlIO_printf( Perl_debug_log,
786e8c11 1580 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1581 (int)depth * 2 + 2, "",
1582 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1583 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1584 (int)depth);
3dab1dad 1585 });
7f69552c
YO
1586
1587 /* Find the node we are going to overwrite */
1588 if ( first == startbranch && OP( last ) != BRANCH ) {
1589 /* whole branch chain */
1590 convert = first;
1591 } else {
1592 /* branch sub-chain */
1593 convert = NEXTOPER( first );
1594 }
1595
a3621e74
YO
1596 /* -- First loop and Setup --
1597
1598 We first traverse the branches and scan each word to determine if it
1599 contains widechars, and how many unique chars there are, this is
1600 important as we have to build a table with at least as many columns as we
1601 have unique chars.
1602
1603 We use an array of integers to represent the character codes 0..255
38a44b82 1604 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1605 native representation of the character value as the key and IV's for the
1606 coded index.
1607
1608 *TODO* If we keep track of how many times each character is used we can
1609 remap the columns so that the table compression later on is more
3b753521 1610 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1611 middle and the least common are on the outside. IMO this would be better
1612 than a most to least common mapping as theres a decent chance the most
1613 common letter will share a node with the least common, meaning the node
486ec47a 1614 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1615 case is when we have the least common nodes twice.
1616
1617 */
1618
a3621e74 1619 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1620 regnode *noper = NEXTOPER( cur );
e1ec3a88 1621 const U8 *uc = (U8*)STRING( noper );
df826430 1622 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1623 STRLEN foldlen = 0;
1624 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1625 STRLEN skiplen = 0;
2af232bd 1626 const U8 *scan = (U8*)NULL;
07be1b83 1627 U32 wordlen = 0; /* required init */
02daf0ab
YO
1628 STRLEN chars = 0;
1629 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1630
3dab1dad 1631 if (OP(noper) == NOTHING) {
df826430
YO
1632 regnode *noper_next= regnext(noper);
1633 if (noper_next != tail && OP(noper_next) == flags) {
1634 noper = noper_next;
1635 uc= (U8*)STRING(noper);
1636 e= uc + STR_LEN(noper);
1637 trie->minlen= STR_LEN(noper);
1638 } else {
1639 trie->minlen= 0;
1640 continue;
1641 }
3dab1dad 1642 }
df826430 1643
fab2782b 1644 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1645 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1646 regardless of encoding */
fab2782b
YO
1647 if (OP( noper ) == EXACTFU_SS) {
1648 /* false positives are ok, so just set this */
1649 TRIE_BITMAP_SET(trie,0xDF);
1650 }
1651 }
a3621e74 1652 for ( ; uc < e ; uc += len ) {
3dab1dad 1653 TRIE_CHARCOUNT(trie)++;
a3621e74 1654 TRIE_READ_CHAR;
3dab1dad 1655 chars++;
a3621e74 1656 if ( uvc < 256 ) {
fab2782b
YO
1657 if ( folder ) {
1658 U8 folded= folder[ (U8) uvc ];
1659 if ( !trie->charmap[ folded ] ) {
1660 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1661 TRIE_STORE_REVCHAR( folded );
1662 }
1663 }
a3621e74
YO
1664 if ( !trie->charmap[ uvc ] ) {
1665 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1666 TRIE_STORE_REVCHAR( uvc );
a3621e74 1667 }
02daf0ab 1668 if ( set_bit ) {
62012aee
KW
1669 /* store the codepoint in the bitmap, and its folded
1670 * equivalent. */
fab2782b 1671 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1672
1673 /* store the folded codepoint */
fab2782b 1674 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1675
1676 if ( !UTF ) {
1677 /* store first byte of utf8 representation of
acdf4139
KW
1678 variant codepoints */
1679 if (! UNI_IS_INVARIANT(uvc)) {
1680 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1681 }
1682 }
02daf0ab
YO
1683 set_bit = 0; /* We've done our bit :-) */
1684 }
a3621e74
YO
1685 } else {
1686 SV** svpp;
55eed653
NC
1687 if ( !widecharmap )
1688 widecharmap = newHV();
a3621e74 1689
55eed653 1690 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1691
1692 if ( !svpp )
e4584336 1693 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1694
1695 if ( !SvTRUE( *svpp ) ) {
1696 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1697 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1698 }
1699 }
1700 }
3dab1dad 1701 if( cur == first ) {
fab2782b
YO
1702 trie->minlen = chars;
1703 trie->maxlen = chars;
3dab1dad 1704 } else if (chars < trie->minlen) {
fab2782b 1705 trie->minlen = chars;
3dab1dad 1706 } else if (chars > trie->maxlen) {
fab2782b
YO
1707 trie->maxlen = chars;
1708 }
1709 if (OP( noper ) == EXACTFU_SS) {
1710 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1711 if (trie->minlen > 1)
1712 trie->minlen= 1;
1713 }
1714 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1715 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1716 * - We assume that any such sequence might match a 2 byte string */
1717 if (trie->minlen > 2 )
1718 trie->minlen= 2;
3dab1dad
YO
1719 }
1720
a3621e74
YO
1721 } /* end first pass */
1722 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1723 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1724 (int)depth * 2 + 2,"",
55eed653 1725 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1726 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1727 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1728 );
a3621e74
YO
1729
1730 /*
1731 We now know what we are dealing with in terms of unique chars and
1732 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1733 representation using a flat table will take. If it's over a reasonable
1734 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1735 conservative but potentially much slower representation using an array
1736 of lists.
1737
1738 At the end we convert both representations into the same compressed
1739 form that will be used in regexec.c for matching with. The latter
1740 is a form that cannot be used to construct with but has memory
1741 properties similar to the list form and access properties similar
1742 to the table form making it both suitable for fast searches and
1743 small enough that its feasable to store for the duration of a program.
1744
1745 See the comment in the code where the compressed table is produced
1746 inplace from the flat tabe representation for an explanation of how
1747 the compression works.
1748
1749 */
1750
1751
2e64971a
DM
1752 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1753 prev_states[1] = 0;
1754
3dab1dad 1755 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1756 /*
1757 Second Pass -- Array Of Lists Representation
1758
1759 Each state will be represented by a list of charid:state records
1760 (reg_trie_trans_le) the first such element holds the CUR and LEN
1761 points of the allocated array. (See defines above).
1762
1763 We build the initial structure using the lists, and then convert
1764 it into the compressed table form which allows faster lookups
1765 (but cant be modified once converted).
a3621e74
YO
1766 */
1767
a3621e74
YO
1768 STRLEN transcount = 1;
1769
1e2e3d02
YO
1770 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1771 "%*sCompiling trie using list compiler\n",
1772 (int)depth * 2 + 2, ""));
686b73d4 1773
c944940b
JH
1774 trie->states = (reg_trie_state *)
1775 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1776 sizeof(reg_trie_state) );
a3621e74
YO
1777 TRIE_LIST_NEW(1);
1778 next_alloc = 2;
1779
1780 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1781
df826430 1782 regnode *noper = NEXTOPER( cur );
c445ea15 1783 U8 *uc = (U8*)STRING( noper );
df826430 1784 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1785 U32 state = 1; /* required init */
1786 U16 charid = 0; /* sanity init */
1787 U8 *scan = (U8*)NULL; /* sanity init */
1788 STRLEN foldlen = 0; /* required init */
07be1b83 1789 U32 wordlen = 0; /* required init */
c445ea15 1790 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1791 STRLEN skiplen = 0;
c445ea15 1792
df826430
YO
1793 if (OP(noper) == NOTHING) {
1794 regnode *noper_next= regnext(noper);
1795 if (noper_next != tail && OP(noper_next) == flags) {
1796 noper = noper_next;
1797 uc= (U8*)STRING(noper);
1798 e= uc + STR_LEN(noper);
1799 }
1800 }
1801
3dab1dad 1802 if (OP(noper) != NOTHING) {
786e8c11 1803 for ( ; uc < e ; uc += len ) {
c445ea15 1804
786e8c11 1805 TRIE_READ_CHAR;
c445ea15 1806
786e8c11
YO
1807 if ( uvc < 256 ) {
1808 charid = trie->charmap[ uvc ];
c445ea15 1809 } else {
55eed653 1810 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1811 if ( !svpp ) {
1812 charid = 0;
1813 } else {
1814 charid=(U16)SvIV( *svpp );
1815 }
c445ea15 1816 }
786e8c11
YO
1817 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1818 if ( charid ) {
a3621e74 1819
786e8c11
YO
1820 U16 check;
1821 U32 newstate = 0;
a3621e74 1822
786e8c11
YO
1823 charid--;
1824 if ( !trie->states[ state ].trans.list ) {
1825 TRIE_LIST_NEW( state );
c445ea15 1826 }
786e8c11
YO
1827 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1828 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1829 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1830 break;
1831 }
1832 }
1833 if ( ! newstate ) {
1834 newstate = next_alloc++;
2e64971a 1835 prev_states[newstate] = state;
786e8c11
YO
1836 TRIE_LIST_PUSH( state, charid, newstate );
1837 transcount++;
1838 }
1839 state = newstate;
1840 } else {
1841 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1842 }
a28509cc 1843 }
c445ea15 1844 }
3dab1dad 1845 TRIE_HANDLE_WORD(state);
a3621e74
YO
1846
1847 } /* end second pass */
1848
1e2e3d02
YO
1849 /* next alloc is the NEXT state to be allocated */
1850 trie->statecount = next_alloc;
c944940b
JH
1851 trie->states = (reg_trie_state *)
1852 PerlMemShared_realloc( trie->states,
1853 next_alloc
1854 * sizeof(reg_trie_state) );
a3621e74 1855
3dab1dad 1856 /* and now dump it out before we compress it */
2b8b4781
NC
1857 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1858 revcharmap, next_alloc,
1859 depth+1)
1e2e3d02 1860 );
a3621e74 1861
c944940b
JH
1862 trie->trans = (reg_trie_trans *)
1863 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1864 {
1865 U32 state;
a3621e74
YO
1866 U32 tp = 0;
1867 U32 zp = 0;
1868
1869
1870 for( state=1 ; state < next_alloc ; state ++ ) {
1871 U32 base=0;
1872
1873 /*
1874 DEBUG_TRIE_COMPILE_MORE_r(
1875 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1876 );
1877 */
1878
1879 if (trie->states[state].trans.list) {
1880 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1881 U16 maxid=minid;
a28509cc 1882 U16 idx;
a3621e74
YO
1883
1884 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1885 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1886 if ( forid < minid ) {
1887 minid=forid;
1888 } else if ( forid > maxid ) {
1889 maxid=forid;
1890 }
a3621e74
YO
1891 }
1892 if ( transcount < tp + maxid - minid + 1) {
1893 transcount *= 2;
c944940b
JH
1894 trie->trans = (reg_trie_trans *)
1895 PerlMemShared_realloc( trie->trans,
446bd890
NC
1896 transcount
1897 * sizeof(reg_trie_trans) );
a3621e74
YO
1898 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1899 }
1900 base = trie->uniquecharcount + tp - minid;
1901 if ( maxid == minid ) {
1902 U32 set = 0;
1903 for ( ; zp < tp ; zp++ ) {
1904 if ( ! trie->trans[ zp ].next ) {
1905 base = trie->uniquecharcount + zp - minid;
1906 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1907 trie->trans[ zp ].check = state;
1908 set = 1;
1909 break;
1910 }
1911 }
1912 if ( !set ) {
1913 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1914 trie->trans[ tp ].check = state;
1915 tp++;
1916 zp = tp;
1917 }
1918 } else {
1919 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1920 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1921 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1922 trie->trans[ tid ].check = state;
1923 }
1924 tp += ( maxid - minid + 1 );
1925 }
1926 Safefree(trie->states[ state ].trans.list);
1927 }
1928 /*
1929 DEBUG_TRIE_COMPILE_MORE_r(
1930 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1931 );
1932 */
1933 trie->states[ state ].trans.base=base;
1934 }
cc601c31 1935 trie->lasttrans = tp + 1;
a3621e74
YO
1936 }
1937 } else {
1938 /*
1939 Second Pass -- Flat Table Representation.
1940
1941 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1942 We know that we will need Charcount+1 trans at most to store the data
1943 (one row per char at worst case) So we preallocate both structures
1944 assuming worst case.
1945
1946 We then construct the trie using only the .next slots of the entry
1947 structs.
1948
3b753521 1949 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1950 make compression both faster and easier by keeping track of how many non
1951 zero fields are in the node.
1952
1953 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1954 transition.
1955
1956 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1957 number representing the first entry of the node, and state as a
1958 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1959 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1960 are 2 entrys per node. eg:
1961
1962 A B A B
1963 1. 2 4 1. 3 7
1964 2. 0 3 3. 0 5
1965 3. 0 0 5. 0 0
1966 4. 0 0 7. 0 0
1967
1968 The table is internally in the right hand, idx form. However as we also
1969 have to deal with the states array which is indexed by nodenum we have to
1970 use TRIE_NODENUM() to convert.
1971
1972 */
1e2e3d02
YO
1973 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1974 "%*sCompiling trie using table compiler\n",
1975 (int)depth * 2 + 2, ""));
3dab1dad 1976
c944940b
JH
1977 trie->trans = (reg_trie_trans *)
1978 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1979 * trie->uniquecharcount + 1,
1980 sizeof(reg_trie_trans) );
1981 trie->states = (reg_trie_state *)
1982 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1983 sizeof(reg_trie_state) );
a3621e74
YO
1984 next_alloc = trie->uniquecharcount + 1;
1985
3dab1dad 1986
a3621e74
YO
1987 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1988
df826430 1989 regnode *noper = NEXTOPER( cur );
a28509cc 1990 const U8 *uc = (U8*)STRING( noper );
df826430 1991 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1992
1993 U32 state = 1; /* required init */
1994
1995 U16 charid = 0; /* sanity init */
1996 U32 accept_state = 0; /* sanity init */
1997 U8 *scan = (U8*)NULL; /* sanity init */
1998
1999 STRLEN foldlen = 0; /* required init */
07be1b83 2000 U32 wordlen = 0; /* required init */
fab2782b 2001 STRLEN skiplen = 0;
a3621e74
YO
2002 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2003
df826430
YO
2004 if (OP(noper) == NOTHING) {
2005 regnode *noper_next= regnext(noper);
2006 if (noper_next != tail && OP(noper_next) == flags) {
2007 noper = noper_next;
2008 uc= (U8*)STRING(noper);
2009 e= uc + STR_LEN(noper);
2010 }
2011 }
fab2782b 2012
3dab1dad 2013 if ( OP(noper) != NOTHING ) {
786e8c11 2014 for ( ; uc < e ; uc += len ) {
a3621e74 2015
786e8c11 2016 TRIE_READ_CHAR;
a3621e74 2017
786e8c11
YO
2018 if ( uvc < 256 ) {
2019 charid = trie->charmap[ uvc ];
2020 } else {
55eed653 2021 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2022 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2023 }
786e8c11
YO
2024 if ( charid ) {
2025 charid--;
2026 if ( !trie->trans[ state + charid ].next ) {
2027 trie->trans[ state + charid ].next = next_alloc;
2028 trie->trans[ state ].check++;
2e64971a
DM
2029 prev_states[TRIE_NODENUM(next_alloc)]
2030 = TRIE_NODENUM(state);
786e8c11
YO
2031 next_alloc += trie->uniquecharcount;
2032 }
2033 state = trie->trans[ state + charid ].next;
2034 } else {
2035 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2036 }
2037 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2038 }
a3621e74 2039 }
3dab1dad
YO
2040 accept_state = TRIE_NODENUM( state );
2041 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2042
2043 } /* end second pass */
2044
3dab1dad 2045 /* and now dump it out before we compress it */
2b8b4781
NC
2046 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2047 revcharmap,
2048 next_alloc, depth+1));
a3621e74 2049
a3621e74
YO
2050 {
2051 /*
2052 * Inplace compress the table.*
2053
2054 For sparse data sets the table constructed by the trie algorithm will
2055 be mostly 0/FAIL transitions or to put it another way mostly empty.
2056 (Note that leaf nodes will not contain any transitions.)
2057
2058 This algorithm compresses the tables by eliminating most such
2059 transitions, at the cost of a modest bit of extra work during lookup:
2060
2061 - Each states[] entry contains a .base field which indicates the
2062 index in the state[] array wheres its transition data is stored.
2063
3b753521 2064 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2065
2066 - If .base is nonzero then charid is added to it to find an entry in
2067 the trans array.
2068
2069 -If trans[states[state].base+charid].check!=state then the
2070 transition is taken to be a 0/Fail transition. Thus if there are fail
2071 transitions at the front of the node then the .base offset will point
2072 somewhere inside the previous nodes data (or maybe even into a node
2073 even earlier), but the .check field determines if the transition is
2074 valid.
2075
786e8c11 2076 XXX - wrong maybe?
a3621e74 2077 The following process inplace converts the table to the compressed
3b753521 2078 table: We first do not compress the root node 1,and mark all its
a3621e74 2079 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2080 allows us to do a DFA construction from the compressed table later,
2081 and ensures that any .base pointers we calculate later are greater
2082 than 0.
a3621e74
YO
2083
2084 - We set 'pos' to indicate the first entry of the second node.
2085
2086 - We then iterate over the columns of the node, finding the first and
2087 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2088 and set the .check pointers accordingly, and advance pos
2089 appropriately and repreat for the next node. Note that when we copy
2090 the next pointers we have to convert them from the original
2091 NODEIDX form to NODENUM form as the former is not valid post
2092 compression.
2093
2094 - If a node has no transitions used we mark its base as 0 and do not
2095 advance the pos pointer.
2096
2097 - If a node only has one transition we use a second pointer into the
2098 structure to fill in allocated fail transitions from other states.
2099 This pointer is independent of the main pointer and scans forward
2100 looking for null transitions that are allocated to a state. When it
2101 finds one it writes the single transition into the "hole". If the
786e8c11 2102 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2103
2104 - Once compressed we can Renew/realloc the structures to release the
2105 excess space.
2106
2107 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2108 specifically Fig 3.47 and the associated pseudocode.
2109
2110 demq
2111 */
a3b680e6 2112 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2113 U32 state, charid;
a3621e74 2114 U32 pos = 0, zp=0;
1e2e3d02 2115 trie->statecount = laststate;
a3621e74
YO
2116
2117 for ( state = 1 ; state < laststate ; state++ ) {
2118 U8 flag = 0;
a28509cc
AL
2119 const U32 stateidx = TRIE_NODEIDX( state );
2120 const U32 o_used = trie->trans[ stateidx ].check;
2121 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2122 trie->trans[ stateidx ].check = 0;
2123
2124 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2125 if ( flag || trie->trans[ stateidx + charid ].next ) {
2126 if ( trie->trans[ stateidx + charid ].next ) {
2127 if (o_used == 1) {
2128 for ( ; zp < pos ; zp++ ) {
2129 if ( ! trie->trans[ zp ].next ) {
2130 break;
2131 }
2132 }
2133 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2134 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2135 trie->trans[ zp ].check = state;
2136 if ( ++zp > pos ) pos = zp;
2137 break;
2138 }
2139 used--;
2140 }
2141 if ( !flag ) {
2142 flag = 1;
2143 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2144 }
2145 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2146 trie->trans[ pos ].check = state;
2147 pos++;
2148 }
2149 }
2150 }
cc601c31 2151 trie->lasttrans = pos + 1;
c944940b
JH
2152 trie->states = (reg_trie_state *)
2153 PerlMemShared_realloc( trie->states, laststate
2154 * sizeof(reg_trie_state) );
a3621e74 2155 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2156 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2157 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2158 (int)depth * 2 + 2,"",
2159 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2160 (IV)next_alloc,
2161 (IV)pos,
a3621e74
YO
2162 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2163 );
2164
2165 } /* end table compress */
2166 }
1e2e3d02
YO
2167 DEBUG_TRIE_COMPILE_MORE_r(
2168 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2169 (int)depth * 2 + 2, "",
2170 (UV)trie->statecount,
2171 (UV)trie->lasttrans)
2172 );
cc601c31 2173 /* resize the trans array to remove unused space */
c944940b
JH
2174 trie->trans = (reg_trie_trans *)
2175 PerlMemShared_realloc( trie->trans, trie->lasttrans
2176 * sizeof(reg_trie_trans) );
a3621e74 2177
3b753521 2178 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2179 U8 nodetype =(U8)(flags & 0xFF);
2180 char *str=NULL;
786e8c11 2181
07be1b83 2182#ifdef DEBUGGING
e62cc96a 2183 regnode *optimize = NULL;
7122b237
YO
2184#ifdef RE_TRACK_PATTERN_OFFSETS
2185
b57a0404
JH
2186 U32 mjd_offset = 0;
2187 U32 mjd_nodelen = 0;
7122b237
YO
2188#endif /* RE_TRACK_PATTERN_OFFSETS */
2189#endif /* DEBUGGING */
a3621e74 2190 /*
3dab1dad
YO
2191 This means we convert either the first branch or the first Exact,
2192 depending on whether the thing following (in 'last') is a branch
2193 or not and whther first is the startbranch (ie is it a sub part of
2194 the alternation or is it the whole thing.)
3b753521 2195 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2196 the whole branch sequence, including the first.
a3621e74 2197 */
3dab1dad 2198 /* Find the node we are going to overwrite */
7f69552c 2199 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2200 /* branch sub-chain */
3dab1dad 2201 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2202#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2203 DEBUG_r({
2204 mjd_offset= Node_Offset((convert));
2205 mjd_nodelen= Node_Length((convert));
2206 });
7122b237 2207#endif
7f69552c 2208 /* whole branch chain */
7122b237
YO
2209 }
2210#ifdef RE_TRACK_PATTERN_OFFSETS
2211 else {
7f69552c
YO
2212 DEBUG_r({
2213 const regnode *nop = NEXTOPER( convert );
2214 mjd_offset= Node_Offset((nop));
2215 mjd_nodelen= Node_Length((nop));
2216 });
07be1b83
YO
2217 }
2218 DEBUG_OPTIMISE_r(
2219 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2220 (int)depth * 2 + 2, "",
786e8c11 2221 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2222 );
7122b237 2223#endif
3dab1dad
YO
2224 /* But first we check to see if there is a common prefix we can
2225 split out as an EXACT and put in front of the TRIE node. */
2226 trie->startstate= 1;
55eed653 2227 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2228 U32 state;
1e2e3d02 2229 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2230 U32 ofs = 0;
8e11feef
RGS
2231 I32 idx = -1;
2232 U32 count = 0;
2233 const U32 base = trie->states[ state ].trans.base;
a3621e74 2234
3dab1dad 2235 if ( trie->states[state].wordnum )
8e11feef 2236 count = 1;
a3621e74 2237
8e11feef 2238 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2239 if ( ( base + ofs >= trie->uniquecharcount ) &&
2240 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2241 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2242 {
3dab1dad 2243 if ( ++count > 1 ) {
2b8b4781 2244 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2245 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2246 if ( state == 1 ) break;
3dab1dad
YO
2247 if ( count == 2 ) {
2248 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2249 DEBUG_OPTIMISE_r(
8e11feef
RGS
2250 PerlIO_printf(Perl_debug_log,
2251 "%*sNew Start State=%"UVuf" Class: [",
2252 (int)depth * 2 + 2, "",
786e8c11 2253 (UV)state));
be8e71aa 2254 if (idx >= 0) {
2b8b4781 2255 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2256 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2257
3dab1dad 2258 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2259 if ( folder )
2260 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2261 DEBUG_OPTIMISE_r(
f1f66076 2262 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2263 );
8e11feef
RGS
2264 }
2265 }
2266 TRIE_BITMAP_SET(trie,*ch);
2267 if ( folder )
2268 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2269 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2270 }
2271 idx = ofs;
2272 }
3dab1dad
YO
2273 }
2274 if ( count == 1 ) {
2b8b4781 2275 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2276 STRLEN len;
2277 char *ch = SvPV( *tmp, len );
de734bd5
A
2278 DEBUG_OPTIMISE_r({
2279 SV *sv=sv_newmortal();
8e11feef
RGS
2280 PerlIO_printf( Perl_debug_log,
2281 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2282 (int)depth * 2 + 2, "",
de734bd5
A
2283 (UV)state, (UV)idx,
2284 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2285 PL_colors[0], PL_colors[1],
2286 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2287 PERL_PV_ESCAPE_FIRSTCHAR
2288 )
2289 );
2290 });
3dab1dad
YO
2291 if ( state==1 ) {
2292 OP( convert ) = nodetype;
2293 str=STRING(convert);
2294 STR_LEN(convert)=0;
2295 }
c490c714
YO
2296 STR_LEN(convert) += len;
2297 while (len--)
de734bd5 2298 *str++ = *ch++;
8e11feef 2299 } else {
f9049ba1 2300#ifdef DEBUGGING
8e11feef
RGS
2301 if (state>1)
2302 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2303#endif
8e11feef
RGS
2304 break;
2305 }
2306 }
2e64971a 2307 trie->prefixlen = (state-1);
3dab1dad 2308 if (str) {
8e11feef 2309 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2310 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2311 trie->startstate = state;
07be1b83
YO
2312 trie->minlen -= (state - 1);
2313 trie->maxlen -= (state - 1);
33809eae
JH
2314#ifdef DEBUGGING
2315 /* At least the UNICOS C compiler choked on this
2316 * being argument to DEBUG_r(), so let's just have
2317 * it right here. */
2318 if (
2319#ifdef PERL_EXT_RE_BUILD
2320 1
2321#else
2322 DEBUG_r_TEST
2323#endif
2324 ) {
2325 regnode *fix = convert;
2326 U32 word = trie->wordcount;
2327 mjd_nodelen++;
2328 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2329 while( ++fix < n ) {
2330 Set_Node_Offset_Length(fix, 0, 0);
2331 }
2332 while (word--) {
2333 SV ** const tmp = av_fetch( trie_words, word, 0 );
2334 if (tmp) {
2335 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2336 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2337 else
2338 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2339 }
2340 }
2341 }
2342#endif
8e11feef
RGS
2343 if (trie->maxlen) {
2344 convert = n;
2345 } else {
3dab1dad 2346 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2347 DEBUG_r(optimize= n);
3dab1dad
YO
2348 }
2349 }
2350 }
a5ca303d
YO
2351 if (!jumper)
2352 jumper = last;
3dab1dad 2353 if ( trie->maxlen ) {
8e11feef
RGS
2354 NEXT_OFF( convert ) = (U16)(tail - convert);
2355 ARG_SET( convert, data_slot );
786e8c11
YO
2356 /* Store the offset to the first unabsorbed branch in
2357 jump[0], which is otherwise unused by the jump logic.
2358 We use this when dumping a trie and during optimisation. */
2359 if (trie->jump)
7f69552c 2360 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2361
6c48061a
YO
2362 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2363 * and there is a bitmap
2364 * and the first "jump target" node we found leaves enough room
2365 * then convert the TRIE node into a TRIEC node, with the bitmap
2366 * embedded inline in the opcode - this is hypothetically faster.
2367 */
2368 if ( !trie->states[trie->startstate].wordnum
2369 && trie->bitmap
2370 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2371 {
2372 OP( convert ) = TRIEC;
2373 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2374 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2375 trie->bitmap= NULL;
2376 } else
2377 OP( convert ) = TRIE;
a3621e74 2378
3dab1dad
YO
2379 /* store the type in the flags */
2380 convert->flags = nodetype;
a5ca303d
YO
2381 DEBUG_r({
2382 optimize = convert
2383 + NODE_STEP_REGNODE
2384 + regarglen[ OP( convert ) ];
2385 });
2386 /* XXX We really should free up the resource in trie now,
2387 as we won't use them - (which resources?) dmq */
3dab1dad 2388 }
a3621e74 2389 /* needed for dumping*/
e62cc96a 2390 DEBUG_r(if (optimize) {
07be1b83 2391 regnode *opt = convert;
bcdf7404 2392
e62cc96a 2393 while ( ++opt < optimize) {
07be1b83
YO
2394 Set_Node_Offset_Length(opt,0,0);
2395 }
786e8c11
YO
2396 /*
2397 Try to clean up some of the debris left after the
2398 optimisation.
a3621e74 2399 */
786e8c11 2400 while( optimize < jumper ) {
07be1b83 2401 mjd_nodelen += Node_Length((optimize));
a3621e74 2402 OP( optimize ) = OPTIMIZED;
07be1b83 2403 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2404 optimize++;
2405 }
07be1b83 2406 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2407 });
2408 } /* end node insert */
2e64971a
DM
2409
2410 /* Finish populating the prev field of the wordinfo array. Walk back
2411 * from each accept state until we find another accept state, and if
2412 * so, point the first word's .prev field at the second word. If the
2413 * second already has a .prev field set, stop now. This will be the
2414 * case either if we've already processed that word's accept state,
3b753521
FN
2415 * or that state had multiple words, and the overspill words were
2416 * already linked up earlier.
2e64971a
DM
2417 */
2418 {
2419 U16 word;
2420 U32 state;
2421 U16 prev;
2422
2423 for (word=1; word <= trie->wordcount; word++) {
2424 prev = 0;
2425 if (trie->wordinfo[word].prev)
2426 continue;
2427 state = trie->wordinfo[word].accept;
2428 while (state) {
2429 state = prev_states[state];
2430 if (!state)
2431 break;
2432 prev = trie->states[state].wordnum;
2433 if (prev)
2434 break;
2435 }
2436 trie->wordinfo[word].prev = prev;
2437 }
2438 Safefree(prev_states);
2439 }
2440
2441
2442 /* and now dump out the compressed format */
2443 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2444
55eed653 2445 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2446#ifdef DEBUGGING
2447 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2448 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2449#else
2450 SvREFCNT_dec(revcharmap);
07be1b83 2451#endif
786e8c11
YO
2452 return trie->jump
2453 ? MADE_JUMP_TRIE
2454 : trie->startstate>1
2455 ? MADE_EXACT_TRIE
2456 : MADE_TRIE;
2457}
2458
2459STATIC void
2460S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2461{
3b753521 2462/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2463
2464 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2465 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2466 ISBN 0-201-10088-6
2467
2468 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2469 suffix of the current state's 'word' that is also a proper prefix of another word in our
2470 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2471 the DFA not to have to restart after its tried and failed a word at a given point, it
2472 simply continues as though it had been matching the other word in the first place.
2473 Consider
2474 'abcdgu'=~/abcdefg|cdgu/
2475 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2476 fail, which would bring us to the state representing 'd' in the second word where we would
2477 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2478 */
2479 /* add a fail transition */
3251b653
NC
2480 const U32 trie_offset = ARG(source);
2481 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2482 U32 *q;
2483 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2484 const U32 numstates = trie->statecount;
786e8c11
YO
2485 const U32 ubound = trie->lasttrans + ucharcount;
2486 U32 q_read = 0;
2487 U32 q_write = 0;
2488 U32 charid;
2489 U32 base = trie->states[ 1 ].trans.base;
2490 U32 *fail;
2491 reg_ac_data *aho;
2492 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2493 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2494
2495 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2496#ifndef DEBUGGING
2497 PERL_UNUSED_ARG(depth);
2498#endif
2499
2500
2501 ARG_SET( stclass, data_slot );
c944940b 2502 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2503 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2504 aho->trie=trie_offset;
446bd890
NC
2505 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2506 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2507 Newxz( q, numstates, U32);
c944940b 2508 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2509 aho->refcount = 1;
2510 fail = aho->fail;
2511 /* initialize fail[0..1] to be 1 so that we always have
2512 a valid final fail state */
2513 fail[ 0 ] = fail[ 1 ] = 1;
2514
2515 for ( charid = 0; charid < ucharcount ; charid++ ) {
2516 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2517 if ( newstate ) {
2518 q[ q_write ] = newstate;
2519 /* set to point at the root */
2520 fail[ q[ q_write++ ] ]=1;
2521 }
2522 }
2523 while ( q_read < q_write) {
2524 const U32 cur = q[ q_read++ % numstates ];
2525 base = trie->states[ cur ].trans.base;
2526
2527 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2528 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2529 if (ch_state) {
2530 U32 fail_state = cur;
2531 U32 fail_base;
2532 do {
2533 fail_state = fail[ fail_state ];
2534 fail_base = aho->states[ fail_state ].trans.base;
2535 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2536
2537 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2538 fail[ ch_state ] = fail_state;
2539 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2540 {
2541 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2542 }
2543 q[ q_write++ % numstates] = ch_state;
2544 }
2545 }
2546 }
2547 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2548 when we fail in state 1, this allows us to use the
2549 charclass scan to find a valid start char. This is based on the principle
2550 that theres a good chance the string being searched contains lots of stuff
2551 that cant be a start char.
2552 */
2553 fail[ 0 ] = fail[ 1 ] = 0;
2554 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2555 PerlIO_printf(Perl_debug_log,
2556 "%*sStclass Failtable (%"UVuf" states): 0",
2557 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2558 );
786e8c11
YO
2559 for( q_read=1; q_read<numstates; q_read++ ) {
2560 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2561 }
2562 PerlIO_printf(Perl_debug_log, "\n");
2563 });
2564 Safefree(q);
2565 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2566}
2567
786e8c11 2568
a3621e74 2569/*
5d1c421c
JH
2570 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2571 * These need to be revisited when a newer toolchain becomes available.
2572 */
2573#if defined(__sparc64__) && defined(__GNUC__)
2574# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2575# undef SPARC64_GCC_WORKAROUND
2576# define SPARC64_GCC_WORKAROUND 1
2577# endif
2578#endif
2579
07be1b83 2580#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2581 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2582 SV * const mysv=sv_newmortal(); \
2583 regnode *Next = regnext(scan); \
2584 regprop(RExC_rx, mysv, scan); \
7f69552c 2585 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2586 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2587 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2588 }});
07be1b83 2589
1de06328 2590
bb914485 2591/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
2592 * one. The regop may be changed if the node(s) contain certain sequences that
2593 * require special handling. The joining is only done if:
bb914485
KW
2594 * 1) there is room in the current conglomerated node to entirely contain the
2595 * next one.
2596 * 2) they are the exact same node type
2597 *
87b8b349 2598 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
2599 * these get optimized out
2600 *
0a982f06
KW
2601 * If a node is to match under /i (folded), the number of characters it matches
2602 * can be different than its character length if it contains a multi-character
2603 * fold. *min_subtract is set to the total delta of the input nodes.
bb914485 2604 *
a0c4c608
KW
2605 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2606 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2607 *
bb914485 2608 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
2609 * multi-character fold sequences. It's been wrong in Perl for a very long
2610 * time. There are three code points in Unicode whose multi-character folds
2611 * were long ago discovered to mess things up. The previous designs for
2612 * dealing with these involved assigning a special node for them. This
2613 * approach doesn't work, as evidenced by this example:
a0c4c608 2614 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
0a982f06
KW
2615 * Both these fold to "sss", but if the pattern is parsed to create a node that
2616 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
2617 * successful match would have to cross the node's boundary. The new approach
2618 * that hopefully generally solves the problem generates an EXACTFU_SS node
2619 * that is "sss".
2620 *
0a982f06
KW
2621 * It turns out that there are problems with all multi-character folds, and not
2622 * just these three. Now the code is general, for all such cases, but the
2623 * three still have some special handling. The approach taken is:
2624 * 1) This routine examines each EXACTFish node that could contain multi-
2625 * character fold sequences. It returns in *min_subtract how much to
9d071ca8 2626 * subtract from the the actual length of the string to get a real minimum
0a982f06
KW
2627 * match length; it is 0 if there are no multi-char folds. This delta is
2628 * used by the caller to adjust the min length of the match, and the delta
2629 * between min and max, so that the optimizer doesn't reject these
2630 * possibilities based on size constraints.
2631 * 2) Certain of these sequences require special handling by the trie code,
2632 * so, if found, this code changes the joined node type to special ops:
2633 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2634 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2635 * is used for an EXACTFU node that contains at least one "ss" sequence in
2636 * it. For non-UTF-8 patterns and strings, this is the only case where
2637 * there is a possible fold length change. That means that a regular
2638 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2639 * with length changes, and so can be processed faster. regexec.c takes
2640 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2641 * pre-folded by regcomp.c. This saves effort in regex matching.
87b8b349 2642 * However, the pre-folding isn't done for non-UTF8 patterns because the
0a982f06
KW
2643 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2644 * down by forcing the pattern into UTF8 unless necessary. Also what
2645 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2646 * possibilities for the non-UTF8 patterns are quite simple, except for
2647 * the sharp s. All the ones that don't involve a UTF-8 target string are
2648 * members of a fold-pair, and arrays are set up for all of them so that
2649 * the other member of the pair can be found quickly. Code elsewhere in
2650 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2651 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2652 * described in the next item.
bb914485
KW
2653 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2654 * 'ss' or not is not knowable at compile time. It will match iff the
2655 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2656 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
6012a526 2657 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
bb914485
KW
2658 * described in item 3). An assumption that the optimizer part of
2659 * regexec.c (probably unwittingly) makes is that a character in the
2660 * pattern corresponds to at most a single character in the target string.
2661 * (And I do mean character, and not byte here, unlike other parts of the
2662 * documentation that have never been updated to account for multibyte
2663 * Unicode.) This assumption is wrong only in this case, as all other
2664 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2665 * virtue of having this file pre-fold UTF-8 patterns. I'm
2666 * reluctant to try to change this assumption, so instead the code punts.
9d071ca8
KW
2667 * This routine examines EXACTF nodes for the sharp s, and returns a
2668 * boolean indicating whether or not the node is an EXACTF node that
2669 * contains a sharp s. When it is true, the caller sets a flag that later
2670 * causes the optimizer in this file to not set values for the floating
2671 * and fixed string lengths, and thus avoids the optimizer code in
2672 * regexec.c that makes the invalid assumption. Thus, there is no
2673 * optimization based on string lengths for EXACTF nodes that contain the
2674 * sharp s. This only happens for /id rules (which means the pattern
2675 * isn't in UTF-8).
bb914485 2676 */
1de06328 2677
9d071ca8 2678#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2679 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2680 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2681
be8e71aa 2682STATIC U32
9d071ca8 2683S_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
2684 /* Merge several consecutive EXACTish nodes into one. */
2685 regnode *n = regnext(scan);
2686 U32 stringok = 1;
2687 regnode *next = scan + NODE_SZ_STR(scan);
2688 U32 merged = 0;
2689 U32 stopnow = 0;
2690#ifdef DEBUGGING
2691 regnode *stop = scan;
72f13be8 2692 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2693#else
d47053eb
RGS
2694 PERL_UNUSED_ARG(depth);
2695#endif
7918f24d
NC
2696
2697 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2698#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2699 PERL_UNUSED_ARG(flags);
2700 PERL_UNUSED_ARG(val);
07be1b83 2701#endif
07be1b83 2702 DEBUG_PEEP("join",scan,depth);
bb914485 2703
3f410cf6
KW
2704 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2705 * EXACT ones that are mergeable to the current one. */
2706 while (n
2707 && (PL_regkind[OP(n)] == NOTHING
2708 || (stringok && OP(n) == OP(scan)))
07be1b83 2709 && NEXT_OFF(n)
3f410cf6
KW
2710 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2711 {
07be1b83
YO
2712
2713 if (OP(n) == TAIL || n > next)
2714 stringok = 0;
2715 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2716 DEBUG_PEEP("skip:",n,depth);
2717 NEXT_OFF(scan) += NEXT_OFF(n);
2718 next = n + NODE_STEP_REGNODE;
2719#ifdef DEBUGGING
2720 if (stringok)
2721 stop = n;
2722#endif
2723 n = regnext(n);
2724 }
2725 else if (stringok) {
786e8c11 2726 const unsigned int oldl = STR_LEN(scan);
07be1b83 2727 regnode * const nnext = regnext(n);
b2230d39 2728
87b8b349
KW
2729 /* XXX I (khw) kind of doubt that this works on platforms where
2730 * U8_MAX is above 255 because of lots of other assumptions */
b2230d39
KW
2731 if (oldl + STR_LEN(n) > U8_MAX)
2732 break;
07be1b83
YO
2733
2734 DEBUG_PEEP("merg",n,depth);
07be1b83 2735 merged++;
b2230d39 2736
07be1b83
YO
2737 NEXT_OFF(scan) += NEXT_OFF(n);
2738 STR_LEN(scan) += STR_LEN(n);
2739 next = n + NODE_SZ_STR(n);
2740 /* Now we can overwrite *n : */
2741 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2742#ifdef DEBUGGING
2743 stop = next - 1;
2744#endif
2745 n = nnext;
2746 if (stopnow) break;
2747 }
2748
d47053eb
RGS
2749#ifdef EXPERIMENTAL_INPLACESCAN
2750 if (flags && !NEXT_OFF(n)) {
2751 DEBUG_PEEP("atch", val, depth);
2752 if (reg_off_by_arg[OP(n)]) {
2753 ARG_SET(n, val - n);
2754 }
2755 else {
2756 NEXT_OFF(n) = val - n;
2757 }
2758 stopnow = 1;
2759 }
07be1b83
YO
2760#endif
2761 }
2c2b7f86 2762
9d071ca8 2763 *min_subtract = 0;
f758bddf 2764 *has_exactf_sharp_s = FALSE;
f646642f 2765
3f410cf6
KW
2766 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2767 * can now analyze for sequences of problematic code points. (Prior to
2768 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2769 * hence missed). The sequences only happen in folding, hence for any
2770 * non-EXACT EXACTish node */
86d6fcad 2771 if (OP(scan) != EXACT) {
0a982f06
KW
2772 const U8 * const s0 = (U8*) STRING(scan);
2773 const U8 * s = s0;
2774 const U8 * const s_end = s0 + STR_LEN(scan);
f758bddf
KW
2775
2776 /* One pass is made over the node's string looking for all the
2777 * possibilities. to avoid some tests in the loop, there are two main
2778 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2779 * non-UTF-8 */
2780 if (UTF) {
86d6fcad 2781
0a982f06
KW
2782 /* Examine the string for a multi-character fold sequence. UTF-8
2783 * patterns have all characters pre-folded by the time this code is
2784 * executed */
2785 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2786 length sequence we are looking for is 2 */
86d6fcad 2787 {
0a982f06
KW
2788 int count = 0;
2789 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2790 if (! len) { /* Not a multi-char fold: get next char */
2791 s += UTF8SKIP(s);
2792 continue;
2793 }
bb914485 2794
0a982f06
KW
2795 /* Nodes with 'ss' require special handling, except for EXACTFL
2796 * and EXACTFA for which there is no multi-char fold to this */
2797 if (len == 2 && *s == 's' && *(s+1) == 's'
2798 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2799 {
2800 count = 2;
2801 OP(scan) = EXACTFU_SS;
2802 s += 2;
2803 }
2804 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2805 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2806 COMBINING_DIAERESIS_UTF8
2807 COMBINING_ACUTE_ACCENT_UTF8,
2808 6)
2809 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2810 COMBINING_DIAERESIS_UTF8
2811 COMBINING_ACUTE_ACCENT_UTF8,
2812 6)))
2813 {
2814 count = 3;
2815
2816 /* These two folds require special handling by trie's, so
2817 * change the node type to indicate this. If EXACTFA and
2818 * EXACTFL were ever to be handled by trie's, this would
2819 * have to be changed. If this node has already been
2820 * changed to EXACTFU_SS in this loop, leave it as is. (I
2821 * (khw) think it doesn't matter in regexec.c for UTF
2822 * patterns, but no need to change it */
2823 if (OP(scan) == EXACTFU) {
2824 OP(scan) = EXACTFU_TRICKYFOLD;
2825 }
2826 s += 6;
2827 }
2828 else { /* Here is a generic multi-char fold. */
2829 const U8* multi_end = s + len;
2830
2831 /* Count how many characters in it. In the case of /l and
2832 * /aa, no folds which contain ASCII code points are
2833 * allowed, so check for those, and skip if found. (In
2834 * EXACTFL, no folds are allowed to any Latin1 code point,
2835 * not just ASCII. But there aren't any of these
2836 * currently, nor ever likely, so don't take the time to
2837 * test for them. The code that generates the
2838 * is_MULTI_foo() macros croaks should one actually get put
2839 * into Unicode .) */
2840 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2841 count = utf8_length(s, multi_end);
2842 s = multi_end;
2843 }
2844 else {
2845 while (s < multi_end) {
2846 if (isASCII(*s)) {
2847 s++;
2848 goto next_iteration;
2849 }
2850 else {
2851 s += UTF8SKIP(s);
2852 }
2853 count++;
2854 }
2855 }
2856 }
f758bddf 2857
0a982f06
KW
2858 /* The delta is how long the sequence is minus 1 (1 is how long
2859 * the character that folds to the sequence is) */
2860 *min_subtract += count - 1;
2861 next_iteration: ;
bb914485
KW
2862 }
2863 }
f758bddf 2864 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
bb914485 2865
0a982f06
KW
2866 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2867 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2868 * nodes can't have multi-char folds to this range (and there are
2db97d41 2869 * no existing ones in the upper latin1 range). In the EXACTF
0a982f06
KW
2870 * case we look also for the sharp s, which can be in the final
2871 * position. Otherwise we can stop looking 1 byte earlier because
2872 * have to find at least two characters for a multi-fold */
f758bddf
KW
2873 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2874
0a982f06
KW
2875 /* The below is perhaps overboard, but this allows us to save a
2876 * test each time through the loop at the expense of a mask. This
2877 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2878 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2879 * are 64. This uses an exclusive 'or' to find that bit and then
2880 * inverts it to form a mask, with just a single 0, in the bit
2881 * position where 'S' and 's' differ. */
2882 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2883 const U8 s_masked = 's' & S_or_s_mask;
2884
2885 while (s < upper) {
40b1ba4f 2886 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
0a982f06
KW
2887 if (! len) { /* Not a multi-char fold. */
2888 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2889 {
2890 *has_exactf_sharp_s = TRUE;
2891 }
2892 s++;
2893 continue;
2894 }
2895
2896 if (len == 2
2897 && ((*s & S_or_s_mask) == s_masked)
2898 && ((*(s+1) & S_or_s_mask) == s_masked))
2899 {
2900
2901 /* EXACTF nodes need to know that the minimum length
2902 * changed so that a sharp s in the string can match this
2903 * ss in the pattern, but they remain EXACTF nodes, as they
2904 * won't match this unless the target string is is UTF-8,
2905 * which we don't know until runtime */
2906 if (OP(scan) != EXACTF) {
2907 OP(scan) = EXACTFU_SS;
2908 }
86d6fcad 2909 }
0a982f06
KW
2910
2911 *min_subtract += len - 1;
2912 s += len;
86d6fcad
KW
2913 }
2914 }
07be1b83 2915 }
3f410cf6 2916
07be1b83 2917#ifdef DEBUGGING
bb789b09
DM
2918 /* Allow dumping but overwriting the collection of skipped
2919 * ops and/or strings with fake optimized ops */
07be1b83
YO
2920 n = scan + NODE_SZ_STR(scan);
2921 while (n <= stop) {
bb789b09
DM
2922 OP(n) = OPTIMIZED;
2923 FLAGS(n) = 0;
2924 NEXT_OFF(n) = 0;
07be1b83
YO
2925 n++;
2926 }
2927#endif
2928 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2929 return stopnow;
2930}
2931
486ec47a 2932/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2933 Finds fixed substrings. */
2934
a0288114 2935/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2936 to the position after last scanned or to NULL. */
2937
40d049e4
YO
2938#define INIT_AND_WITHP \
2939 assert(!and_withp); \
2940 Newx(and_withp,1,struct regnode_charclass_class); \
2941 SAVEFREEPV(and_withp)
07be1b83 2942
b515a41d 2943/* this is a chain of data about sub patterns we are processing that
486ec47a 2944 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2945 we can simulate recursion without losing state. */
2946struct scan_frame;
2947typedef struct scan_frame {
2948 regnode *last; /* last node to process in this frame */
2949 regnode *next; /* next node to process when last is reached */
2950 struct scan_frame *prev; /*previous frame*/
2951 I32 stop; /* what stopparen do we use */
2952} scan_frame;
2953
304ee84b
YO
2954
2955#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2956
e1d1eefb
YO
2957#define CASE_SYNST_FNC(nAmE) \
2958case nAmE: \
2959 if (flags & SCF_DO_STCLASS_AND) { \
2960 for (value = 0; value < 256; value++) \
2961 if (!is_ ## nAmE ## _cp(value)) \
2962 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2963 } \
2964 else { \
2965 for (value = 0; value < 256; value++) \
2966 if (is_ ## nAmE ## _cp(value)) \
2967 ANYOF_BITMAP_SET(data->start_class, value); \
2968 } \
2969 break; \
2970case N ## nAmE: \
2971 if (flags & SCF_DO_STCLASS_AND) { \
2972 for (value = 0; value < 256; value++) \
2973 if (is_ ## nAmE ## _cp(value)) \
2974 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2975 } \
2976 else { \
2977 for (value = 0; value < 256; value++) \
2978 if (!is_ ## nAmE ## _cp(value)) \
2979 ANYOF_BITMAP_SET(data->start_class, value); \
2980 } \
2981 break
2982
2983
2984
76e3520e 2985STATIC I32
40d049e4 2986S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 2987 I32 *minlenp, I32 *deltap,
40d049e4
YO
2988 regnode *last,
2989 scan_data_t *data,
2990 I32 stopparen,
2991 U8* recursed,
2992 struct regnode_charclass_class *and_withp,
2993 U32 flags, U32 depth)
c277df42
IZ
2994 /* scanp: Start here (read-write). */
2995 /* deltap: Write maxlen-minlen here. */
2996 /* last: Stop before this one. */
40d049e4
YO
2997 /* data: string data about the pattern */
2998 /* stopparen: treat close N as END */
2999 /* recursed: which subroutines have we recursed into */
3000 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 3001{
97aff369 3002 dVAR;
2d608413
KW
3003 I32 min = 0; /* There must be at least this number of characters to match */
3004 I32 pars = 0, code;
c277df42
IZ
3005 regnode *scan = *scanp, *next;
3006 I32 delta = 0;
3007 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 3008 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
3009 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3010 scan_data_t data_fake;
a3621e74 3011 SV *re_trie_maxbuff = NULL;
786e8c11 3012 regnode *first_non_open = scan;
e2e6a0f1 3013 I32 stopmin = I32_MAX;
8aa23a47 3014 scan_frame *frame = NULL;
a3621e74 3015 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 3016
7918f24d
NC
3017 PERL_ARGS_ASSERT_STUDY_CHUNK;
3018
13a24bad 3019#ifdef DEBUGGING
40d049e4 3020 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 3021#endif
40d049e4 3022
786e8c11 3023 if ( depth == 0 ) {
40d049e4 3024 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
3025 first_non_open=regnext(first_non_open);
3026 }
3027
b81d288d 3028
8aa23a47
YO
3029 fake_study_recurse:
3030 while ( scan && OP(scan) != END && scan < last ){
2d608413
KW
3031 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3032 node length to get a real minimum (because
3033 the folded version may be shorter) */
f758bddf 3034 bool has_exactf_sharp_s = FALSE;
8aa23a47 3035 /* Peephole optimizer: */
304ee84b 3036 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3037 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3038
3039 /* Its not clear to khw or hv why this is done here, and not in the
3040 * clauses that deal with EXACT nodes. khw's guess is that it's
3041 * because of a previous design */
9d071ca8 3042 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3043
3044 /* Follow the next-chain of the current node and optimize
3045 away all the NOTHINGs from it. */
3046 if (OP(scan) != CURLYX) {
3047 const int max = (reg_off_by_arg[OP(scan)]
3048 ? I32_MAX
3049 /* I32 may be smaller than U16 on CRAYs! */
3050 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3051 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3052 int noff;
3053 regnode *n = scan;
686b73d4 3054
8aa23a47
YO
3055 /* Skip NOTHING and LONGJMP. */
3056 while ((n = regnext(n))
3057 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3058 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3059 && off + noff < max)
3060 off += noff;
3061 if (reg_off_by_arg[OP(scan)])
3062 ARG(scan) = off;
3063 else
3064 NEXT_OFF(scan) = off;
3065 }
a3621e74 3066
c277df42 3067
8aa23a47
YO
3068
3069 /* The principal pseudo-switch. Cannot be a switch, since we
3070 look into several different things. */
3071 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3072 || OP(scan) == IFTHEN) {
3073 next = regnext(scan);
3074 code = OP(scan);
3075 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3076
8aa23a47
YO
3077 if (OP(next) == code || code == IFTHEN) {
3078 /* NOTE - There is similar code to this block below for handling
3079 TRIE nodes on a re-study. If you change stuff here check there
3080 too. */
3081 I32 max1 = 0, min1 = I32_MAX, num = 0;
3082 struct regnode_charclass_class accum;
3083 regnode * const startbranch=scan;
686b73d4 3084
8aa23a47 3085 if (flags & SCF_DO_SUBSTR)
304ee84b 3086 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3087 if (flags & SCF_DO_STCLASS)
e755fd73 3088 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3089
3090 while (OP(scan) == code) {
3091 I32 deltanext, minnext, f = 0, fake;
3092 struct regnode_charclass_class this_class;
3093
3094 num++;
3095 data_fake.flags = 0;
3096 if (data) {
3097 data_fake.whilem_c = data->whilem_c;
3098 data_fake.last_closep = data->last_closep;
3099 }
3100 else
3101 data_fake.last_closep = &fake;
58e23c8d
YO
3102
3103 data_fake.pos_delta = delta;
8aa23a47
YO
3104 next = regnext(scan);
3105 scan = NEXTOPER(scan);
3106 if (code != BRANCH)
c277df42 3107 scan = NEXTOPER(scan);
8aa23a47 3108 if (flags & SCF_DO_STCLASS) {
e755fd73 3109 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3110 data_fake.start_class = &this_class;
3111 f = SCF_DO_STCLASS_AND;
58e23c8d 3112 }
8aa23a47
YO
3113 if (flags & SCF_WHILEM_VISITED_POS)
3114 f |= SCF_WHILEM_VISITED_POS;
3115
3116 /* we suppose the run is continuous, last=next...*/
3117 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3118 next, &data_fake,
3119 stopparen, recursed, NULL, f,depth+1);
3120 if (min1 > minnext)
3121 min1 = minnext;
3122 if (max1 < minnext + deltanext)
3123 max1 = minnext + deltanext;
3124 if (deltanext == I32_MAX)
3125 is_inf = is_inf_internal = 1;
3126 scan = next;
3127 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3128 pars++;
3129 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3130 if ( stopmin > minnext)
3131 stopmin = min + min1;
3132 flags &= ~SCF_DO_SUBSTR;
3133 if (data)
3134 data->flags |= SCF_SEEN_ACCEPT;
3135 }
3136 if (data) {
3137 if (data_fake.flags & SF_HAS_EVAL)
3138 data->flags |= SF_HAS_EVAL;
3139 data->whilem_c = data_fake.whilem_c;
3dab1dad 3140 }
8aa23a47 3141 if (flags & SCF_DO_STCLASS)
3fffb88a 3142 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3143 }
3144 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3145 min1 = 0;
3146 if (flags & SCF_DO_SUBSTR) {
3147 data->pos_min += min1;
3148 data->pos_delta += max1 - min1;
3149 if (max1 != min1 || is_inf)
3150 data->longest = &(data->longest_float);
3151 }
3152 min += min1;
3153 delta += max1 - min1;
3154 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3155 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3156 if (min1) {
3157 cl_and(data->start_class, and_withp);
3158 flags &= ~SCF_DO_STCLASS;
653099ff 3159 }
8aa23a47
YO
3160 }
3161 else if (flags & SCF_DO_STCLASS_AND) {
3162 if (min1) {
3163 cl_and(data->start_class, &accum);
3164 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3165 }
8aa23a47
YO
3166 else {
3167 /* Switch to OR mode: cache the old value of
3168 * data->start_class */
3169 INIT_AND_WITHP;
3170 StructCopy(data->start_class, and_withp,
3171 struct regnode_charclass_class);
3172 flags &= ~SCF_DO_STCLASS_AND;
3173 StructCopy(&accum, data->start_class,
3174 struct regnode_charclass_class);
3175 flags |= SCF_DO_STCLASS_OR;
3176 data->start_class->flags |= ANYOF_EOS;
de0c8cb8 3177 }
8aa23a47 3178 }
a3621e74 3179
8aa23a47
YO
3180 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3181 /* demq.
a3621e74 3182
8aa23a47
YO
3183 Assuming this was/is a branch we are dealing with: 'scan' now
3184 points at the item that follows the branch sequence, whatever
3185 it is. We now start at the beginning of the sequence and look
3186 for subsequences of
a3621e74 3187
8aa23a47
YO
3188 BRANCH->EXACT=>x1
3189 BRANCH->EXACT=>x2
3190 tail
a3621e74 3191
8aa23a47 3192 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3193
486ec47a 3194 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3195 element into a trie and then add the subsequent branch exact
3196 strings to the trie.
a3621e74 3197
8aa23a47 3198 We have two cases
a3621e74 3199
3b753521 3200 1. patterns where the whole set of branches can be converted.
a3621e74 3201
8aa23a47 3202 2. patterns where only a subset can be converted.
a3621e74 3203
8aa23a47
YO
3204 In case 1 we can replace the whole set with a single regop
3205 for the trie. In case 2 we need to keep the start and end
3b753521 3206 branches so
a3621e74 3207
8aa23a47
YO
3208 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3209 becomes BRANCH TRIE; BRANCH X;
786e8c11 3210
8aa23a47
YO
3211 There is an additional case, that being where there is a
3212 common prefix, which gets split out into an EXACT like node
3213 preceding the TRIE node.
a3621e74 3214
8aa23a47
YO
3215 If x(1..n)==tail then we can do a simple trie, if not we make
3216 a "jump" trie, such that when we match the appropriate word
486ec47a 3217 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3218 a nested if into a case structure of sorts.
b515a41d 3219
8aa23a47 3220 */
686b73d4 3221
8aa23a47
YO
3222 int made=0;
3223 if (!re_trie_maxbuff) {
3224 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3225 if (!SvIOK(re_trie_maxbuff))
3226 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3227 }
3228 if ( SvIV(re_trie_maxbuff)>=0 ) {
3229 regnode *cur;
3230 regnode *first = (regnode *)NULL;
3231 regnode *last = (regnode *)NULL;
3232 regnode *tail = scan;
fab2782b 3233 U8 trietype = 0;
8aa23a47 3234 U32 count=0;
a3621e74
YO
3235
3236#ifdef DEBUGGING
8aa23a47 3237 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3238#endif
8aa23a47
YO
3239 /* var tail is used because there may be a TAIL
3240 regop in the way. Ie, the exacts will point to the
3241 thing following the TAIL, but the last branch will
3242 point at the TAIL. So we advance tail. If we
3243 have nested (?:) we may have to move through several
3244 tails.
3245 */
3246
3247 while ( OP( tail ) == TAIL ) {
3248 /* this is the TAIL generated by (?:) */
3249 tail = regnext( tail );
3250 }
a3621e74 3251
8aa23a47 3252
df826430 3253 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3254 regprop(RExC_rx, mysv, tail );
3255 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3256 (int)depth * 2 + 2, "",
3257 "Looking for TRIE'able sequences. Tail node is: ",
3258 SvPV_nolen_const( mysv )
3259 );
3260 });
3261
3262 /*
3263
fab2782b
YO
3264 Step through the branches
3265 cur represents each branch,
3266 noper is the first thing to be matched as part of that branch
3267 noper_next is the regnext() of that node.
3268
3269 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3270 via a "jump trie" but we also support building with NOJUMPTRIE,
3271 which restricts the trie logic to structures like /FOO|BAR/.
3272
3273 If noper is a trieable nodetype then the branch is a possible optimization
3274 target. If we are building under NOJUMPTRIE then we require that noper_next
3275 is the same as scan (our current position in the regex program).
3276
3277 Once we have two or more consecutive such branches we can create a
3278 trie of the EXACT's contents and stitch it in place into the program.
3279
3280 If the sequence represents all of the branches in the alternation we
3281 replace the entire thing with a single TRIE node.
3282
3283 Otherwise when it is a subsequence we need to stitch it in place and
3284 replace only the relevant branches. This means the first branch has
3285 to remain as it is used by the alternation logic, and its next pointer,
3286 and needs to be repointed at the item on the branch chain following
3287 the last branch we have optimized away.
3288
3289 This could be either a BRANCH, in which case the subsequence is internal,
3290 or it could be the item following the branch sequence in which case the
3291 subsequence is at the end (which does not necessarily mean the first node
3292 is the start of the alternation).
3293
3294 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3295
3296 optype | trietype
3297 ----------------+-----------
3298 NOTHING | NOTHING
3299 EXACT | EXACT
3300 EXACTFU | EXACTFU
3301 EXACTFU_SS | EXACTFU
3302 EXACTFU_TRICKYFOLD | EXACTFU
3303 EXACTFA | 0
3304
8aa23a47
YO
3305
3306 */
fab2782b
YO
3307#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3308 ( EXACT == (X) ) ? EXACT : \
3309 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3310 0 )
8aa23a47
YO
3311
3312 /* dont use tail as the end marker for this traverse */
3313 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3314 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3315 U8 noper_type = OP( noper );
3316 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3317#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3318 regnode * const noper_next = regnext( noper );
df826430
YO
3319 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3320 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3321#endif
3322
df826430 3323 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3324 regprop(RExC_rx, mysv, cur);
3325 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3326 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3327
3328 regprop(RExC_rx, mysv, noper);
3329 PerlIO_printf( Perl_debug_log, " -> %s",
3330 SvPV_nolen_const(mysv));
3331
3332 if ( noper_next ) {
3333 regprop(RExC_rx, mysv, noper_next );
3334 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3335 SvPV_nolen_const(mysv));
3336 }
df826430
YO
3337 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3338 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3339 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3340 );
8aa23a47 3341 });
fab2782b
YO
3342
3343 /* Is noper a trieable nodetype that can be merged with the
3344 * current trie (if there is one)? */
3345 if ( noper_trietype
3346 &&
3347 (
df826430
YO
3348 ( noper_trietype == NOTHING)
3349 || ( trietype == NOTHING )
a40630bf 3350 || ( trietype == noper_trietype )
fab2782b 3351 )
786e8c11 3352#ifdef NOJUMPTRIE
8aa23a47 3353 && noper_next == tail
786e8c11 3354#endif
8aa23a47
YO
3355 && count < U16_MAX)
3356 {
fab2782b
YO
3357 /* Handle mergable triable node
3358 * Either we are the first node in a new trieable sequence,
3359 * in which case we do some bookkeeping, otherwise we update
3360 * the end pointer. */
fab2782b 3361 if ( !first ) {
3b6759a6 3362 first = cur;
df826430
YO
3363 if ( noper_trietype == NOTHING ) {
3364#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3365 regnode * const noper_next = regnext( noper );
3b6759a6 3366 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3367 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3368#endif
3369
190c1910 3370 if ( noper_next_trietype ) {
df826430 3371 trietype = noper_next_trietype;
190c1910
YO
3372 } else if (noper_next_type) {
3373 /* a NOTHING regop is 1 regop wide. We need at least two
3374 * for a trie so we can't merge this in */
3375 first = NULL;
3376 }
3377 } else {
3378 trietype = noper_trietype;
3b6759a6 3379 }
8aa23a47 3380 } else {
fab2782b
YO
3381 if ( trietype == NOTHING )
3382 trietype = noper_trietype;
8aa23a47
YO
3383 last = cur;
3384 }
df826430
YO
3385 if (first)
3386 count++;
fab2782b
YO
3387 } /* end handle mergable triable node */
3388 else {
3389 /* handle unmergable node -
3390 * noper may either be a triable node which can not be tried
3391 * together with the current trie, or a non triable node */
729aaeb5
YO
3392 if ( last ) {
3393 /* If last is set and trietype is not NOTHING then we have found
3394 * at least two triable branch sequences in a row of a similar
3395 * trietype so we can turn them into a trie. If/when we
3396 * allow NOTHING to start a trie sequence this condition will be
3397 * required, and it isn't expensive so we leave it in for now. */
e6351b37 3398 if ( trietype && trietype != NOTHING )
729aaeb5
YO
3399 make_trie( pRExC_state,
3400 startbranch, first, cur, tail, count,
3401 trietype, depth+1 );
fab2782b 3402 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3403 }
fab2782b 3404 if ( noper_trietype
786e8c11 3405#ifdef NOJUMPTRIE
8aa23a47 3406 && noper_next == tail
786e8c11 3407#endif
8aa23a47 3408 ){
fab2782b 3409 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3410 count = 1;
3411 first = cur;
fab2782b
YO
3412 trietype = noper_trietype;
3413 } else if (first) {
3414 /* if we already saw a first but the current node is not triable then we have
3415 * to reset the first information. */
8aa23a47
YO
3416 count = 0;
3417 first = NULL;
fab2782b 3418 trietype = 0;
8aa23a47 3419 }
fab2782b
YO
3420 } /* end handle unmergable node */
3421 } /* loop over branches */
df826430 3422 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3423 regprop(RExC_rx, mysv, cur);
3424 PerlIO_printf( Perl_debug_log,
3425 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3426 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3427
3428 });
e6351b37 3429 if ( last && trietype ) {
3b6759a6
YO
3430 if ( trietype != NOTHING ) {
3431 /* the last branch of the sequence was part of a trie,
3432 * so we have to construct it here outside of the loop
3433 */
3434 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3435#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3436 if ( ((made == MADE_EXACT_TRIE &&
3437 startbranch == first)
3438 || ( first_non_open == first )) &&
3439 depth==0 ) {
3440 flags |= SCF_TRIE_RESTUDY;
3441 if ( startbranch == first
3442 && scan == tail )
3443 {
3444 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3445 }
8aa23a47 3446 }
3dab1dad 3447#endif
3b6759a6
YO
3448 } else {
3449 /* at this point we know whatever we have is a NOTHING sequence/branch
3450 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3451 */
3452 if ( startbranch == first ) {
3453 regnode *opt;
3454 /* the entire thing is a NOTHING sequence, something like this:
3455 * (?:|) So we can turn it into a plain NOTHING op. */
3456 DEBUG_TRIE_COMPILE_r({
3457 regprop(RExC_rx, mysv, cur);
3458 PerlIO_printf( Perl_debug_log,
3459 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3460 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3461
3462 });
3463 OP(startbranch)= NOTHING;
3464 NEXT_OFF(startbranch)= tail - startbranch;
3465 for ( opt= startbranch + 1; opt < tail ; opt++ )
3466 OP(opt)= OPTIMIZED;
3467 }
3468 }
fab2782b
YO
3469 } /* end if ( last) */
3470 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3471
3472 } /* do trie */
3473
653099ff 3474 }
8aa23a47
YO
3475 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3476 scan = NEXTOPER(NEXTOPER(scan));
3477 } else /* single branch is optimized. */
3478 scan = NEXTOPER(scan);
3479 continue;
3480 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3481 scan_frame *newframe = NULL;
3482 I32 paren;
3483 regnode *start;
3484 regnode *end;
3485
3486 if (OP(scan) != SUSPEND) {
3487 /* set the pointer */
3488 if (OP(scan) == GOSUB) {
3489 paren = ARG(scan);
3490 RExC_recurse[ARG2L(scan)] = scan;
3491 start = RExC_open_parens[paren-1];
3492 end = RExC_close_parens[paren-1];
3493 } else {
3494 paren = 0;
f8fc2ecf 3495 start = RExC_rxi->program + 1;
8aa23a47
YO
3496 end = RExC_opend;
3497 }
3498 if (!recursed) {
3499 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3500 SAVEFREEPV(recursed);
3501 }
3502 if (!PAREN_TEST(recursed,paren+1)) {
3503 PAREN_SET(recursed,paren+1);
3504 Newx(newframe,1,scan_frame);
3505 } else {
3506 if (flags & SCF_DO_SUBSTR) {
304ee84b 3507 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3508 data->longest = &(data->longest_float);
3509 }
3510 is_inf = is_inf_internal = 1;
3511 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3512 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3513 flags &= ~SCF_DO_STCLASS;
3514 }
3515 } else {
3516 Newx(newframe,1,scan_frame);
3517 paren = stopparen;
3518 start = scan+2;
3519 end = regnext(scan);
3520 }
3521 if (newframe) {
3522 assert(start);
3523 assert(end);
3524 SAVEFREEPV(newframe);
3525 newframe->next = regnext(scan);
3526 newframe->last = last;
3527 newframe->stop = stopparen;
3528 newframe->prev = frame;
3529
3530 frame = newframe;
3531 scan = start;
3532 stopparen = paren;
3533 last = end;
3534
3535 continue;
3536 }
3537 }
3538 else if (OP(scan) == EXACT) {
3539 I32 l = STR_LEN(scan);
3540 UV uc;
3541 if (UTF) {
3542 const U8 * const s = (U8*)STRING(scan);
4b88fb76 3543 uc = utf8_to_uvchr_buf(s, s + l, NULL);
8aa23a47 3544 l = utf8_length(s, s + l);
8aa23a47
YO
3545 } else {
3546 uc = *((U8*)STRING(scan));
3547 }
3548 min += l;
3549 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3550 /* The code below prefers earlier match for fixed
3551 offset, later match for variable offset. */
3552 if (data->last_end == -1) { /* Update the start info. */
3553 data->last_start_min = data->pos_min;
3554 data->last_start_max = is_inf
3555 ? I32_MAX : data->pos_min + data->pos_delta;
b515a41d 3556 }
8aa23a47
YO
3557 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3558 if (UTF)
3559 SvUTF8_on(data->last_found);
3560 {
3561 SV * const sv = data->last_found;
3562 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3563 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3564 if (mg && mg->mg_len >= 0)
3565 mg->mg_len += utf8_length((U8*)STRING(scan),
3566 (U8*)STRING(scan)+STR_LEN(scan));
b515a41d 3567 }
8aa23a47
YO
3568 data->last_end = data->pos_min + l;
3569 data->pos_min += l; /* As in the first entry. */
3570 data->flags &= ~SF_BEFORE_EOL;
3571 }
3572 if (flags & SCF_DO_STCLASS_AND) {
3573 /* Check whether it is compatible with what we know already! */
3574 int compat = 1;
3575
54251c2e 3576
486ec47a 3577 /* If compatible, we or it in below. It is compatible if is
54251c2e
KW
3578 * in the bitmp and either 1) its bit or its fold is set, or 2)
3579 * it's for a locale. Even if there isn't unicode semantics
3580 * here, at runtime there may be because of matching against a
3581 * utf8 string, so accept a possible false positive for
3582 * latin1-range folds */
8aa23a47
YO
3583 if (uc >= 0x100 ||
3584 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))