This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
New COW mechanism
[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) \
288b8c02 452 SAVEDESTRUCTOR_X(clear_re,(void*)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) \
288b8c02 483 SAVEDESTRUCTOR_X(clear_re,(void*)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) \
288b8c02 501 SAVEDESTRUCTOR_X(clear_re,(void*)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) \
288b8c02 520 SAVEDESTRUCTOR_X(clear_re,(void*)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
acfe0abc 698static void clear_re(pTHX_ void *r);
4327152a 699
653099ff 700/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 701 Update the longest found anchored substring and the longest found
653099ff
GS
702 floating substrings if needed. */
703
4327152a 704STATIC void
304ee84b 705S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 706{
e1ec3a88
AL
707 const STRLEN l = CHR_SVLEN(data->last_found);
708 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 709 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 710
7918f24d
NC
711 PERL_ARGS_ASSERT_SCAN_COMMIT;
712
c277df42 713 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 714 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
715 if (*data->longest == data->longest_fixed) {
716 data->offset_fixed = l ? data->last_start_min : data->pos_min;
717 if (data->flags & SF_BEFORE_EOL)
b81d288d 718 data->flags
c277df42
IZ
719 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
720 else
721 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 722 data->minlen_fixed=minlenp;
1de06328 723 data->lookbehind_fixed=0;
a0ed51b3 724 }
304ee84b 725 else { /* *data->longest == data->longest_float */
c277df42 726 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
727 data->offset_float_max = (l
728 ? data->last_start_max
c277df42 729 : data->pos_min + data->pos_delta);
304ee84b 730 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 731 data->offset_float_max = I32_MAX;
c277df42 732 if (data->flags & SF_BEFORE_EOL)
b81d288d 733 data->flags
c277df42
IZ
734 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
735 else
736 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
737 data->minlen_float=minlenp;
738 data->lookbehind_float=0;
c277df42
IZ
739 }
740 }
741 SvCUR_set(data->last_found, 0);
0eda9292 742 {
a28509cc 743 SV * const sv = data->last_found;
097eb12c
AL
744 if (SvUTF8(sv) && SvMAGICAL(sv)) {
745 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
746 if (mg)
747 mg->mg_len = 0;
748 }
0eda9292 749 }
c277df42
IZ
750 data->last_end = -1;
751 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 752 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
753}
754
653099ff
GS
755/* Can match anything (initialization) */
756STATIC void
3fffb88a 757S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 758{
7918f24d
NC
759 PERL_ARGS_ASSERT_CL_ANYTHING;
760
f8bef550 761 ANYOF_BITMAP_SETALL(cl);
dd58aee1 762 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
ae6bdd37 763 |ANYOF_NON_UTF8_LATIN1_ALL;
3fffb88a
KW
764
765 /* If any portion of the regex is to operate under locale rules,
766 * initialization includes it. The reason this isn't done for all regexes
767 * is that the optimizer was written under the assumption that locale was
768 * all-or-nothing. Given the complexity and lack of documentation in the
769 * optimizer, and that there are inadequate test cases for locale, so many
770 * parts of it may not work properly, it is safest to avoid locale unless
771 * necessary. */
772 if (RExC_contains_locale) {
9d7a1e63 773 ANYOF_CLASS_SETALL(cl); /* /l uses class */
ae6bdd37 774 cl->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3fffb88a 775 }
9d7a1e63
KW
776 else {
777 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
778 }
653099ff
GS
779}
780
781/* Can match anything (initialization) */
782STATIC int
5f66b61c 783S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
784{
785 int value;
786
7918f24d
NC
787 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
788
aaa51d5e 789 for (value = 0; value <= ANYOF_MAX; value += 2)
653099ff
GS
790 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
791 return 1;
1aa99e6b
IH
792 if (!(cl->flags & ANYOF_UNICODE_ALL))
793 return 0;
10edeb5d 794 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 795 return 0;
653099ff
GS
796 return 1;
797}
798
799/* Can match anything (initialization) */
800STATIC void
e755fd73 801S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 802{
7918f24d
NC
803 PERL_ARGS_ASSERT_CL_INIT;
804
8ecf7187 805 Zero(cl, 1, struct regnode_charclass_class);
653099ff 806 cl->type = ANYOF;
3fffb88a 807 cl_anything(pRExC_state, cl);
1411dba4 808 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
809}
810
1051e1c4
KW
811/* These two functions currently do the exact same thing */
812#define cl_init_zero S_cl_init
653099ff 813
dd58aee1
KW
814/* 'AND' a given class with another one. Can create false positives. 'cl'
815 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
816 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 817STATIC void
5f66b61c 818S_cl_and(struct regnode_charclass_class *cl,
a28509cc 819 const struct regnode_charclass_class *and_with)
653099ff 820{
7918f24d 821 PERL_ARGS_ASSERT_CL_AND;
40d049e4
YO
822
823 assert(and_with->type == ANYOF);
1e6ade67 824
c6b76537 825 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
826 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
827 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 828 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
829 && !(and_with->flags & ANYOF_LOC_FOLD)
830 && !(cl->flags & ANYOF_LOC_FOLD)) {
653099ff
GS
831 int i;
832
833 if (and_with->flags & ANYOF_INVERT)
834 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
835 cl->bitmap[i] &= ~and_with->bitmap[i];
836 else
837 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
838 cl->bitmap[i] &= and_with->bitmap[i];
839 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 840
c6b76537 841 if (and_with->flags & ANYOF_INVERT) {
8951c461 842
c6b76537
KW
843 /* Here, the and'ed node is inverted. Get the AND of the flags that
844 * aren't affected by the inversion. Those that are affected are
845 * handled individually below */
846 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
847 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
848 cl->flags |= affected_flags;
849
850 /* We currently don't know how to deal with things that aren't in the
851 * bitmap, but we know that the intersection is no greater than what
852 * is already in cl, so let there be false positives that get sorted
853 * out after the synthetic start class succeeds, and the node is
854 * matched for real. */
855
856 /* The inversion of these two flags indicate that the resulting
857 * intersection doesn't have them */
858 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
859 cl->flags &= ~ANYOF_UNICODE_ALL;
860 }
c6b76537
KW
861 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
862 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 863 }
1aa99e6b 864 }
c6b76537 865 else { /* and'd node is not inverted */
3ad98780
KW
866 U8 outside_bitmap_but_not_utf8; /* Temp variable */
867
137165a6 868 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
869
870 /* Here 'and_with' doesn't match anything outside the bitmap
871 * (except possibly ANYOF_UNICODE_ALL), which means the
872 * intersection can't either, except for ANYOF_UNICODE_ALL, in
873 * which case we don't know what the intersection is, but it's no
874 * greater than what cl already has, so can just leave it alone,
875 * with possible false positives */
876 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
877 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 878 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 879 }
137165a6 880 }
c6b76537
KW
881 else if (! ANYOF_NONBITMAP(cl)) {
882
883 /* Here, 'and_with' does match something outside the bitmap, and cl
884 * doesn't have a list of things to match outside the bitmap. If
885 * cl can match all code points above 255, the intersection will
3ad98780
KW
886 * be those above-255 code points that 'and_with' matches. If cl
887 * can't match all Unicode code points, it means that it can't
888 * match anything outside the bitmap (since the 'if' that got us
889 * into this block tested for that), so we leave the bitmap empty.
890 */
c6b76537
KW
891 if (cl->flags & ANYOF_UNICODE_ALL) {
892 ARG_SET(cl, ARG(and_with));
3ad98780
KW
893
894 /* and_with's ARG may match things that don't require UTF8.
895 * And now cl's will too, in spite of this being an 'and'. See
896 * the comments below about the kludge */
897 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
898 }
899 }
900 else {
901 /* Here, both 'and_with' and cl match something outside the
902 * bitmap. Currently we do not do the intersection, so just match
903 * whatever cl had at the beginning. */
904 }
905
906
3ad98780
KW
907 /* Take the intersection of the two sets of flags. However, the
908 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
909 * kludge around the fact that this flag is not treated like the others
910 * which are initialized in cl_anything(). The way the optimizer works
911 * is that the synthetic start class (SSC) is initialized to match
912 * anything, and then the first time a real node is encountered, its
913 * values are AND'd with the SSC's with the result being the values of
914 * the real node. However, there are paths through the optimizer where
915 * the AND never gets called, so those initialized bits are set
916 * inappropriately, which is not usually a big deal, as they just cause
917 * false positives in the SSC, which will just mean a probably
918 * imperceptible slow down in execution. However this bit has a
919 * higher false positive consequence in that it can cause utf8.pm,
920 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
921 * bigger slowdown and also causes significant extra memory to be used.
922 * In order to prevent this, the code now takes a different tack. The
923 * bit isn't set unless some part of the regular expression needs it,
924 * but once set it won't get cleared. This means that these extra
925 * modules won't get loaded unless there was some path through the
926 * pattern that would have required them anyway, and so any false
927 * positives that occur by not ANDing them out when they could be
928 * aren't as severe as they would be if we treated this bit like all
929 * the others */
930 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
931 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 932 cl->flags &= and_with->flags;
3ad98780 933 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 934 }
653099ff
GS
935}
936
dd58aee1
KW
937/* 'OR' a given class with another one. Can create false positives. 'cl'
938 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
939 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 940STATIC void
3fffb88a 941S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 942{
7918f24d
NC
943 PERL_ARGS_ASSERT_CL_OR;
944
653099ff 945 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
946
947 /* Here, the or'd node is to be inverted. This means we take the
948 * complement of everything not in the bitmap, but currently we don't
949 * know what that is, so give up and match anything */
950 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 951 cl_anything(pRExC_state, cl);
c6b76537 952 }
653099ff
GS
953 /* We do not use
954 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
955 * <= (B1 | !B2) | (CL1 | !CL2)
956 * which is wasteful if CL2 is small, but we ignore CL2:
957 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
958 * XXXX Can we handle case-fold? Unclear:
959 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
960 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
961 */
c6b76537 962 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
963 && !(or_with->flags & ANYOF_LOC_FOLD)
964 && !(cl->flags & ANYOF_LOC_FOLD) ) {
653099ff
GS
965 int i;
966
967 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
968 cl->bitmap[i] |= ~or_with->bitmap[i];
969 } /* XXXX: logic is complicated otherwise */
970 else {
3fffb88a 971 cl_anything(pRExC_state, cl);
653099ff 972 }
c6b76537
KW
973
974 /* And, we can just take the union of the flags that aren't affected
975 * by the inversion */
976 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
977
978 /* For the remaining flags:
979 ANYOF_UNICODE_ALL and inverted means to not match anything above
980 255, which means that the union with cl should just be
981 what cl has in it, so can ignore this flag
982 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
983 is 127-255 to match them, but then invert that, so the
984 union with cl should just be what cl has in it, so can
985 ignore this flag
986 */
987 } else { /* 'or_with' is not inverted */
653099ff
GS
988 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
989 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
990 && (!(or_with->flags & ANYOF_LOC_FOLD)
991 || (cl->flags & ANYOF_LOC_FOLD)) ) {
653099ff
GS
992 int i;
993
994 /* OR char bitmap and class bitmap separately */
995 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
996 cl->bitmap[i] |= or_with->bitmap[i];
1e6ade67 997 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
653099ff
GS
998 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
999 cl->classflags[i] |= or_with->classflags[i];
1000 cl->flags |= ANYOF_CLASS;
1001 }
1002 }
1003 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 1004 cl_anything(pRExC_state, cl);
653099ff 1005 }
9826f543 1006
c6b76537
KW
1007 if (ANYOF_NONBITMAP(or_with)) {
1008
1009 /* Use the added node's outside-the-bit-map match if there isn't a
1010 * conflict. If there is a conflict (both nodes match something
1011 * outside the bitmap, but what they match outside is not the same
1012 * pointer, and hence not easily compared until XXX we extend
1013 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
1014 * match everything outside the bitmap. If that stuff is all above
1015 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1016 if (! ANYOF_NONBITMAP(cl)) {
1017 ARG_SET(cl, ARG(or_with));
1018 }
1019 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1020
1021 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1022 cl_anything(pRExC_state, cl);
1023 }
1024 else {
1025 cl->flags |= ANYOF_UNICODE_ALL;
1026 }
c6b76537 1027 }
4c34a693 1028 }
0b9668ee
KW
1029
1030 /* Take the union */
1031 cl->flags |= or_with->flags;
1aa99e6b 1032 }
653099ff
GS
1033}
1034
a3621e74
YO
1035#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1036#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1037#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1038#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1039
3dab1dad
YO
1040
1041#ifdef DEBUGGING
07be1b83 1042/*
2b8b4781
NC
1043 dump_trie(trie,widecharmap,revcharmap)
1044 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1045 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1046
1047 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1048 The _interim_ variants are used for debugging the interim
1049 tables that are used to generate the final compressed
1050 representation which is what dump_trie expects.
1051
486ec47a 1052 Part of the reason for their existence is to provide a form
3dab1dad 1053 of documentation as to how the different representations function.
07be1b83
YO
1054
1055*/
3dab1dad
YO
1056
1057/*
3dab1dad
YO
1058 Dumps the final compressed table form of the trie to Perl_debug_log.
1059 Used for debugging make_trie().
1060*/
b9a59e08 1061
3dab1dad 1062STATIC void
2b8b4781
NC
1063S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1064 AV *revcharmap, U32 depth)
3dab1dad
YO
1065{
1066 U32 state;
ab3bbdeb 1067 SV *sv=sv_newmortal();
55eed653 1068 int colwidth= widecharmap ? 6 : 4;
2e64971a 1069 U16 word;
3dab1dad
YO
1070 GET_RE_DEBUG_FLAGS_DECL;
1071
7918f24d 1072 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1073
3dab1dad
YO
1074 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1075 (int)depth * 2 + 2,"",
1076 "Match","Base","Ofs" );
1077
1078 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1079 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1080 if ( tmp ) {
ab3bbdeb
YO
1081 PerlIO_printf( Perl_debug_log, "%*s",
1082 colwidth,
ddc5bc0f 1083 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1084 PL_colors[0], PL_colors[1],
1085 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1086 PERL_PV_ESCAPE_FIRSTCHAR
1087 )
1088 );
3dab1dad
YO
1089 }
1090 }
1091 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1092 (int)depth * 2 + 2,"");
1093
1094 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1095 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1096 PerlIO_printf( Perl_debug_log, "\n");
1097
1e2e3d02 1098 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1099 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1100
1101 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1102
1103 if ( trie->states[ state ].wordnum ) {
1104 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1105 } else {
1106 PerlIO_printf( Perl_debug_log, "%6s", "" );
1107 }
1108
1109 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1110
1111 if ( base ) {
1112 U32 ofs = 0;
1113
1114 while( ( base + ofs < trie->uniquecharcount ) ||
1115 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1116 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1117 ofs++;
1118
1119 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1120
1121 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1122 if ( ( base + ofs >= trie->uniquecharcount ) &&
1123 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1124 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1125 {
ab3bbdeb
YO
1126 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1127 colwidth,
3dab1dad
YO
1128 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1129 } else {
ab3bbdeb 1130 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1131 }
1132 }
1133
1134 PerlIO_printf( Perl_debug_log, "]");
1135
1136 }
1137 PerlIO_printf( Perl_debug_log, "\n" );
1138 }
2e64971a
DM
1139 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1140 for (word=1; word <= trie->wordcount; word++) {
1141 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1142 (int)word, (int)(trie->wordinfo[word].prev),
1143 (int)(trie->wordinfo[word].len));
1144 }
1145 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1146}
1147/*
3dab1dad
YO
1148 Dumps a fully constructed but uncompressed trie in list form.
1149 List tries normally only are used for construction when the number of
1150 possible chars (trie->uniquecharcount) is very high.
1151 Used for debugging make_trie().
1152*/
1153STATIC void
55eed653 1154S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1155 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1156 U32 depth)
3dab1dad
YO
1157{
1158 U32 state;
ab3bbdeb 1159 SV *sv=sv_newmortal();
55eed653 1160 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1161 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1162
1163 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1164
3dab1dad 1165 /* print out the table precompression. */
ab3bbdeb
YO
1166 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1167 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1168 "------:-----+-----------------\n" );
3dab1dad
YO
1169
1170 for( state=1 ; state < next_alloc ; state ++ ) {
1171 U16 charid;
1172
ab3bbdeb 1173 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1174 (int)depth * 2 + 2,"", (UV)state );
1175 if ( ! trie->states[ state ].wordnum ) {
1176 PerlIO_printf( Perl_debug_log, "%5s| ","");
1177 } else {
1178 PerlIO_printf( Perl_debug_log, "W%4x| ",
1179 trie->states[ state ].wordnum
1180 );
1181 }
1182 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1183 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1184 if ( tmp ) {
1185 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1186 colwidth,
ddc5bc0f 1187 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1188 PL_colors[0], PL_colors[1],
1189 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1190 PERL_PV_ESCAPE_FIRSTCHAR
1191 ) ,
1e2e3d02
YO
1192 TRIE_LIST_ITEM(state,charid).forid,
1193 (UV)TRIE_LIST_ITEM(state,charid).newstate
1194 );
1195 if (!(charid % 10))
664e119d
RGS
1196 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1197 (int)((depth * 2) + 14), "");
1e2e3d02 1198 }
ab3bbdeb
YO
1199 }
1200 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1201 }
1202}
1203
1204/*
3dab1dad
YO
1205 Dumps a fully constructed but uncompressed trie in table form.
1206 This is the normal DFA style state transition table, with a few
1207 twists to facilitate compression later.
1208 Used for debugging make_trie().
1209*/
1210STATIC void
55eed653 1211S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1212 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1213 U32 depth)
3dab1dad
YO
1214{
1215 U32 state;
1216 U16 charid;
ab3bbdeb 1217 SV *sv=sv_newmortal();
55eed653 1218 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1219 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1220
1221 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1222
1223 /*
1224 print out the table precompression so that we can do a visual check
1225 that they are identical.
1226 */
1227
1228 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1229
1230 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1231 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1232 if ( tmp ) {
ab3bbdeb
YO
1233 PerlIO_printf( Perl_debug_log, "%*s",
1234 colwidth,
ddc5bc0f 1235 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1236 PL_colors[0], PL_colors[1],
1237 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1238 PERL_PV_ESCAPE_FIRSTCHAR
1239 )
1240 );
3dab1dad
YO
1241 }
1242 }
1243
1244 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1245
1246 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1247 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1248 }
1249
1250 PerlIO_printf( Perl_debug_log, "\n" );
1251
1252 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1253
1254 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1255 (int)depth * 2 + 2,"",
1256 (UV)TRIE_NODENUM( state ) );
1257
1258 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1259 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1260 if (v)
1261 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1262 else
1263 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1264 }
1265 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1266 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1267 } else {
1268 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1269 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1270 }
1271 }
07be1b83 1272}
3dab1dad
YO
1273
1274#endif
1275
2e64971a 1276
786e8c11
YO
1277/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1278 startbranch: the first branch in the whole branch sequence
1279 first : start branch of sequence of branch-exact nodes.
1280 May be the same as startbranch
1281 last : Thing following the last branch.
1282 May be the same as tail.
1283 tail : item following the branch sequence
1284 count : words in the sequence
1285 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1286 depth : indent depth
3dab1dad 1287
786e8c11 1288Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1289
786e8c11
YO
1290A trie is an N'ary tree where the branches are determined by digital
1291decomposition of the key. IE, at the root node you look up the 1st character and
1292follow that branch repeat until you find the end of the branches. Nodes can be
1293marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1294
786e8c11 1295 /he|she|his|hers/
72f13be8 1296
786e8c11
YO
1297would convert into the following structure. Numbers represent states, letters
1298following numbers represent valid transitions on the letter from that state, if
1299the number is in square brackets it represents an accepting state, otherwise it
1300will be in parenthesis.
07be1b83 1301
786e8c11
YO
1302 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1303 | |
1304 | (2)
1305 | |
1306 (1) +-i->(6)-+-s->[7]
1307 |
1308 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1309
786e8c11
YO
1310 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1311
1312This shows that when matching against the string 'hers' we will begin at state 1
1313read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1314then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1315is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1316single traverse. We store a mapping from accepting to state to which word was
1317matched, and then when we have multiple possibilities we try to complete the
1318rest of the regex in the order in which they occured in the alternation.
1319
1320The only prior NFA like behaviour that would be changed by the TRIE support is
1321the silent ignoring of duplicate alternations which are of the form:
1322
1323 / (DUPE|DUPE) X? (?{ ... }) Y /x
1324
4b714af6 1325Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1326and without the optimisation. With the optimisations dupes will be silently
486ec47a 1327ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1328the following demonstrates:
1329
1330 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1331
1332which prints out 'word' three times, but
1333
1334 'words'=~/(word|word|word)(?{ print $1 })S/
1335
1336which doesnt print it out at all. This is due to other optimisations kicking in.
1337
1338Example of what happens on a structural level:
1339
486ec47a 1340The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1341
1342 1: CURLYM[1] {1,32767}(18)
1343 5: BRANCH(8)
1344 6: EXACT <ac>(16)
1345 8: BRANCH(11)
1346 9: EXACT <ad>(16)
1347 11: BRANCH(14)
1348 12: EXACT <ab>(16)
1349 16: SUCCEED(0)
1350 17: NOTHING(18)
1351 18: END(0)
1352
1353This would be optimizable with startbranch=5, first=5, last=16, tail=16
1354and should turn into:
1355
1356 1: CURLYM[1] {1,32767}(18)
1357 5: TRIE(16)
1358 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1359 <ac>
1360 <ad>
1361 <ab>
1362 16: SUCCEED(0)
1363 17: NOTHING(18)
1364 18: END(0)
1365
1366Cases where tail != last would be like /(?foo|bar)baz/:
1367
1368 1: BRANCH(4)
1369 2: EXACT <foo>(8)
1370 4: BRANCH(7)
1371 5: EXACT <bar>(8)
1372 7: TAIL(8)
1373 8: EXACT <baz>(10)
1374 10: END(0)
1375
1376which would be optimizable with startbranch=1, first=1, last=7, tail=8
1377and would end up looking like:
1378
1379 1: TRIE(8)
1380 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1381 <foo>
1382 <bar>
1383 7: TAIL(8)
1384 8: EXACT <baz>(10)
1385 10: END(0)
1386
1387 d = uvuni_to_utf8_flags(d, uv, 0);
1388
1389is the recommended Unicode-aware way of saying
1390
1391 *(d++) = uv;
1392*/
1393
fab2782b 1394#define TRIE_STORE_REVCHAR(val) \
786e8c11 1395 STMT_START { \
73031816 1396 if (UTF) { \
fab2782b 1397 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1398 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
fab2782b 1399 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
73031816
NC
1400 SvCUR_set(zlopp, kapow - flrbbbbb); \
1401 SvPOK_on(zlopp); \
1402 SvUTF8_on(zlopp); \
1403 av_push(revcharmap, zlopp); \
1404 } else { \
fab2782b 1405 char ooooff = (char)val; \
73031816
NC
1406 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1407 } \
1408 } STMT_END
786e8c11 1409
fab2782b
YO
1410#define TRIE_READ_CHAR STMT_START { \
1411 wordlen++; \
1412 if ( UTF ) { \
1413 /* if it is UTF then it is either already folded, or does not need folding */ \
1414 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1415 } \
1416 else if (folder == PL_fold_latin1) { \
1417 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1418 if ( foldlen > 0 ) { \
1419 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1420 foldlen -= len; \
1421 scan += len; \
1422 len = 0; \
1423 } else { \
1424 len = 1; \
1425 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1426 skiplen = UNISKIP(uvc); \
1427 foldlen -= skiplen; \
1428 scan = foldbuf + skiplen; \
1429 } \
1430 } else { \
1431 /* raw data, will be folded later if needed */ \
1432 uvc = (U32)*uc; \
1433 len = 1; \
1434 } \
786e8c11
YO
1435} STMT_END
1436
1437
1438
1439#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1440 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1441 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1442 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1443 } \
1444 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1445 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1446 TRIE_LIST_CUR( state )++; \
1447} STMT_END
07be1b83 1448
786e8c11
YO
1449#define TRIE_LIST_NEW(state) STMT_START { \
1450 Newxz( trie->states[ state ].trans.list, \
1451 4, reg_trie_trans_le ); \
1452 TRIE_LIST_CUR( state ) = 1; \
1453 TRIE_LIST_LEN( state ) = 4; \
1454} STMT_END
07be1b83 1455
786e8c11
YO
1456#define TRIE_HANDLE_WORD(state) STMT_START { \
1457 U16 dupe= trie->states[ state ].wordnum; \
1458 regnode * const noper_next = regnext( noper ); \
1459 \
786e8c11
YO
1460 DEBUG_r({ \
1461 /* store the word for dumping */ \
1462 SV* tmp; \
1463 if (OP(noper) != NOTHING) \
740cce10 1464 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1465 else \
740cce10 1466 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1467 av_push( trie_words, tmp ); \
786e8c11
YO
1468 }); \
1469 \
1470 curword++; \
2e64971a
DM
1471 trie->wordinfo[curword].prev = 0; \
1472 trie->wordinfo[curword].len = wordlen; \
1473 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1474 \
1475 if ( noper_next < tail ) { \
1476 if (!trie->jump) \
c944940b 1477 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1478 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1479 if (!jumper) \
1480 jumper = noper_next; \
1481 if (!nextbranch) \
1482 nextbranch= regnext(cur); \
1483 } \
1484 \
1485 if ( dupe ) { \
2e64971a
DM
1486 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1487 /* chain, so that when the bits of chain are later */\
1488 /* linked together, the dups appear in the chain */\
1489 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1490 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1491 } else { \
1492 /* we haven't inserted this word yet. */ \
1493 trie->states[ state ].wordnum = curword; \
1494 } \
1495} STMT_END
07be1b83 1496
3dab1dad 1497
786e8c11
YO
1498#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1499 ( ( base + charid >= ucharcount \
1500 && base + charid < ubound \
1501 && state == trie->trans[ base - ucharcount + charid ].check \
1502 && trie->trans[ base - ucharcount + charid ].next ) \
1503 ? trie->trans[ base - ucharcount + charid ].next \
1504 : ( state==1 ? special : 0 ) \
1505 )
3dab1dad 1506
786e8c11
YO
1507#define MADE_TRIE 1
1508#define MADE_JUMP_TRIE 2
1509#define MADE_EXACT_TRIE 4
3dab1dad 1510
a3621e74 1511STATIC I32
786e8c11 1512S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1513{
27da23d5 1514 dVAR;
a3621e74
YO
1515 /* first pass, loop through and scan words */
1516 reg_trie_data *trie;
55eed653 1517 HV *widecharmap = NULL;
2b8b4781 1518 AV *revcharmap = newAV();
a3621e74 1519 regnode *cur;
9f7f3913 1520 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1521 STRLEN len = 0;
1522 UV uvc = 0;
1523 U16 curword = 0;
1524 U32 next_alloc = 0;
786e8c11
YO
1525 regnode *jumper = NULL;
1526 regnode *nextbranch = NULL;
7f69552c 1527 regnode *convert = NULL;
2e64971a 1528 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1529 /* we just use folder as a flag in utf8 */
1e696034 1530 const U8 * folder = NULL;
a3621e74 1531
2b8b4781
NC
1532#ifdef DEBUGGING
1533 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1534 AV *trie_words = NULL;
1535 /* along with revcharmap, this only used during construction but both are
1536 * useful during debugging so we store them in the struct when debugging.
8e11feef 1537 */
2b8b4781
NC
1538#else
1539 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1540 STRLEN trie_charcount=0;
3dab1dad 1541#endif
2b8b4781 1542 SV *re_trie_maxbuff;
a3621e74 1543 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1544
1545 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1546#ifndef DEBUGGING
1547 PERL_UNUSED_ARG(depth);
1548#endif
a3621e74 1549
1e696034 1550 switch (flags) {
c46d03cf 1551 case EXACT: break;
2f7f8cb1 1552 case EXACTFA:
fab2782b
YO
1553 case EXACTFU_SS:
1554 case EXACTFU_TRICKYFOLD:
1e696034
KW
1555 case EXACTFU: folder = PL_fold_latin1; break;
1556 case EXACTF: folder = PL_fold; break;
1557 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1558 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1559 }
1560
c944940b 1561 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1562 trie->refcount = 1;
3dab1dad 1563 trie->startstate = 1;
786e8c11 1564 trie->wordcount = word_count;
f8fc2ecf 1565 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1566 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1567 if (flags == EXACT)
c944940b 1568 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1569 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1570 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1571
a3621e74 1572 DEBUG_r({
2b8b4781 1573 trie_words = newAV();
a3621e74 1574 });
a3621e74 1575
0111c4fd 1576 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1577 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1578 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1579 }
df826430 1580 DEBUG_TRIE_COMPILE_r({
3dab1dad 1581 PerlIO_printf( Perl_debug_log,
786e8c11 1582 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1583 (int)depth * 2 + 2, "",
1584 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1585 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1586 (int)depth);
3dab1dad 1587 });
7f69552c
YO
1588
1589 /* Find the node we are going to overwrite */
1590 if ( first == startbranch && OP( last ) != BRANCH ) {
1591 /* whole branch chain */
1592 convert = first;
1593 } else {
1594 /* branch sub-chain */
1595 convert = NEXTOPER( first );
1596 }
1597
a3621e74
YO
1598 /* -- First loop and Setup --
1599
1600 We first traverse the branches and scan each word to determine if it
1601 contains widechars, and how many unique chars there are, this is
1602 important as we have to build a table with at least as many columns as we
1603 have unique chars.
1604
1605 We use an array of integers to represent the character codes 0..255
38a44b82 1606 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1607 native representation of the character value as the key and IV's for the
1608 coded index.
1609
1610 *TODO* If we keep track of how many times each character is used we can
1611 remap the columns so that the table compression later on is more
3b753521 1612 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1613 middle and the least common are on the outside. IMO this would be better
1614 than a most to least common mapping as theres a decent chance the most
1615 common letter will share a node with the least common, meaning the node
486ec47a 1616 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1617 case is when we have the least common nodes twice.
1618
1619 */
1620
a3621e74 1621 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1622 regnode *noper = NEXTOPER( cur );
e1ec3a88 1623 const U8 *uc = (U8*)STRING( noper );
df826430 1624 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1625 STRLEN foldlen = 0;
1626 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1627 STRLEN skiplen = 0;
2af232bd 1628 const U8 *scan = (U8*)NULL;
07be1b83 1629 U32 wordlen = 0; /* required init */
02daf0ab
YO
1630 STRLEN chars = 0;
1631 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1632
3dab1dad 1633 if (OP(noper) == NOTHING) {
df826430
YO
1634 regnode *noper_next= regnext(noper);
1635 if (noper_next != tail && OP(noper_next) == flags) {
1636 noper = noper_next;
1637 uc= (U8*)STRING(noper);
1638 e= uc + STR_LEN(noper);
1639 trie->minlen= STR_LEN(noper);
1640 } else {
1641 trie->minlen= 0;
1642 continue;
1643 }
3dab1dad 1644 }
df826430 1645
fab2782b 1646 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1647 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1648 regardless of encoding */
fab2782b
YO
1649 if (OP( noper ) == EXACTFU_SS) {
1650 /* false positives are ok, so just set this */
1651 TRIE_BITMAP_SET(trie,0xDF);
1652 }
1653 }
a3621e74 1654 for ( ; uc < e ; uc += len ) {
3dab1dad 1655 TRIE_CHARCOUNT(trie)++;
a3621e74 1656 TRIE_READ_CHAR;
3dab1dad 1657 chars++;
a3621e74 1658 if ( uvc < 256 ) {
fab2782b
YO
1659 if ( folder ) {
1660 U8 folded= folder[ (U8) uvc ];
1661 if ( !trie->charmap[ folded ] ) {
1662 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1663 TRIE_STORE_REVCHAR( folded );
1664 }
1665 }
a3621e74
YO
1666 if ( !trie->charmap[ uvc ] ) {
1667 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1668 TRIE_STORE_REVCHAR( uvc );
a3621e74 1669 }
02daf0ab 1670 if ( set_bit ) {
62012aee
KW
1671 /* store the codepoint in the bitmap, and its folded
1672 * equivalent. */
fab2782b 1673 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1674
1675 /* store the folded codepoint */
fab2782b 1676 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1677
1678 if ( !UTF ) {
1679 /* store first byte of utf8 representation of
acdf4139
KW
1680 variant codepoints */
1681 if (! UNI_IS_INVARIANT(uvc)) {
1682 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1683 }
1684 }
02daf0ab
YO
1685 set_bit = 0; /* We've done our bit :-) */
1686 }
a3621e74
YO
1687 } else {
1688 SV** svpp;
55eed653
NC
1689 if ( !widecharmap )
1690 widecharmap = newHV();
a3621e74 1691
55eed653 1692 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1693
1694 if ( !svpp )
e4584336 1695 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1696
1697 if ( !SvTRUE( *svpp ) ) {
1698 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1699 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1700 }
1701 }
1702 }
3dab1dad 1703 if( cur == first ) {
fab2782b
YO
1704 trie->minlen = chars;
1705 trie->maxlen = chars;
3dab1dad 1706 } else if (chars < trie->minlen) {
fab2782b 1707 trie->minlen = chars;
3dab1dad 1708 } else if (chars > trie->maxlen) {
fab2782b
YO
1709 trie->maxlen = chars;
1710 }
1711 if (OP( noper ) == EXACTFU_SS) {
1712 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1713 if (trie->minlen > 1)
1714 trie->minlen= 1;
1715 }
1716 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1717 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1718 * - We assume that any such sequence might match a 2 byte string */
1719 if (trie->minlen > 2 )
1720 trie->minlen= 2;
3dab1dad
YO
1721 }
1722
a3621e74
YO
1723 } /* end first pass */
1724 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1725 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1726 (int)depth * 2 + 2,"",
55eed653 1727 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1728 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1729 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1730 );
a3621e74
YO
1731
1732 /*
1733 We now know what we are dealing with in terms of unique chars and
1734 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1735 representation using a flat table will take. If it's over a reasonable
1736 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1737 conservative but potentially much slower representation using an array
1738 of lists.
1739
1740 At the end we convert both representations into the same compressed
1741 form that will be used in regexec.c for matching with. The latter
1742 is a form that cannot be used to construct with but has memory
1743 properties similar to the list form and access properties similar
1744 to the table form making it both suitable for fast searches and
1745 small enough that its feasable to store for the duration of a program.
1746
1747 See the comment in the code where the compressed table is produced
1748 inplace from the flat tabe representation for an explanation of how
1749 the compression works.
1750
1751 */
1752
1753
2e64971a
DM
1754 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1755 prev_states[1] = 0;
1756
3dab1dad 1757 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1758 /*
1759 Second Pass -- Array Of Lists Representation
1760
1761 Each state will be represented by a list of charid:state records
1762 (reg_trie_trans_le) the first such element holds the CUR and LEN
1763 points of the allocated array. (See defines above).
1764
1765 We build the initial structure using the lists, and then convert
1766 it into the compressed table form which allows faster lookups
1767 (but cant be modified once converted).
a3621e74
YO
1768 */
1769
a3621e74
YO
1770 STRLEN transcount = 1;
1771
1e2e3d02
YO
1772 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1773 "%*sCompiling trie using list compiler\n",
1774 (int)depth * 2 + 2, ""));
686b73d4 1775
c944940b
JH
1776 trie->states = (reg_trie_state *)
1777 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1778 sizeof(reg_trie_state) );
a3621e74
YO
1779 TRIE_LIST_NEW(1);
1780 next_alloc = 2;
1781
1782 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1783
df826430 1784 regnode *noper = NEXTOPER( cur );
c445ea15 1785 U8 *uc = (U8*)STRING( noper );
df826430 1786 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1787 U32 state = 1; /* required init */
1788 U16 charid = 0; /* sanity init */
1789 U8 *scan = (U8*)NULL; /* sanity init */
1790 STRLEN foldlen = 0; /* required init */
07be1b83 1791 U32 wordlen = 0; /* required init */
c445ea15 1792 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1793 STRLEN skiplen = 0;
c445ea15 1794
df826430
YO
1795 if (OP(noper) == NOTHING) {
1796 regnode *noper_next= regnext(noper);
1797 if (noper_next != tail && OP(noper_next) == flags) {
1798 noper = noper_next;
1799 uc= (U8*)STRING(noper);
1800 e= uc + STR_LEN(noper);
1801 }
1802 }
1803
3dab1dad 1804 if (OP(noper) != NOTHING) {
786e8c11 1805 for ( ; uc < e ; uc += len ) {
c445ea15 1806
786e8c11 1807 TRIE_READ_CHAR;
c445ea15 1808
786e8c11
YO
1809 if ( uvc < 256 ) {
1810 charid = trie->charmap[ uvc ];
c445ea15 1811 } else {
55eed653 1812 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1813 if ( !svpp ) {
1814 charid = 0;
1815 } else {
1816 charid=(U16)SvIV( *svpp );
1817 }
c445ea15 1818 }
786e8c11
YO
1819 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1820 if ( charid ) {
a3621e74 1821
786e8c11
YO
1822 U16 check;
1823 U32 newstate = 0;
a3621e74 1824
786e8c11
YO
1825 charid--;
1826 if ( !trie->states[ state ].trans.list ) {
1827 TRIE_LIST_NEW( state );
c445ea15 1828 }
786e8c11
YO
1829 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1830 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1831 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1832 break;
1833 }
1834 }
1835 if ( ! newstate ) {
1836 newstate = next_alloc++;
2e64971a 1837 prev_states[newstate] = state;
786e8c11
YO
1838 TRIE_LIST_PUSH( state, charid, newstate );
1839 transcount++;
1840 }
1841 state = newstate;
1842 } else {
1843 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1844 }
a28509cc 1845 }
c445ea15 1846 }
3dab1dad 1847 TRIE_HANDLE_WORD(state);
a3621e74
YO
1848
1849 } /* end second pass */
1850
1e2e3d02
YO
1851 /* next alloc is the NEXT state to be allocated */
1852 trie->statecount = next_alloc;
c944940b
JH
1853 trie->states = (reg_trie_state *)
1854 PerlMemShared_realloc( trie->states,
1855 next_alloc
1856 * sizeof(reg_trie_state) );
a3621e74 1857
3dab1dad 1858 /* and now dump it out before we compress it */
2b8b4781
NC
1859 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1860 revcharmap, next_alloc,
1861 depth+1)
1e2e3d02 1862 );
a3621e74 1863
c944940b
JH
1864 trie->trans = (reg_trie_trans *)
1865 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1866 {
1867 U32 state;
a3621e74
YO
1868 U32 tp = 0;
1869 U32 zp = 0;
1870
1871
1872 for( state=1 ; state < next_alloc ; state ++ ) {
1873 U32 base=0;
1874
1875 /*
1876 DEBUG_TRIE_COMPILE_MORE_r(
1877 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1878 );
1879 */
1880
1881 if (trie->states[state].trans.list) {
1882 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1883 U16 maxid=minid;
a28509cc 1884 U16 idx;
a3621e74
YO
1885
1886 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1887 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1888 if ( forid < minid ) {
1889 minid=forid;
1890 } else if ( forid > maxid ) {
1891 maxid=forid;
1892 }
a3621e74
YO
1893 }
1894 if ( transcount < tp + maxid - minid + 1) {
1895 transcount *= 2;
c944940b
JH
1896 trie->trans = (reg_trie_trans *)
1897 PerlMemShared_realloc( trie->trans,
446bd890
NC
1898 transcount
1899 * sizeof(reg_trie_trans) );
a3621e74
YO
1900 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1901 }
1902 base = trie->uniquecharcount + tp - minid;
1903 if ( maxid == minid ) {
1904 U32 set = 0;
1905 for ( ; zp < tp ; zp++ ) {
1906 if ( ! trie->trans[ zp ].next ) {
1907 base = trie->uniquecharcount + zp - minid;
1908 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1909 trie->trans[ zp ].check = state;
1910 set = 1;
1911 break;
1912 }
1913 }
1914 if ( !set ) {
1915 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1916 trie->trans[ tp ].check = state;
1917 tp++;
1918 zp = tp;
1919 }
1920 } else {
1921 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1922 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1923 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1924 trie->trans[ tid ].check = state;
1925 }
1926 tp += ( maxid - minid + 1 );
1927 }
1928 Safefree(trie->states[ state ].trans.list);
1929 }
1930 /*
1931 DEBUG_TRIE_COMPILE_MORE_r(
1932 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1933 );
1934 */
1935 trie->states[ state ].trans.base=base;
1936 }
cc601c31 1937 trie->lasttrans = tp + 1;
a3621e74
YO
1938 }
1939 } else {
1940 /*
1941 Second Pass -- Flat Table Representation.
1942
1943 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1944 We know that we will need Charcount+1 trans at most to store the data
1945 (one row per char at worst case) So we preallocate both structures
1946 assuming worst case.
1947
1948 We then construct the trie using only the .next slots of the entry
1949 structs.
1950
3b753521 1951 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1952 make compression both faster and easier by keeping track of how many non
1953 zero fields are in the node.
1954
1955 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1956 transition.
1957
1958 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1959 number representing the first entry of the node, and state as a
1960 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1961 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1962 are 2 entrys per node. eg:
1963
1964 A B A B
1965 1. 2 4 1. 3 7
1966 2. 0 3 3. 0 5
1967 3. 0 0 5. 0 0
1968 4. 0 0 7. 0 0
1969
1970 The table is internally in the right hand, idx form. However as we also
1971 have to deal with the states array which is indexed by nodenum we have to
1972 use TRIE_NODENUM() to convert.
1973
1974 */
1e2e3d02
YO
1975 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1976 "%*sCompiling trie using table compiler\n",
1977 (int)depth * 2 + 2, ""));
3dab1dad 1978
c944940b
JH
1979 trie->trans = (reg_trie_trans *)
1980 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1981 * trie->uniquecharcount + 1,
1982 sizeof(reg_trie_trans) );
1983 trie->states = (reg_trie_state *)
1984 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1985 sizeof(reg_trie_state) );
a3621e74
YO
1986 next_alloc = trie->uniquecharcount + 1;
1987
3dab1dad 1988
a3621e74
YO
1989 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1990
df826430 1991 regnode *noper = NEXTOPER( cur );
a28509cc 1992 const U8 *uc = (U8*)STRING( noper );
df826430 1993 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1994
1995 U32 state = 1; /* required init */
1996
1997 U16 charid = 0; /* sanity init */
1998 U32 accept_state = 0; /* sanity init */
1999 U8 *scan = (U8*)NULL; /* sanity init */
2000
2001 STRLEN foldlen = 0; /* required init */
07be1b83 2002 U32 wordlen = 0; /* required init */
fab2782b 2003 STRLEN skiplen = 0;
a3621e74
YO
2004 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2005
df826430
YO
2006 if (OP(noper) == NOTHING) {
2007 regnode *noper_next= regnext(noper);
2008 if (noper_next != tail && OP(noper_next) == flags) {
2009 noper = noper_next;
2010 uc= (U8*)STRING(noper);
2011 e= uc + STR_LEN(noper);
2012 }
2013 }
fab2782b 2014
3dab1dad 2015 if ( OP(noper) != NOTHING ) {
786e8c11 2016 for ( ; uc < e ; uc += len ) {
a3621e74 2017
786e8c11 2018 TRIE_READ_CHAR;
a3621e74 2019
786e8c11
YO
2020 if ( uvc < 256 ) {
2021 charid = trie->charmap[ uvc ];
2022 } else {
55eed653 2023 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2024 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2025 }
786e8c11
YO
2026 if ( charid ) {
2027 charid--;
2028 if ( !trie->trans[ state + charid ].next ) {
2029 trie->trans[ state + charid ].next = next_alloc;
2030 trie->trans[ state ].check++;
2e64971a
DM
2031 prev_states[TRIE_NODENUM(next_alloc)]
2032 = TRIE_NODENUM(state);
786e8c11
YO
2033 next_alloc += trie->uniquecharcount;
2034 }
2035 state = trie->trans[ state + charid ].next;
2036 } else {
2037 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2038 }
2039 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2040 }
a3621e74 2041 }
3dab1dad
YO
2042 accept_state = TRIE_NODENUM( state );
2043 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2044
2045 } /* end second pass */
2046
3dab1dad 2047 /* and now dump it out before we compress it */
2b8b4781
NC
2048 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2049 revcharmap,
2050 next_alloc, depth+1));
a3621e74 2051
a3621e74
YO
2052 {
2053 /*
2054 * Inplace compress the table.*
2055
2056 For sparse data sets the table constructed by the trie algorithm will
2057 be mostly 0/FAIL transitions or to put it another way mostly empty.
2058 (Note that leaf nodes will not contain any transitions.)
2059
2060 This algorithm compresses the tables by eliminating most such
2061 transitions, at the cost of a modest bit of extra work during lookup:
2062
2063 - Each states[] entry contains a .base field which indicates the
2064 index in the state[] array wheres its transition data is stored.
2065
3b753521 2066 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2067
2068 - If .base is nonzero then charid is added to it to find an entry in
2069 the trans array.
2070
2071 -If trans[states[state].base+charid].check!=state then the
2072 transition is taken to be a 0/Fail transition. Thus if there are fail
2073 transitions at the front of the node then the .base offset will point
2074 somewhere inside the previous nodes data (or maybe even into a node
2075 even earlier), but the .check field determines if the transition is
2076 valid.
2077
786e8c11 2078 XXX - wrong maybe?
a3621e74 2079 The following process inplace converts the table to the compressed
3b753521 2080 table: We first do not compress the root node 1,and mark all its
a3621e74 2081 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2082 allows us to do a DFA construction from the compressed table later,
2083 and ensures that any .base pointers we calculate later are greater
2084 than 0.
a3621e74
YO
2085
2086 - We set 'pos' to indicate the first entry of the second node.
2087
2088 - We then iterate over the columns of the node, finding the first and
2089 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2090 and set the .check pointers accordingly, and advance pos
2091 appropriately and repreat for the next node. Note that when we copy
2092 the next pointers we have to convert them from the original
2093 NODEIDX form to NODENUM form as the former is not valid post
2094 compression.
2095
2096 - If a node has no transitions used we mark its base as 0 and do not
2097 advance the pos pointer.
2098
2099 - If a node only has one transition we use a second pointer into the
2100 structure to fill in allocated fail transitions from other states.
2101 This pointer is independent of the main pointer and scans forward
2102 looking for null transitions that are allocated to a state. When it
2103 finds one it writes the single transition into the "hole". If the
786e8c11 2104 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2105
2106 - Once compressed we can Renew/realloc the structures to release the
2107 excess space.
2108
2109 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2110 specifically Fig 3.47 and the associated pseudocode.
2111
2112 demq
2113 */
a3b680e6 2114 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2115 U32 state, charid;
a3621e74 2116 U32 pos = 0, zp=0;
1e2e3d02 2117 trie->statecount = laststate;
a3621e74
YO
2118
2119 for ( state = 1 ; state < laststate ; state++ ) {
2120 U8 flag = 0;
a28509cc
AL
2121 const U32 stateidx = TRIE_NODEIDX( state );
2122 const U32 o_used = trie->trans[ stateidx ].check;
2123 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2124 trie->trans[ stateidx ].check = 0;
2125
2126 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2127 if ( flag || trie->trans[ stateidx + charid ].next ) {
2128 if ( trie->trans[ stateidx + charid ].next ) {
2129 if (o_used == 1) {
2130 for ( ; zp < pos ; zp++ ) {
2131 if ( ! trie->trans[ zp ].next ) {
2132 break;
2133 }
2134 }
2135 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2136 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2137 trie->trans[ zp ].check = state;
2138 if ( ++zp > pos ) pos = zp;
2139 break;
2140 }
2141 used--;
2142 }
2143 if ( !flag ) {
2144 flag = 1;
2145 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2146 }
2147 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2148 trie->trans[ pos ].check = state;
2149 pos++;
2150 }
2151 }
2152 }
cc601c31 2153 trie->lasttrans = pos + 1;
c944940b
JH
2154 trie->states = (reg_trie_state *)
2155 PerlMemShared_realloc( trie->states, laststate
2156 * sizeof(reg_trie_state) );
a3621e74 2157 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2158 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2159 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2160 (int)depth * 2 + 2,"",
2161 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2162 (IV)next_alloc,
2163 (IV)pos,
a3621e74
YO
2164 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2165 );
2166
2167 } /* end table compress */
2168 }
1e2e3d02
YO
2169 DEBUG_TRIE_COMPILE_MORE_r(
2170 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2171 (int)depth * 2 + 2, "",
2172 (UV)trie->statecount,
2173 (UV)trie->lasttrans)
2174 );
cc601c31 2175 /* resize the trans array to remove unused space */
c944940b
JH
2176 trie->trans = (reg_trie_trans *)
2177 PerlMemShared_realloc( trie->trans, trie->lasttrans
2178 * sizeof(reg_trie_trans) );
a3621e74 2179
3b753521 2180 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2181 U8 nodetype =(U8)(flags & 0xFF);
2182 char *str=NULL;
786e8c11 2183
07be1b83 2184#ifdef DEBUGGING
e62cc96a 2185 regnode *optimize = NULL;
7122b237
YO
2186#ifdef RE_TRACK_PATTERN_OFFSETS
2187
b57a0404
JH
2188 U32 mjd_offset = 0;
2189 U32 mjd_nodelen = 0;
7122b237
YO
2190#endif /* RE_TRACK_PATTERN_OFFSETS */
2191#endif /* DEBUGGING */
a3621e74 2192 /*
3dab1dad
YO
2193 This means we convert either the first branch or the first Exact,
2194 depending on whether the thing following (in 'last') is a branch
2195 or not and whther first is the startbranch (ie is it a sub part of
2196 the alternation or is it the whole thing.)
3b753521 2197 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2198 the whole branch sequence, including the first.
a3621e74 2199 */
3dab1dad 2200 /* Find the node we are going to overwrite */
7f69552c 2201 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2202 /* branch sub-chain */
3dab1dad 2203 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2204#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2205 DEBUG_r({
2206 mjd_offset= Node_Offset((convert));
2207 mjd_nodelen= Node_Length((convert));
2208 });
7122b237 2209#endif
7f69552c 2210 /* whole branch chain */
7122b237
YO
2211 }
2212#ifdef RE_TRACK_PATTERN_OFFSETS
2213 else {
7f69552c
YO
2214 DEBUG_r({
2215 const regnode *nop = NEXTOPER( convert );
2216 mjd_offset= Node_Offset((nop));
2217 mjd_nodelen= Node_Length((nop));
2218 });
07be1b83
YO
2219 }
2220 DEBUG_OPTIMISE_r(
2221 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2222 (int)depth * 2 + 2, "",
786e8c11 2223 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2224 );
7122b237 2225#endif
3dab1dad
YO
2226 /* But first we check to see if there is a common prefix we can
2227 split out as an EXACT and put in front of the TRIE node. */
2228 trie->startstate= 1;
55eed653 2229 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2230 U32 state;
1e2e3d02 2231 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2232 U32 ofs = 0;
8e11feef
RGS
2233 I32 idx = -1;
2234 U32 count = 0;
2235 const U32 base = trie->states[ state ].trans.base;
a3621e74 2236
3dab1dad 2237 if ( trie->states[state].wordnum )
8e11feef 2238 count = 1;
a3621e74 2239
8e11feef 2240 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2241 if ( ( base + ofs >= trie->uniquecharcount ) &&
2242 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2243 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2244 {
3dab1dad 2245 if ( ++count > 1 ) {
2b8b4781 2246 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2247 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2248 if ( state == 1 ) break;
3dab1dad
YO
2249 if ( count == 2 ) {
2250 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2251 DEBUG_OPTIMISE_r(
8e11feef
RGS
2252 PerlIO_printf(Perl_debug_log,
2253 "%*sNew Start State=%"UVuf" Class: [",
2254 (int)depth * 2 + 2, "",
786e8c11 2255 (UV)state));
be8e71aa 2256 if (idx >= 0) {
2b8b4781 2257 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2258 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2259
3dab1dad 2260 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2261 if ( folder )
2262 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2263 DEBUG_OPTIMISE_r(
f1f66076 2264 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2265 );
8e11feef
RGS
2266 }
2267 }
2268 TRIE_BITMAP_SET(trie,*ch);
2269 if ( folder )
2270 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2271 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2272 }
2273 idx = ofs;
2274 }
3dab1dad
YO
2275 }
2276 if ( count == 1 ) {
2b8b4781 2277 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2278 STRLEN len;
2279 char *ch = SvPV( *tmp, len );
de734bd5
A
2280 DEBUG_OPTIMISE_r({
2281 SV *sv=sv_newmortal();
8e11feef
RGS
2282 PerlIO_printf( Perl_debug_log,
2283 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2284 (int)depth * 2 + 2, "",
de734bd5
A
2285 (UV)state, (UV)idx,
2286 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2287 PL_colors[0], PL_colors[1],
2288 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2289 PERL_PV_ESCAPE_FIRSTCHAR
2290 )
2291 );
2292 });
3dab1dad
YO
2293 if ( state==1 ) {
2294 OP( convert ) = nodetype;
2295 str=STRING(convert);
2296 STR_LEN(convert)=0;
2297 }
c490c714
YO
2298 STR_LEN(convert) += len;
2299 while (len--)
de734bd5 2300 *str++ = *ch++;
8e11feef 2301 } else {
f9049ba1 2302#ifdef DEBUGGING
8e11feef
RGS
2303 if (state>1)
2304 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2305#endif
8e11feef
RGS
2306 break;
2307 }
2308 }
2e64971a 2309 trie->prefixlen = (state-1);
3dab1dad 2310 if (str) {
8e11feef 2311 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2312 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2313 trie->startstate = state;
07be1b83
YO
2314 trie->minlen -= (state - 1);
2315 trie->maxlen -= (state - 1);
33809eae
JH
2316#ifdef DEBUGGING
2317 /* At least the UNICOS C compiler choked on this
2318 * being argument to DEBUG_r(), so let's just have
2319 * it right here. */
2320 if (
2321#ifdef PERL_EXT_RE_BUILD
2322 1
2323#else
2324 DEBUG_r_TEST
2325#endif
2326 ) {
2327 regnode *fix = convert;
2328 U32 word = trie->wordcount;
2329 mjd_nodelen++;
2330 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2331 while( ++fix < n ) {
2332 Set_Node_Offset_Length(fix, 0, 0);
2333 }
2334 while (word--) {
2335 SV ** const tmp = av_fetch( trie_words, word, 0 );
2336 if (tmp) {
2337 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2338 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2339 else
2340 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2341 }
2342 }
2343 }
2344#endif
8e11feef
RGS
2345 if (trie->maxlen) {
2346 convert = n;
2347 } else {
3dab1dad 2348 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2349 DEBUG_r(optimize= n);
3dab1dad
YO
2350 }
2351 }
2352 }
a5ca303d
YO
2353 if (!jumper)
2354 jumper = last;
3dab1dad 2355 if ( trie->maxlen ) {
8e11feef
RGS
2356 NEXT_OFF( convert ) = (U16)(tail - convert);
2357 ARG_SET( convert, data_slot );
786e8c11
YO
2358 /* Store the offset to the first unabsorbed branch in
2359 jump[0], which is otherwise unused by the jump logic.
2360 We use this when dumping a trie and during optimisation. */
2361 if (trie->jump)
7f69552c 2362 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2363
6c48061a
YO
2364 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2365 * and there is a bitmap
2366 * and the first "jump target" node we found leaves enough room
2367 * then convert the TRIE node into a TRIEC node, with the bitmap
2368 * embedded inline in the opcode - this is hypothetically faster.
2369 */
2370 if ( !trie->states[trie->startstate].wordnum
2371 && trie->bitmap
2372 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2373 {
2374 OP( convert ) = TRIEC;
2375 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2376 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2377 trie->bitmap= NULL;
2378 } else
2379 OP( convert ) = TRIE;
a3621e74 2380
3dab1dad
YO
2381 /* store the type in the flags */
2382 convert->flags = nodetype;
a5ca303d
YO
2383 DEBUG_r({
2384 optimize = convert
2385 + NODE_STEP_REGNODE
2386 + regarglen[ OP( convert ) ];
2387 });
2388 /* XXX We really should free up the resource in trie now,
2389 as we won't use them - (which resources?) dmq */
3dab1dad 2390 }
a3621e74 2391 /* needed for dumping*/
e62cc96a 2392 DEBUG_r(if (optimize) {
07be1b83 2393 regnode *opt = convert;
bcdf7404 2394
e62cc96a 2395 while ( ++opt < optimize) {
07be1b83
YO
2396 Set_Node_Offset_Length(opt,0,0);
2397 }
786e8c11
YO
2398 /*
2399 Try to clean up some of the debris left after the
2400 optimisation.
a3621e74 2401 */
786e8c11 2402 while( optimize < jumper ) {
07be1b83 2403 mjd_nodelen += Node_Length((optimize));
a3621e74 2404 OP( optimize ) = OPTIMIZED;
07be1b83 2405 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2406 optimize++;
2407 }
07be1b83 2408 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2409 });
2410 } /* end node insert */
2e64971a
DM
2411
2412 /* Finish populating the prev field of the wordinfo array. Walk back
2413 * from each accept state until we find another accept state, and if
2414 * so, point the first word's .prev field at the second word. If the
2415 * second already has a .prev field set, stop now. This will be the
2416 * case either if we've already processed that word's accept state,
3b753521
FN
2417 * or that state had multiple words, and the overspill words were
2418 * already linked up earlier.
2e64971a
DM
2419 */
2420 {
2421 U16 word;
2422 U32 state;
2423 U16 prev;
2424
2425 for (word=1; word <= trie->wordcount; word++) {
2426 prev = 0;
2427 if (trie->wordinfo[word].prev)
2428 continue;
2429 state = trie->wordinfo[word].accept;
2430 while (state) {
2431 state = prev_states[state];
2432 if (!state)
2433 break;
2434 prev = trie->states[state].wordnum;
2435 if (prev)
2436 break;
2437 }
2438 trie->wordinfo[word].prev = prev;
2439 }
2440 Safefree(prev_states);
2441 }
2442
2443
2444 /* and now dump out the compressed format */
2445 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2446
55eed653 2447 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2448#ifdef DEBUGGING
2449 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2450 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2451#else
2452 SvREFCNT_dec(revcharmap);
07be1b83 2453#endif
786e8c11
YO
2454 return trie->jump
2455 ? MADE_JUMP_TRIE
2456 : trie->startstate>1
2457 ? MADE_EXACT_TRIE
2458 : MADE_TRIE;
2459}
2460
2461STATIC void
2462S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2463{
3b753521 2464/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2465
2466 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2467 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2468 ISBN 0-201-10088-6
2469
2470 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2471 suffix of the current state's 'word' that is also a proper prefix of another word in our
2472 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2473 the DFA not to have to restart after its tried and failed a word at a given point, it
2474 simply continues as though it had been matching the other word in the first place.
2475 Consider
2476 'abcdgu'=~/abcdefg|cdgu/
2477 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2478 fail, which would bring us to the state representing 'd' in the second word where we would
2479 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2480 */
2481 /* add a fail transition */
3251b653
NC
2482 const U32 trie_offset = ARG(source);
2483 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2484 U32 *q;
2485 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2486 const U32 numstates = trie->statecount;
786e8c11
YO
2487 const U32 ubound = trie->lasttrans + ucharcount;
2488 U32 q_read = 0;
2489 U32 q_write = 0;
2490 U32 charid;
2491 U32 base = trie->states[ 1 ].trans.base;
2492 U32 *fail;
2493 reg_ac_data *aho;
2494 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2495 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2496
2497 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2498#ifndef DEBUGGING
2499 PERL_UNUSED_ARG(depth);
2500#endif
2501
2502
2503 ARG_SET( stclass, data_slot );
c944940b 2504 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2505 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2506 aho->trie=trie_offset;
446bd890
NC
2507 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2508 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2509 Newxz( q, numstates, U32);
c944940b 2510 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2511 aho->refcount = 1;
2512 fail = aho->fail;
2513 /* initialize fail[0..1] to be 1 so that we always have
2514 a valid final fail state */
2515 fail[ 0 ] = fail[ 1 ] = 1;
2516
2517 for ( charid = 0; charid < ucharcount ; charid++ ) {
2518 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2519 if ( newstate ) {
2520 q[ q_write ] = newstate;
2521 /* set to point at the root */
2522 fail[ q[ q_write++ ] ]=1;
2523 }
2524 }
2525 while ( q_read < q_write) {
2526 const U32 cur = q[ q_read++ % numstates ];
2527 base = trie->states[ cur ].trans.base;
2528
2529 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2530 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2531 if (ch_state) {
2532 U32 fail_state = cur;
2533 U32 fail_base;
2534 do {
2535 fail_state = fail[ fail_state ];
2536 fail_base = aho->states[ fail_state ].trans.base;
2537 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2538
2539 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2540 fail[ ch_state ] = fail_state;
2541 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2542 {
2543 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2544 }
2545 q[ q_write++ % numstates] = ch_state;
2546 }
2547 }
2548 }
2549 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2550 when we fail in state 1, this allows us to use the
2551 charclass scan to find a valid start char. This is based on the principle
2552 that theres a good chance the string being searched contains lots of stuff
2553 that cant be a start char.
2554 */
2555 fail[ 0 ] = fail[ 1 ] = 0;
2556 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2557 PerlIO_printf(Perl_debug_log,
2558 "%*sStclass Failtable (%"UVuf" states): 0",
2559 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2560 );
786e8c11
YO
2561 for( q_read=1; q_read<numstates; q_read++ ) {
2562 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2563 }
2564 PerlIO_printf(Perl_debug_log, "\n");
2565 });
2566 Safefree(q);
2567 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2568}
2569
786e8c11 2570
a3621e74 2571/*
5d1c421c
JH
2572 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2573 * These need to be revisited when a newer toolchain becomes available.
2574 */
2575#if defined(__sparc64__) && defined(__GNUC__)
2576# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2577# undef SPARC64_GCC_WORKAROUND
2578# define SPARC64_GCC_WORKAROUND 1
2579# endif
2580#endif
2581
07be1b83 2582#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2583 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2584 SV * const mysv=sv_newmortal(); \
2585 regnode *Next = regnext(scan); \
2586 regprop(RExC_rx, mysv, scan); \
7f69552c 2587 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2588 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2589 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2590 }});
07be1b83 2591
1de06328 2592
bb914485 2593/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
2594 * one. The regop may be changed if the node(s) contain certain sequences that
2595 * require special handling. The joining is only done if:
bb914485
KW
2596 * 1) there is room in the current conglomerated node to entirely contain the
2597 * next one.
2598 * 2) they are the exact same node type
2599 *
87b8b349 2600 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
2601 * these get optimized out
2602 *
0a982f06
KW
2603 * If a node is to match under /i (folded), the number of characters it matches
2604 * can be different than its character length if it contains a multi-character
2605 * fold. *min_subtract is set to the total delta of the input nodes.
bb914485 2606 *
a0c4c608
KW
2607 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2608 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2609 *
bb914485 2610 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
2611 * multi-character fold sequences. It's been wrong in Perl for a very long
2612 * time. There are three code points in Unicode whose multi-character folds
2613 * were long ago discovered to mess things up. The previous designs for
2614 * dealing with these involved assigning a special node for them. This
2615 * approach doesn't work, as evidenced by this example:
a0c4c608 2616 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
0a982f06
KW
2617 * Both these fold to "sss", but if the pattern is parsed to create a node that
2618 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
2619 * successful match would have to cross the node's boundary. The new approach
2620 * that hopefully generally solves the problem generates an EXACTFU_SS node
2621 * that is "sss".
2622 *
0a982f06
KW
2623 * It turns out that there are problems with all multi-character folds, and not
2624 * just these three. Now the code is general, for all such cases, but the
2625 * three still have some special handling. The approach taken is:
2626 * 1) This routine examines each EXACTFish node that could contain multi-
2627 * character fold sequences. It returns in *min_subtract how much to
9d071ca8 2628 * subtract from the the actual length of the string to get a real minimum
0a982f06
KW
2629 * match length; it is 0 if there are no multi-char folds. This delta is
2630 * used by the caller to adjust the min length of the match, and the delta
2631 * between min and max, so that the optimizer doesn't reject these
2632 * possibilities based on size constraints.
2633 * 2) Certain of these sequences require special handling by the trie code,
2634 * so, if found, this code changes the joined node type to special ops:
2635 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2636 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2637 * is used for an EXACTFU node that contains at least one "ss" sequence in
2638 * it. For non-UTF-8 patterns and strings, this is the only case where
2639 * there is a possible fold length change. That means that a regular
2640 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2641 * with length changes, and so can be processed faster. regexec.c takes
2642 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2643 * pre-folded by regcomp.c. This saves effort in regex matching.
87b8b349 2644 * However, the pre-folding isn't done for non-UTF8 patterns because the
0a982f06
KW
2645 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2646 * down by forcing the pattern into UTF8 unless necessary. Also what
2647 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2648 * possibilities for the non-UTF8 patterns are quite simple, except for
2649 * the sharp s. All the ones that don't involve a UTF-8 target string are
2650 * members of a fold-pair, and arrays are set up for all of them so that
2651 * the other member of the pair can be found quickly. Code elsewhere in
2652 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2653 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2654 * described in the next item.
bb914485
KW
2655 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2656 * 'ss' or not is not knowable at compile time. It will match iff the
2657 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2658 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
6012a526 2659 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
bb914485
KW
2660 * described in item 3). An assumption that the optimizer part of
2661 * regexec.c (probably unwittingly) makes is that a character in the
2662 * pattern corresponds to at most a single character in the target string.
2663 * (And I do mean character, and not byte here, unlike other parts of the
2664 * documentation that have never been updated to account for multibyte
2665 * Unicode.) This assumption is wrong only in this case, as all other
2666 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2667 * virtue of having this file pre-fold UTF-8 patterns. I'm
2668 * reluctant to try to change this assumption, so instead the code punts.
9d071ca8
KW
2669 * This routine examines EXACTF nodes for the sharp s, and returns a
2670 * boolean indicating whether or not the node is an EXACTF node that
2671 * contains a sharp s. When it is true, the caller sets a flag that later
2672 * causes the optimizer in this file to not set values for the floating
2673 * and fixed string lengths, and thus avoids the optimizer code in
2674 * regexec.c that makes the invalid assumption. Thus, there is no
2675 * optimization based on string lengths for EXACTF nodes that contain the
2676 * sharp s. This only happens for /id rules (which means the pattern
2677 * isn't in UTF-8).
bb914485 2678 */
1de06328 2679
9d071ca8 2680#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2681 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2682 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2683
be8e71aa 2684STATIC U32
9d071ca8 2685S_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
2686 /* Merge several consecutive EXACTish nodes into one. */
2687 regnode *n = regnext(scan);
2688 U32 stringok = 1;
2689 regnode *next = scan + NODE_SZ_STR(scan);
2690 U32 merged = 0;
2691 U32 stopnow = 0;
2692#ifdef DEBUGGING
2693 regnode *stop = scan;
72f13be8 2694 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2695#else
d47053eb
RGS
2696 PERL_UNUSED_ARG(depth);
2697#endif
7918f24d
NC
2698
2699 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2700#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2701 PERL_UNUSED_ARG(flags);
2702 PERL_UNUSED_ARG(val);
07be1b83 2703#endif
07be1b83 2704 DEBUG_PEEP("join",scan,depth);
bb914485 2705
3f410cf6
KW
2706 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2707 * EXACT ones that are mergeable to the current one. */
2708 while (n
2709 && (PL_regkind[OP(n)] == NOTHING
2710 || (stringok && OP(n) == OP(scan)))
07be1b83 2711 && NEXT_OFF(n)
3f410cf6
KW
2712 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2713 {
07be1b83
YO
2714
2715 if (OP(n) == TAIL || n > next)
2716 stringok = 0;
2717 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2718 DEBUG_PEEP("skip:",n,depth);
2719 NEXT_OFF(scan) += NEXT_OFF(n);
2720 next = n + NODE_STEP_REGNODE;
2721#ifdef DEBUGGING
2722 if (stringok)
2723 stop = n;
2724#endif
2725 n = regnext(n);
2726 }
2727 else if (stringok) {
786e8c11 2728 const unsigned int oldl = STR_LEN(scan);
07be1b83 2729 regnode * const nnext = regnext(n);
b2230d39 2730
87b8b349
KW
2731 /* XXX I (khw) kind of doubt that this works on platforms where
2732 * U8_MAX is above 255 because of lots of other assumptions */
b2230d39
KW
2733 if (oldl + STR_LEN(n) > U8_MAX)
2734 break;
07be1b83
YO
2735
2736 DEBUG_PEEP("merg",n,depth);
07be1b83 2737 merged++;
b2230d39 2738
07be1b83
YO
2739 NEXT_OFF(scan) += NEXT_OFF(n);
2740 STR_LEN(scan) += STR_LEN(n);
2741 next = n + NODE_SZ_STR(n);
2742 /* Now we can overwrite *n : */
2743 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2744#ifdef DEBUGGING
2745 stop = next - 1;
2746#endif
2747 n = nnext;
2748 if (stopnow) break;
2749 }
2750
d47053eb
RGS
2751#ifdef EXPERIMENTAL_INPLACESCAN
2752 if (flags && !NEXT_OFF(n)) {
2753 DEBUG_PEEP("atch", val, depth);
2754 if (reg_off_by_arg[OP(n)]) {
2755 ARG_SET(n, val - n);
2756 }
2757 else {
2758 NEXT_OFF(n) = val - n;
2759 }
2760 stopnow = 1;
2761 }
07be1b83
YO
2762#endif
2763 }
2c2b7f86 2764
9d071ca8 2765 *min_subtract = 0;
f758bddf 2766 *has_exactf_sharp_s = FALSE;
f646642f 2767
3f410cf6
KW
2768 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2769 * can now analyze for sequences of problematic code points. (Prior to
2770 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2771 * hence missed). The sequences only happen in folding, hence for any
2772 * non-EXACT EXACTish node */
86d6fcad 2773 if (OP(scan) != EXACT) {
0a982f06
KW
2774 const U8 * const s0 = (U8*) STRING(scan);
2775 const U8 * s = s0;
2776 const U8 * const s_end = s0 + STR_LEN(scan);
f758bddf
KW
2777
2778 /* One pass is made over the node's string looking for all the
2779 * possibilities. to avoid some tests in the loop, there are two main
2780 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2781 * non-UTF-8 */
2782 if (UTF) {
86d6fcad 2783
0a982f06
KW
2784 /* Examine the string for a multi-character fold sequence. UTF-8
2785 * patterns have all characters pre-folded by the time this code is
2786 * executed */
2787 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2788 length sequence we are looking for is 2 */
86d6fcad 2789 {
0a982f06
KW
2790 int count = 0;
2791 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2792 if (! len) { /* Not a multi-char fold: get next char */
2793 s += UTF8SKIP(s);
2794 continue;
2795 }
bb914485 2796
0a982f06
KW
2797 /* Nodes with 'ss' require special handling, except for EXACTFL
2798 * and EXACTFA for which there is no multi-char fold to this */
2799 if (len == 2 && *s == 's' && *(s+1) == 's'
2800 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2801 {
2802 count = 2;
2803 OP(scan) = EXACTFU_SS;
2804 s += 2;
2805 }
2806 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2807 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2808 COMBINING_DIAERESIS_UTF8
2809 COMBINING_ACUTE_ACCENT_UTF8,
2810 6)
2811 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2812 COMBINING_DIAERESIS_UTF8
2813 COMBINING_ACUTE_ACCENT_UTF8,
2814 6)))
2815 {
2816 count = 3;
2817
2818 /* These two folds require special handling by trie's, so
2819 * change the node type to indicate this. If EXACTFA and
2820 * EXACTFL were ever to be handled by trie's, this would
2821 * have to be changed. If this node has already been
2822 * changed to EXACTFU_SS in this loop, leave it as is. (I
2823 * (khw) think it doesn't matter in regexec.c for UTF
2824 * patterns, but no need to change it */
2825 if (OP(scan) == EXACTFU) {
2826 OP(scan) = EXACTFU_TRICKYFOLD;
2827 }
2828 s += 6;
2829 }
2830 else { /* Here is a generic multi-char fold. */
2831 const U8* multi_end = s + len;
2832
2833 /* Count how many characters in it. In the case of /l and
2834 * /aa, no folds which contain ASCII code points are
2835 * allowed, so check for those, and skip if found. (In
2836 * EXACTFL, no folds are allowed to any Latin1 code point,
2837 * not just ASCII. But there aren't any of these
2838 * currently, nor ever likely, so don't take the time to
2839 * test for them. The code that generates the
2840 * is_MULTI_foo() macros croaks should one actually get put
2841 * into Unicode .) */
2842 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2843 count = utf8_length(s, multi_end);
2844 s = multi_end;
2845 }
2846 else {
2847 while (s < multi_end) {
2848 if (isASCII(*s)) {
2849 s++;
2850 goto next_iteration;
2851 }
2852 else {
2853 s += UTF8SKIP(s);
2854 }
2855 count++;
2856 }
2857 }
2858 }
f758bddf 2859
0a982f06
KW
2860 /* The delta is how long the sequence is minus 1 (1 is how long
2861 * the character that folds to the sequence is) */
2862 *min_subtract += count - 1;
2863 next_iteration: ;
bb914485
KW
2864 }
2865 }
f758bddf 2866 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
bb914485 2867
0a982f06
KW
2868 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2869 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2870 * nodes can't have multi-char folds to this range (and there are
2db97d41 2871 * no existing ones in the upper latin1 range). In the EXACTF
0a982f06
KW
2872 * case we look also for the sharp s, which can be in the final
2873 * position. Otherwise we can stop looking 1 byte earlier because
2874 * have to find at least two characters for a multi-fold */
f758bddf
KW
2875 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2876
0a982f06
KW
2877 /* The below is perhaps overboard, but this allows us to save a
2878 * test each time through the loop at the expense of a mask. This
2879 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2880 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2881 * are 64. This uses an exclusive 'or' to find that bit and then
2882 * inverts it to form a mask, with just a single 0, in the bit
2883 * position where 'S' and 's' differ. */
2884 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2885 const U8 s_masked = 's' & S_or_s_mask;
2886
2887 while (s < upper) {
40b1ba4f 2888 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
0a982f06
KW
2889 if (! len) { /* Not a multi-char fold. */
2890 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2891 {
2892 *has_exactf_sharp_s = TRUE;
2893 }
2894 s++;
2895 continue;
2896 }
2897
2898 if (len == 2
2899 && ((*s & S_or_s_mask) == s_masked)
2900 && ((*(s+1) & S_or_s_mask) == s_masked))
2901 {
2902
2903 /* EXACTF nodes need to know that the minimum length
2904 * changed so that a sharp s in the string can match this
2905 * ss in the pattern, but they remain EXACTF nodes, as they
2906 * won't match this unless the target string is is UTF-8,
2907 * which we don't know until runtime */
2908 if (OP(scan) != EXACTF) {
2909 OP(scan) = EXACTFU_SS;
2910 }
86d6fcad 2911 }
0a982f06
KW
2912
2913 *min_subtract += len - 1;
2914 s += len;
86d6fcad
KW
2915 }
2916 }
07be1b83 2917 }
3f410cf6 2918
07be1b83 2919#ifdef DEBUGGING
bb789b09
DM
2920 /* Allow dumping but overwriting the collection of skipped
2921 * ops and/or strings with fake optimized ops */
07be1b83
YO
2922 n = scan + NODE_SZ_STR(scan);
2923 while (n <= stop) {
bb789b09
DM
2924 OP(n) = OPTIMIZED;
2925 FLAGS(n) = 0;
2926 NEXT_OFF(n) = 0;
07be1b83
YO
2927 n++;
2928 }
2929#endif
2930 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2931 return stopnow;
2932}
2933
486ec47a 2934/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2935 Finds fixed substrings. */
2936
a0288114 2937/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2938 to the position after last scanned or to NULL. */
2939
40d049e4
YO
2940#define INIT_AND_WITHP \
2941 assert(!and_withp); \
2942 Newx(and_withp,1,struct regnode_charclass_class); \
2943 SAVEFREEPV(and_withp)
07be1b83 2944
b515a41d 2945/* this is a chain of data about sub patterns we are processing that
486ec47a 2946 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2947 we can simulate recursion without losing state. */
2948struct scan_frame;
2949typedef struct scan_frame {
2950 regnode *last; /* last node to process in this frame */
2951 regnode *next; /* next node to process when last is reached */
2952 struct scan_frame *prev; /*previous frame*/
2953 I32 stop; /* what stopparen do we use */
2954} scan_frame;
2955
304ee84b
YO
2956
2957#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2958
e1d1eefb
YO
2959#define CASE_SYNST_FNC(nAmE) \
2960case nAmE: \
2961 if (flags & SCF_DO_STCLASS_AND) { \
2962 for (value = 0; value < 256; value++) \
2963 if (!is_ ## nAmE ## _cp(value)) \
2964 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2965 } \
2966 else { \
2967 for (value = 0; value < 256; value++) \
2968 if (is_ ## nAmE ## _cp(value)) \
2969 ANYOF_BITMAP_SET(data->start_class, value); \
2970 } \
2971 break; \
2972case N ## nAmE: \
2973 if (flags & SCF_DO_STCLASS_AND) { \
2974 for (value = 0; value < 256; value++) \
2975 if (is_ ## nAmE ## _cp(value)) \
2976 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2977 } \
2978 else { \
2979 for (value = 0; value < 256; value++) \
2980 if (!is_ ## nAmE ## _cp(value)) \
2981 ANYOF_BITMAP_SET(data->start_class, value); \
2982 } \
2983 break
2984
2985
2986
76e3520e 2987STATIC I32
40d049e4 2988S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 2989 I32 *minlenp, I32 *deltap,
40d049e4
YO
2990 regnode *last,
2991 scan_data_t *data,
2992 I32 stopparen,
2993 U8* recursed,
2994 struct regnode_charclass_class *and_withp,
2995 U32 flags, U32 depth)
c277df42
IZ
2996 /* scanp: Start here (read-write). */
2997 /* deltap: Write maxlen-minlen here. */
2998 /* last: Stop before this one. */
40d049e4
YO
2999 /* data: string data about the pattern */
3000 /* stopparen: treat close N as END */
3001 /* recursed: which subroutines have we recursed into */
3002 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 3003{
97aff369 3004 dVAR;
2d608413
KW
3005 I32 min = 0; /* There must be at least this number of characters to match */
3006 I32 pars = 0, code;
c277df42
IZ
3007 regnode *scan = *scanp, *next;
3008 I32 delta = 0;
3009 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 3010 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
3011 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3012 scan_data_t data_fake;
a3621e74 3013 SV *re_trie_maxbuff = NULL;
786e8c11 3014 regnode *first_non_open = scan;
e2e6a0f1 3015 I32 stopmin = I32_MAX;
8aa23a47 3016 scan_frame *frame = NULL;
a3621e74 3017 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 3018
7918f24d
NC
3019 PERL_ARGS_ASSERT_STUDY_CHUNK;
3020
13a24bad 3021#ifdef DEBUGGING
40d049e4 3022 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 3023#endif
40d049e4 3024
786e8c11 3025 if ( depth == 0 ) {
40d049e4 3026 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
3027 first_non_open=regnext(first_non_open);
3028 }
3029
b81d288d 3030
8aa23a47
YO
3031 fake_study_recurse:
3032 while ( scan && OP(scan) != END && scan < last ){
2d608413
KW
3033 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3034 node length to get a real minimum (because
3035 the folded version may be shorter) */
f758bddf 3036 bool has_exactf_sharp_s = FALSE;
8aa23a47 3037 /* Peephole optimizer: */
304ee84b 3038 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3039 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3040
3041 /* Its not clear to khw or hv why this is done here, and not in the
3042 * clauses that deal with EXACT nodes. khw's guess is that it's
3043 * because of a previous design */
9d071ca8 3044 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3045
3046 /* Follow the next-chain of the current node and optimize
3047 away all the NOTHINGs from it. */
3048 if (OP(scan) != CURLYX) {
3049 const int max = (reg_off_by_arg[OP(scan)]
3050 ? I32_MAX
3051 /* I32 may be smaller than U16 on CRAYs! */
3052 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3053 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3054 int noff;
3055 regnode *n = scan;
686b73d4 3056
8aa23a47
YO
3057 /* Skip NOTHING and LONGJMP. */
3058 while ((n = regnext(n))
3059 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3060 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3061 && off + noff < max)
3062 off += noff;
3063 if (reg_off_by_arg[OP(scan)])
3064 ARG(scan) = off;
3065 else
3066 NEXT_OFF(scan) = off;
3067 }
a3621e74 3068
c277df42 3069
8aa23a47
YO
3070
3071 /* The principal pseudo-switch. Cannot be a switch, since we
3072 look into several different things. */
3073 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3074 || OP(scan) == IFTHEN) {
3075 next = regnext(scan);
3076 code = OP(scan);
3077 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3078
8aa23a47
YO
3079 if (OP(next) == code || code == IFTHEN) {
3080 /* NOTE - There is similar code to this block below for handling
3081 TRIE nodes on a re-study. If you change stuff here check there
3082 too. */
3083 I32 max1 = 0, min1 = I32_MAX, num = 0;
3084 struct regnode_charclass_class accum;
3085 regnode * const startbranch=scan;
686b73d4 3086
8aa23a47 3087 if (flags & SCF_DO_SUBSTR)
304ee84b 3088 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3089 if (flags & SCF_DO_STCLASS)
e755fd73 3090 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3091
3092 while (OP(scan) == code) {
3093 I32 deltanext, minnext, f = 0, fake;
3094 struct regnode_charclass_class this_class;
3095
3096 num++;
3097 data_fake.flags = 0;
3098 if (data) {
3099 data_fake.whilem_c = data->whilem_c;
3100 data_fake.last_closep = data->last_closep;
3101 }
3102 else
3103 data_fake.last_closep = &fake;
58e23c8d
YO
3104
3105 data_fake.pos_delta = delta;
8aa23a47
YO
3106 next = regnext(scan);
3107 scan = NEXTOPER(scan);
3108 if (code != BRANCH)
c277df42 3109 scan = NEXTOPER(scan);
8aa23a47 3110 if (flags & SCF_DO_STCLASS) {
e755fd73 3111 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3112 data_fake.start_class = &this_class;
3113 f = SCF_DO_STCLASS_AND;
58e23c8d 3114 }
8aa23a47
YO
3115 if (flags & SCF_WHILEM_VISITED_POS)
3116 f |= SCF_WHILEM_VISITED_POS;
3117
3118 /* we suppose the run is continuous, last=next...*/
3119 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3120 next, &data_fake,
3121 stopparen, recursed, NULL, f,depth+1);
3122 if (min1 > minnext)
3123 min1 = minnext;
3124 if (max1 < minnext + deltanext)
3125 max1 = minnext + deltanext;
3126 if (deltanext == I32_MAX)
3127 is_inf = is_inf_internal = 1;
3128 scan = next;
3129 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3130 pars++;
3131 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3132 if ( stopmin > minnext)
3133 stopmin = min + min1;
3134 flags &= ~SCF_DO_SUBSTR;
3135 if (data)
3136 data->flags |= SCF_SEEN_ACCEPT;
3137 }
3138 if (data) {
3139 if (data_fake.flags & SF_HAS_EVAL)
3140 data->flags |= SF_HAS_EVAL;
3141 data->whilem_c = data_fake.whilem_c;
3dab1dad 3142 }
8aa23a47 3143 if (flags & SCF_DO_STCLASS)
3fffb88a 3144 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3145 }
3146 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3147 min1 = 0;
3148 if (flags & SCF_DO_SUBSTR) {
3149 data->pos_min += min1;
3150 data->pos_delta += max1 - min1;
3151 if (max1 != min1 || is_inf)
3152 data->longest = &(data->longest_float);
3153 }
3154 min += min1;
3155 delta += max1 - min1;
3156 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3157 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3158 if (min1) {
3159 cl_and(data->start_class, and_withp);
3160 flags &= ~SCF_DO_STCLASS;
653099ff 3161 }
8aa23a47
YO
3162 }
3163 else if (flags & SCF_DO_STCLASS_AND) {
3164 if (min1) {
3165 cl_and(data->start_class, &accum);
3166 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3167 }
8aa23a47
YO
3168 else {
3169 /* Switch to OR mode: cache the old value of
3170 * data->start_class */
3171 INIT_AND_WITHP;
3172 StructCopy(data->start_class, and_withp,
3173 struct regnode_charclass_class);
3174 flags &= ~SCF_DO_STCLASS_AND;
3175 StructCopy(&accum, data->start_class,
3176 struct regnode_charclass_class);
3177 flags |= SCF_DO_STCLASS_OR;
3178 data->start_class->flags |= ANYOF_EOS;
de0c8cb8 3179 }
8aa23a47 3180 }
a3621e74 3181
8aa23a47
YO
3182 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3183 /* demq.
a3621e74 3184
8aa23a47
YO
3185 Assuming this was/is a branch we are dealing with: 'scan' now
3186 points at the item that follows the branch sequence, whatever
3187 it is. We now start at the beginning of the sequence and look
3188 for subsequences of
a3621e74 3189
8aa23a47
YO
3190 BRANCH->EXACT=>x1
3191 BRANCH->EXACT=>x2
3192 tail
a3621e74 3193
8aa23a47 3194 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3195
486ec47a 3196 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3197 element into a trie and then add the subsequent branch exact
3198 strings to the trie.
a3621e74 3199
8aa23a47 3200 We have two cases
a3621e74 3201
3b753521 3202 1. patterns where the whole set of branches can be converted.
a3621e74 3203
8aa23a47 3204 2. patterns where only a subset can be converted.
a3621e74 3205
8aa23a47
YO
3206 In case 1 we can replace the whole set with a single regop
3207 for the trie. In case 2 we need to keep the start and end
3b753521 3208 branches so
a3621e74 3209
8aa23a47
YO
3210 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3211 becomes BRANCH TRIE; BRANCH X;
786e8c11 3212
8aa23a47
YO
3213 There is an additional case, that being where there is a
3214 common prefix, which gets split out into an EXACT like node
3215 preceding the TRIE node.
a3621e74 3216
8aa23a47
YO
3217 If x(1..n)==tail then we can do a simple trie, if not we make
3218 a "jump" trie, such that when we match the appropriate word
486ec47a 3219 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3220 a nested if into a case structure of sorts.
b515a41d 3221
8aa23a47 3222 */
686b73d4 3223
8aa23a47
YO
3224 int made=0;
3225 if (!re_trie_maxbuff) {
3226 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3227 if (!SvIOK(re_trie_maxbuff))
3228 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3229 }
3230 if ( SvIV(re_trie_maxbuff)>=0 ) {
3231 regnode *cur;
3232 regnode *first = (regnode *)NULL;
3233 regnode *last = (regnode *)NULL;
3234 regnode *tail = scan;
fab2782b 3235 U8 trietype = 0;
8aa23a47 3236 U32 count=0;
a3621e74
YO
3237
3238#ifdef DEBUGGING
8aa23a47 3239 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3240#endif
8aa23a47
YO
3241 /* var tail is used because there may be a TAIL
3242 regop in the way. Ie, the exacts will point to the
3243 thing following the TAIL, but the last branch will
3244 point at the TAIL. So we advance tail. If we
3245 have nested (?:) we may have to move through several
3246 tails.
3247 */
3248
3249 while ( OP( tail ) == TAIL ) {
3250 /* this is the TAIL generated by (?:) */
3251 tail = regnext( tail );
3252 }
a3621e74 3253
8aa23a47 3254
df826430 3255 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3256 regprop(RExC_rx, mysv, tail );
3257 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3258 (int)depth * 2 + 2, "",
3259 "Looking for TRIE'able sequences. Tail node is: ",
3260 SvPV_nolen_const( mysv )
3261 );
3262 });
3263
3264 /*
3265
fab2782b
YO
3266 Step through the branches
3267 cur represents each branch,
3268 noper is the first thing to be matched as part of that branch
3269 noper_next is the regnext() of that node.
3270
3271 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3272 via a "jump trie" but we also support building with NOJUMPTRIE,
3273 which restricts the trie logic to structures like /FOO|BAR/.
3274
3275 If noper is a trieable nodetype then the branch is a possible optimization
3276 target. If we are building under NOJUMPTRIE then we require that noper_next
3277 is the same as scan (our current position in the regex program).
3278
3279 Once we have two or more consecutive such branches we can create a
3280 trie of the EXACT's contents and stitch it in place into the program.
3281
3282 If the sequence represents all of the branches in the alternation we
3283 replace the entire thing with a single TRIE node.
3284
3285 Otherwise when it is a subsequence we need to stitch it in place and
3286 replace only the relevant branches. This means the first branch has
3287 to remain as it is used by the alternation logic, and its next pointer,
3288 and needs to be repointed at the item on the branch chain following
3289 the last branch we have optimized away.
3290
3291 This could be either a BRANCH, in which case the subsequence is internal,
3292 or it could be the item following the branch sequence in which case the
3293 subsequence is at the end (which does not necessarily mean the first node
3294 is the start of the alternation).
3295
3296 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3297
3298 optype | trietype
3299 ----------------+-----------
3300 NOTHING | NOTHING
3301 EXACT | EXACT
3302 EXACTFU | EXACTFU
3303 EXACTFU_SS | EXACTFU
3304 EXACTFU_TRICKYFOLD | EXACTFU
3305 EXACTFA | 0
3306
8aa23a47
YO
3307
3308 */
fab2782b
YO
3309#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3310 ( EXACT == (X) ) ? EXACT : \
3311 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3312 0 )
8aa23a47
YO
3313
3314 /* dont use tail as the end marker for this traverse */
3315 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3316 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3317 U8 noper_type = OP( noper );
3318 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3319#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3320 regnode * const noper_next = regnext( noper );
df826430
YO
3321 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3322 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3323#endif
3324
df826430 3325 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3326 regprop(RExC_rx, mysv, cur);
3327 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3328 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3329
3330 regprop(RExC_rx, mysv, noper);
3331 PerlIO_printf( Perl_debug_log, " -> %s",
3332 SvPV_nolen_const(mysv));
3333
3334 if ( noper_next ) {
3335 regprop(RExC_rx, mysv, noper_next );
3336 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3337 SvPV_nolen_const(mysv));
3338 }
df826430
YO
3339 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3340 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3341 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3342 );
8aa23a47 3343 });
fab2782b
YO
3344
3345 /* Is noper a trieable nodetype that can be merged with the
3346 * current trie (if there is one)? */
3347 if ( noper_trietype
3348 &&
3349 (
df826430
YO
3350 ( noper_trietype == NOTHING)
3351 || ( trietype == NOTHING )
a40630bf 3352 || ( trietype == noper_trietype )
fab2782b 3353 )
786e8c11 3354#ifdef NOJUMPTRIE
8aa23a47 3355 && noper_next == tail
786e8c11 3356#endif
8aa23a47
YO
3357 && count < U16_MAX)
3358 {
fab2782b
YO
3359 /* Handle mergable triable node
3360 * Either we are the first node in a new trieable sequence,
3361 * in which case we do some bookkeeping, otherwise we update
3362 * the end pointer. */
fab2782b 3363 if ( !first ) {
3b6759a6 3364 first = cur;
df826430
YO
3365 if ( noper_trietype == NOTHING ) {
3366#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3367 regnode * const noper_next = regnext( noper );
3b6759a6 3368 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3369 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3370#endif
3371
190c1910 3372 if ( noper_next_trietype ) {
df826430 3373 trietype = noper_next_trietype;
190c1910
YO
3374 } else if (noper_next_type) {
3375 /* a NOTHING regop is 1 regop wide. We need at least two
3376 * for a trie so we can't merge this in */
3377 first = NULL;
3378 }
3379 } else {
3380 trietype = noper_trietype;
3b6759a6 3381 }
8aa23a47 3382 } else {
fab2782b
YO
3383 if ( trietype == NOTHING )
3384 trietype = noper_trietype;
8aa23a47
YO
3385 last = cur;
3386 }
df826430
YO
3387 if (first)
3388 count++;
fab2782b
YO
3389 } /* end handle mergable triable node */
3390 else {
3391 /* handle unmergable node -
3392 * noper may either be a triable node which can not be tried
3393 * together with the current trie, or a non triable node */
729aaeb5
YO
3394 if ( last ) {
3395 /* If last is set and trietype is not NOTHING then we have found
3396 * at least two triable branch sequences in a row of a similar
3397 * trietype so we can turn them into a trie. If/when we
3398 * allow NOTHING to start a trie sequence this condition will be
3399 * required, and it isn't expensive so we leave it in for now. */
e6351b37 3400 if ( trietype && trietype != NOTHING )
729aaeb5
YO
3401 make_trie( pRExC_state,
3402 startbranch, first, cur, tail, count,
3403 trietype, depth+1 );
fab2782b 3404 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3405 }
fab2782b 3406 if ( noper_trietype
786e8c11 3407#ifdef NOJUMPTRIE
8aa23a47 3408 && noper_next == tail
786e8c11 3409#endif
8aa23a47 3410 ){
fab2782b 3411 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3412 count = 1;
3413 first = cur;
fab2782b
YO
3414 trietype = noper_trietype;
3415 } else if (first) {
3416 /* if we already saw a first but the current node is not triable then we have
3417 * to reset the first information. */
8aa23a47
YO
3418 count = 0;
3419 first = NULL;
fab2782b 3420 trietype = 0;
8aa23a47 3421 }
fab2782b
YO
3422 } /* end handle unmergable node */
3423 } /* loop over branches */
df826430 3424 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3425 regprop(RExC_rx, mysv, cur);
3426 PerlIO_printf( Perl_debug_log,
3427 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3428 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3429
3430 });
e6351b37 3431 if ( last && trietype ) {
3b6759a6
YO
3432 if ( trietype != NOTHING ) {
3433 /* the last branch of the sequence was part of a trie,
3434 * so we have to construct it here outside of the loop
3435 */
3436 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3437#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3438 if ( ((made == MADE_EXACT_TRIE &&
3439 startbranch == first)
3440 || ( first_non_open == first )) &&
3441 depth==0 ) {
3442 flags |= SCF_TRIE_RESTUDY;
3443 if ( startbranch == first
3444 && scan == tail )
3445 {
3446 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3447 }
8aa23a47 3448 }
3dab1dad 3449#endif
3b6759a6
YO
3450 } else {
3451 /* at this point we know whatever we have is a NOTHING sequence/branch
3452 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3453 */
3454 if ( startbranch == first ) {
3455 regnode *opt;
3456 /* the entire thing is a NOTHING sequence, something like this:
3457 * (?:|) So we can turn it into a plain NOTHING op. */
3458 DEBUG_TRIE_COMPILE_r({
3459 regprop(RExC_rx, mysv, cur);
3460 PerlIO_printf( Perl_debug_log,
3461 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3462 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3463
3464 });
3465 OP(startbranch)= NOTHING;
3466 NEXT_OFF(startbranch)= tail - startbranch;
3467 for ( opt= startbranch + 1; opt < tail ; opt++ )
3468 OP(opt)= OPTIMIZED;
3469 }
3470 }
fab2782b
YO
3471 } /* end if ( last) */
3472 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3473
3474 } /* do trie */
3475
653099ff 3476 }
8aa23a47
YO
3477 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3478 scan = NEXTOPER(NEXTOPER(scan));
3479 } else /* single branch is optimized. */
3480 scan = NEXTOPER(scan);
3481 continue;
3482 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3483 scan_frame *newframe = NULL;
3484 I32 paren;
3485 regnode *start;
3486 regnode *end;
3487
3488 if (OP(scan) != SUSPEND) {
3489 /* set the pointer */
3490 if (OP(scan) == GOSUB) {
3491 paren = ARG(scan);
3492 RExC_recurse[ARG2L(scan)] = scan;
3493 start = RExC_open_parens[paren-1];
3494 end = RExC_close_parens[paren-1];
3495 } else {
3496 paren = 0;
f8fc2ecf 3497 start = RExC_rxi->program + 1;
8aa23a47
YO
3498 end = RExC_opend;
3499 }
3500 if (!recursed) {
3501 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3502 SAVEFREEPV(recursed);
3503 }
3504 if (!PAREN_TEST(recursed,paren+1)) {
3505 PAREN_SET(recursed,paren+1);
3506 Newx(newframe,1,scan_frame);
3507 } else {
3508 if (flags & SCF_DO_SUBSTR) {
304ee84b 3509 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3510 data->longest = &(data->longest_float);
3511 }
3512 is_inf = is_inf_internal = 1;
3513 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3514 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3515 flags &= ~SCF_DO_STCLASS;
3516 }
3517 } else {
3518 Newx(newframe,1,scan_frame);
3519 paren = stopparen;
3520 start = scan+2;
3521 end = regnext(scan);
3522 }
3523 if (newframe) {
3524 assert(start);
3525 assert(end);
3526 SAVEFREEPV(newframe);
3527 newframe->next = regnext(scan);
3528 newframe->last = last;
3529 newframe->stop = stopparen;
3530 newframe->prev = frame;
3531
3532 frame = newframe;
3533 scan = start;
3534 stopparen = paren;
3535 last = end;
3536
3537 continue;
3538 }
3539 }
3540 else if (OP(scan) == EXACT) {
3541 I32 l = STR_LEN(scan);
3542 UV uc;
3543 if (UTF) {
3544 const U8 * const s = (U8*)STRING(scan);
4b88fb76 3545 uc = utf8_to_uvchr_buf(s, s + l, NULL);
8aa23a47 3546 l = utf8_length(s, s + l);
8aa23a47
YO
3547 } else {
3548 uc = *((U8*)STRING(scan));
3549 }
3550 min += l;
3551 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3552 /* The code below prefers earlier match for fixed
3553 offset, later match for variable offset. */
3554 if (data->last_end == -1) { /* Update the start info. */
3555 data->last_start_min = data->pos_min;
3556 data->last_start_max = is_inf
3557 ? I32_MAX : data->pos_min + data->pos_delta;
b515a41d 3558 }
8aa23a47
YO
3559 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3560 if (UTF)
3561 SvUTF8_on(data->last_found);
3562 {
3563 SV * const sv = data->last_found;
3564 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3565 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3566 if (mg && mg->mg_len >= 0)
3567 mg->mg_len += utf8_length((U8*)STRING(scan),
3568 (U8*)STRING(scan)+STR_LEN(scan));
b515a41d 3569 }
8aa23a47
YO
3570 data->last_end = data->pos_min + l;
3571 data->pos_min += l; /* As in the first entry. */
3572 data->flags &= ~SF_BEFORE_EOL;
3573 }
3574 if (flags & SCF_DO_STCLASS_AND) {
3575 /* Check whether it is compatible with what we know already! */
3576 int compat = 1;
3577
54251c2e 3578
486ec47a 3579 /* If compatible, we or it in below. It is compatible if is
54251c2e
KW
3580 * in the bitmp and either 1) its bit or its fold is set, or 2)
3581 * it's for a locale. Even if there isn't unicode semantics
3582 * here, at runtime there may be because of matching against a
3583 * utf8 string, so accept a possible false positive for