This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Silence valgrind warning
[perl5.git] / regcomp.c
CommitLineData
a0d0e21e
LW
1/* regcomp.c
2 */
3
4/*
4ac71550
TC
5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
6 *
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
a0d0e21e
LW
8 */
9
61296642
DM
10/* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
166f8a29 12 * a regular expression.
e4a054ea
DM
13 *
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
166f8a29
DM
18 */
19
a687059c
LW
20/* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
22 */
23
24/* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
27 */
28
e50aee73 29/* The names of the functions have been changed from regcomp and
3b753521 30 * regexec to pregcomp and pregexec in order to avoid conflicts
e50aee73
AD
31 * with the POSIX routines of the same names.
32*/
33
b9d5759e 34#ifdef PERL_EXT_RE_BUILD
54df2634 35#include "re_top.h"
b81d288d 36#endif
56953603 37
a687059c 38/*
e50aee73 39 * pregcomp and pregexec -- regsub and regerror are not used in perl
a687059c
LW
40 *
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
43 *
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
47 *
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
50 * from defects in it.
51 *
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
54 *
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
57 *
58 *
59 **** Alterations to Henry's code are...
60 ****
4bb101f2 61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
1129b882
NC
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
a687059c 64 ****
9ef589d8
LW
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
67
a687059c
LW
68 *
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
72 */
73#include "EXTERN.h"
864dbfa3 74#define PERL_IN_REGCOMP_C
a687059c 75#include "perl.h"
d06ea78c 76
acfe0abc 77#ifndef PERL_IN_XSUB_RE
d06ea78c
GS
78# include "INTERN.h"
79#endif
c277df42
IZ
80
81#define REG_COMP_C
54df2634
NC
82#ifdef PERL_IN_XSUB_RE
83# include "re_comp.h"
84#else
85# include "regcomp.h"
86#endif
a687059c 87
04e98a4d 88#include "dquote_static.c"
370b8f2f
TC
89#ifndef PERL_IN_XSUB_RE
90# include "charclass_invlists.h"
91#endif
04e98a4d 92
94dc5c2d
KW
93#define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
94
d4cce5f1 95#ifdef op
11343788 96#undef op
d4cce5f1 97#endif /* op */
11343788 98
fe14fcc3 99#ifdef MSDOS
7e4e8c89 100# if defined(BUGGY_MSC6)
fe14fcc3 101 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
7e4e8c89 102# pragma optimize("a",off)
fe14fcc3 103 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
7e4e8c89
NC
104# pragma optimize("w",on )
105# endif /* BUGGY_MSC6 */
fe14fcc3
LW
106#endif /* MSDOS */
107
a687059c
LW
108#ifndef STATIC
109#define STATIC static
110#endif
111
b1603ef8 112
830247a4 113typedef struct RExC_state_t {
514a91f1
DM
114 U32 flags; /* RXf_* are we folding, multilining? */
115 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 116 char *precomp; /* uncompiled string. */
288b8c02 117 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf
YO
118 regexp *rx; /* perl core regexp structure */
119 regexp_internal *rxi; /* internal data for regexp object pprivate field */
fac92740 120 char *start; /* Start of input for compile */
830247a4
IZ
121 char *end; /* End of input for compile */
122 char *parse; /* Input-scan pointer. */
123 I32 whilem_seen; /* number of WHILEM in this expr */
fac92740 124 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 125 regnode *emit_bound; /* First regnode outside of the allocated space */
ffc61ed2 126 regnode *emit; /* Code-emit pointer; &regdummy = don't = compiling */
830247a4
IZ
127 I32 naughty; /* How bad is this pattern? */
128 I32 sawback; /* Did we see \1, ...? */
129 U32 seen;
130 I32 size; /* Code size. */
c74340f9
YO
131 I32 npar; /* Capture buffer count, (OPEN). */
132 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 133 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
134 I32 extralen;
135 I32 seen_zerolen;
136 I32 seen_evals;
40d049e4
YO
137 regnode **open_parens; /* pointers to open parens */
138 regnode **close_parens; /* pointers to close parens */
139 regnode *opend; /* END node in program */
02daf0ab
YO
140 I32 utf8; /* whether the pattern is utf8 or not */
141 I32 orig_utf8; /* whether the pattern was originally in utf8 */
142 /* XXX use this for future optimisation of case
143 * where pattern must be upgraded to utf8. */
e40e74fe
KW
144 I32 uni_semantics; /* If a d charset modifier should use unicode
145 rules, even if the pattern is not in
146 utf8 */
81714fb9 147 HV *paren_names; /* Paren names */
1f1031fe 148
40d049e4
YO
149 regnode **recurse; /* Recurse regops */
150 I32 recurse_count; /* Number of recurse regops */
b57e4118 151 I32 in_lookbehind;
4624b182 152 I32 contains_locale;
bb3f3ed2 153 I32 override_recoding;
3d2bd50a 154 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 155 within pattern */
b1603ef8
DM
156 int num_code_blocks; /* size of code_blocks[] */
157 int code_index; /* next code_blocks[] slot */
830247a4
IZ
158#if ADD_TO_REGEXEC
159 char *starttry; /* -Dr: where regtry was called. */
160#define RExC_starttry (pRExC_state->starttry)
161#endif
3dab1dad 162#ifdef DEBUGGING
be8e71aa 163 const char *lastparse;
3dab1dad 164 I32 lastnum;
1f1031fe 165 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
166#define RExC_lastparse (pRExC_state->lastparse)
167#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 168#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 169#endif
830247a4
IZ
170} RExC_state_t;
171
e2509266 172#define RExC_flags (pRExC_state->flags)
514a91f1 173#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 174#define RExC_precomp (pRExC_state->precomp)
288b8c02 175#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 176#define RExC_rx (pRExC_state->rx)
f8fc2ecf 177#define RExC_rxi (pRExC_state->rxi)
fac92740 178#define RExC_start (pRExC_state->start)
830247a4
IZ
179#define RExC_end (pRExC_state->end)
180#define RExC_parse (pRExC_state->parse)
181#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
182#ifdef RE_TRACK_PATTERN_OFFSETS
183#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
184#endif
830247a4 185#define RExC_emit (pRExC_state->emit)
fac92740 186#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 187#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
188#define RExC_naughty (pRExC_state->naughty)
189#define RExC_sawback (pRExC_state->sawback)
190#define RExC_seen (pRExC_state->seen)
191#define RExC_size (pRExC_state->size)
192#define RExC_npar (pRExC_state->npar)
e2e6a0f1 193#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
194#define RExC_extralen (pRExC_state->extralen)
195#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
196#define RExC_seen_evals (pRExC_state->seen_evals)
1aa99e6b 197#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 198#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 199#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
200#define RExC_open_parens (pRExC_state->open_parens)
201#define RExC_close_parens (pRExC_state->close_parens)
202#define RExC_opend (pRExC_state->opend)
81714fb9 203#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
204#define RExC_recurse (pRExC_state->recurse)
205#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 206#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 207#define RExC_contains_locale (pRExC_state->contains_locale)
bb3f3ed2 208#define RExC_override_recoding (pRExC_state->override_recoding)
830247a4 209
cde0cee5 210
a687059c
LW
211#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
212#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
213 ((*s) == '{' && regcurly(s)))
a687059c 214
35c8bce7
LW
215#ifdef SPSTART
216#undef SPSTART /* dratted cpp namespace... */
217#endif
a687059c
LW
218/*
219 * Flags to be passed up and down.
220 */
a687059c 221#define WORST 0 /* Worst case. */
a3b492c3 222#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee
KW
223
224/* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
cf8c372d
KW
225 * character, and if utf8, must be invariant. Note that this is not the same
226 * thing as REGNODE_SIMPLE */
fda99bee 227#define SIMPLE 0x02
a3b492c3
YO
228#define SPSTART 0x04 /* Starts with * or +. */
229#define TRYAGAIN 0x08 /* Weeded out a declaration. */
230#define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
a687059c 231
3dab1dad
YO
232#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
233
07be1b83
YO
234/* whether trie related optimizations are enabled */
235#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
236#define TRIE_STUDY_OPT
786e8c11 237#define FULL_TRIE_STUDY
07be1b83
YO
238#define TRIE_STCLASS
239#endif
1de06328
YO
240
241
40d049e4
YO
242
243#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
244#define PBITVAL(paren) (1 << ((paren) & 7))
245#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
246#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
247#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
248
bbd61b5f
KW
249/* If not already in utf8, do a longjmp back to the beginning */
250#define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
251#define REQUIRE_UTF8 STMT_START { \
252 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
253 } STMT_END
40d049e4 254
1de06328
YO
255/* About scan_data_t.
256
257 During optimisation we recurse through the regexp program performing
258 various inplace (keyhole style) optimisations. In addition study_chunk
259 and scan_commit populate this data structure with information about
260 what strings MUST appear in the pattern. We look for the longest
3b753521 261 string that must appear at a fixed location, and we look for the
1de06328
YO
262 longest string that may appear at a floating location. So for instance
263 in the pattern:
264
265 /FOO[xX]A.*B[xX]BAR/
266
267 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
268 strings (because they follow a .* construct). study_chunk will identify
269 both FOO and BAR as being the longest fixed and floating strings respectively.
270
271 The strings can be composites, for instance
272
273 /(f)(o)(o)/
274
275 will result in a composite fixed substring 'foo'.
276
277 For each string some basic information is maintained:
278
279 - offset or min_offset
280 This is the position the string must appear at, or not before.
281 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
282 characters must match before the string we are searching for.
283 Likewise when combined with minlenp and the length of the string it
1de06328
YO
284 tells us how many characters must appear after the string we have
285 found.
286
287 - max_offset
288 Only used for floating strings. This is the rightmost point that
3b753521 289 the string can appear at. If set to I32 max it indicates that the
1de06328
YO
290 string can occur infinitely far to the right.
291
292 - minlenp
293 A pointer to the minimum length of the pattern that the string
294 was found inside. This is important as in the case of positive
295 lookahead or positive lookbehind we can have multiple patterns
296 involved. Consider
297
298 /(?=FOO).*F/
299
300 The minimum length of the pattern overall is 3, the minimum length
301 of the lookahead part is 3, but the minimum length of the part that
302 will actually match is 1. So 'FOO's minimum length is 3, but the
303 minimum length for the F is 1. This is important as the minimum length
304 is used to determine offsets in front of and behind the string being
305 looked for. Since strings can be composites this is the length of the
486ec47a 306 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
307 the length is calculated by study_chunk, so that the minimum lengths
308 are not known until the full pattern has been compiled, thus the
309 pointer to the value.
310
311 - lookbehind
312
313 In the case of lookbehind the string being searched for can be
314 offset past the start point of the final matching string.
315 If this value was just blithely removed from the min_offset it would
316 invalidate some of the calculations for how many chars must match
317 before or after (as they are derived from min_offset and minlen and
318 the length of the string being searched for).
319 When the final pattern is compiled and the data is moved from the
320 scan_data_t structure into the regexp structure the information
321 about lookbehind is factored in, with the information that would
322 have been lost precalculated in the end_shift field for the
323 associated string.
324
325 The fields pos_min and pos_delta are used to store the minimum offset
326 and the delta to the maximum offset at the current point in the pattern.
327
328*/
2c2d71f5
JH
329
330typedef struct scan_data_t {
1de06328
YO
331 /*I32 len_min; unused */
332 /*I32 len_delta; unused */
2c2d71f5
JH
333 I32 pos_min;
334 I32 pos_delta;
335 SV *last_found;
1de06328 336 I32 last_end; /* min value, <0 unless valid. */
2c2d71f5
JH
337 I32 last_start_min;
338 I32 last_start_max;
1de06328
YO
339 SV **longest; /* Either &l_fixed, or &l_float. */
340 SV *longest_fixed; /* longest fixed string found in pattern */
341 I32 offset_fixed; /* offset where it starts */
486ec47a 342 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
343 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
344 SV *longest_float; /* longest floating string found in pattern */
345 I32 offset_float_min; /* earliest point in string it can appear */
346 I32 offset_float_max; /* latest point in string it can appear */
486ec47a 347 I32 *minlen_float; /* pointer to the minlen relevant to the string */
1de06328 348 I32 lookbehind_float; /* is the position of the string modified by LB */
2c2d71f5
JH
349 I32 flags;
350 I32 whilem_c;
cb434fcc 351 I32 *last_closep;
653099ff 352 struct regnode_charclass_class *start_class;
2c2d71f5
JH
353} scan_data_t;
354
a687059c 355/*
e50aee73 356 * Forward declarations for pregcomp()'s friends.
a687059c 357 */
a0d0e21e 358
27da23d5 359static const scan_data_t zero_scan_data =
1de06328 360 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
361
362#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
363#define SF_BEFORE_SEOL 0x0001
364#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
365#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
366#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
367
09b7f37c
CB
368#ifdef NO_UNARY_PLUS
369# define SF_FIX_SHIFT_EOL (0+2)
370# define SF_FL_SHIFT_EOL (0+4)
371#else
372# define SF_FIX_SHIFT_EOL (+2)
373# define SF_FL_SHIFT_EOL (+4)
374#endif
c277df42
IZ
375
376#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
377#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
378
379#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
380#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
381#define SF_IS_INF 0x0040
382#define SF_HAS_PAR 0x0080
383#define SF_IN_PAR 0x0100
384#define SF_HAS_EVAL 0x0200
385#define SCF_DO_SUBSTR 0x0400
653099ff
GS
386#define SCF_DO_STCLASS_AND 0x0800
387#define SCF_DO_STCLASS_OR 0x1000
388#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 389#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 390
786e8c11 391#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 392#define SCF_SEEN_ACCEPT 0x8000
07be1b83 393
43fead97 394#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
395
396/* The enums for all these are ordered so things work out correctly */
a62b1201 397#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
cfaf538b 398#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
00b27cfc 399#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
400#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
401#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1
KW
402#define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
403#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a62b1201 404
43fead97 405#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 406
ffc61ed2 407#define OOB_UNICODE 12345678
93733859 408#define OOB_NAMEDCLASS -1
b8c5462f 409
a0ed51b3
LW
410#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
411#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
412
8615cb43 413
b45f050a
JF
414/* length of regex to show in messages that don't mark a position within */
415#define RegexLengthToShowInErrorMessages 127
416
417/*
418 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
419 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
420 * op/pragma/warn/regcomp.
421 */
7253e4e3
RK
422#define MARKER1 "<-- HERE" /* marker as it appears in the description */
423#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 424
7253e4e3 425#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
426
427/*
428 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
429 * arg. Show regex, up to a maximum length. If it's too long, chop and add
430 * "...".
431 */
58e23c8d 432#define _FAIL(code) STMT_START { \
bfed75c6 433 const char *ellipses = ""; \
ccb2c380
MP
434 IV len = RExC_end - RExC_precomp; \
435 \
436 if (!SIZE_ONLY) \
288b8c02 437 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
438 if (len > RegexLengthToShowInErrorMessages) { \
439 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
440 len = RegexLengthToShowInErrorMessages - 10; \
441 ellipses = "..."; \
442 } \
58e23c8d 443 code; \
ccb2c380 444} STMT_END
8615cb43 445
58e23c8d
YO
446#define FAIL(msg) _FAIL( \
447 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
448 msg, (int)len, RExC_precomp, ellipses))
449
450#define FAIL2(msg,arg) _FAIL( \
451 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
452 arg, (int)len, RExC_precomp, ellipses))
453
b45f050a 454/*
b45f050a
JF
455 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
456 */
ccb2c380 457#define Simple_vFAIL(m) STMT_START { \
a28509cc 458 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
459 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
460 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
461} STMT_END
b45f050a
JF
462
463/*
464 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
465 */
ccb2c380
MP
466#define vFAIL(m) STMT_START { \
467 if (!SIZE_ONLY) \
288b8c02 468 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
469 Simple_vFAIL(m); \
470} STMT_END
b45f050a
JF
471
472/*
473 * Like Simple_vFAIL(), but accepts two arguments.
474 */
ccb2c380 475#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 476 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
477 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
478 (int)offset, RExC_precomp, RExC_precomp + offset); \
479} STMT_END
b45f050a
JF
480
481/*
482 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
483 */
ccb2c380
MP
484#define vFAIL2(m,a1) STMT_START { \
485 if (!SIZE_ONLY) \
288b8c02 486 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
487 Simple_vFAIL2(m, a1); \
488} STMT_END
b45f050a
JF
489
490
491/*
492 * Like Simple_vFAIL(), but accepts three arguments.
493 */
ccb2c380 494#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 495 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
496 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
497 (int)offset, RExC_precomp, RExC_precomp + offset); \
498} STMT_END
b45f050a
JF
499
500/*
501 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
502 */
ccb2c380
MP
503#define vFAIL3(m,a1,a2) STMT_START { \
504 if (!SIZE_ONLY) \
288b8c02 505 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
506 Simple_vFAIL3(m, a1, a2); \
507} STMT_END
b45f050a
JF
508
509/*
510 * Like Simple_vFAIL(), but accepts four arguments.
511 */
ccb2c380 512#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 513 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
514 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
515 (int)offset, RExC_precomp, RExC_precomp + offset); \
516} STMT_END
b45f050a 517
668c081a 518#define ckWARNreg(loc,m) STMT_START { \
a28509cc 519 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
520 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
521 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
522} STMT_END
523
668c081a 524#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 525 const IV offset = loc - RExC_precomp; \
d1d15184 526 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
527 m REPORT_LOCATION, \
528 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
529} STMT_END
530
2335b3d3
KW
531#define ckWARN2regdep(loc,m, a1) STMT_START { \
532 const IV offset = loc - RExC_precomp; \
533 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
534 m REPORT_LOCATION, \
535 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
536} STMT_END
537
668c081a 538#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 539 const IV offset = loc - RExC_precomp; \
668c081a 540 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
541 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
542} STMT_END
543
544#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 545 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
546 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
547 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
548} STMT_END
549
668c081a
NC
550#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
551 const IV offset = loc - RExC_precomp; \
552 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
553 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
554} STMT_END
555
ccb2c380 556#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 557 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
558 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
559 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
560} STMT_END
561
668c081a
NC
562#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
563 const IV offset = loc - RExC_precomp; \
564 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
565 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
566} STMT_END
567
ccb2c380 568#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 569 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
570 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
571 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
572} STMT_END
9d1d55b5 573
8615cb43 574
cd439c50 575/* Allow for side effects in s */
ccb2c380
MP
576#define REGC(c,s) STMT_START { \
577 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
578} STMT_END
cd439c50 579
fac92740
MJD
580/* Macros for recording node offsets. 20001227 mjd@plover.com
581 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
582 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
583 * Element 0 holds the number n.
07be1b83 584 * Position is 1 indexed.
fac92740 585 */
7122b237
YO
586#ifndef RE_TRACK_PATTERN_OFFSETS
587#define Set_Node_Offset_To_R(node,byte)
588#define Set_Node_Offset(node,byte)
589#define Set_Cur_Node_Offset
590#define Set_Node_Length_To_R(node,len)
591#define Set_Node_Length(node,len)
592#define Set_Node_Cur_Length(node)
593#define Node_Offset(n)
594#define Node_Length(n)
595#define Set_Node_Offset_Length(node,offset,len)
596#define ProgLen(ri) ri->u.proglen
597#define SetProgLen(ri,x) ri->u.proglen = x
598#else
599#define ProgLen(ri) ri->u.offsets[0]
600#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
601#define Set_Node_Offset_To_R(node,byte) STMT_START { \
602 if (! SIZE_ONLY) { \
603 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 604 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 605 if((node) < 0) { \
551405c4 606 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
607 } else { \
608 RExC_offsets[2*(node)-1] = (byte); \
609 } \
610 } \
611} STMT_END
612
613#define Set_Node_Offset(node,byte) \
614 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
615#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
616
617#define Set_Node_Length_To_R(node,len) STMT_START { \
618 if (! SIZE_ONLY) { \
619 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 620 __LINE__, (int)(node), (int)(len))); \
ccb2c380 621 if((node) < 0) { \
551405c4 622 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
623 } else { \
624 RExC_offsets[2*(node)] = (len); \
625 } \
626 } \
627} STMT_END
628
629#define Set_Node_Length(node,len) \
630 Set_Node_Length_To_R((node)-RExC_emit_start, len)
631#define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
632#define Set_Node_Cur_Length(node) \
633 Set_Node_Length(node, RExC_parse - parse_start)
fac92740
MJD
634
635/* Get offsets and lengths */
636#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
637#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
638
07be1b83
YO
639#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
640 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
641 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
642} STMT_END
7122b237 643#endif
07be1b83
YO
644
645#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
646#define EXPERIMENTAL_INPLACESCAN
f427392e 647#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 648
304ee84b
YO
649#define DEBUG_STUDYDATA(str,data,depth) \
650DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 651 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
652 "%*s" str "Pos:%"IVdf"/%"IVdf \
653 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
654 (int)(depth)*2, "", \
655 (IV)((data)->pos_min), \
656 (IV)((data)->pos_delta), \
304ee84b 657 (UV)((data)->flags), \
1de06328 658 (IV)((data)->whilem_c), \
304ee84b
YO
659 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
660 is_inf ? "INF " : "" \
1de06328
YO
661 ); \
662 if ((data)->last_found) \
663 PerlIO_printf(Perl_debug_log, \
664 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
665 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
666 SvPVX_const((data)->last_found), \
667 (IV)((data)->last_end), \
668 (IV)((data)->last_start_min), \
669 (IV)((data)->last_start_max), \
670 ((data)->longest && \
671 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
672 SvPVX_const((data)->longest_fixed), \
673 (IV)((data)->offset_fixed), \
674 ((data)->longest && \
675 (data)->longest==&((data)->longest_float)) ? "*" : "", \
676 SvPVX_const((data)->longest_float), \
677 (IV)((data)->offset_float_min), \
678 (IV)((data)->offset_float_max) \
679 ); \
680 PerlIO_printf(Perl_debug_log,"\n"); \
681});
682
acfe0abc 683static void clear_re(pTHX_ void *r);
4327152a 684
653099ff 685/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 686 Update the longest found anchored substring and the longest found
653099ff
GS
687 floating substrings if needed. */
688
4327152a 689STATIC void
304ee84b 690S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 691{
e1ec3a88
AL
692 const STRLEN l = CHR_SVLEN(data->last_found);
693 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 694 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 695
7918f24d
NC
696 PERL_ARGS_ASSERT_SCAN_COMMIT;
697
c277df42 698 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 699 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
700 if (*data->longest == data->longest_fixed) {
701 data->offset_fixed = l ? data->last_start_min : data->pos_min;
702 if (data->flags & SF_BEFORE_EOL)
b81d288d 703 data->flags
c277df42
IZ
704 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
705 else
706 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 707 data->minlen_fixed=minlenp;
1de06328 708 data->lookbehind_fixed=0;
a0ed51b3 709 }
304ee84b 710 else { /* *data->longest == data->longest_float */
c277df42 711 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
712 data->offset_float_max = (l
713 ? data->last_start_max
c277df42 714 : data->pos_min + data->pos_delta);
304ee84b 715 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 716 data->offset_float_max = I32_MAX;
c277df42 717 if (data->flags & SF_BEFORE_EOL)
b81d288d 718 data->flags
c277df42
IZ
719 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
720 else
721 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
722 data->minlen_float=minlenp;
723 data->lookbehind_float=0;
c277df42
IZ
724 }
725 }
726 SvCUR_set(data->last_found, 0);
0eda9292 727 {
a28509cc 728 SV * const sv = data->last_found;
097eb12c
AL
729 if (SvUTF8(sv) && SvMAGICAL(sv)) {
730 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
731 if (mg)
732 mg->mg_len = 0;
733 }
0eda9292 734 }
c277df42
IZ
735 data->last_end = -1;
736 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 737 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
738}
739
653099ff
GS
740/* Can match anything (initialization) */
741STATIC void
3fffb88a 742S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 743{
7918f24d
NC
744 PERL_ARGS_ASSERT_CL_ANYTHING;
745
f8bef550 746 ANYOF_BITMAP_SETALL(cl);
dd58aee1 747 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
3ad98780 748 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
3fffb88a
KW
749
750 /* If any portion of the regex is to operate under locale rules,
751 * initialization includes it. The reason this isn't done for all regexes
752 * is that the optimizer was written under the assumption that locale was
753 * all-or-nothing. Given the complexity and lack of documentation in the
754 * optimizer, and that there are inadequate test cases for locale, so many
755 * parts of it may not work properly, it is safest to avoid locale unless
756 * necessary. */
757 if (RExC_contains_locale) {
9d7a1e63 758 ANYOF_CLASS_SETALL(cl); /* /l uses class */
3fffb88a
KW
759 cl->flags |= ANYOF_LOCALE;
760 }
9d7a1e63
KW
761 else {
762 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
763 }
653099ff
GS
764}
765
766/* Can match anything (initialization) */
767STATIC int
5f66b61c 768S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
769{
770 int value;
771
7918f24d
NC
772 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
773
aaa51d5e 774 for (value = 0; value <= ANYOF_MAX; value += 2)
653099ff
GS
775 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
776 return 1;
1aa99e6b
IH
777 if (!(cl->flags & ANYOF_UNICODE_ALL))
778 return 0;
10edeb5d 779 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 780 return 0;
653099ff
GS
781 return 1;
782}
783
784/* Can match anything (initialization) */
785STATIC void
e755fd73 786S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 787{
7918f24d
NC
788 PERL_ARGS_ASSERT_CL_INIT;
789
8ecf7187 790 Zero(cl, 1, struct regnode_charclass_class);
653099ff 791 cl->type = ANYOF;
3fffb88a 792 cl_anything(pRExC_state, cl);
1411dba4 793 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
794}
795
1051e1c4
KW
796/* These two functions currently do the exact same thing */
797#define cl_init_zero S_cl_init
653099ff 798
dd58aee1
KW
799/* 'AND' a given class with another one. Can create false positives. 'cl'
800 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
801 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 802STATIC void
5f66b61c 803S_cl_and(struct regnode_charclass_class *cl,
a28509cc 804 const struct regnode_charclass_class *and_with)
653099ff 805{
7918f24d 806 PERL_ARGS_ASSERT_CL_AND;
40d049e4
YO
807
808 assert(and_with->type == ANYOF);
1e6ade67 809
c6b76537 810 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
811 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
812 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 813 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
814 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
815 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
653099ff
GS
816 int i;
817
818 if (and_with->flags & ANYOF_INVERT)
819 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
820 cl->bitmap[i] &= ~and_with->bitmap[i];
821 else
822 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
823 cl->bitmap[i] &= and_with->bitmap[i];
824 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 825
c6b76537 826 if (and_with->flags & ANYOF_INVERT) {
8951c461 827
c6b76537
KW
828 /* Here, the and'ed node is inverted. Get the AND of the flags that
829 * aren't affected by the inversion. Those that are affected are
830 * handled individually below */
831 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
832 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
833 cl->flags |= affected_flags;
834
835 /* We currently don't know how to deal with things that aren't in the
836 * bitmap, but we know that the intersection is no greater than what
837 * is already in cl, so let there be false positives that get sorted
838 * out after the synthetic start class succeeds, and the node is
839 * matched for real. */
840
841 /* The inversion of these two flags indicate that the resulting
842 * intersection doesn't have them */
843 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
844 cl->flags &= ~ANYOF_UNICODE_ALL;
845 }
c6b76537
KW
846 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
847 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 848 }
1aa99e6b 849 }
c6b76537 850 else { /* and'd node is not inverted */
3ad98780
KW
851 U8 outside_bitmap_but_not_utf8; /* Temp variable */
852
137165a6 853 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
854
855 /* Here 'and_with' doesn't match anything outside the bitmap
856 * (except possibly ANYOF_UNICODE_ALL), which means the
857 * intersection can't either, except for ANYOF_UNICODE_ALL, in
858 * which case we don't know what the intersection is, but it's no
859 * greater than what cl already has, so can just leave it alone,
860 * with possible false positives */
861 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
862 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 863 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 864 }
137165a6 865 }
c6b76537
KW
866 else if (! ANYOF_NONBITMAP(cl)) {
867
868 /* Here, 'and_with' does match something outside the bitmap, and cl
869 * doesn't have a list of things to match outside the bitmap. If
870 * cl can match all code points above 255, the intersection will
3ad98780
KW
871 * be those above-255 code points that 'and_with' matches. If cl
872 * can't match all Unicode code points, it means that it can't
873 * match anything outside the bitmap (since the 'if' that got us
874 * into this block tested for that), so we leave the bitmap empty.
875 */
c6b76537
KW
876 if (cl->flags & ANYOF_UNICODE_ALL) {
877 ARG_SET(cl, ARG(and_with));
3ad98780
KW
878
879 /* and_with's ARG may match things that don't require UTF8.
880 * And now cl's will too, in spite of this being an 'and'. See
881 * the comments below about the kludge */
882 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
883 }
884 }
885 else {
886 /* Here, both 'and_with' and cl match something outside the
887 * bitmap. Currently we do not do the intersection, so just match
888 * whatever cl had at the beginning. */
889 }
890
891
3ad98780
KW
892 /* Take the intersection of the two sets of flags. However, the
893 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
894 * kludge around the fact that this flag is not treated like the others
895 * which are initialized in cl_anything(). The way the optimizer works
896 * is that the synthetic start class (SSC) is initialized to match
897 * anything, and then the first time a real node is encountered, its
898 * values are AND'd with the SSC's with the result being the values of
899 * the real node. However, there are paths through the optimizer where
900 * the AND never gets called, so those initialized bits are set
901 * inappropriately, which is not usually a big deal, as they just cause
902 * false positives in the SSC, which will just mean a probably
903 * imperceptible slow down in execution. However this bit has a
904 * higher false positive consequence in that it can cause utf8.pm,
905 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
906 * bigger slowdown and also causes significant extra memory to be used.
907 * In order to prevent this, the code now takes a different tack. The
908 * bit isn't set unless some part of the regular expression needs it,
909 * but once set it won't get cleared. This means that these extra
910 * modules won't get loaded unless there was some path through the
911 * pattern that would have required them anyway, and so any false
912 * positives that occur by not ANDing them out when they could be
913 * aren't as severe as they would be if we treated this bit like all
914 * the others */
915 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
916 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 917 cl->flags &= and_with->flags;
3ad98780 918 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 919 }
653099ff
GS
920}
921
dd58aee1
KW
922/* 'OR' a given class with another one. Can create false positives. 'cl'
923 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
924 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 925STATIC void
3fffb88a 926S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 927{
7918f24d
NC
928 PERL_ARGS_ASSERT_CL_OR;
929
653099ff 930 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
931
932 /* Here, the or'd node is to be inverted. This means we take the
933 * complement of everything not in the bitmap, but currently we don't
934 * know what that is, so give up and match anything */
935 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 936 cl_anything(pRExC_state, cl);
c6b76537 937 }
653099ff
GS
938 /* We do not use
939 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
940 * <= (B1 | !B2) | (CL1 | !CL2)
941 * which is wasteful if CL2 is small, but we ignore CL2:
942 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
943 * XXXX Can we handle case-fold? Unclear:
944 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
945 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
946 */
c6b76537 947 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
948 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
949 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
653099ff
GS
950 int i;
951
952 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
953 cl->bitmap[i] |= ~or_with->bitmap[i];
954 } /* XXXX: logic is complicated otherwise */
955 else {
3fffb88a 956 cl_anything(pRExC_state, cl);
653099ff 957 }
c6b76537
KW
958
959 /* And, we can just take the union of the flags that aren't affected
960 * by the inversion */
961 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
962
963 /* For the remaining flags:
964 ANYOF_UNICODE_ALL and inverted means to not match anything above
965 255, which means that the union with cl should just be
966 what cl has in it, so can ignore this flag
967 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
968 is 127-255 to match them, but then invert that, so the
969 union with cl should just be what cl has in it, so can
970 ignore this flag
971 */
972 } else { /* 'or_with' is not inverted */
653099ff
GS
973 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
974 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
975 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
976 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
653099ff
GS
977 int i;
978
979 /* OR char bitmap and class bitmap separately */
980 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
981 cl->bitmap[i] |= or_with->bitmap[i];
1e6ade67 982 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
653099ff
GS
983 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
984 cl->classflags[i] |= or_with->classflags[i];
985 cl->flags |= ANYOF_CLASS;
986 }
987 }
988 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 989 cl_anything(pRExC_state, cl);
653099ff 990 }
9826f543 991
c6b76537
KW
992 if (ANYOF_NONBITMAP(or_with)) {
993
994 /* Use the added node's outside-the-bit-map match if there isn't a
995 * conflict. If there is a conflict (both nodes match something
996 * outside the bitmap, but what they match outside is not the same
997 * pointer, and hence not easily compared until XXX we extend
998 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
999 * match everything outside the bitmap. If that stuff is all above
1000 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1001 if (! ANYOF_NONBITMAP(cl)) {
1002 ARG_SET(cl, ARG(or_with));
1003 }
1004 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1005
1006 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1007 cl_anything(pRExC_state, cl);
1008 }
1009 else {
1010 cl->flags |= ANYOF_UNICODE_ALL;
1011 }
c6b76537 1012 }
4c34a693 1013 }
0b9668ee
KW
1014
1015 /* Take the union */
1016 cl->flags |= or_with->flags;
1aa99e6b 1017 }
653099ff
GS
1018}
1019
a3621e74
YO
1020#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1021#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1022#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1023#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1024
3dab1dad
YO
1025
1026#ifdef DEBUGGING
07be1b83 1027/*
2b8b4781
NC
1028 dump_trie(trie,widecharmap,revcharmap)
1029 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1030 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1031
1032 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1033 The _interim_ variants are used for debugging the interim
1034 tables that are used to generate the final compressed
1035 representation which is what dump_trie expects.
1036
486ec47a 1037 Part of the reason for their existence is to provide a form
3dab1dad 1038 of documentation as to how the different representations function.
07be1b83
YO
1039
1040*/
3dab1dad
YO
1041
1042/*
3dab1dad
YO
1043 Dumps the final compressed table form of the trie to Perl_debug_log.
1044 Used for debugging make_trie().
1045*/
b9a59e08 1046
3dab1dad 1047STATIC void
2b8b4781
NC
1048S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1049 AV *revcharmap, U32 depth)
3dab1dad
YO
1050{
1051 U32 state;
ab3bbdeb 1052 SV *sv=sv_newmortal();
55eed653 1053 int colwidth= widecharmap ? 6 : 4;
2e64971a 1054 U16 word;
3dab1dad
YO
1055 GET_RE_DEBUG_FLAGS_DECL;
1056
7918f24d 1057 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1058
3dab1dad
YO
1059 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1060 (int)depth * 2 + 2,"",
1061 "Match","Base","Ofs" );
1062
1063 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1064 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1065 if ( tmp ) {
ab3bbdeb
YO
1066 PerlIO_printf( Perl_debug_log, "%*s",
1067 colwidth,
ddc5bc0f 1068 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1069 PL_colors[0], PL_colors[1],
1070 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1071 PERL_PV_ESCAPE_FIRSTCHAR
1072 )
1073 );
3dab1dad
YO
1074 }
1075 }
1076 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1077 (int)depth * 2 + 2,"");
1078
1079 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1080 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1081 PerlIO_printf( Perl_debug_log, "\n");
1082
1e2e3d02 1083 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1084 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1085
1086 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1087
1088 if ( trie->states[ state ].wordnum ) {
1089 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1090 } else {
1091 PerlIO_printf( Perl_debug_log, "%6s", "" );
1092 }
1093
1094 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1095
1096 if ( base ) {
1097 U32 ofs = 0;
1098
1099 while( ( base + ofs < trie->uniquecharcount ) ||
1100 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1101 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1102 ofs++;
1103
1104 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1105
1106 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1107 if ( ( base + ofs >= trie->uniquecharcount ) &&
1108 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1109 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1110 {
ab3bbdeb
YO
1111 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1112 colwidth,
3dab1dad
YO
1113 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1114 } else {
ab3bbdeb 1115 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1116 }
1117 }
1118
1119 PerlIO_printf( Perl_debug_log, "]");
1120
1121 }
1122 PerlIO_printf( Perl_debug_log, "\n" );
1123 }
2e64971a
DM
1124 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1125 for (word=1; word <= trie->wordcount; word++) {
1126 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1127 (int)word, (int)(trie->wordinfo[word].prev),
1128 (int)(trie->wordinfo[word].len));
1129 }
1130 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1131}
1132/*
3dab1dad
YO
1133 Dumps a fully constructed but uncompressed trie in list form.
1134 List tries normally only are used for construction when the number of
1135 possible chars (trie->uniquecharcount) is very high.
1136 Used for debugging make_trie().
1137*/
1138STATIC void
55eed653 1139S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1140 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1141 U32 depth)
3dab1dad
YO
1142{
1143 U32 state;
ab3bbdeb 1144 SV *sv=sv_newmortal();
55eed653 1145 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1146 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1147
1148 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1149
3dab1dad 1150 /* print out the table precompression. */
ab3bbdeb
YO
1151 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1152 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1153 "------:-----+-----------------\n" );
3dab1dad
YO
1154
1155 for( state=1 ; state < next_alloc ; state ++ ) {
1156 U16 charid;
1157
ab3bbdeb 1158 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1159 (int)depth * 2 + 2,"", (UV)state );
1160 if ( ! trie->states[ state ].wordnum ) {
1161 PerlIO_printf( Perl_debug_log, "%5s| ","");
1162 } else {
1163 PerlIO_printf( Perl_debug_log, "W%4x| ",
1164 trie->states[ state ].wordnum
1165 );
1166 }
1167 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1168 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1169 if ( tmp ) {
1170 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1171 colwidth,
ddc5bc0f 1172 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1173 PL_colors[0], PL_colors[1],
1174 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1175 PERL_PV_ESCAPE_FIRSTCHAR
1176 ) ,
1e2e3d02
YO
1177 TRIE_LIST_ITEM(state,charid).forid,
1178 (UV)TRIE_LIST_ITEM(state,charid).newstate
1179 );
1180 if (!(charid % 10))
664e119d
RGS
1181 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1182 (int)((depth * 2) + 14), "");
1e2e3d02 1183 }
ab3bbdeb
YO
1184 }
1185 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1186 }
1187}
1188
1189/*
3dab1dad
YO
1190 Dumps a fully constructed but uncompressed trie in table form.
1191 This is the normal DFA style state transition table, with a few
1192 twists to facilitate compression later.
1193 Used for debugging make_trie().
1194*/
1195STATIC void
55eed653 1196S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1197 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1198 U32 depth)
3dab1dad
YO
1199{
1200 U32 state;
1201 U16 charid;
ab3bbdeb 1202 SV *sv=sv_newmortal();
55eed653 1203 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1204 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1205
1206 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1207
1208 /*
1209 print out the table precompression so that we can do a visual check
1210 that they are identical.
1211 */
1212
1213 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1214
1215 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1216 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1217 if ( tmp ) {
ab3bbdeb
YO
1218 PerlIO_printf( Perl_debug_log, "%*s",
1219 colwidth,
ddc5bc0f 1220 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1221 PL_colors[0], PL_colors[1],
1222 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1223 PERL_PV_ESCAPE_FIRSTCHAR
1224 )
1225 );
3dab1dad
YO
1226 }
1227 }
1228
1229 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1230
1231 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1232 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1233 }
1234
1235 PerlIO_printf( Perl_debug_log, "\n" );
1236
1237 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1238
1239 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1240 (int)depth * 2 + 2,"",
1241 (UV)TRIE_NODENUM( state ) );
1242
1243 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1244 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1245 if (v)
1246 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1247 else
1248 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1249 }
1250 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1251 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1252 } else {
1253 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1254 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1255 }
1256 }
07be1b83 1257}
3dab1dad
YO
1258
1259#endif
1260
2e64971a 1261
786e8c11
YO
1262/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1263 startbranch: the first branch in the whole branch sequence
1264 first : start branch of sequence of branch-exact nodes.
1265 May be the same as startbranch
1266 last : Thing following the last branch.
1267 May be the same as tail.
1268 tail : item following the branch sequence
1269 count : words in the sequence
1270 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1271 depth : indent depth
3dab1dad 1272
786e8c11 1273Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1274
786e8c11
YO
1275A trie is an N'ary tree where the branches are determined by digital
1276decomposition of the key. IE, at the root node you look up the 1st character and
1277follow that branch repeat until you find the end of the branches. Nodes can be
1278marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1279
786e8c11 1280 /he|she|his|hers/
72f13be8 1281
786e8c11
YO
1282would convert into the following structure. Numbers represent states, letters
1283following numbers represent valid transitions on the letter from that state, if
1284the number is in square brackets it represents an accepting state, otherwise it
1285will be in parenthesis.
07be1b83 1286
786e8c11
YO
1287 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1288 | |
1289 | (2)
1290 | |
1291 (1) +-i->(6)-+-s->[7]
1292 |
1293 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1294
786e8c11
YO
1295 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1296
1297This shows that when matching against the string 'hers' we will begin at state 1
1298read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1299then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1300is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1301single traverse. We store a mapping from accepting to state to which word was
1302matched, and then when we have multiple possibilities we try to complete the
1303rest of the regex in the order in which they occured in the alternation.
1304
1305The only prior NFA like behaviour that would be changed by the TRIE support is
1306the silent ignoring of duplicate alternations which are of the form:
1307
1308 / (DUPE|DUPE) X? (?{ ... }) Y /x
1309
4b714af6 1310Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1311and without the optimisation. With the optimisations dupes will be silently
486ec47a 1312ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1313the following demonstrates:
1314
1315 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1316
1317which prints out 'word' three times, but
1318
1319 'words'=~/(word|word|word)(?{ print $1 })S/
1320
1321which doesnt print it out at all. This is due to other optimisations kicking in.
1322
1323Example of what happens on a structural level:
1324
486ec47a 1325The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1326
1327 1: CURLYM[1] {1,32767}(18)
1328 5: BRANCH(8)
1329 6: EXACT <ac>(16)
1330 8: BRANCH(11)
1331 9: EXACT <ad>(16)
1332 11: BRANCH(14)
1333 12: EXACT <ab>(16)
1334 16: SUCCEED(0)
1335 17: NOTHING(18)
1336 18: END(0)
1337
1338This would be optimizable with startbranch=5, first=5, last=16, tail=16
1339and should turn into:
1340
1341 1: CURLYM[1] {1,32767}(18)
1342 5: TRIE(16)
1343 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1344 <ac>
1345 <ad>
1346 <ab>
1347 16: SUCCEED(0)
1348 17: NOTHING(18)
1349 18: END(0)
1350
1351Cases where tail != last would be like /(?foo|bar)baz/:
1352
1353 1: BRANCH(4)
1354 2: EXACT <foo>(8)
1355 4: BRANCH(7)
1356 5: EXACT <bar>(8)
1357 7: TAIL(8)
1358 8: EXACT <baz>(10)
1359 10: END(0)
1360
1361which would be optimizable with startbranch=1, first=1, last=7, tail=8
1362and would end up looking like:
1363
1364 1: TRIE(8)
1365 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1366 <foo>
1367 <bar>
1368 7: TAIL(8)
1369 8: EXACT <baz>(10)
1370 10: END(0)
1371
1372 d = uvuni_to_utf8_flags(d, uv, 0);
1373
1374is the recommended Unicode-aware way of saying
1375
1376 *(d++) = uv;
1377*/
1378
fab2782b 1379#define TRIE_STORE_REVCHAR(val) \
786e8c11 1380 STMT_START { \
73031816 1381 if (UTF) { \
fab2782b 1382 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1383 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
fab2782b 1384 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
73031816
NC
1385 SvCUR_set(zlopp, kapow - flrbbbbb); \
1386 SvPOK_on(zlopp); \
1387 SvUTF8_on(zlopp); \
1388 av_push(revcharmap, zlopp); \
1389 } else { \
fab2782b 1390 char ooooff = (char)val; \
73031816
NC
1391 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1392 } \
1393 } STMT_END
786e8c11 1394
fab2782b
YO
1395#define TRIE_READ_CHAR STMT_START { \
1396 wordlen++; \
1397 if ( UTF ) { \
1398 /* if it is UTF then it is either already folded, or does not need folding */ \
1399 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1400 } \
1401 else if (folder == PL_fold_latin1) { \
1402 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1403 if ( foldlen > 0 ) { \
1404 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1405 foldlen -= len; \
1406 scan += len; \
1407 len = 0; \
1408 } else { \
1409 len = 1; \
1410 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1411 skiplen = UNISKIP(uvc); \
1412 foldlen -= skiplen; \
1413 scan = foldbuf + skiplen; \
1414 } \
1415 } else { \
1416 /* raw data, will be folded later if needed */ \
1417 uvc = (U32)*uc; \
1418 len = 1; \
1419 } \
786e8c11
YO
1420} STMT_END
1421
1422
1423
1424#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1425 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1426 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1427 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1428 } \
1429 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1430 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1431 TRIE_LIST_CUR( state )++; \
1432} STMT_END
07be1b83 1433
786e8c11
YO
1434#define TRIE_LIST_NEW(state) STMT_START { \
1435 Newxz( trie->states[ state ].trans.list, \
1436 4, reg_trie_trans_le ); \
1437 TRIE_LIST_CUR( state ) = 1; \
1438 TRIE_LIST_LEN( state ) = 4; \
1439} STMT_END
07be1b83 1440
786e8c11
YO
1441#define TRIE_HANDLE_WORD(state) STMT_START { \
1442 U16 dupe= trie->states[ state ].wordnum; \
1443 regnode * const noper_next = regnext( noper ); \
1444 \
786e8c11
YO
1445 DEBUG_r({ \
1446 /* store the word for dumping */ \
1447 SV* tmp; \
1448 if (OP(noper) != NOTHING) \
740cce10 1449 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1450 else \
740cce10 1451 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1452 av_push( trie_words, tmp ); \
786e8c11
YO
1453 }); \
1454 \
1455 curword++; \
2e64971a
DM
1456 trie->wordinfo[curword].prev = 0; \
1457 trie->wordinfo[curword].len = wordlen; \
1458 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1459 \
1460 if ( noper_next < tail ) { \
1461 if (!trie->jump) \
c944940b 1462 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1463 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1464 if (!jumper) \
1465 jumper = noper_next; \
1466 if (!nextbranch) \
1467 nextbranch= regnext(cur); \
1468 } \
1469 \
1470 if ( dupe ) { \
2e64971a
DM
1471 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1472 /* chain, so that when the bits of chain are later */\
1473 /* linked together, the dups appear in the chain */\
1474 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1475 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1476 } else { \
1477 /* we haven't inserted this word yet. */ \
1478 trie->states[ state ].wordnum = curword; \
1479 } \
1480} STMT_END
07be1b83 1481
3dab1dad 1482
786e8c11
YO
1483#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1484 ( ( base + charid >= ucharcount \
1485 && base + charid < ubound \
1486 && state == trie->trans[ base - ucharcount + charid ].check \
1487 && trie->trans[ base - ucharcount + charid ].next ) \
1488 ? trie->trans[ base - ucharcount + charid ].next \
1489 : ( state==1 ? special : 0 ) \
1490 )
3dab1dad 1491
786e8c11
YO
1492#define MADE_TRIE 1
1493#define MADE_JUMP_TRIE 2
1494#define MADE_EXACT_TRIE 4
3dab1dad 1495
a3621e74 1496STATIC I32
786e8c11 1497S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1498{
27da23d5 1499 dVAR;
a3621e74
YO
1500 /* first pass, loop through and scan words */
1501 reg_trie_data *trie;
55eed653 1502 HV *widecharmap = NULL;
2b8b4781 1503 AV *revcharmap = newAV();
a3621e74 1504 regnode *cur;
9f7f3913 1505 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1506 STRLEN len = 0;
1507 UV uvc = 0;
1508 U16 curword = 0;
1509 U32 next_alloc = 0;
786e8c11
YO
1510 regnode *jumper = NULL;
1511 regnode *nextbranch = NULL;
7f69552c 1512 regnode *convert = NULL;
2e64971a 1513 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1514 /* we just use folder as a flag in utf8 */
1e696034 1515 const U8 * folder = NULL;
a3621e74 1516
2b8b4781
NC
1517#ifdef DEBUGGING
1518 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1519 AV *trie_words = NULL;
1520 /* along with revcharmap, this only used during construction but both are
1521 * useful during debugging so we store them in the struct when debugging.
8e11feef 1522 */
2b8b4781
NC
1523#else
1524 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1525 STRLEN trie_charcount=0;
3dab1dad 1526#endif
2b8b4781 1527 SV *re_trie_maxbuff;
a3621e74 1528 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1529
1530 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1531#ifndef DEBUGGING
1532 PERL_UNUSED_ARG(depth);
1533#endif
a3621e74 1534
1e696034 1535 switch (flags) {
c46d03cf 1536 case EXACT: break;
2f7f8cb1 1537 case EXACTFA:
fab2782b
YO
1538 case EXACTFU_SS:
1539 case EXACTFU_TRICKYFOLD:
1e696034
KW
1540 case EXACTFU: folder = PL_fold_latin1; break;
1541 case EXACTF: folder = PL_fold; break;
1542 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1543 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1544 }
1545
c944940b 1546 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1547 trie->refcount = 1;
3dab1dad 1548 trie->startstate = 1;
786e8c11 1549 trie->wordcount = word_count;
f8fc2ecf 1550 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1551 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1552 if (flags == EXACT)
c944940b 1553 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1554 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1555 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1556
a3621e74 1557 DEBUG_r({
2b8b4781 1558 trie_words = newAV();
a3621e74 1559 });
a3621e74 1560
0111c4fd 1561 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1562 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1563 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1564 }
df826430 1565 DEBUG_TRIE_COMPILE_r({
3dab1dad 1566 PerlIO_printf( Perl_debug_log,
786e8c11 1567 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1568 (int)depth * 2 + 2, "",
1569 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1570 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1571 (int)depth);
3dab1dad 1572 });
7f69552c
YO
1573
1574 /* Find the node we are going to overwrite */
1575 if ( first == startbranch && OP( last ) != BRANCH ) {
1576 /* whole branch chain */
1577 convert = first;
1578 } else {
1579 /* branch sub-chain */
1580 convert = NEXTOPER( first );
1581 }
1582
a3621e74
YO
1583 /* -- First loop and Setup --
1584
1585 We first traverse the branches and scan each word to determine if it
1586 contains widechars, and how many unique chars there are, this is
1587 important as we have to build a table with at least as many columns as we
1588 have unique chars.
1589
1590 We use an array of integers to represent the character codes 0..255
38a44b82 1591 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1592 native representation of the character value as the key and IV's for the
1593 coded index.
1594
1595 *TODO* If we keep track of how many times each character is used we can
1596 remap the columns so that the table compression later on is more
3b753521 1597 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1598 middle and the least common are on the outside. IMO this would be better
1599 than a most to least common mapping as theres a decent chance the most
1600 common letter will share a node with the least common, meaning the node
486ec47a 1601 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1602 case is when we have the least common nodes twice.
1603
1604 */
1605
a3621e74 1606 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1607 regnode *noper = NEXTOPER( cur );
e1ec3a88 1608 const U8 *uc = (U8*)STRING( noper );
df826430 1609 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1610 STRLEN foldlen = 0;
1611 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1612 STRLEN skiplen = 0;
2af232bd 1613 const U8 *scan = (U8*)NULL;
07be1b83 1614 U32 wordlen = 0; /* required init */
02daf0ab
YO
1615 STRLEN chars = 0;
1616 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1617
3dab1dad 1618 if (OP(noper) == NOTHING) {
df826430
YO
1619 regnode *noper_next= regnext(noper);
1620 if (noper_next != tail && OP(noper_next) == flags) {
1621 noper = noper_next;
1622 uc= (U8*)STRING(noper);
1623 e= uc + STR_LEN(noper);
1624 trie->minlen= STR_LEN(noper);
1625 } else {
1626 trie->minlen= 0;
1627 continue;
1628 }
3dab1dad 1629 }
df826430 1630
fab2782b 1631 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1632 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1633 regardless of encoding */
fab2782b
YO
1634 if (OP( noper ) == EXACTFU_SS) {
1635 /* false positives are ok, so just set this */
1636 TRIE_BITMAP_SET(trie,0xDF);
1637 }
1638 }
a3621e74 1639 for ( ; uc < e ; uc += len ) {
3dab1dad 1640 TRIE_CHARCOUNT(trie)++;
a3621e74 1641 TRIE_READ_CHAR;
3dab1dad 1642 chars++;
a3621e74 1643 if ( uvc < 256 ) {
fab2782b
YO
1644 if ( folder ) {
1645 U8 folded= folder[ (U8) uvc ];
1646 if ( !trie->charmap[ folded ] ) {
1647 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1648 TRIE_STORE_REVCHAR( folded );
1649 }
1650 }
a3621e74
YO
1651 if ( !trie->charmap[ uvc ] ) {
1652 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1653 TRIE_STORE_REVCHAR( uvc );
a3621e74 1654 }
02daf0ab 1655 if ( set_bit ) {
62012aee
KW
1656 /* store the codepoint in the bitmap, and its folded
1657 * equivalent. */
fab2782b 1658 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1659
1660 /* store the folded codepoint */
fab2782b 1661 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1662
1663 if ( !UTF ) {
1664 /* store first byte of utf8 representation of
acdf4139
KW
1665 variant codepoints */
1666 if (! UNI_IS_INVARIANT(uvc)) {
1667 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1668 }
1669 }
02daf0ab
YO
1670 set_bit = 0; /* We've done our bit :-) */
1671 }
a3621e74
YO
1672 } else {
1673 SV** svpp;
55eed653
NC
1674 if ( !widecharmap )
1675 widecharmap = newHV();
a3621e74 1676
55eed653 1677 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1678
1679 if ( !svpp )
e4584336 1680 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1681
1682 if ( !SvTRUE( *svpp ) ) {
1683 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1684 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1685 }
1686 }
1687 }
3dab1dad 1688 if( cur == first ) {
fab2782b
YO
1689 trie->minlen = chars;
1690 trie->maxlen = chars;
3dab1dad 1691 } else if (chars < trie->minlen) {
fab2782b 1692 trie->minlen = chars;
3dab1dad 1693 } else if (chars > trie->maxlen) {
fab2782b
YO
1694 trie->maxlen = chars;
1695 }
1696 if (OP( noper ) == EXACTFU_SS) {
1697 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1698 if (trie->minlen > 1)
1699 trie->minlen= 1;
1700 }
1701 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1702 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1703 * - We assume that any such sequence might match a 2 byte string */
1704 if (trie->minlen > 2 )
1705 trie->minlen= 2;
3dab1dad
YO
1706 }
1707
a3621e74
YO
1708 } /* end first pass */
1709 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1710 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1711 (int)depth * 2 + 2,"",
55eed653 1712 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1713 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1714 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1715 );
a3621e74
YO
1716
1717 /*
1718 We now know what we are dealing with in terms of unique chars and
1719 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1720 representation using a flat table will take. If it's over a reasonable
1721 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1722 conservative but potentially much slower representation using an array
1723 of lists.
1724
1725 At the end we convert both representations into the same compressed
1726 form that will be used in regexec.c for matching with. The latter
1727 is a form that cannot be used to construct with but has memory
1728 properties similar to the list form and access properties similar
1729 to the table form making it both suitable for fast searches and
1730 small enough that its feasable to store for the duration of a program.
1731
1732 See the comment in the code where the compressed table is produced
1733 inplace from the flat tabe representation for an explanation of how
1734 the compression works.
1735
1736 */
1737
1738
2e64971a
DM
1739 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1740 prev_states[1] = 0;
1741
3dab1dad 1742 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1743 /*
1744 Second Pass -- Array Of Lists Representation
1745
1746 Each state will be represented by a list of charid:state records
1747 (reg_trie_trans_le) the first such element holds the CUR and LEN
1748 points of the allocated array. (See defines above).
1749
1750 We build the initial structure using the lists, and then convert
1751 it into the compressed table form which allows faster lookups
1752 (but cant be modified once converted).
a3621e74
YO
1753 */
1754
a3621e74
YO
1755 STRLEN transcount = 1;
1756
1e2e3d02
YO
1757 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1758 "%*sCompiling trie using list compiler\n",
1759 (int)depth * 2 + 2, ""));
686b73d4 1760
c944940b
JH
1761 trie->states = (reg_trie_state *)
1762 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1763 sizeof(reg_trie_state) );
a3621e74
YO
1764 TRIE_LIST_NEW(1);
1765 next_alloc = 2;
1766
1767 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1768
df826430 1769 regnode *noper = NEXTOPER( cur );
c445ea15 1770 U8 *uc = (U8*)STRING( noper );
df826430 1771 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1772 U32 state = 1; /* required init */
1773 U16 charid = 0; /* sanity init */
1774 U8 *scan = (U8*)NULL; /* sanity init */
1775 STRLEN foldlen = 0; /* required init */
07be1b83 1776 U32 wordlen = 0; /* required init */
c445ea15 1777 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1778 STRLEN skiplen = 0;
c445ea15 1779
df826430
YO
1780 if (OP(noper) == NOTHING) {
1781 regnode *noper_next= regnext(noper);
1782 if (noper_next != tail && OP(noper_next) == flags) {
1783 noper = noper_next;
1784 uc= (U8*)STRING(noper);
1785 e= uc + STR_LEN(noper);
1786 }
1787 }
1788
3dab1dad 1789 if (OP(noper) != NOTHING) {
786e8c11 1790 for ( ; uc < e ; uc += len ) {
c445ea15 1791
786e8c11 1792 TRIE_READ_CHAR;
c445ea15 1793
786e8c11
YO
1794 if ( uvc < 256 ) {
1795 charid = trie->charmap[ uvc ];
c445ea15 1796 } else {
55eed653 1797 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1798 if ( !svpp ) {
1799 charid = 0;
1800 } else {
1801 charid=(U16)SvIV( *svpp );
1802 }
c445ea15 1803 }
786e8c11
YO
1804 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1805 if ( charid ) {
a3621e74 1806
786e8c11
YO
1807 U16 check;
1808 U32 newstate = 0;
a3621e74 1809
786e8c11
YO
1810 charid--;
1811 if ( !trie->states[ state ].trans.list ) {
1812 TRIE_LIST_NEW( state );
c445ea15 1813 }
786e8c11
YO
1814 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1815 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1816 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1817 break;
1818 }
1819 }
1820 if ( ! newstate ) {
1821 newstate = next_alloc++;
2e64971a 1822 prev_states[newstate] = state;
786e8c11
YO
1823 TRIE_LIST_PUSH( state, charid, newstate );
1824 transcount++;
1825 }
1826 state = newstate;
1827 } else {
1828 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1829 }
a28509cc 1830 }
c445ea15 1831 }
3dab1dad 1832 TRIE_HANDLE_WORD(state);
a3621e74
YO
1833
1834 } /* end second pass */
1835
1e2e3d02
YO
1836 /* next alloc is the NEXT state to be allocated */
1837 trie->statecount = next_alloc;
c944940b
JH
1838 trie->states = (reg_trie_state *)
1839 PerlMemShared_realloc( trie->states,
1840 next_alloc
1841 * sizeof(reg_trie_state) );
a3621e74 1842
3dab1dad 1843 /* and now dump it out before we compress it */
2b8b4781
NC
1844 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1845 revcharmap, next_alloc,
1846 depth+1)
1e2e3d02 1847 );
a3621e74 1848
c944940b
JH
1849 trie->trans = (reg_trie_trans *)
1850 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1851 {
1852 U32 state;
a3621e74
YO
1853 U32 tp = 0;
1854 U32 zp = 0;
1855
1856
1857 for( state=1 ; state < next_alloc ; state ++ ) {
1858 U32 base=0;
1859
1860 /*
1861 DEBUG_TRIE_COMPILE_MORE_r(
1862 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1863 );
1864 */
1865
1866 if (trie->states[state].trans.list) {
1867 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1868 U16 maxid=minid;
a28509cc 1869 U16 idx;
a3621e74
YO
1870
1871 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1872 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1873 if ( forid < minid ) {
1874 minid=forid;
1875 } else if ( forid > maxid ) {
1876 maxid=forid;
1877 }
a3621e74
YO
1878 }
1879 if ( transcount < tp + maxid - minid + 1) {
1880 transcount *= 2;
c944940b
JH
1881 trie->trans = (reg_trie_trans *)
1882 PerlMemShared_realloc( trie->trans,
446bd890
NC
1883 transcount
1884 * sizeof(reg_trie_trans) );
a3621e74
YO
1885 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1886 }
1887 base = trie->uniquecharcount + tp - minid;
1888 if ( maxid == minid ) {
1889 U32 set = 0;
1890 for ( ; zp < tp ; zp++ ) {
1891 if ( ! trie->trans[ zp ].next ) {
1892 base = trie->uniquecharcount + zp - minid;
1893 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1894 trie->trans[ zp ].check = state;
1895 set = 1;
1896 break;
1897 }
1898 }
1899 if ( !set ) {
1900 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1901 trie->trans[ tp ].check = state;
1902 tp++;
1903 zp = tp;
1904 }
1905 } else {
1906 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1907 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1908 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1909 trie->trans[ tid ].check = state;
1910 }
1911 tp += ( maxid - minid + 1 );
1912 }
1913 Safefree(trie->states[ state ].trans.list);
1914 }
1915 /*
1916 DEBUG_TRIE_COMPILE_MORE_r(
1917 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1918 );
1919 */
1920 trie->states[ state ].trans.base=base;
1921 }
cc601c31 1922 trie->lasttrans = tp + 1;
a3621e74
YO
1923 }
1924 } else {
1925 /*
1926 Second Pass -- Flat Table Representation.
1927
1928 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1929 We know that we will need Charcount+1 trans at most to store the data
1930 (one row per char at worst case) So we preallocate both structures
1931 assuming worst case.
1932
1933 We then construct the trie using only the .next slots of the entry
1934 structs.
1935
3b753521 1936 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1937 make compression both faster and easier by keeping track of how many non
1938 zero fields are in the node.
1939
1940 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1941 transition.
1942
1943 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1944 number representing the first entry of the node, and state as a
1945 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1946 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1947 are 2 entrys per node. eg:
1948
1949 A B A B
1950 1. 2 4 1. 3 7
1951 2. 0 3 3. 0 5
1952 3. 0 0 5. 0 0
1953 4. 0 0 7. 0 0
1954
1955 The table is internally in the right hand, idx form. However as we also
1956 have to deal with the states array which is indexed by nodenum we have to
1957 use TRIE_NODENUM() to convert.
1958
1959 */
1e2e3d02
YO
1960 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1961 "%*sCompiling trie using table compiler\n",
1962 (int)depth * 2 + 2, ""));
3dab1dad 1963
c944940b
JH
1964 trie->trans = (reg_trie_trans *)
1965 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1966 * trie->uniquecharcount + 1,
1967 sizeof(reg_trie_trans) );
1968 trie->states = (reg_trie_state *)
1969 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1970 sizeof(reg_trie_state) );
a3621e74
YO
1971 next_alloc = trie->uniquecharcount + 1;
1972
3dab1dad 1973
a3621e74
YO
1974 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1975
df826430 1976 regnode *noper = NEXTOPER( cur );
a28509cc 1977 const U8 *uc = (U8*)STRING( noper );
df826430 1978 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1979
1980 U32 state = 1; /* required init */
1981
1982 U16 charid = 0; /* sanity init */
1983 U32 accept_state = 0; /* sanity init */
1984 U8 *scan = (U8*)NULL; /* sanity init */
1985
1986 STRLEN foldlen = 0; /* required init */
07be1b83 1987 U32 wordlen = 0; /* required init */
fab2782b 1988 STRLEN skiplen = 0;
a3621e74
YO
1989 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1990
df826430
YO
1991 if (OP(noper) == NOTHING) {
1992 regnode *noper_next= regnext(noper);
1993 if (noper_next != tail && OP(noper_next) == flags) {
1994 noper = noper_next;
1995 uc= (U8*)STRING(noper);
1996 e= uc + STR_LEN(noper);
1997 }
1998 }
fab2782b 1999
3dab1dad 2000 if ( OP(noper) != NOTHING ) {
786e8c11 2001 for ( ; uc < e ; uc += len ) {
a3621e74 2002
786e8c11 2003 TRIE_READ_CHAR;
a3621e74 2004
786e8c11
YO
2005 if ( uvc < 256 ) {
2006 charid = trie->charmap[ uvc ];
2007 } else {
55eed653 2008 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2009 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2010 }
786e8c11
YO
2011 if ( charid ) {
2012 charid--;
2013 if ( !trie->trans[ state + charid ].next ) {
2014 trie->trans[ state + charid ].next = next_alloc;
2015 trie->trans[ state ].check++;
2e64971a
DM
2016 prev_states[TRIE_NODENUM(next_alloc)]
2017 = TRIE_NODENUM(state);
786e8c11
YO
2018 next_alloc += trie->uniquecharcount;
2019 }
2020 state = trie->trans[ state + charid ].next;
2021 } else {
2022 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2023 }
2024 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2025 }
a3621e74 2026 }
3dab1dad
YO
2027 accept_state = TRIE_NODENUM( state );
2028 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2029
2030 } /* end second pass */
2031
3dab1dad 2032 /* and now dump it out before we compress it */
2b8b4781
NC
2033 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2034 revcharmap,
2035 next_alloc, depth+1));
a3621e74 2036
a3621e74
YO
2037 {
2038 /*
2039 * Inplace compress the table.*
2040
2041 For sparse data sets the table constructed by the trie algorithm will
2042 be mostly 0/FAIL transitions or to put it another way mostly empty.
2043 (Note that leaf nodes will not contain any transitions.)
2044
2045 This algorithm compresses the tables by eliminating most such
2046 transitions, at the cost of a modest bit of extra work during lookup:
2047
2048 - Each states[] entry contains a .base field which indicates the
2049 index in the state[] array wheres its transition data is stored.
2050
3b753521 2051 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2052
2053 - If .base is nonzero then charid is added to it to find an entry in
2054 the trans array.
2055
2056 -If trans[states[state].base+charid].check!=state then the
2057 transition is taken to be a 0/Fail transition. Thus if there are fail
2058 transitions at the front of the node then the .base offset will point
2059 somewhere inside the previous nodes data (or maybe even into a node
2060 even earlier), but the .check field determines if the transition is
2061 valid.
2062
786e8c11 2063 XXX - wrong maybe?
a3621e74 2064 The following process inplace converts the table to the compressed
3b753521 2065 table: We first do not compress the root node 1,and mark all its
a3621e74 2066 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2067 allows us to do a DFA construction from the compressed table later,
2068 and ensures that any .base pointers we calculate later are greater
2069 than 0.
a3621e74
YO
2070
2071 - We set 'pos' to indicate the first entry of the second node.
2072
2073 - We then iterate over the columns of the node, finding the first and
2074 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2075 and set the .check pointers accordingly, and advance pos
2076 appropriately and repreat for the next node. Note that when we copy
2077 the next pointers we have to convert them from the original
2078 NODEIDX form to NODENUM form as the former is not valid post
2079 compression.
2080
2081 - If a node has no transitions used we mark its base as 0 and do not
2082 advance the pos pointer.
2083
2084 - If a node only has one transition we use a second pointer into the
2085 structure to fill in allocated fail transitions from other states.
2086 This pointer is independent of the main pointer and scans forward
2087 looking for null transitions that are allocated to a state. When it
2088 finds one it writes the single transition into the "hole". If the
786e8c11 2089 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2090
2091 - Once compressed we can Renew/realloc the structures to release the
2092 excess space.
2093
2094 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2095 specifically Fig 3.47 and the associated pseudocode.
2096
2097 demq
2098 */
a3b680e6 2099 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2100 U32 state, charid;
a3621e74 2101 U32 pos = 0, zp=0;
1e2e3d02 2102 trie->statecount = laststate;
a3621e74
YO
2103
2104 for ( state = 1 ; state < laststate ; state++ ) {
2105 U8 flag = 0;
a28509cc
AL
2106 const U32 stateidx = TRIE_NODEIDX( state );
2107 const U32 o_used = trie->trans[ stateidx ].check;
2108 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2109 trie->trans[ stateidx ].check = 0;
2110
2111 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2112 if ( flag || trie->trans[ stateidx + charid ].next ) {
2113 if ( trie->trans[ stateidx + charid ].next ) {
2114 if (o_used == 1) {
2115 for ( ; zp < pos ; zp++ ) {
2116 if ( ! trie->trans[ zp ].next ) {
2117 break;
2118 }
2119 }
2120 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2121 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2122 trie->trans[ zp ].check = state;
2123 if ( ++zp > pos ) pos = zp;
2124 break;
2125 }
2126 used--;
2127 }
2128 if ( !flag ) {
2129 flag = 1;
2130 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2131 }
2132 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2133 trie->trans[ pos ].check = state;
2134 pos++;
2135 }
2136 }
2137 }
cc601c31 2138 trie->lasttrans = pos + 1;
c944940b
JH
2139 trie->states = (reg_trie_state *)
2140 PerlMemShared_realloc( trie->states, laststate
2141 * sizeof(reg_trie_state) );
a3621e74 2142 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2143 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2144 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2145 (int)depth * 2 + 2,"",
2146 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2147 (IV)next_alloc,
2148 (IV)pos,
a3621e74
YO
2149 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2150 );
2151
2152 } /* end table compress */
2153 }
1e2e3d02
YO
2154 DEBUG_TRIE_COMPILE_MORE_r(
2155 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2156 (int)depth * 2 + 2, "",
2157 (UV)trie->statecount,
2158 (UV)trie->lasttrans)
2159 );
cc601c31 2160 /* resize the trans array to remove unused space */
c944940b
JH
2161 trie->trans = (reg_trie_trans *)
2162 PerlMemShared_realloc( trie->trans, trie->lasttrans
2163 * sizeof(reg_trie_trans) );
a3621e74 2164
3b753521 2165 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2166 U8 nodetype =(U8)(flags & 0xFF);
2167 char *str=NULL;
786e8c11 2168
07be1b83 2169#ifdef DEBUGGING
e62cc96a 2170 regnode *optimize = NULL;
7122b237
YO
2171#ifdef RE_TRACK_PATTERN_OFFSETS
2172
b57a0404
JH
2173 U32 mjd_offset = 0;
2174 U32 mjd_nodelen = 0;
7122b237
YO
2175#endif /* RE_TRACK_PATTERN_OFFSETS */
2176#endif /* DEBUGGING */
a3621e74 2177 /*
3dab1dad
YO
2178 This means we convert either the first branch or the first Exact,
2179 depending on whether the thing following (in 'last') is a branch
2180 or not and whther first is the startbranch (ie is it a sub part of
2181 the alternation or is it the whole thing.)
3b753521 2182 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2183 the whole branch sequence, including the first.
a3621e74 2184 */
3dab1dad 2185 /* Find the node we are going to overwrite */
7f69552c 2186 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2187 /* branch sub-chain */
3dab1dad 2188 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2189#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2190 DEBUG_r({
2191 mjd_offset= Node_Offset((convert));
2192 mjd_nodelen= Node_Length((convert));
2193 });
7122b237 2194#endif
7f69552c 2195 /* whole branch chain */
7122b237
YO
2196 }
2197#ifdef RE_TRACK_PATTERN_OFFSETS
2198 else {
7f69552c
YO
2199 DEBUG_r({
2200 const regnode *nop = NEXTOPER( convert );
2201 mjd_offset= Node_Offset((nop));
2202 mjd_nodelen= Node_Length((nop));
2203 });
07be1b83
YO
2204 }
2205 DEBUG_OPTIMISE_r(
2206 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2207 (int)depth * 2 + 2, "",
786e8c11 2208 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2209 );
7122b237 2210#endif
3dab1dad
YO
2211 /* But first we check to see if there is a common prefix we can
2212 split out as an EXACT and put in front of the TRIE node. */
2213 trie->startstate= 1;
55eed653 2214 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2215 U32 state;
1e2e3d02 2216 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2217 U32 ofs = 0;
8e11feef
RGS
2218 I32 idx = -1;
2219 U32 count = 0;
2220 const U32 base = trie->states[ state ].trans.base;
a3621e74 2221
3dab1dad 2222 if ( trie->states[state].wordnum )
8e11feef 2223 count = 1;
a3621e74 2224
8e11feef 2225 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2226 if ( ( base + ofs >= trie->uniquecharcount ) &&
2227 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2228 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2229 {
3dab1dad 2230 if ( ++count > 1 ) {
2b8b4781 2231 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2232 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2233 if ( state == 1 ) break;
3dab1dad
YO
2234 if ( count == 2 ) {
2235 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2236 DEBUG_OPTIMISE_r(
8e11feef
RGS
2237 PerlIO_printf(Perl_debug_log,
2238 "%*sNew Start State=%"UVuf" Class: [",
2239 (int)depth * 2 + 2, "",
786e8c11 2240 (UV)state));
be8e71aa 2241 if (idx >= 0) {
2b8b4781 2242 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2243 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2244
3dab1dad 2245 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2246 if ( folder )
2247 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2248 DEBUG_OPTIMISE_r(
f1f66076 2249 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2250 );
8e11feef
RGS
2251 }
2252 }
2253 TRIE_BITMAP_SET(trie,*ch);
2254 if ( folder )
2255 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2256 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2257 }
2258 idx = ofs;
2259 }
3dab1dad
YO
2260 }
2261 if ( count == 1 ) {
2b8b4781 2262 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2263 STRLEN len;
2264 char *ch = SvPV( *tmp, len );
de734bd5
A
2265 DEBUG_OPTIMISE_r({
2266 SV *sv=sv_newmortal();
8e11feef
RGS
2267 PerlIO_printf( Perl_debug_log,
2268 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2269 (int)depth * 2 + 2, "",
de734bd5
A
2270 (UV)state, (UV)idx,
2271 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2272 PL_colors[0], PL_colors[1],
2273 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2274 PERL_PV_ESCAPE_FIRSTCHAR
2275 )
2276 );
2277 });
3dab1dad
YO
2278 if ( state==1 ) {
2279 OP( convert ) = nodetype;
2280 str=STRING(convert);
2281 STR_LEN(convert)=0;
2282 }
c490c714
YO
2283 STR_LEN(convert) += len;
2284 while (len--)
de734bd5 2285 *str++ = *ch++;
8e11feef 2286 } else {
f9049ba1 2287#ifdef DEBUGGING
8e11feef
RGS
2288 if (state>1)
2289 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2290#endif
8e11feef
RGS
2291 break;
2292 }
2293 }
2e64971a 2294 trie->prefixlen = (state-1);
3dab1dad 2295 if (str) {
8e11feef 2296 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2297 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2298 trie->startstate = state;
07be1b83
YO
2299 trie->minlen -= (state - 1);
2300 trie->maxlen -= (state - 1);
33809eae
JH
2301#ifdef DEBUGGING
2302 /* At least the UNICOS C compiler choked on this
2303 * being argument to DEBUG_r(), so let's just have
2304 * it right here. */
2305 if (
2306#ifdef PERL_EXT_RE_BUILD
2307 1
2308#else
2309 DEBUG_r_TEST
2310#endif
2311 ) {
2312 regnode *fix = convert;
2313 U32 word = trie->wordcount;
2314 mjd_nodelen++;
2315 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2316 while( ++fix < n ) {
2317 Set_Node_Offset_Length(fix, 0, 0);
2318 }
2319 while (word--) {
2320 SV ** const tmp = av_fetch( trie_words, word, 0 );
2321 if (tmp) {
2322 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2323 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2324 else
2325 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2326 }
2327 }
2328 }
2329#endif
8e11feef
RGS
2330 if (trie->maxlen) {
2331 convert = n;
2332 } else {
3dab1dad 2333 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2334 DEBUG_r(optimize= n);
3dab1dad
YO
2335 }
2336 }
2337 }
a5ca303d
YO
2338 if (!jumper)
2339 jumper = last;
3dab1dad 2340 if ( trie->maxlen ) {
8e11feef
RGS
2341 NEXT_OFF( convert ) = (U16)(tail - convert);
2342 ARG_SET( convert, data_slot );
786e8c11
YO
2343 /* Store the offset to the first unabsorbed branch in
2344 jump[0], which is otherwise unused by the jump logic.
2345 We use this when dumping a trie and during optimisation. */
2346 if (trie->jump)
7f69552c 2347 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2348
6c48061a
YO
2349 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2350 * and there is a bitmap
2351 * and the first "jump target" node we found leaves enough room
2352 * then convert the TRIE node into a TRIEC node, with the bitmap
2353 * embedded inline in the opcode - this is hypothetically faster.
2354 */
2355 if ( !trie->states[trie->startstate].wordnum
2356 && trie->bitmap
2357 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2358 {
2359 OP( convert ) = TRIEC;
2360 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2361 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2362 trie->bitmap= NULL;
2363 } else
2364 OP( convert ) = TRIE;
a3621e74 2365
3dab1dad
YO
2366 /* store the type in the flags */
2367 convert->flags = nodetype;
a5ca303d
YO
2368 DEBUG_r({
2369 optimize = convert
2370 + NODE_STEP_REGNODE
2371 + regarglen[ OP( convert ) ];
2372 });
2373 /* XXX We really should free up the resource in trie now,
2374 as we won't use them - (which resources?) dmq */
3dab1dad 2375 }
a3621e74 2376 /* needed for dumping*/
e62cc96a 2377 DEBUG_r(if (optimize) {
07be1b83 2378 regnode *opt = convert;
bcdf7404 2379
e62cc96a 2380 while ( ++opt < optimize) {
07be1b83
YO
2381 Set_Node_Offset_Length(opt,0,0);
2382 }
786e8c11
YO
2383 /*
2384 Try to clean up some of the debris left after the
2385 optimisation.
a3621e74 2386 */
786e8c11 2387 while( optimize < jumper ) {
07be1b83 2388 mjd_nodelen += Node_Length((optimize));
a3621e74 2389 OP( optimize ) = OPTIMIZED;
07be1b83 2390 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2391 optimize++;
2392 }
07be1b83 2393 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2394 });
2395 } /* end node insert */
2e64971a
DM
2396
2397 /* Finish populating the prev field of the wordinfo array. Walk back
2398 * from each accept state until we find another accept state, and if
2399 * so, point the first word's .prev field at the second word. If the
2400 * second already has a .prev field set, stop now. This will be the
2401 * case either if we've already processed that word's accept state,
3b753521
FN
2402 * or that state had multiple words, and the overspill words were
2403 * already linked up earlier.
2e64971a
DM
2404 */
2405 {
2406 U16 word;
2407 U32 state;
2408 U16 prev;
2409
2410 for (word=1; word <= trie->wordcount; word++) {
2411 prev = 0;
2412 if (trie->wordinfo[word].prev)
2413 continue;
2414 state = trie->wordinfo[word].accept;
2415 while (state) {
2416 state = prev_states[state];
2417 if (!state)
2418 break;
2419 prev = trie->states[state].wordnum;
2420 if (prev)
2421 break;
2422 }
2423 trie->wordinfo[word].prev = prev;
2424 }
2425 Safefree(prev_states);
2426 }
2427
2428
2429 /* and now dump out the compressed format */
2430 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2431
55eed653 2432 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2433#ifdef DEBUGGING
2434 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2435 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2436#else
2437 SvREFCNT_dec(revcharmap);
07be1b83 2438#endif
786e8c11
YO
2439 return trie->jump
2440 ? MADE_JUMP_TRIE
2441 : trie->startstate>1
2442 ? MADE_EXACT_TRIE
2443 : MADE_TRIE;
2444}
2445
2446STATIC void
2447S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2448{
3b753521 2449/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2450
2451 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2452 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2453 ISBN 0-201-10088-6
2454
2455 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2456 suffix of the current state's 'word' that is also a proper prefix of another word in our
2457 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2458 the DFA not to have to restart after its tried and failed a word at a given point, it
2459 simply continues as though it had been matching the other word in the first place.
2460 Consider
2461 'abcdgu'=~/abcdefg|cdgu/
2462 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2463 fail, which would bring us to the state representing 'd' in the second word where we would
2464 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2465 */
2466 /* add a fail transition */
3251b653
NC
2467 const U32 trie_offset = ARG(source);
2468 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2469 U32 *q;
2470 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2471 const U32 numstates = trie->statecount;
786e8c11
YO
2472 const U32 ubound = trie->lasttrans + ucharcount;
2473 U32 q_read = 0;
2474 U32 q_write = 0;
2475 U32 charid;
2476 U32 base = trie->states[ 1 ].trans.base;
2477 U32 *fail;
2478 reg_ac_data *aho;
2479 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2480 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2481
2482 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2483#ifndef DEBUGGING
2484 PERL_UNUSED_ARG(depth);
2485#endif
2486
2487
2488 ARG_SET( stclass, data_slot );
c944940b 2489 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2490 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2491 aho->trie=trie_offset;
446bd890
NC
2492 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2493 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2494 Newxz( q, numstates, U32);
c944940b 2495 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2496 aho->refcount = 1;
2497 fail = aho->fail;
2498 /* initialize fail[0..1] to be 1 so that we always have
2499 a valid final fail state */
2500 fail[ 0 ] = fail[ 1 ] = 1;
2501
2502 for ( charid = 0; charid < ucharcount ; charid++ ) {
2503 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2504 if ( newstate ) {
2505 q[ q_write ] = newstate;
2506 /* set to point at the root */
2507 fail[ q[ q_write++ ] ]=1;
2508 }
2509 }
2510 while ( q_read < q_write) {
2511 const U32 cur = q[ q_read++ % numstates ];
2512 base = trie->states[ cur ].trans.base;
2513
2514 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2515 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2516 if (ch_state) {
2517 U32 fail_state = cur;
2518 U32 fail_base;
2519 do {
2520 fail_state = fail[ fail_state ];
2521 fail_base = aho->states[ fail_state ].trans.base;
2522 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2523
2524 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2525 fail[ ch_state ] = fail_state;
2526 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2527 {
2528 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2529 }
2530 q[ q_write++ % numstates] = ch_state;
2531 }
2532 }
2533 }
2534 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2535 when we fail in state 1, this allows us to use the
2536 charclass scan to find a valid start char. This is based on the principle
2537 that theres a good chance the string being searched contains lots of stuff
2538 that cant be a start char.
2539 */
2540 fail[ 0 ] = fail[ 1 ] = 0;
2541 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2542 PerlIO_printf(Perl_debug_log,
2543 "%*sStclass Failtable (%"UVuf" states): 0",
2544 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2545 );
786e8c11
YO
2546 for( q_read=1; q_read<numstates; q_read++ ) {
2547 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2548 }
2549 PerlIO_printf(Perl_debug_log, "\n");
2550 });
2551 Safefree(q);
2552 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2553}
2554
786e8c11 2555
a3621e74 2556/*
5d1c421c
JH
2557 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2558 * These need to be revisited when a newer toolchain becomes available.
2559 */
2560#if defined(__sparc64__) && defined(__GNUC__)
2561# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2562# undef SPARC64_GCC_WORKAROUND
2563# define SPARC64_GCC_WORKAROUND 1
2564# endif
2565#endif
2566
07be1b83 2567#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2568 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2569 SV * const mysv=sv_newmortal(); \
2570 regnode *Next = regnext(scan); \
2571 regprop(RExC_rx, mysv, scan); \
7f69552c 2572 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2573 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2574 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2575 }});
07be1b83 2576
1de06328 2577
bb914485
KW
2578/* The below joins as many adjacent EXACTish nodes as possible into a single
2579 * one, and looks for problematic sequences of characters whose folds vs.
2580 * non-folds have sufficiently different lengths, that the optimizer would be
2581 * fooled into rejecting legitimate matches of them, and the trie construction
2582 * code can't cope with them. The joining is only done if:
2583 * 1) there is room in the current conglomerated node to entirely contain the
2584 * next one.
2585 * 2) they are the exact same node type
2586 *
2587 * The adjacent nodes actually may be separated by NOTHING kind nodes, and
2588 * these get optimized out
2589 *
9d071ca8
KW
2590 * If there are problematic code sequences, *min_subtract is set to the delta
2591 * that the minimum size of the node can be less than its actual size. And,
2592 * the node type of the result is changed to reflect that it contains these
bb914485
KW
2593 * sequences.
2594 *
a0c4c608
KW
2595 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2596 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2597 *
bb914485
KW
2598 * This is as good a place as any to discuss the design of handling these
2599 * problematic sequences. It's been wrong in Perl for a very long time. There
2600 * are three code points in Unicode whose folded lengths differ so much from
2601 * the un-folded lengths that it causes problems for the optimizer and trie
2602 * construction. Why only these are problematic, and not others where lengths
2603 * also differ is something I (khw) do not understand. New versions of Unicode
2604 * might add more such code points. Hopefully the logic in fold_grind.t that
287722f3 2605 * figures out what to test (in part by verifying that each size-combination
bb914485 2606 * gets tested) will catch any that do come along, so they can be added to the
287722f3
KW
2607 * special handling below. The chances of new ones are actually rather small,
2608 * as most, if not all, of the world's scripts that have casefolding have
2609 * already been encoded by Unicode. Also, a number of Unicode's decisions were
2610 * made to allow compatibility with pre-existing standards, and almost all of
2611 * those have already been dealt with. These would otherwise be the most
2612 * likely candidates for generating further tricky sequences. In other words,
2613 * Unicode by itself is unlikely to add new ones unless it is for compatibility
a0c4c608 2614 * with pre-existing standards, and there aren't many of those left.
bb914485
KW
2615 *
2616 * The previous designs for dealing with these involved assigning a special
2617 * node for them. This approach doesn't work, as evidenced by this example:
a0c4c608 2618 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
bb914485
KW
2619 * Both these fold to "sss", but if the pattern is parsed to create a node of
2620 * that would match just the \xDF, it won't be able to handle the case where a
2621 * successful match would have to cross the node's boundary. The new approach
2622 * that hopefully generally solves the problem generates an EXACTFU_SS node
2623 * that is "sss".
2624 *
2625 * There are a number of components to the approach (a lot of work for just
2626 * three code points!):
2627 * 1) This routine examines each EXACTFish node that could contain the
9d071ca8
KW
2628 * problematic sequences. It returns in *min_subtract how much to
2629 * subtract from the the actual length of the string to get a real minimum
2630 * for one that could match it. This number is usually 0 except for the
2631 * problematic sequences. This delta is used by the caller to adjust the
2632 * min length of the match, and the delta between min and max, so that the
2633 * optimizer doesn't reject these possibilities based on size constraints.
bb914485
KW
2634 * 2) These sequences are not currently correctly handled by the trie code
2635 * either, so it changes the joined node type to ops that are not handled
fab2782b 2636 * by trie's, those new ops being EXACTFU_SS and EXACTFU_TRICKYFOLD.
bb914485
KW
2637 * 3) This is sufficient for the two Greek sequences (described below), but
2638 * the one involving the Sharp s (\xDF) needs more. The node type
2639 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2640 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2641 * case where there is a possible fold length change. That means that a
2642 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2643 * itself with length changes, and so can be processed faster. regexec.c
2644 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2645 * is pre-folded by regcomp.c. This saves effort in regex matching.
2646 * However, probably mostly for historical reasons, the pre-folding isn't
a0c4c608
KW
2647 * done for non-UTF8 patterns (and it can't be for EXACTF and EXACTFL
2648 * nodes, as what they fold to isn't known until runtime.) The fold
2649 * possibilities for the non-UTF8 patterns are quite simple, except for
2650 * the sharp s. All the ones that don't involve a UTF-8 target string
2651 * are members of a fold-pair, and arrays are set up for all of them
2652 * that quickly find the other member of the pair. It might actually
2653 * be faster to pre-fold these, but it isn't currently done, except for
2654 * the sharp s. Code elsewhere in this file makes sure that it gets
2655 * folded to 'ss', even if the pattern isn't UTF-8. This avoids the
2656 * issues described in the next item.
bb914485
KW
2657 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2658 * 'ss' or not is not knowable at compile time. It will match iff the
2659 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2660 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2661 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2662 * described in item 3). An assumption that the optimizer part of
2663 * regexec.c (probably unwittingly) makes is that a character in the
2664 * pattern corresponds to at most a single character in the target string.
2665 * (And I do mean character, and not byte here, unlike other parts of the
2666 * documentation that have never been updated to account for multibyte
2667 * Unicode.) This assumption is wrong only in this case, as all other
2668 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2669 * virtue of having this file pre-fold UTF-8 patterns. I'm
2670 * reluctant to try to change this assumption, so instead the code punts.
9d071ca8
KW
2671 * This routine examines EXACTF nodes for the sharp s, and returns a
2672 * boolean indicating whether or not the node is an EXACTF node that
2673 * contains a sharp s. When it is true, the caller sets a flag that later
2674 * causes the optimizer in this file to not set values for the floating
2675 * and fixed string lengths, and thus avoids the optimizer code in
2676 * regexec.c that makes the invalid assumption. Thus, there is no
2677 * optimization based on string lengths for EXACTF nodes that contain the
2678 * sharp s. This only happens for /id rules (which means the pattern
2679 * isn't in UTF-8).
bb914485 2680 */
1de06328 2681
9d071ca8 2682#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2683 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2684 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2685
be8e71aa 2686STATIC U32
9d071ca8 2687S_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
2688 /* Merge several consecutive EXACTish nodes into one. */
2689 regnode *n = regnext(scan);
2690 U32 stringok = 1;
2691 regnode *next = scan + NODE_SZ_STR(scan);
2692 U32 merged = 0;
2693 U32 stopnow = 0;
2694#ifdef DEBUGGING
2695 regnode *stop = scan;
72f13be8 2696 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2697#else
d47053eb
RGS
2698 PERL_UNUSED_ARG(depth);
2699#endif
7918f24d
NC
2700
2701 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2702#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2703 PERL_UNUSED_ARG(flags);
2704 PERL_UNUSED_ARG(val);
07be1b83 2705#endif
07be1b83 2706 DEBUG_PEEP("join",scan,depth);
bb914485 2707
3f410cf6
KW
2708 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2709 * EXACT ones that are mergeable to the current one. */
2710 while (n
2711 && (PL_regkind[OP(n)] == NOTHING
2712 || (stringok && OP(n) == OP(scan)))
07be1b83 2713 && NEXT_OFF(n)
3f410cf6
KW
2714 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2715 {
07be1b83
YO
2716
2717 if (OP(n) == TAIL || n > next)
2718 stringok = 0;
2719 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2720 DEBUG_PEEP("skip:",n,depth);
2721 NEXT_OFF(scan) += NEXT_OFF(n);
2722 next = n + NODE_STEP_REGNODE;
2723#ifdef DEBUGGING
2724 if (stringok)
2725 stop = n;
2726#endif
2727 n = regnext(n);
2728 }
2729 else if (stringok) {
786e8c11 2730 const unsigned int oldl = STR_LEN(scan);
07be1b83 2731 regnode * const nnext = regnext(n);
b2230d39
KW
2732
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) {
f758bddf
KW
2774 U8 *s;
2775 U8 * s0 = (U8*) STRING(scan);
2776 U8 * const s_end = s0 + STR_LEN(scan);
2777
2778 /* The below is perhaps overboard, but this allows us to save a test
2779 * each time through the loop at the expense of a mask. This is
2780 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2781 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2782 * This uses an exclusive 'or' to find that bit and then inverts it to
2783 * form a mask, with just a single 0, in the bit position where 'S' and
2784 * 's' differ. */
dbeb8947 2785 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
f758bddf
KW
2786 const U8 s_masked = 's' & S_or_s_mask;
2787
2788 /* One pass is made over the node's string looking for all the
2789 * possibilities. to avoid some tests in the loop, there are two main
2790 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2791 * non-UTF-8 */
2792 if (UTF) {
86d6fcad 2793
f758bddf
KW
2794 /* There are two problematic Greek code points in Unicode
2795 * casefolding
86d6fcad
KW
2796 *
2797 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2798 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2799 *
2800 * which casefold to
2801 *
2802 * Unicode UTF-8
2803 *
2804 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2805 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2806 *
2807 * This means that in case-insensitive matching (or "loose
2808 * matching", as Unicode calls it), an EXACTF of length six (the
2809 * UTF-8 encoded byte length of the above casefolded versions) can
2810 * match a target string of length two (the byte length of UTF-8
2811 * encoded U+0390 or U+03B0). This would rather mess up the
2812 * minimum length computation. (there are other code points that
2813 * also fold to these two sequences, but the delta is smaller)
2814 *
f758bddf
KW
2815 * If these sequences are found, the minimum length is decreased by
2816 * four (six minus two).
86d6fcad 2817 *
f758bddf
KW
2818 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2819 * LETTER SHARP S. We decrease the min length by 1 for each
2820 * occurrence of 'ss' found */
3f410cf6 2821
e294cc5d 2822#ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
f758bddf
KW
2823# define U390_first_byte 0xb4
2824 const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
2825# define U3B0_first_byte 0xb5
2826 const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
e294cc5d 2827#else
f758bddf
KW
2828# define U390_first_byte 0xce
2829 const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
2830# define U3B0_first_byte 0xcf
2831 const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
e294cc5d 2832#endif
f758bddf
KW
2833 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2834 yields a net of 0 */
2835 /* Examine the string for one of the problematic sequences */
2836 for (s = s0;
2837 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2838 * sequence we are looking for is 2 */
2839 s += UTF8SKIP(s))
86d6fcad 2840 {
bb914485 2841
f758bddf
KW
2842 /* Look for the first byte in each problematic sequence */
2843 switch (*s) {
2844 /* We don't have to worry about other things that fold to
2845 * 's' (such as the long s, U+017F), as all above-latin1
2846 * code points have been pre-folded */
2847 case 's':
2848 case 'S':
2849
a0c4c608
KW
2850 /* Current character is an 's' or 'S'. If next one is
2851 * as well, we have the dreaded sequence */
f758bddf
KW
2852 if (((*(s+1) & S_or_s_mask) == s_masked)
2853 /* These two node types don't have special handling
2854 * for 'ss' */
2855 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2856 {
9d071ca8 2857 *min_subtract += 1;
f758bddf
KW
2858 OP(scan) = EXACTFU_SS;
2859 s++; /* No need to look at this character again */
2860 }
2861 break;
2862
2863 case U390_first_byte:
2864 if (s_end - s >= len
2865
2866 /* The 1's are because are skipping comparing the
2867 * first byte */
2868 && memEQ(s + 1, U390_tail, len - 1))
2869 {
2870 goto greek_sequence;
2871 }
2872 break;
2873
2874 case U3B0_first_byte:
2875 if (! (s_end - s >= len
2876 && memEQ(s + 1, U3B0_tail, len - 1)))
2877 {
2878 break;
2879 }
2880 greek_sequence:
9d071ca8 2881 *min_subtract += 4;
f758bddf
KW
2882
2883 /* This can't currently be handled by trie's, so change
2884 * the node type to indicate this. If EXACTFA and
2885 * EXACTFL were ever to be handled by trie's, this
2886 * would have to be changed. If this node has already
2887 * been changed to EXACTFU_SS in this loop, leave it as
2888 * is. (I (khw) think it doesn't matter in regexec.c
2889 * for UTF patterns, but no need to change it */
2890 if (OP(scan) == EXACTFU) {
fab2782b 2891 OP(scan) = EXACTFU_TRICKYFOLD;
f758bddf
KW
2892 }
2893 s += 6; /* We already know what this sequence is. Skip
2894 the rest of it */
2895 break;
bb914485
KW
2896 }
2897 }
2898 }
f758bddf 2899 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
bb914485 2900
f758bddf
KW
2901 /* Here, the pattern is not UTF-8. We need to look only for the
2902 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2903 * in the final position. Otherwise we can stop looking 1 byte
2904 * earlier because have to find both the first and second 's' */
2905 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2906
2907 for (s = s0; s < upper; s++) {
2908 switch (*s) {
2909 case 'S':
2910 case 's':
2911 if (s_end - s > 1
2912 && ((*(s+1) & S_or_s_mask) == s_masked))
2913 {
9d071ca8 2914 *min_subtract += 1;
f758bddf
KW
2915
2916 /* EXACTF nodes need to know that the minimum
2917 * length changed so that a sharp s in the string
2918 * can match this ss in the pattern, but they
2919 * remain EXACTF nodes, as they are not trie'able,
2920 * so don't have to invent a new node type to
2921 * exclude them from the trie code */
2922 if (OP(scan) != EXACTF) {
2923 OP(scan) = EXACTFU_SS;
2924 }
2925 s++;
2926 }
2927 break;
2928 case LATIN_SMALL_LETTER_SHARP_S:
2929 if (OP(scan) == EXACTF) {
2930 *has_exactf_sharp_s = TRUE;
2931 }
2932 break;
86d6fcad
KW
2933 }
2934 }
2935 }
07be1b83 2936 }
3f410cf6 2937
07be1b83 2938#ifdef DEBUGGING
bb789b09
DM
2939 /* Allow dumping but overwriting the collection of skipped
2940 * ops and/or strings with fake optimized ops */
07be1b83
YO
2941 n = scan + NODE_SZ_STR(scan);
2942 while (n <= stop) {
bb789b09
DM
2943 OP(n) = OPTIMIZED;
2944 FLAGS(n) = 0;
2945 NEXT_OFF(n) = 0;
07be1b83
YO
2946 n++;
2947 }
2948#endif
2949 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2950 return stopnow;
2951}
2952
486ec47a 2953/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2954 Finds fixed substrings. */
2955
a0288114 2956/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2957 to the position after last scanned or to NULL. */
2958
40d049e4
YO
2959#define INIT_AND_WITHP \
2960 assert(!and_withp); \
2961 Newx(and_withp,1,struct regnode_charclass_class); \
2962 SAVEFREEPV(and_withp)
07be1b83 2963
b515a41d 2964/* this is a chain of data about sub patterns we are processing that
486ec47a 2965 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2966 we can simulate recursion without losing state. */
2967struct scan_frame;
2968typedef struct scan_frame {
2969 regnode *last; /* last node to process in this frame */
2970 regnode *next; /* next node to process when last is reached */
2971 struct scan_frame *prev; /*previous frame*/
2972 I32 stop; /* what stopparen do we use */
2973} scan_frame;
2974
304ee84b
YO
2975
2976#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2977
e1d1eefb
YO
2978#define CASE_SYNST_FNC(nAmE) \
2979case nAmE: \
2980 if (flags & SCF_DO_STCLASS_AND) { \
2981 for (value = 0; value < 256; value++) \
2982 if (!is_ ## nAmE ## _cp(value)) \
2983 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2984 } \
2985 else { \
2986 for (value = 0; value < 256; value++) \
2987 if (is_ ## nAmE ## _cp(value)) \
2988 ANYOF_BITMAP_SET(data->start_class, value); \
2989 } \
2990 break; \
2991case N ## nAmE: \
2992 if (flags & SCF_DO_STCLASS_AND) { \
2993 for (value = 0; value < 256; value++) \
2994 if (is_ ## nAmE ## _cp(value)) \
2995 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2996 } \
2997 else { \
2998 for (value = 0; value < 256; value++) \
2999 if (!is_ ## nAmE ## _cp(value)) \
3000 ANYOF_BITMAP_SET(data->start_class, value); \
3001 } \
3002 break
3003
3004
3005
76e3520e 3006STATIC I32
40d049e4 3007S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 3008 I32 *minlenp, I32 *deltap,
40d049e4
YO
3009 regnode *last,
3010 scan_data_t *data,
3011 I32 stopparen,
3012 U8* recursed,
3013 struct regnode_charclass_class *and_withp,
3014 U32 flags, U32 depth)
c277df42
IZ
3015 /* scanp: Start here (read-write). */
3016 /* deltap: Write maxlen-minlen here. */
3017 /* last: Stop before this one. */
40d049e4
YO
3018 /* data: string data about the pattern */
3019 /* stopparen: treat close N as END */
3020 /* recursed: which subroutines have we recursed into */
3021 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 3022{
97aff369 3023 dVAR;
c277df42
IZ
3024 I32 min = 0, pars = 0, code;
3025 regnode *scan = *scanp, *next;
3026 I32 delta = 0;
3027 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 3028 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
3029 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3030 scan_data_t data_fake;
a3621e74 3031 SV *re_trie_maxbuff = NULL;
786e8c11 3032 regnode *first_non_open = scan;
e2e6a0f1 3033 I32 stopmin = I32_MAX;
8aa23a47 3034 scan_frame *frame = NULL;
a3621e74 3035 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 3036
7918f24d
NC
3037 PERL_ARGS_ASSERT_STUDY_CHUNK;
3038
13a24bad 3039#ifdef DEBUGGING
40d049e4 3040 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 3041#endif
40d049e4 3042
786e8c11 3043 if ( depth == 0 ) {
40d049e4 3044 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
3045 first_non_open=regnext(first_non_open);
3046 }
3047
b81d288d 3048
8aa23a47
YO
3049 fake_study_recurse:
3050 while ( scan && OP(scan) != END && scan < last ){
9d071ca8
KW
3051 UV min_subtract = 0; /* How much to subtract from the minimum node
3052 length to get a real minimum (because the
3053 folded version may be shorter) */
f758bddf 3054 bool has_exactf_sharp_s = FALSE;
8aa23a47 3055 /* Peephole optimizer: */
304ee84b 3056 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3057 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3058
3059 /* Its not clear to khw or hv why this is done here, and not in the
3060 * clauses that deal with EXACT nodes. khw's guess is that it's
3061 * because of a previous design */
9d071ca8 3062 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3063
3064 /* Follow the next-chain of the current node and optimize
3065 away all the NOTHINGs from it. */
3066 if (OP(scan) != CURLYX) {
3067 const int max = (reg_off_by_arg[OP(scan)]
3068 ? I32_MAX
3069 /* I32 may be smaller than U16 on CRAYs! */
3070 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3071 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3072 int noff;
3073 regnode *n = scan;
686b73d4 3074
8aa23a47
YO
3075 /* Skip NOTHING and LONGJMP. */
3076 while ((n = regnext(n))
3077 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3078 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3079 && off + noff < max)
3080 off += noff;
3081 if (reg_off_by_arg[OP(scan)])
3082 ARG(scan) = off;
3083 else
3084 NEXT_OFF(scan) = off;
3085 }
a3621e74 3086
c277df42 3087
8aa23a47
YO
3088
3089 /* The principal pseudo-switch. Cannot be a switch, since we
3090 look into several different things. */
3091 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3092 || OP(scan) == IFTHEN) {
3093 next = regnext(scan);
3094 code = OP(scan);
3095 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3096
8aa23a47
YO
3097 if (OP(next) == code || code == IFTHEN) {
3098 /* NOTE - There is similar code to this block below for handling
3099 TRIE nodes on a re-study. If you change stuff here check there
3100 too. */
3101 I32 max1 = 0, min1 = I32_MAX, num = 0;
3102 struct regnode_charclass_class accum;
3103 regnode * const startbranch=scan;
686b73d4 3104
8aa23a47 3105 if (flags & SCF_DO_SUBSTR)
304ee84b 3106 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3107 if (flags & SCF_DO_STCLASS)
e755fd73 3108 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3109
3110 while (OP(scan) == code) {
3111 I32 deltanext, minnext, f = 0, fake;
3112 struct regnode_charclass_class this_class;
3113
3114 num++;
3115 data_fake.flags = 0;
3116 if (data) {
3117 data_fake.whilem_c = data->whilem_c;
3118 data_fake.last_closep = data->last_closep;
3119 }
3120 else
3121 data_fake.last_closep = &fake;
58e23c8d
YO
3122
3123 data_fake.pos_delta = delta;
8aa23a47
YO
3124 next = regnext(scan);
3125 scan = NEXTOPER(scan);
3126 if (code != BRANCH)
c277df42 3127 scan = NEXTOPER(scan);
8aa23a47 3128 if (flags & SCF_DO_STCLASS) {
e755fd73 3129 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3130 data_fake.start_class = &this_class;
3131 f = SCF_DO_STCLASS_AND;
58e23c8d 3132 }
8aa23a47
YO
3133 if (flags & SCF_WHILEM_VISITED_POS)
3134 f |= SCF_WHILEM_VISITED_POS;
3135
3136 /* we suppose the run is continuous, last=next...*/
3137 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3138 next, &data_fake,
3139 stopparen, recursed, NULL, f,depth+1);
3140 if (min1 > minnext)
3141 min1 = minnext;
3142 if (max1 < minnext + deltanext)
3143 max1 = minnext + deltanext;
3144 if (deltanext == I32_MAX)
3145 is_inf = is_inf_internal = 1;
3146 scan = next;
3147 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3148 pars++;
3149 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3150 if ( stopmin > minnext)
3151 stopmin = min + min1;
3152 flags &= ~SCF_DO_SUBSTR;
3153 if (data)
3154 data->flags |= SCF_SEEN_ACCEPT;
3155 }
3156 if (data) {
3157 if (data_fake.flags & SF_HAS_EVAL)
3158 data->flags |= SF_HAS_EVAL;
3159 data->whilem_c = data_fake.whilem_c;
3dab1dad 3160 }
8aa23a47 3161 if (flags & SCF_DO_STCLASS)
3fffb88a 3162 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3163 }
3164 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3165 min1 = 0;
3166 if (flags & SCF_DO_SUBSTR) {
3167 data->pos_min += min1;
3168 data->pos_delta += max1 - min1;
3169 if (max1 != min1 || is_inf)
3170 data->longest = &(data->longest_float);
3171 }
3172 min += min1;
3173 delta += max1 - min1;
3174 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3175 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3176 if (min1) {
3177 cl_and(data->start_class, and_withp);
3178 flags &= ~SCF_DO_STCLASS;
653099ff 3179 }
8aa23a47
YO
3180 }
3181 else if (flags & SCF_DO_STCLASS_AND) {
3182 if (min1) {
3183 cl_and(data->start_class, &accum);
3184 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3185 }
8aa23a47
YO
3186 else {
3187 /* Switch to OR mode: cache the old value of
3188 * data->start_class */
3189 INIT_AND_WITHP;
3190 StructCopy(data->start_class, and_withp,
3191 struct regnode_charclass_class);
3192 flags &= ~SCF_DO_STCLASS_AND;
3193 StructCopy(&accum, data->start_class,
3194 struct regnode_charclass_class);
3195 flags |= SCF_DO_STCLASS_OR;
3196 data->start_class->flags |= ANYOF_EOS;
de0c8cb8 3197 }
8aa23a47 3198 }
a3621e74 3199
8aa23a47
YO
3200 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3201 /* demq.
a3621e74 3202
8aa23a47
YO
3203 Assuming this was/is a branch we are dealing with: 'scan' now
3204 points at the item that follows the branch sequence, whatever
3205 it is. We now start at the beginning of the sequence and look
3206 for subsequences of
a3621e74 3207
8aa23a47
YO
3208 BRANCH->EXACT=>x1
3209 BRANCH->EXACT=>x2
3210 tail
a3621e74 3211
8aa23a47 3212 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3213
486ec47a 3214 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3215 element into a trie and then add the subsequent branch exact
3216 strings to the trie.
a3621e74 3217
8aa23a47 3218 We have two cases
a3621e74 3219
3b753521 3220 1. patterns where the whole set of branches can be converted.
a3621e74 3221
8aa23a47 3222 2. patterns where only a subset can be converted.
a3621e74 3223
8aa23a47
YO
3224 In case 1 we can replace the whole set with a single regop
3225 for the trie. In case 2 we need to keep the start and end
3b753521 3226 branches so
a3621e74 3227
8aa23a47
YO
3228 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3229 becomes BRANCH TRIE; BRANCH X;
786e8c11 3230
8aa23a47
YO
3231 There is an additional case, that being where there is a
3232 common prefix, which gets split out into an EXACT like node
3233 preceding the TRIE node.
a3621e74 3234
8aa23a47
YO
3235 If x(1..n)==tail then we can do a simple trie, if not we make
3236 a "jump" trie, such that when we match the appropriate word
486ec47a 3237 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3238 a nested if into a case structure of sorts.
b515a41d 3239
8aa23a47 3240 */
686b73d4 3241
8aa23a47
YO
3242 int made=0;
3243 if (!re_trie_maxbuff) {
3244 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3245 if (!SvIOK(re_trie_maxbuff))
3246 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3247 }
3248 if ( SvIV(re_trie_maxbuff)>=0 ) {
3249 regnode *cur;
3250 regnode *first = (regnode *)NULL;
3251 regnode *last = (regnode *)NULL;
3252 regnode *tail = scan;
fab2782b 3253 U8 trietype = 0;
8aa23a47 3254 U32 count=0;
a3621e74
YO
3255
3256#ifdef DEBUGGING
8aa23a47 3257 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3258#endif
8aa23a47
YO
3259 /* var tail is used because there may be a TAIL
3260 regop in the way. Ie, the exacts will point to the
3261 thing following the TAIL, but the last branch will
3262 point at the TAIL. So we advance tail. If we
3263 have nested (?:) we may have to move through several
3264 tails.
3265 */
3266
3267 while ( OP( tail ) == TAIL ) {
3268 /* this is the TAIL generated by (?:) */
3269 tail = regnext( tail );
3270 }
a3621e74 3271
8aa23a47 3272
df826430 3273 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3274 regprop(RExC_rx, mysv, tail );
3275 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3276 (int)depth * 2 + 2, "",
3277 "Looking for TRIE'able sequences. Tail node is: ",
3278 SvPV_nolen_const( mysv )
3279 );
3280 });
3281
3282 /*
3283
fab2782b
YO
3284 Step through the branches
3285 cur represents each branch,
3286 noper is the first thing to be matched as part of that branch
3287 noper_next is the regnext() of that node.
3288
3289 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3290 via a "jump trie" but we also support building with NOJUMPTRIE,
3291 which restricts the trie logic to structures like /FOO|BAR/.
3292
3293 If noper is a trieable nodetype then the branch is a possible optimization
3294 target. If we are building under NOJUMPTRIE then we require that noper_next
3295 is the same as scan (our current position in the regex program).
3296
3297 Once we have two or more consecutive such branches we can create a
3298 trie of the EXACT's contents and stitch it in place into the program.
3299
3300 If the sequence represents all of the branches in the alternation we
3301 replace the entire thing with a single TRIE node.
3302
3303 Otherwise when it is a subsequence we need to stitch it in place and
3304 replace only the relevant branches. This means the first branch has
3305 to remain as it is used by the alternation logic, and its next pointer,
3306 and needs to be repointed at the item on the branch chain following
3307 the last branch we have optimized away.
3308
3309 This could be either a BRANCH, in which case the subsequence is internal,
3310 or it could be the item following the branch sequence in which case the
3311 subsequence is at the end (which does not necessarily mean the first node
3312 is the start of the alternation).
3313
3314 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3315
3316 optype | trietype
3317 ----------------+-----------
3318 NOTHING | NOTHING
3319 EXACT | EXACT
3320 EXACTFU | EXACTFU
3321 EXACTFU_SS | EXACTFU
3322 EXACTFU_TRICKYFOLD | EXACTFU
3323 EXACTFA | 0
3324
8aa23a47
YO
3325
3326 */
fab2782b
YO
3327#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3328 ( EXACT == (X) ) ? EXACT : \
3329 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3330 0 )
8aa23a47
YO
3331
3332 /* dont use tail as the end marker for this traverse */
3333 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3334 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3335 U8 noper_type = OP( noper );
3336 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3337#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3338 regnode * const noper_next = regnext( noper );
df826430
YO
3339 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3340 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3341#endif
3342
df826430 3343 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3344 regprop(RExC_rx, mysv, cur);
3345 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3346 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3347
3348 regprop(RExC_rx, mysv, noper);
3349 PerlIO_printf( Perl_debug_log, " -> %s",
3350 SvPV_nolen_const(mysv));
3351
3352 if ( noper_next ) {
3353 regprop(RExC_rx, mysv, noper_next );
3354 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3355 SvPV_nolen_const(mysv));
3356 }
df826430
YO
3357 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3358 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3359 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3360 );
8aa23a47 3361 });
fab2782b
YO
3362
3363 /* Is noper a trieable nodetype that can be merged with the
3364 * current trie (if there is one)? */
3365 if ( noper_trietype
3366 &&
3367 (
df826430
YO
3368 ( noper_trietype == NOTHING)
3369 || ( trietype == NOTHING )
a40630bf 3370 || ( trietype == noper_trietype )
fab2782b 3371 )
786e8c11 3372#ifdef NOJUMPTRIE
8aa23a47 3373 && noper_next == tail
786e8c11 3374#endif
8aa23a47
YO
3375 && count < U16_MAX)
3376 {
fab2782b
YO
3377 /* Handle mergable triable node
3378 * Either we are the first node in a new trieable sequence,
3379 * in which case we do some bookkeeping, otherwise we update
3380 * the end pointer. */
fab2782b 3381 if ( !first ) {
3b6759a6
YO
3382 first = cur;
3383 trietype = noper_trietype;
df826430
YO
3384 if ( noper_trietype == NOTHING ) {
3385#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3386 regnode * const noper_next = regnext( noper );
3b6759a6 3387 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3388 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3389#endif
3390
3b6759a6 3391 if ( noper_next_trietype )
df826430 3392 trietype = noper_next_trietype;
3b6759a6 3393 }
8aa23a47 3394 } else {
fab2782b
YO
3395 if ( trietype == NOTHING )
3396 trietype = noper_trietype;
8aa23a47
YO
3397 last = cur;
3398 }
df826430
YO
3399 if (first)
3400 count++;
fab2782b
YO
3401 } /* end handle mergable triable node */
3402 else {
3403 /* handle unmergable node -
3404 * noper may either be a triable node which can not be tried
3405 * together with the current trie, or a non triable node */
729aaeb5
YO
3406 if ( last ) {
3407 /* If last is set and trietype is not NOTHING then we have found
3408 * at least two triable branch sequences in a row of a similar
3409 * trietype so we can turn them into a trie. If/when we
3410 * allow NOTHING to start a trie sequence this condition will be
3411 * required, and it isn't expensive so we leave it in for now. */
3412 if ( trietype != NOTHING )
3413 make_trie( pRExC_state,
3414 startbranch, first, cur, tail, count,
3415 trietype, depth+1 );
fab2782b 3416 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3417 }
fab2782b 3418 if ( noper_trietype
786e8c11 3419#ifdef NOJUMPTRIE
8aa23a47 3420 && noper_next == tail
786e8c11 3421#endif
8aa23a47 3422 ){
fab2782b 3423 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3424 count = 1;
3425 first = cur;
fab2782b
YO
3426 trietype = noper_trietype;
3427 } else if (first) {
3428 /* if we already saw a first but the current node is not triable then we have
3429 * to reset the first information. */
8aa23a47
YO
3430 count = 0;
3431 first = NULL;
fab2782b 3432 trietype = 0;
8aa23a47 3433 }
fab2782b
YO
3434 } /* end handle unmergable node */
3435 } /* loop over branches */
df826430 3436 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3437 regprop(RExC_rx, mysv, cur);
3438 PerlIO_printf( Perl_debug_log,
3439 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3440 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3441
3442 });
3b6759a6
YO
3443 if ( last ) {
3444 if ( trietype != NOTHING ) {
3445 /* the last branch of the sequence was part of a trie,
3446 * so we have to construct it here outside of the loop
3447 */
3448 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3449#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3450 if ( ((made == MADE_EXACT_TRIE &&
3451 startbranch == first)
3452 || ( first_non_open == first )) &&
3453 depth==0 ) {
3454 flags |= SCF_TRIE_RESTUDY;
3455 if ( startbranch == first
3456 && scan == tail )
3457 {
3458 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3459 }
8aa23a47 3460 }
3dab1dad 3461#endif
3b6759a6
YO
3462 } else {
3463 /* at this point we know whatever we have is a NOTHING sequence/branch
3464 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3465 */
3466 if ( startbranch == first ) {
3467 regnode *opt;
3468 /* the entire thing is a NOTHING sequence, something like this:
3469 * (?:|) So we can turn it into a plain NOTHING op. */
3470 DEBUG_TRIE_COMPILE_r({
3471 regprop(RExC_rx, mysv, cur);
3472 PerlIO_printf( Perl_debug_log,
3473 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3474 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3475
3476 });
3477 OP(startbranch)= NOTHING;
3478 NEXT_OFF(startbranch)= tail - startbranch;
3479 for ( opt= startbranch + 1; opt < tail ; opt++ )
3480 OP(opt)= OPTIMIZED;
3481 }
3482 }
fab2782b
YO
3483 } /* end if ( last) */
3484 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3485
3486 } /* do trie */
3487
653099ff 3488 }
8aa23a47
YO
3489 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3490 scan = NEXTOPER(NEXTOPER(scan));
3491 } else /* single branch is optimized. */
3492 scan = NEXTOPER(scan);
3493 continue;
3494 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3495 scan_frame *newframe = NULL;
3496 I32 paren;
3497 regnode *start;
3498 regnode *end;
3499
3500 if (OP(scan) != SUSPEND) {
3501 /* set the pointer */
3502 if (OP(scan) == GOSUB) {
3503 paren = ARG(scan);
3504 RExC_recurse[ARG2L(scan)] = scan;
3505 start = RExC_open_parens[paren-1];
3506 end = RExC_close_parens[paren-1];
3507 } else {
3508 paren = 0;
f8fc2ecf 3509 start = RExC_rxi->program + 1;
8aa23a47
YO
3510 end = RExC_opend;
3511 }
3512 if (!recursed) {
3513 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3514 SAVEFREEPV(recursed);
3515 }
3516 if (!PAREN_TEST(recursed,paren+1)) {
3517 PAREN_SET(recursed,paren+1);
3518 Newx(newframe,1,scan_frame);
3519 } else {
3520 if (flags & SCF_DO_SUBSTR) {
304ee84b 3521 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3522 data->longest = &(data->longest_float);
3523 }
3524 is_inf = is_inf_internal = 1;
3525 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3526 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3527 flags &= ~SCF_DO_STCLASS;
3528 }
3529 } else {
3530 Newx(newframe,1,scan_frame);
3531 paren = stopparen;
3532 start = scan+2;
3533 end = regnext(scan);
3534 }
3535 if (newframe) {
3536 assert(start);
3537 assert(end);
3538 SAVEFREEPV(newframe);
3539 newframe->next = regnext(scan);
3540 newframe->last = last;
3541 newframe->stop = stopparen;
3542 newframe->prev = frame;
3543
3544 frame = newframe;
3545 scan = start;
3546 stopparen = paren;
3547 last = end;
3548
3549 continue;
3550 }
3551 }
3552 else if (OP(scan) == EXACT) {
3553 I32 l = STR_LEN(scan);
3554 UV uc;
3555 if (UTF) {
3556 const U8 * const s = (U8*)STRING(scan);
4b88fb76 3557 uc = utf8_to_uvchr_buf(s, s + l, NULL);
8aa23a47 3558 l = utf8_length(s, s + l);
8aa23a47
YO
3559 } else {
3560 uc = *((U8*)STRING(scan));
3561 }
3562 min += l;
3563 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3564 /* The code below prefers earlier match for fixed
3565 offset, later match for variable offset. */
3566 if (data->last_end == -1) { /* Update the start info. */
3567 data->last_start_min = data->pos_min;
3568 data->last_start_max = is_inf
3569 ? I32_MAX : data->pos_min + data->pos_delta;
b515a41d 3570 }
8aa23a47
YO
3571 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3572 if (UTF)
3573 SvUTF8_on(data->last_found);
3574 {
3575 SV * const sv = data->last_found;
3576 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3577 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3578 if (mg && mg->mg_len >= 0)
3579 mg->mg_len += utf8_length((U8*)STRING(scan),
3580 (U8*)STRING(scan)+STR_LEN(scan));
b515a41d 3581 }
8aa23a47
YO
3582 data->last_end = data->pos_min + l;
3583 data->pos_min += l; /* As in the first entry. */
3584 data->flags &= ~SF_BEFORE_EOL;
3585 }
3586 if (flags & SCF_DO_STCLASS_AND) {
3587 /* Check whether it is compatible with what we know already! */
3588 int compat = 1;
3589
54251c2e 3590
486ec47a 3591 /* If compatible, we or it in below. It is compatible if is
54251c2e
KW
3592 * in the bitmp and either 1) its bit or its fold is set, or 2)
3593 * it's for a locale. Even if there isn't unicode semantics
3594 * here, at runtime there may be because of matching against a
3595 * utf8 string, so accept a possible false positive for
3596 * latin1-range folds */
8aa23a47
YO
3597 if (uc >= 0x100 ||
3598 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3599 && !ANYOF_BITMAP_TEST(data->start_class, uc)