This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
revert a trailing whitespace removal
[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
830247a4 112typedef struct RExC_state_t {
e2509266 113 U32 flags; /* are we folding, multilining? */
830247a4 114 char *precomp; /* uncompiled string. */
288b8c02 115 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf
YO
116 regexp *rx; /* perl core regexp structure */
117 regexp_internal *rxi; /* internal data for regexp object pprivate field */
fac92740 118 char *start; /* Start of input for compile */
830247a4
IZ
119 char *end; /* End of input for compile */
120 char *parse; /* Input-scan pointer. */
121 I32 whilem_seen; /* number of WHILEM in this expr */
fac92740 122 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 123 regnode *emit_bound; /* First regnode outside of the allocated space */
ffc61ed2 124 regnode *emit; /* Code-emit pointer; &regdummy = don't = compiling */
830247a4
IZ
125 I32 naughty; /* How bad is this pattern? */
126 I32 sawback; /* Did we see \1, ...? */
127 U32 seen;
128 I32 size; /* Code size. */
c74340f9
YO
129 I32 npar; /* Capture buffer count, (OPEN). */
130 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 131 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
132 I32 extralen;
133 I32 seen_zerolen;
134 I32 seen_evals;
40d049e4
YO
135 regnode **open_parens; /* pointers to open parens */
136 regnode **close_parens; /* pointers to close parens */
137 regnode *opend; /* END node in program */
02daf0ab
YO
138 I32 utf8; /* whether the pattern is utf8 or not */
139 I32 orig_utf8; /* whether the pattern was originally in utf8 */
140 /* XXX use this for future optimisation of case
141 * where pattern must be upgraded to utf8. */
e40e74fe
KW
142 I32 uni_semantics; /* If a d charset modifier should use unicode
143 rules, even if the pattern is not in
144 utf8 */
81714fb9 145 HV *paren_names; /* Paren names */
1f1031fe 146
40d049e4
YO
147 regnode **recurse; /* Recurse regops */
148 I32 recurse_count; /* Number of recurse regops */
b57e4118 149 I32 in_lookbehind;
4624b182 150 I32 contains_locale;
bb3f3ed2 151 I32 override_recoding;
830247a4
IZ
152#if ADD_TO_REGEXEC
153 char *starttry; /* -Dr: where regtry was called. */
154#define RExC_starttry (pRExC_state->starttry)
155#endif
3dab1dad 156#ifdef DEBUGGING
be8e71aa 157 const char *lastparse;
3dab1dad 158 I32 lastnum;
1f1031fe 159 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
160#define RExC_lastparse (pRExC_state->lastparse)
161#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 162#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 163#endif
830247a4
IZ
164} RExC_state_t;
165
e2509266 166#define RExC_flags (pRExC_state->flags)
830247a4 167#define RExC_precomp (pRExC_state->precomp)
288b8c02 168#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 169#define RExC_rx (pRExC_state->rx)
f8fc2ecf 170#define RExC_rxi (pRExC_state->rxi)
fac92740 171#define RExC_start (pRExC_state->start)
830247a4
IZ
172#define RExC_end (pRExC_state->end)
173#define RExC_parse (pRExC_state->parse)
174#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
175#ifdef RE_TRACK_PATTERN_OFFSETS
176#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
177#endif
830247a4 178#define RExC_emit (pRExC_state->emit)
fac92740 179#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 180#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
181#define RExC_naughty (pRExC_state->naughty)
182#define RExC_sawback (pRExC_state->sawback)
183#define RExC_seen (pRExC_state->seen)
184#define RExC_size (pRExC_state->size)
185#define RExC_npar (pRExC_state->npar)
e2e6a0f1 186#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
187#define RExC_extralen (pRExC_state->extralen)
188#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
189#define RExC_seen_evals (pRExC_state->seen_evals)
1aa99e6b 190#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 191#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 192#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
193#define RExC_open_parens (pRExC_state->open_parens)
194#define RExC_close_parens (pRExC_state->close_parens)
195#define RExC_opend (pRExC_state->opend)
81714fb9 196#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
197#define RExC_recurse (pRExC_state->recurse)
198#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 199#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 200#define RExC_contains_locale (pRExC_state->contains_locale)
bb3f3ed2 201#define RExC_override_recoding (pRExC_state->override_recoding)
830247a4 202
cde0cee5 203
a687059c
LW
204#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
205#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
206 ((*s) == '{' && regcurly(s)))
a687059c 207
35c8bce7
LW
208#ifdef SPSTART
209#undef SPSTART /* dratted cpp namespace... */
210#endif
a687059c
LW
211/*
212 * Flags to be passed up and down.
213 */
a687059c 214#define WORST 0 /* Worst case. */
a3b492c3 215#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee
KW
216
217/* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
cf8c372d
KW
218 * character, and if utf8, must be invariant. Note that this is not the same
219 * thing as REGNODE_SIMPLE */
fda99bee 220#define SIMPLE 0x02
a3b492c3
YO
221#define SPSTART 0x04 /* Starts with * or +. */
222#define TRYAGAIN 0x08 /* Weeded out a declaration. */
223#define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
a687059c 224
3dab1dad
YO
225#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
226
07be1b83
YO
227/* whether trie related optimizations are enabled */
228#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
229#define TRIE_STUDY_OPT
786e8c11 230#define FULL_TRIE_STUDY
07be1b83
YO
231#define TRIE_STCLASS
232#endif
1de06328
YO
233
234
40d049e4
YO
235
236#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
237#define PBITVAL(paren) (1 << ((paren) & 7))
238#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
239#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
240#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
241
bbd61b5f
KW
242/* If not already in utf8, do a longjmp back to the beginning */
243#define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
244#define REQUIRE_UTF8 STMT_START { \
245 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
246 } STMT_END
40d049e4 247
1de06328
YO
248/* About scan_data_t.
249
250 During optimisation we recurse through the regexp program performing
251 various inplace (keyhole style) optimisations. In addition study_chunk
252 and scan_commit populate this data structure with information about
253 what strings MUST appear in the pattern. We look for the longest
3b753521 254 string that must appear at a fixed location, and we look for the
1de06328
YO
255 longest string that may appear at a floating location. So for instance
256 in the pattern:
257
258 /FOO[xX]A.*B[xX]BAR/
259
260 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
261 strings (because they follow a .* construct). study_chunk will identify
262 both FOO and BAR as being the longest fixed and floating strings respectively.
263
264 The strings can be composites, for instance
265
266 /(f)(o)(o)/
267
268 will result in a composite fixed substring 'foo'.
269
270 For each string some basic information is maintained:
271
272 - offset or min_offset
273 This is the position the string must appear at, or not before.
274 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
275 characters must match before the string we are searching for.
276 Likewise when combined with minlenp and the length of the string it
1de06328
YO
277 tells us how many characters must appear after the string we have
278 found.
279
280 - max_offset
281 Only used for floating strings. This is the rightmost point that
3b753521 282 the string can appear at. If set to I32 max it indicates that the
1de06328
YO
283 string can occur infinitely far to the right.
284
285 - minlenp
286 A pointer to the minimum length of the pattern that the string
287 was found inside. This is important as in the case of positive
288 lookahead or positive lookbehind we can have multiple patterns
289 involved. Consider
290
291 /(?=FOO).*F/
292
293 The minimum length of the pattern overall is 3, the minimum length
294 of the lookahead part is 3, but the minimum length of the part that
295 will actually match is 1. So 'FOO's minimum length is 3, but the
296 minimum length for the F is 1. This is important as the minimum length
297 is used to determine offsets in front of and behind the string being
298 looked for. Since strings can be composites this is the length of the
486ec47a 299 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
300 the length is calculated by study_chunk, so that the minimum lengths
301 are not known until the full pattern has been compiled, thus the
302 pointer to the value.
303
304 - lookbehind
305
306 In the case of lookbehind the string being searched for can be
307 offset past the start point of the final matching string.
308 If this value was just blithely removed from the min_offset it would
309 invalidate some of the calculations for how many chars must match
310 before or after (as they are derived from min_offset and minlen and
311 the length of the string being searched for).
312 When the final pattern is compiled and the data is moved from the
313 scan_data_t structure into the regexp structure the information
314 about lookbehind is factored in, with the information that would
315 have been lost precalculated in the end_shift field for the
316 associated string.
317
318 The fields pos_min and pos_delta are used to store the minimum offset
319 and the delta to the maximum offset at the current point in the pattern.
320
321*/
2c2d71f5
JH
322
323typedef struct scan_data_t {
1de06328
YO
324 /*I32 len_min; unused */
325 /*I32 len_delta; unused */
2c2d71f5
JH
326 I32 pos_min;
327 I32 pos_delta;
328 SV *last_found;
1de06328 329 I32 last_end; /* min value, <0 unless valid. */
2c2d71f5
JH
330 I32 last_start_min;
331 I32 last_start_max;
1de06328
YO
332 SV **longest; /* Either &l_fixed, or &l_float. */
333 SV *longest_fixed; /* longest fixed string found in pattern */
334 I32 offset_fixed; /* offset where it starts */
486ec47a 335 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
336 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
337 SV *longest_float; /* longest floating string found in pattern */
338 I32 offset_float_min; /* earliest point in string it can appear */
339 I32 offset_float_max; /* latest point in string it can appear */
486ec47a 340 I32 *minlen_float; /* pointer to the minlen relevant to the string */
1de06328 341 I32 lookbehind_float; /* is the position of the string modified by LB */
2c2d71f5
JH
342 I32 flags;
343 I32 whilem_c;
cb434fcc 344 I32 *last_closep;
653099ff 345 struct regnode_charclass_class *start_class;
2c2d71f5
JH
346} scan_data_t;
347
a687059c 348/*
e50aee73 349 * Forward declarations for pregcomp()'s friends.
a687059c 350 */
a0d0e21e 351
27da23d5 352static const scan_data_t zero_scan_data =
1de06328 353 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
354
355#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
356#define SF_BEFORE_SEOL 0x0001
357#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
358#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
359#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
360
09b7f37c
CB
361#ifdef NO_UNARY_PLUS
362# define SF_FIX_SHIFT_EOL (0+2)
363# define SF_FL_SHIFT_EOL (0+4)
364#else
365# define SF_FIX_SHIFT_EOL (+2)
366# define SF_FL_SHIFT_EOL (+4)
367#endif
c277df42
IZ
368
369#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
370#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
371
372#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
373#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
374#define SF_IS_INF 0x0040
375#define SF_HAS_PAR 0x0080
376#define SF_IN_PAR 0x0100
377#define SF_HAS_EVAL 0x0200
378#define SCF_DO_SUBSTR 0x0400
653099ff
GS
379#define SCF_DO_STCLASS_AND 0x0800
380#define SCF_DO_STCLASS_OR 0x1000
381#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 382#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 383
786e8c11 384#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 385#define SCF_SEEN_ACCEPT 0x8000
07be1b83 386
43fead97 387#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
388
389/* The enums for all these are ordered so things work out correctly */
a62b1201 390#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
cfaf538b 391#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
00b27cfc 392#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
393#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
394#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1
KW
395#define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
396#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a62b1201 397
43fead97 398#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 399
ffc61ed2 400#define OOB_UNICODE 12345678
93733859 401#define OOB_NAMEDCLASS -1
b8c5462f 402
a0ed51b3
LW
403#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
404#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
405
8615cb43 406
b45f050a
JF
407/* length of regex to show in messages that don't mark a position within */
408#define RegexLengthToShowInErrorMessages 127
409
410/*
411 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
412 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
413 * op/pragma/warn/regcomp.
414 */
7253e4e3
RK
415#define MARKER1 "<-- HERE" /* marker as it appears in the description */
416#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 417
7253e4e3 418#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
419
420/*
421 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
422 * arg. Show regex, up to a maximum length. If it's too long, chop and add
423 * "...".
424 */
58e23c8d 425#define _FAIL(code) STMT_START { \
bfed75c6 426 const char *ellipses = ""; \
ccb2c380
MP
427 IV len = RExC_end - RExC_precomp; \
428 \
429 if (!SIZE_ONLY) \
288b8c02 430 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
431 if (len > RegexLengthToShowInErrorMessages) { \
432 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
433 len = RegexLengthToShowInErrorMessages - 10; \
434 ellipses = "..."; \
435 } \
58e23c8d 436 code; \
ccb2c380 437} STMT_END
8615cb43 438
58e23c8d
YO
439#define FAIL(msg) _FAIL( \
440 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
441 msg, (int)len, RExC_precomp, ellipses))
442
443#define FAIL2(msg,arg) _FAIL( \
444 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
445 arg, (int)len, RExC_precomp, ellipses))
446
b45f050a 447/*
b45f050a
JF
448 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
449 */
ccb2c380 450#define Simple_vFAIL(m) STMT_START { \
a28509cc 451 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
452 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
453 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
454} STMT_END
b45f050a
JF
455
456/*
457 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
458 */
ccb2c380
MP
459#define vFAIL(m) STMT_START { \
460 if (!SIZE_ONLY) \
288b8c02 461 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
462 Simple_vFAIL(m); \
463} STMT_END
b45f050a
JF
464
465/*
466 * Like Simple_vFAIL(), but accepts two arguments.
467 */
ccb2c380 468#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 469 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
470 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
471 (int)offset, RExC_precomp, RExC_precomp + offset); \
472} STMT_END
b45f050a
JF
473
474/*
475 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
476 */
ccb2c380
MP
477#define vFAIL2(m,a1) STMT_START { \
478 if (!SIZE_ONLY) \
288b8c02 479 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
480 Simple_vFAIL2(m, a1); \
481} STMT_END
b45f050a
JF
482
483
484/*
485 * Like Simple_vFAIL(), but accepts three arguments.
486 */
ccb2c380 487#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 488 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
489 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
490 (int)offset, RExC_precomp, RExC_precomp + offset); \
491} STMT_END
b45f050a
JF
492
493/*
494 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
495 */
ccb2c380
MP
496#define vFAIL3(m,a1,a2) STMT_START { \
497 if (!SIZE_ONLY) \
288b8c02 498 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
499 Simple_vFAIL3(m, a1, a2); \
500} STMT_END
b45f050a
JF
501
502/*
503 * Like Simple_vFAIL(), but accepts four arguments.
504 */
ccb2c380 505#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 506 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
507 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
508 (int)offset, RExC_precomp, RExC_precomp + offset); \
509} STMT_END
b45f050a 510
668c081a 511#define ckWARNreg(loc,m) STMT_START { \
a28509cc 512 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
513 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
514 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
515} STMT_END
516
668c081a 517#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 518 const IV offset = loc - RExC_precomp; \
d1d15184 519 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
520 m REPORT_LOCATION, \
521 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
522} STMT_END
523
2335b3d3
KW
524#define ckWARN2regdep(loc,m, a1) STMT_START { \
525 const IV offset = loc - RExC_precomp; \
526 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
527 m REPORT_LOCATION, \
528 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
529} STMT_END
530
668c081a 531#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 532 const IV offset = loc - RExC_precomp; \
668c081a 533 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
534 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
535} STMT_END
536
537#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 538 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
539 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
540 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
541} STMT_END
542
668c081a
NC
543#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
544 const IV offset = loc - RExC_precomp; \
545 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
546 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
547} STMT_END
548
ccb2c380 549#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 550 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
551 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
552 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
553} STMT_END
554
668c081a
NC
555#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
556 const IV offset = loc - RExC_precomp; \
557 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
558 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
559} STMT_END
560
ccb2c380 561#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 562 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
563 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
564 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
565} STMT_END
9d1d55b5 566
8615cb43 567
cd439c50 568/* Allow for side effects in s */
ccb2c380
MP
569#define REGC(c,s) STMT_START { \
570 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
571} STMT_END
cd439c50 572
fac92740
MJD
573/* Macros for recording node offsets. 20001227 mjd@plover.com
574 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
575 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
576 * Element 0 holds the number n.
07be1b83 577 * Position is 1 indexed.
fac92740 578 */
7122b237
YO
579#ifndef RE_TRACK_PATTERN_OFFSETS
580#define Set_Node_Offset_To_R(node,byte)
581#define Set_Node_Offset(node,byte)
582#define Set_Cur_Node_Offset
583#define Set_Node_Length_To_R(node,len)
584#define Set_Node_Length(node,len)
585#define Set_Node_Cur_Length(node)
586#define Node_Offset(n)
587#define Node_Length(n)
588#define Set_Node_Offset_Length(node,offset,len)
589#define ProgLen(ri) ri->u.proglen
590#define SetProgLen(ri,x) ri->u.proglen = x
591#else
592#define ProgLen(ri) ri->u.offsets[0]
593#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
594#define Set_Node_Offset_To_R(node,byte) STMT_START { \
595 if (! SIZE_ONLY) { \
596 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 597 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 598 if((node) < 0) { \
551405c4 599 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
600 } else { \
601 RExC_offsets[2*(node)-1] = (byte); \
602 } \
603 } \
604} STMT_END
605
606#define Set_Node_Offset(node,byte) \
607 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
608#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
609
610#define Set_Node_Length_To_R(node,len) STMT_START { \
611 if (! SIZE_ONLY) { \
612 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 613 __LINE__, (int)(node), (int)(len))); \
ccb2c380 614 if((node) < 0) { \
551405c4 615 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
616 } else { \
617 RExC_offsets[2*(node)] = (len); \
618 } \
619 } \
620} STMT_END
621
622#define Set_Node_Length(node,len) \
623 Set_Node_Length_To_R((node)-RExC_emit_start, len)
624#define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
625#define Set_Node_Cur_Length(node) \
626 Set_Node_Length(node, RExC_parse - parse_start)
fac92740
MJD
627
628/* Get offsets and lengths */
629#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
630#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
631
07be1b83
YO
632#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
633 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
634 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
635} STMT_END
7122b237 636#endif
07be1b83
YO
637
638#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
639#define EXPERIMENTAL_INPLACESCAN
f427392e 640#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 641
304ee84b
YO
642#define DEBUG_STUDYDATA(str,data,depth) \
643DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 644 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
645 "%*s" str "Pos:%"IVdf"/%"IVdf \
646 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
647 (int)(depth)*2, "", \
648 (IV)((data)->pos_min), \
649 (IV)((data)->pos_delta), \
304ee84b 650 (UV)((data)->flags), \
1de06328 651 (IV)((data)->whilem_c), \
304ee84b
YO
652 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
653 is_inf ? "INF " : "" \
1de06328
YO
654 ); \
655 if ((data)->last_found) \
656 PerlIO_printf(Perl_debug_log, \
657 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
658 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
659 SvPVX_const((data)->last_found), \
660 (IV)((data)->last_end), \
661 (IV)((data)->last_start_min), \
662 (IV)((data)->last_start_max), \
663 ((data)->longest && \
664 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
665 SvPVX_const((data)->longest_fixed), \
666 (IV)((data)->offset_fixed), \
667 ((data)->longest && \
668 (data)->longest==&((data)->longest_float)) ? "*" : "", \
669 SvPVX_const((data)->longest_float), \
670 (IV)((data)->offset_float_min), \
671 (IV)((data)->offset_float_max) \
672 ); \
673 PerlIO_printf(Perl_debug_log,"\n"); \
674});
675
acfe0abc 676static void clear_re(pTHX_ void *r);
4327152a 677
653099ff 678/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 679 Update the longest found anchored substring and the longest found
653099ff
GS
680 floating substrings if needed. */
681
4327152a 682STATIC void
304ee84b 683S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 684{
e1ec3a88
AL
685 const STRLEN l = CHR_SVLEN(data->last_found);
686 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 687 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 688
7918f24d
NC
689 PERL_ARGS_ASSERT_SCAN_COMMIT;
690
c277df42 691 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 692 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
693 if (*data->longest == data->longest_fixed) {
694 data->offset_fixed = l ? data->last_start_min : data->pos_min;
695 if (data->flags & SF_BEFORE_EOL)
b81d288d 696 data->flags
c277df42
IZ
697 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
698 else
699 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 700 data->minlen_fixed=minlenp;
1de06328 701 data->lookbehind_fixed=0;
a0ed51b3 702 }
304ee84b 703 else { /* *data->longest == data->longest_float */
c277df42 704 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
705 data->offset_float_max = (l
706 ? data->last_start_max
c277df42 707 : data->pos_min + data->pos_delta);
304ee84b 708 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 709 data->offset_float_max = I32_MAX;
c277df42 710 if (data->flags & SF_BEFORE_EOL)
b81d288d 711 data->flags
c277df42
IZ
712 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
713 else
714 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
715 data->minlen_float=minlenp;
716 data->lookbehind_float=0;
c277df42
IZ
717 }
718 }
719 SvCUR_set(data->last_found, 0);
0eda9292 720 {
a28509cc 721 SV * const sv = data->last_found;
097eb12c
AL
722 if (SvUTF8(sv) && SvMAGICAL(sv)) {
723 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
724 if (mg)
725 mg->mg_len = 0;
726 }
0eda9292 727 }
c277df42
IZ
728 data->last_end = -1;
729 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 730 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
731}
732
653099ff
GS
733/* Can match anything (initialization) */
734STATIC void
3fffb88a 735S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 736{
7918f24d
NC
737 PERL_ARGS_ASSERT_CL_ANYTHING;
738
f8bef550 739 ANYOF_BITMAP_SETALL(cl);
dd58aee1 740 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
3ad98780 741 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
3fffb88a
KW
742
743 /* If any portion of the regex is to operate under locale rules,
744 * initialization includes it. The reason this isn't done for all regexes
745 * is that the optimizer was written under the assumption that locale was
746 * all-or-nothing. Given the complexity and lack of documentation in the
747 * optimizer, and that there are inadequate test cases for locale, so many
748 * parts of it may not work properly, it is safest to avoid locale unless
749 * necessary. */
750 if (RExC_contains_locale) {
9d7a1e63 751 ANYOF_CLASS_SETALL(cl); /* /l uses class */
3fffb88a
KW
752 cl->flags |= ANYOF_LOCALE;
753 }
9d7a1e63
KW
754 else {
755 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
756 }
653099ff
GS
757}
758
759/* Can match anything (initialization) */
760STATIC int
5f66b61c 761S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
762{
763 int value;
764
7918f24d
NC
765 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
766
aaa51d5e 767 for (value = 0; value <= ANYOF_MAX; value += 2)
653099ff
GS
768 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
769 return 1;
1aa99e6b
IH
770 if (!(cl->flags & ANYOF_UNICODE_ALL))
771 return 0;
10edeb5d 772 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 773 return 0;
653099ff
GS
774 return 1;
775}
776
777/* Can match anything (initialization) */
778STATIC void
e755fd73 779S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 780{
7918f24d
NC
781 PERL_ARGS_ASSERT_CL_INIT;
782
8ecf7187 783 Zero(cl, 1, struct regnode_charclass_class);
653099ff 784 cl->type = ANYOF;
3fffb88a 785 cl_anything(pRExC_state, cl);
1411dba4 786 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
787}
788
1051e1c4
KW
789/* These two functions currently do the exact same thing */
790#define cl_init_zero S_cl_init
653099ff 791
dd58aee1
KW
792/* 'AND' a given class with another one. Can create false positives. 'cl'
793 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
794 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 795STATIC void
5f66b61c 796S_cl_and(struct regnode_charclass_class *cl,
a28509cc 797 const struct regnode_charclass_class *and_with)
653099ff 798{
7918f24d 799 PERL_ARGS_ASSERT_CL_AND;
40d049e4
YO
800
801 assert(and_with->type == ANYOF);
1e6ade67 802
c6b76537 803 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
804 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
805 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 806 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
807 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
808 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
653099ff
GS
809 int i;
810
811 if (and_with->flags & ANYOF_INVERT)
812 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
813 cl->bitmap[i] &= ~and_with->bitmap[i];
814 else
815 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
816 cl->bitmap[i] &= and_with->bitmap[i];
817 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 818
c6b76537 819 if (and_with->flags & ANYOF_INVERT) {
8951c461 820
c6b76537
KW
821 /* Here, the and'ed node is inverted. Get the AND of the flags that
822 * aren't affected by the inversion. Those that are affected are
823 * handled individually below */
824 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
825 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
826 cl->flags |= affected_flags;
827
828 /* We currently don't know how to deal with things that aren't in the
829 * bitmap, but we know that the intersection is no greater than what
830 * is already in cl, so let there be false positives that get sorted
831 * out after the synthetic start class succeeds, and the node is
832 * matched for real. */
833
834 /* The inversion of these two flags indicate that the resulting
835 * intersection doesn't have them */
836 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
837 cl->flags &= ~ANYOF_UNICODE_ALL;
838 }
c6b76537
KW
839 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
840 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 841 }
1aa99e6b 842 }
c6b76537 843 else { /* and'd node is not inverted */
3ad98780
KW
844 U8 outside_bitmap_but_not_utf8; /* Temp variable */
845
137165a6 846 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
847
848 /* Here 'and_with' doesn't match anything outside the bitmap
849 * (except possibly ANYOF_UNICODE_ALL), which means the
850 * intersection can't either, except for ANYOF_UNICODE_ALL, in
851 * which case we don't know what the intersection is, but it's no
852 * greater than what cl already has, so can just leave it alone,
853 * with possible false positives */
854 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
855 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 856 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 857 }
137165a6 858 }
c6b76537
KW
859 else if (! ANYOF_NONBITMAP(cl)) {
860
861 /* Here, 'and_with' does match something outside the bitmap, and cl
862 * doesn't have a list of things to match outside the bitmap. If
863 * cl can match all code points above 255, the intersection will
3ad98780
KW
864 * be those above-255 code points that 'and_with' matches. If cl
865 * can't match all Unicode code points, it means that it can't
866 * match anything outside the bitmap (since the 'if' that got us
867 * into this block tested for that), so we leave the bitmap empty.
868 */
c6b76537
KW
869 if (cl->flags & ANYOF_UNICODE_ALL) {
870 ARG_SET(cl, ARG(and_with));
3ad98780
KW
871
872 /* and_with's ARG may match things that don't require UTF8.
873 * And now cl's will too, in spite of this being an 'and'. See
874 * the comments below about the kludge */
875 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
876 }
877 }
878 else {
879 /* Here, both 'and_with' and cl match something outside the
880 * bitmap. Currently we do not do the intersection, so just match
881 * whatever cl had at the beginning. */
882 }
883
884
3ad98780
KW
885 /* Take the intersection of the two sets of flags. However, the
886 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
887 * kludge around the fact that this flag is not treated like the others
888 * which are initialized in cl_anything(). The way the optimizer works
889 * is that the synthetic start class (SSC) is initialized to match
890 * anything, and then the first time a real node is encountered, its
891 * values are AND'd with the SSC's with the result being the values of
892 * the real node. However, there are paths through the optimizer where
893 * the AND never gets called, so those initialized bits are set
894 * inappropriately, which is not usually a big deal, as they just cause
895 * false positives in the SSC, which will just mean a probably
896 * imperceptible slow down in execution. However this bit has a
897 * higher false positive consequence in that it can cause utf8.pm,
898 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
899 * bigger slowdown and also causes significant extra memory to be used.
900 * In order to prevent this, the code now takes a different tack. The
901 * bit isn't set unless some part of the regular expression needs it,
902 * but once set it won't get cleared. This means that these extra
903 * modules won't get loaded unless there was some path through the
904 * pattern that would have required them anyway, and so any false
905 * positives that occur by not ANDing them out when they could be
906 * aren't as severe as they would be if we treated this bit like all
907 * the others */
908 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
909 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 910 cl->flags &= and_with->flags;
3ad98780 911 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 912 }
653099ff
GS
913}
914
dd58aee1
KW
915/* 'OR' a given class with another one. Can create false positives. 'cl'
916 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
917 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 918STATIC void
3fffb88a 919S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 920{
7918f24d
NC
921 PERL_ARGS_ASSERT_CL_OR;
922
653099ff 923 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
924
925 /* Here, the or'd node is to be inverted. This means we take the
926 * complement of everything not in the bitmap, but currently we don't
927 * know what that is, so give up and match anything */
928 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 929 cl_anything(pRExC_state, cl);
c6b76537 930 }
653099ff
GS
931 /* We do not use
932 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
933 * <= (B1 | !B2) | (CL1 | !CL2)
934 * which is wasteful if CL2 is small, but we ignore CL2:
935 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
936 * XXXX Can we handle case-fold? Unclear:
937 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
938 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
939 */
c6b76537 940 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
941 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
942 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
653099ff
GS
943 int i;
944
945 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
946 cl->bitmap[i] |= ~or_with->bitmap[i];
947 } /* XXXX: logic is complicated otherwise */
948 else {
3fffb88a 949 cl_anything(pRExC_state, cl);
653099ff 950 }
c6b76537
KW
951
952 /* And, we can just take the union of the flags that aren't affected
953 * by the inversion */
954 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
955
956 /* For the remaining flags:
957 ANYOF_UNICODE_ALL and inverted means to not match anything above
958 255, which means that the union with cl should just be
959 what cl has in it, so can ignore this flag
960 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
961 is 127-255 to match them, but then invert that, so the
962 union with cl should just be what cl has in it, so can
963 ignore this flag
964 */
965 } else { /* 'or_with' is not inverted */
653099ff
GS
966 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
967 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
968 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
969 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
653099ff
GS
970 int i;
971
972 /* OR char bitmap and class bitmap separately */
973 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
974 cl->bitmap[i] |= or_with->bitmap[i];
1e6ade67 975 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
653099ff
GS
976 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
977 cl->classflags[i] |= or_with->classflags[i];
978 cl->flags |= ANYOF_CLASS;
979 }
980 }
981 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 982 cl_anything(pRExC_state, cl);
653099ff 983 }
9826f543 984
c6b76537
KW
985 if (ANYOF_NONBITMAP(or_with)) {
986
987 /* Use the added node's outside-the-bit-map match if there isn't a
988 * conflict. If there is a conflict (both nodes match something
989 * outside the bitmap, but what they match outside is not the same
990 * pointer, and hence not easily compared until XXX we extend
991 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
992 * match everything outside the bitmap. If that stuff is all above
993 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
994 if (! ANYOF_NONBITMAP(cl)) {
995 ARG_SET(cl, ARG(or_with));
996 }
997 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
998
999 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1000 cl_anything(pRExC_state, cl);
1001 }
1002 else {
1003 cl->flags |= ANYOF_UNICODE_ALL;
1004 }
c6b76537 1005 }
4c34a693 1006 }
0b9668ee
KW
1007
1008 /* Take the union */
1009 cl->flags |= or_with->flags;
1aa99e6b 1010 }
653099ff
GS
1011}
1012
a3621e74
YO
1013#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1014#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1015#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1016#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1017
3dab1dad
YO
1018
1019#ifdef DEBUGGING
07be1b83 1020/*
2b8b4781
NC
1021 dump_trie(trie,widecharmap,revcharmap)
1022 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1023 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1024
1025 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1026 The _interim_ variants are used for debugging the interim
1027 tables that are used to generate the final compressed
1028 representation which is what dump_trie expects.
1029
486ec47a 1030 Part of the reason for their existence is to provide a form
3dab1dad 1031 of documentation as to how the different representations function.
07be1b83
YO
1032
1033*/
3dab1dad
YO
1034
1035/*
3dab1dad
YO
1036 Dumps the final compressed table form of the trie to Perl_debug_log.
1037 Used for debugging make_trie().
1038*/
b9a59e08 1039
3dab1dad 1040STATIC void
2b8b4781
NC
1041S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1042 AV *revcharmap, U32 depth)
3dab1dad
YO
1043{
1044 U32 state;
ab3bbdeb 1045 SV *sv=sv_newmortal();
55eed653 1046 int colwidth= widecharmap ? 6 : 4;
2e64971a 1047 U16 word;
3dab1dad
YO
1048 GET_RE_DEBUG_FLAGS_DECL;
1049
7918f24d 1050 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1051
3dab1dad
YO
1052 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1053 (int)depth * 2 + 2,"",
1054 "Match","Base","Ofs" );
1055
1056 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1057 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1058 if ( tmp ) {
ab3bbdeb
YO
1059 PerlIO_printf( Perl_debug_log, "%*s",
1060 colwidth,
ddc5bc0f 1061 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1062 PL_colors[0], PL_colors[1],
1063 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1064 PERL_PV_ESCAPE_FIRSTCHAR
1065 )
1066 );
3dab1dad
YO
1067 }
1068 }
1069 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1070 (int)depth * 2 + 2,"");
1071
1072 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1073 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1074 PerlIO_printf( Perl_debug_log, "\n");
1075
1e2e3d02 1076 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1077 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1078
1079 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1080
1081 if ( trie->states[ state ].wordnum ) {
1082 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1083 } else {
1084 PerlIO_printf( Perl_debug_log, "%6s", "" );
1085 }
1086
1087 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1088
1089 if ( base ) {
1090 U32 ofs = 0;
1091
1092 while( ( base + ofs < trie->uniquecharcount ) ||
1093 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1094 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1095 ofs++;
1096
1097 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1098
1099 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1100 if ( ( base + ofs >= trie->uniquecharcount ) &&
1101 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1102 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1103 {
ab3bbdeb
YO
1104 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1105 colwidth,
3dab1dad
YO
1106 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1107 } else {
ab3bbdeb 1108 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1109 }
1110 }
1111
1112 PerlIO_printf( Perl_debug_log, "]");
1113
1114 }
1115 PerlIO_printf( Perl_debug_log, "\n" );
1116 }
2e64971a
DM
1117 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1118 for (word=1; word <= trie->wordcount; word++) {
1119 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1120 (int)word, (int)(trie->wordinfo[word].prev),
1121 (int)(trie->wordinfo[word].len));
1122 }
1123 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1124}
1125/*
3dab1dad
YO
1126 Dumps a fully constructed but uncompressed trie in list form.
1127 List tries normally only are used for construction when the number of
1128 possible chars (trie->uniquecharcount) is very high.
1129 Used for debugging make_trie().
1130*/
1131STATIC void
55eed653 1132S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1133 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1134 U32 depth)
3dab1dad
YO
1135{
1136 U32 state;
ab3bbdeb 1137 SV *sv=sv_newmortal();
55eed653 1138 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1139 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1140
1141 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1142
3dab1dad 1143 /* print out the table precompression. */
ab3bbdeb
YO
1144 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1145 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1146 "------:-----+-----------------\n" );
3dab1dad
YO
1147
1148 for( state=1 ; state < next_alloc ; state ++ ) {
1149 U16 charid;
1150
ab3bbdeb 1151 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1152 (int)depth * 2 + 2,"", (UV)state );
1153 if ( ! trie->states[ state ].wordnum ) {
1154 PerlIO_printf( Perl_debug_log, "%5s| ","");
1155 } else {
1156 PerlIO_printf( Perl_debug_log, "W%4x| ",
1157 trie->states[ state ].wordnum
1158 );
1159 }
1160 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1161 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1162 if ( tmp ) {
1163 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1164 colwidth,
ddc5bc0f 1165 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1166 PL_colors[0], PL_colors[1],
1167 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1168 PERL_PV_ESCAPE_FIRSTCHAR
1169 ) ,
1e2e3d02
YO
1170 TRIE_LIST_ITEM(state,charid).forid,
1171 (UV)TRIE_LIST_ITEM(state,charid).newstate
1172 );
1173 if (!(charid % 10))
664e119d
RGS
1174 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1175 (int)((depth * 2) + 14), "");
1e2e3d02 1176 }
ab3bbdeb
YO
1177 }
1178 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1179 }
1180}
1181
1182/*
3dab1dad
YO
1183 Dumps a fully constructed but uncompressed trie in table form.
1184 This is the normal DFA style state transition table, with a few
1185 twists to facilitate compression later.
1186 Used for debugging make_trie().
1187*/
1188STATIC void
55eed653 1189S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1190 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1191 U32 depth)
3dab1dad
YO
1192{
1193 U32 state;
1194 U16 charid;
ab3bbdeb 1195 SV *sv=sv_newmortal();
55eed653 1196 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1197 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1198
1199 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1200
1201 /*
1202 print out the table precompression so that we can do a visual check
1203 that they are identical.
1204 */
1205
1206 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1207
1208 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1209 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1210 if ( tmp ) {
ab3bbdeb
YO
1211 PerlIO_printf( Perl_debug_log, "%*s",
1212 colwidth,
ddc5bc0f 1213 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1214 PL_colors[0], PL_colors[1],
1215 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1216 PERL_PV_ESCAPE_FIRSTCHAR
1217 )
1218 );
3dab1dad
YO
1219 }
1220 }
1221
1222 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1223
1224 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1225 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1226 }
1227
1228 PerlIO_printf( Perl_debug_log, "\n" );
1229
1230 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1231
1232 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1233 (int)depth * 2 + 2,"",
1234 (UV)TRIE_NODENUM( state ) );
1235
1236 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1237 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1238 if (v)
1239 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1240 else
1241 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1242 }
1243 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1244 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1245 } else {
1246 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1247 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1248 }
1249 }
07be1b83 1250}
3dab1dad
YO
1251
1252#endif
1253
2e64971a 1254
786e8c11
YO
1255/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1256 startbranch: the first branch in the whole branch sequence
1257 first : start branch of sequence of branch-exact nodes.
1258 May be the same as startbranch
1259 last : Thing following the last branch.
1260 May be the same as tail.
1261 tail : item following the branch sequence
1262 count : words in the sequence
1263 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1264 depth : indent depth
3dab1dad 1265
786e8c11 1266Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1267
786e8c11
YO
1268A trie is an N'ary tree where the branches are determined by digital
1269decomposition of the key. IE, at the root node you look up the 1st character and
1270follow that branch repeat until you find the end of the branches. Nodes can be
1271marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1272
786e8c11 1273 /he|she|his|hers/
72f13be8 1274
786e8c11
YO
1275would convert into the following structure. Numbers represent states, letters
1276following numbers represent valid transitions on the letter from that state, if
1277the number is in square brackets it represents an accepting state, otherwise it
1278will be in parenthesis.
07be1b83 1279
786e8c11
YO
1280 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1281 | |
1282 | (2)
1283 | |
1284 (1) +-i->(6)-+-s->[7]
1285 |
1286 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1287
786e8c11
YO
1288 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1289
1290This shows that when matching against the string 'hers' we will begin at state 1
1291read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1292then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1293is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1294single traverse. We store a mapping from accepting to state to which word was
1295matched, and then when we have multiple possibilities we try to complete the
1296rest of the regex in the order in which they occured in the alternation.
1297
1298The only prior NFA like behaviour that would be changed by the TRIE support is
1299the silent ignoring of duplicate alternations which are of the form:
1300
1301 / (DUPE|DUPE) X? (?{ ... }) Y /x
1302
4b714af6 1303Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1304and without the optimisation. With the optimisations dupes will be silently
486ec47a 1305ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1306the following demonstrates:
1307
1308 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1309
1310which prints out 'word' three times, but
1311
1312 'words'=~/(word|word|word)(?{ print $1 })S/
1313
1314which doesnt print it out at all. This is due to other optimisations kicking in.
1315
1316Example of what happens on a structural level:
1317
486ec47a 1318The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1319
1320 1: CURLYM[1] {1,32767}(18)
1321 5: BRANCH(8)
1322 6: EXACT <ac>(16)
1323 8: BRANCH(11)
1324 9: EXACT <ad>(16)
1325 11: BRANCH(14)
1326 12: EXACT <ab>(16)
1327 16: SUCCEED(0)
1328 17: NOTHING(18)
1329 18: END(0)
1330
1331This would be optimizable with startbranch=5, first=5, last=16, tail=16
1332and should turn into:
1333
1334 1: CURLYM[1] {1,32767}(18)
1335 5: TRIE(16)
1336 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1337 <ac>
1338 <ad>
1339 <ab>
1340 16: SUCCEED(0)
1341 17: NOTHING(18)
1342 18: END(0)
1343
1344Cases where tail != last would be like /(?foo|bar)baz/:
1345
1346 1: BRANCH(4)
1347 2: EXACT <foo>(8)
1348 4: BRANCH(7)
1349 5: EXACT <bar>(8)
1350 7: TAIL(8)
1351 8: EXACT <baz>(10)
1352 10: END(0)
1353
1354which would be optimizable with startbranch=1, first=1, last=7, tail=8
1355and would end up looking like:
1356
1357 1: TRIE(8)
1358 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1359 <foo>
1360 <bar>
1361 7: TAIL(8)
1362 8: EXACT <baz>(10)
1363 10: END(0)
1364
1365 d = uvuni_to_utf8_flags(d, uv, 0);
1366
1367is the recommended Unicode-aware way of saying
1368
1369 *(d++) = uv;
1370*/
1371
fab2782b 1372#define TRIE_STORE_REVCHAR(val) \
786e8c11 1373 STMT_START { \
73031816 1374 if (UTF) { \
fab2782b 1375 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1376 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
fab2782b 1377 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
73031816
NC
1378 SvCUR_set(zlopp, kapow - flrbbbbb); \
1379 SvPOK_on(zlopp); \
1380 SvUTF8_on(zlopp); \
1381 av_push(revcharmap, zlopp); \
1382 } else { \
fab2782b 1383 char ooooff = (char)val; \
73031816
NC
1384 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1385 } \
1386 } STMT_END
786e8c11 1387
fab2782b
YO
1388#define TRIE_READ_CHAR STMT_START { \
1389 wordlen++; \
1390 if ( UTF ) { \
1391 /* if it is UTF then it is either already folded, or does not need folding */ \
1392 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1393 } \
1394 else if (folder == PL_fold_latin1) { \
1395 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1396 if ( foldlen > 0 ) { \
1397 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1398 foldlen -= len; \
1399 scan += len; \
1400 len = 0; \
1401 } else { \
1402 len = 1; \
1403 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1404 skiplen = UNISKIP(uvc); \
1405 foldlen -= skiplen; \
1406 scan = foldbuf + skiplen; \
1407 } \
1408 } else { \
1409 /* raw data, will be folded later if needed */ \
1410 uvc = (U32)*uc; \
1411 len = 1; \
1412 } \
786e8c11
YO
1413} STMT_END
1414
1415
1416
1417#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1418 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1419 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1420 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1421 } \
1422 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1423 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1424 TRIE_LIST_CUR( state )++; \
1425} STMT_END
07be1b83 1426
786e8c11
YO
1427#define TRIE_LIST_NEW(state) STMT_START { \
1428 Newxz( trie->states[ state ].trans.list, \
1429 4, reg_trie_trans_le ); \
1430 TRIE_LIST_CUR( state ) = 1; \
1431 TRIE_LIST_LEN( state ) = 4; \
1432} STMT_END
07be1b83 1433
786e8c11
YO
1434#define TRIE_HANDLE_WORD(state) STMT_START { \
1435 U16 dupe= trie->states[ state ].wordnum; \
1436 regnode * const noper_next = regnext( noper ); \
1437 \
786e8c11
YO
1438 DEBUG_r({ \
1439 /* store the word for dumping */ \
1440 SV* tmp; \
1441 if (OP(noper) != NOTHING) \
740cce10 1442 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1443 else \
740cce10 1444 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1445 av_push( trie_words, tmp ); \
786e8c11
YO
1446 }); \
1447 \
1448 curword++; \
2e64971a
DM
1449 trie->wordinfo[curword].prev = 0; \
1450 trie->wordinfo[curword].len = wordlen; \
1451 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1452 \
1453 if ( noper_next < tail ) { \
1454 if (!trie->jump) \
c944940b 1455 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1456 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1457 if (!jumper) \
1458 jumper = noper_next; \
1459 if (!nextbranch) \
1460 nextbranch= regnext(cur); \
1461 } \
1462 \
1463 if ( dupe ) { \
2e64971a
DM
1464 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1465 /* chain, so that when the bits of chain are later */\
1466 /* linked together, the dups appear in the chain */\
1467 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1468 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1469 } else { \
1470 /* we haven't inserted this word yet. */ \
1471 trie->states[ state ].wordnum = curword; \
1472 } \
1473} STMT_END
07be1b83 1474
3dab1dad 1475
786e8c11
YO
1476#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1477 ( ( base + charid >= ucharcount \
1478 && base + charid < ubound \
1479 && state == trie->trans[ base - ucharcount + charid ].check \
1480 && trie->trans[ base - ucharcount + charid ].next ) \
1481 ? trie->trans[ base - ucharcount + charid ].next \
1482 : ( state==1 ? special : 0 ) \
1483 )
3dab1dad 1484
786e8c11
YO
1485#define MADE_TRIE 1
1486#define MADE_JUMP_TRIE 2
1487#define MADE_EXACT_TRIE 4
3dab1dad 1488
a3621e74 1489STATIC I32
786e8c11 1490S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1491{
27da23d5 1492 dVAR;
a3621e74
YO
1493 /* first pass, loop through and scan words */
1494 reg_trie_data *trie;
55eed653 1495 HV *widecharmap = NULL;
2b8b4781 1496 AV *revcharmap = newAV();
a3621e74 1497 regnode *cur;
9f7f3913 1498 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1499 STRLEN len = 0;
1500 UV uvc = 0;
1501 U16 curword = 0;
1502 U32 next_alloc = 0;
786e8c11
YO
1503 regnode *jumper = NULL;
1504 regnode *nextbranch = NULL;
7f69552c 1505 regnode *convert = NULL;
2e64971a 1506 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1507 /* we just use folder as a flag in utf8 */
1e696034 1508 const U8 * folder = NULL;
a3621e74 1509
2b8b4781
NC
1510#ifdef DEBUGGING
1511 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1512 AV *trie_words = NULL;
1513 /* along with revcharmap, this only used during construction but both are
1514 * useful during debugging so we store them in the struct when debugging.
8e11feef 1515 */
2b8b4781
NC
1516#else
1517 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1518 STRLEN trie_charcount=0;
3dab1dad 1519#endif
2b8b4781 1520 SV *re_trie_maxbuff;
a3621e74 1521 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1522
1523 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1524#ifndef DEBUGGING
1525 PERL_UNUSED_ARG(depth);
1526#endif
a3621e74 1527
1e696034 1528 switch (flags) {
c46d03cf 1529 case EXACT: break;
2f7f8cb1 1530 case EXACTFA:
fab2782b
YO
1531 case EXACTFU_SS:
1532 case EXACTFU_TRICKYFOLD:
1e696034
KW
1533 case EXACTFU: folder = PL_fold_latin1; break;
1534 case EXACTF: folder = PL_fold; break;
1535 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1536 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1537 }
1538
c944940b 1539 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1540 trie->refcount = 1;
3dab1dad 1541 trie->startstate = 1;
786e8c11 1542 trie->wordcount = word_count;
f8fc2ecf 1543 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1544 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1545 if (flags == EXACT)
c944940b 1546 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1547 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1548 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1549
a3621e74 1550 DEBUG_r({
2b8b4781 1551 trie_words = newAV();
a3621e74 1552 });
a3621e74 1553
0111c4fd 1554 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1555 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1556 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1557 }
df826430 1558 DEBUG_TRIE_COMPILE_r({
3dab1dad 1559 PerlIO_printf( Perl_debug_log,
786e8c11 1560 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1561 (int)depth * 2 + 2, "",
1562 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1563 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1564 (int)depth);
3dab1dad 1565 });
7f69552c
YO
1566
1567 /* Find the node we are going to overwrite */
1568 if ( first == startbranch && OP( last ) != BRANCH ) {
1569 /* whole branch chain */
1570 convert = first;
1571 } else {
1572 /* branch sub-chain */
1573 convert = NEXTOPER( first );
1574 }
1575
a3621e74
YO
1576 /* -- First loop and Setup --
1577
1578 We first traverse the branches and scan each word to determine if it
1579 contains widechars, and how many unique chars there are, this is
1580 important as we have to build a table with at least as many columns as we
1581 have unique chars.
1582
1583 We use an array of integers to represent the character codes 0..255
38a44b82 1584 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1585 native representation of the character value as the key and IV's for the
1586 coded index.
1587
1588 *TODO* If we keep track of how many times each character is used we can
1589 remap the columns so that the table compression later on is more
3b753521 1590 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1591 middle and the least common are on the outside. IMO this would be better
1592 than a most to least common mapping as theres a decent chance the most
1593 common letter will share a node with the least common, meaning the node
486ec47a 1594 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1595 case is when we have the least common nodes twice.
1596
1597 */
1598
a3621e74 1599 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1600 regnode *noper = NEXTOPER( cur );
e1ec3a88 1601 const U8 *uc = (U8*)STRING( noper );
df826430 1602 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1603 STRLEN foldlen = 0;
1604 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1605 STRLEN skiplen = 0;
2af232bd 1606 const U8 *scan = (U8*)NULL;
07be1b83 1607 U32 wordlen = 0; /* required init */
02daf0ab
YO
1608 STRLEN chars = 0;
1609 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1610
3dab1dad 1611 if (OP(noper) == NOTHING) {
df826430
YO
1612 regnode *noper_next= regnext(noper);
1613 if (noper_next != tail && OP(noper_next) == flags) {
1614 noper = noper_next;
1615 uc= (U8*)STRING(noper);
1616 e= uc + STR_LEN(noper);
1617 trie->minlen= STR_LEN(noper);
1618 } else {
1619 trie->minlen= 0;
1620 continue;
1621 }
3dab1dad 1622 }
df826430 1623
fab2782b 1624 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1625 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1626 regardless of encoding */
fab2782b
YO
1627 if (OP( noper ) == EXACTFU_SS) {
1628 /* false positives are ok, so just set this */
1629 TRIE_BITMAP_SET(trie,0xDF);
1630 }
1631 }
a3621e74 1632 for ( ; uc < e ; uc += len ) {
3dab1dad 1633 TRIE_CHARCOUNT(trie)++;
a3621e74 1634 TRIE_READ_CHAR;
3dab1dad 1635 chars++;
a3621e74 1636 if ( uvc < 256 ) {
fab2782b
YO
1637 if ( folder ) {
1638 U8 folded= folder[ (U8) uvc ];
1639 if ( !trie->charmap[ folded ] ) {
1640 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1641 TRIE_STORE_REVCHAR( folded );
1642 }
1643 }
a3621e74
YO
1644 if ( !trie->charmap[ uvc ] ) {
1645 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1646 TRIE_STORE_REVCHAR( uvc );
a3621e74 1647 }
02daf0ab 1648 if ( set_bit ) {
62012aee
KW
1649 /* store the codepoint in the bitmap, and its folded
1650 * equivalent. */
fab2782b 1651 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1652
1653 /* store the folded codepoint */
fab2782b 1654 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1655
1656 if ( !UTF ) {
1657 /* store first byte of utf8 representation of
acdf4139
KW
1658 variant codepoints */
1659 if (! UNI_IS_INVARIANT(uvc)) {
1660 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1661 }
1662 }
02daf0ab
YO
1663 set_bit = 0; /* We've done our bit :-) */
1664 }
a3621e74
YO
1665 } else {
1666 SV** svpp;
55eed653
NC
1667 if ( !widecharmap )
1668 widecharmap = newHV();
a3621e74 1669
55eed653 1670 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1671
1672 if ( !svpp )
e4584336 1673 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1674
1675 if ( !SvTRUE( *svpp ) ) {
1676 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1677 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1678 }
1679 }
1680 }
3dab1dad 1681 if( cur == first ) {
fab2782b
YO
1682 trie->minlen = chars;
1683 trie->maxlen = chars;
3dab1dad 1684 } else if (chars < trie->minlen) {
fab2782b 1685 trie->minlen = chars;
3dab1dad 1686 } else if (chars > trie->maxlen) {
fab2782b
YO
1687 trie->maxlen = chars;
1688 }
1689 if (OP( noper ) == EXACTFU_SS) {
1690 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1691 if (trie->minlen > 1)
1692 trie->minlen= 1;
1693 }
1694 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1695 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1696 * - We assume that any such sequence might match a 2 byte string */
1697 if (trie->minlen > 2 )
1698 trie->minlen= 2;
3dab1dad
YO
1699 }
1700
a3621e74
YO
1701 } /* end first pass */
1702 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1703 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1704 (int)depth * 2 + 2,"",
55eed653 1705 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1706 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1707 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1708 );
a3621e74
YO
1709
1710 /*
1711 We now know what we are dealing with in terms of unique chars and
1712 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1713 representation using a flat table will take. If it's over a reasonable
1714 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1715 conservative but potentially much slower representation using an array
1716 of lists.
1717
1718 At the end we convert both representations into the same compressed
1719 form that will be used in regexec.c for matching with. The latter
1720 is a form that cannot be used to construct with but has memory
1721 properties similar to the list form and access properties similar
1722 to the table form making it both suitable for fast searches and
1723 small enough that its feasable to store for the duration of a program.
1724
1725 See the comment in the code where the compressed table is produced
1726 inplace from the flat tabe representation for an explanation of how
1727 the compression works.
1728
1729 */
1730
1731
2e64971a
DM
1732 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1733 prev_states[1] = 0;
1734
3dab1dad 1735 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1736 /*
1737 Second Pass -- Array Of Lists Representation
1738
1739 Each state will be represented by a list of charid:state records
1740 (reg_trie_trans_le) the first such element holds the CUR and LEN
1741 points of the allocated array. (See defines above).
1742
1743 We build the initial structure using the lists, and then convert
1744 it into the compressed table form which allows faster lookups
1745 (but cant be modified once converted).
a3621e74
YO
1746 */
1747
a3621e74
YO
1748 STRLEN transcount = 1;
1749
1e2e3d02
YO
1750 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1751 "%*sCompiling trie using list compiler\n",
1752 (int)depth * 2 + 2, ""));
686b73d4 1753
c944940b
JH
1754 trie->states = (reg_trie_state *)
1755 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1756 sizeof(reg_trie_state) );
a3621e74
YO
1757 TRIE_LIST_NEW(1);
1758 next_alloc = 2;
1759
1760 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1761
df826430 1762 regnode *noper = NEXTOPER( cur );
c445ea15 1763 U8 *uc = (U8*)STRING( noper );
df826430 1764 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1765 U32 state = 1; /* required init */
1766 U16 charid = 0; /* sanity init */
1767 U8 *scan = (U8*)NULL; /* sanity init */
1768 STRLEN foldlen = 0; /* required init */
07be1b83 1769 U32 wordlen = 0; /* required init */
c445ea15 1770 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1771 STRLEN skiplen = 0;
c445ea15 1772
df826430
YO
1773 if (OP(noper) == NOTHING) {
1774 regnode *noper_next= regnext(noper);
1775 if (noper_next != tail && OP(noper_next) == flags) {
1776 noper = noper_next;
1777 uc= (U8*)STRING(noper);
1778 e= uc + STR_LEN(noper);
1779 }
1780 }
1781
3dab1dad 1782 if (OP(noper) != NOTHING) {
786e8c11 1783 for ( ; uc < e ; uc += len ) {
c445ea15 1784
786e8c11 1785 TRIE_READ_CHAR;
c445ea15 1786
786e8c11
YO
1787 if ( uvc < 256 ) {
1788 charid = trie->charmap[ uvc ];
c445ea15 1789 } else {
55eed653 1790 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1791 if ( !svpp ) {
1792 charid = 0;
1793 } else {
1794 charid=(U16)SvIV( *svpp );
1795 }
c445ea15 1796 }
786e8c11
YO
1797 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1798 if ( charid ) {
a3621e74 1799
786e8c11
YO
1800 U16 check;
1801 U32 newstate = 0;
a3621e74 1802
786e8c11
YO
1803 charid--;
1804 if ( !trie->states[ state ].trans.list ) {
1805 TRIE_LIST_NEW( state );
c445ea15 1806 }
786e8c11
YO
1807 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1808 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1809 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1810 break;
1811 }
1812 }
1813 if ( ! newstate ) {
1814 newstate = next_alloc++;
2e64971a 1815 prev_states[newstate] = state;
786e8c11
YO
1816 TRIE_LIST_PUSH( state, charid, newstate );
1817 transcount++;
1818 }
1819 state = newstate;
1820 } else {
1821 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1822 }
a28509cc 1823 }
c445ea15 1824 }
3dab1dad 1825 TRIE_HANDLE_WORD(state);
a3621e74
YO
1826
1827 } /* end second pass */
1828
1e2e3d02
YO
1829 /* next alloc is the NEXT state to be allocated */
1830 trie->statecount = next_alloc;
c944940b
JH
1831 trie->states = (reg_trie_state *)
1832 PerlMemShared_realloc( trie->states,
1833 next_alloc
1834 * sizeof(reg_trie_state) );
a3621e74 1835
3dab1dad 1836 /* and now dump it out before we compress it */
2b8b4781
NC
1837 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1838 revcharmap, next_alloc,
1839 depth+1)
1e2e3d02 1840 );
a3621e74 1841
c944940b
JH
1842 trie->trans = (reg_trie_trans *)
1843 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1844 {
1845 U32 state;
a3621e74
YO
1846 U32 tp = 0;
1847 U32 zp = 0;
1848
1849
1850 for( state=1 ; state < next_alloc ; state ++ ) {
1851 U32 base=0;
1852
1853 /*
1854 DEBUG_TRIE_COMPILE_MORE_r(
1855 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1856 );
1857 */
1858
1859 if (trie->states[state].trans.list) {
1860 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1861 U16 maxid=minid;
a28509cc 1862 U16 idx;
a3621e74
YO
1863
1864 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1865 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1866 if ( forid < minid ) {
1867 minid=forid;
1868 } else if ( forid > maxid ) {
1869 maxid=forid;
1870 }
a3621e74
YO
1871 }
1872 if ( transcount < tp + maxid - minid + 1) {
1873 transcount *= 2;
c944940b
JH
1874 trie->trans = (reg_trie_trans *)
1875 PerlMemShared_realloc( trie->trans,
446bd890
NC
1876 transcount
1877 * sizeof(reg_trie_trans) );
a3621e74
YO
1878 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1879 }
1880 base = trie->uniquecharcount + tp - minid;
1881 if ( maxid == minid ) {
1882 U32 set = 0;
1883 for ( ; zp < tp ; zp++ ) {
1884 if ( ! trie->trans[ zp ].next ) {
1885 base = trie->uniquecharcount + zp - minid;
1886 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1887 trie->trans[ zp ].check = state;
1888 set = 1;
1889 break;
1890 }
1891 }
1892 if ( !set ) {
1893 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1894 trie->trans[ tp ].check = state;
1895 tp++;
1896 zp = tp;
1897 }
1898 } else {
1899 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1900 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1901 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1902 trie->trans[ tid ].check = state;
1903 }
1904 tp += ( maxid - minid + 1 );
1905 }
1906 Safefree(trie->states[ state ].trans.list);
1907 }
1908 /*
1909 DEBUG_TRIE_COMPILE_MORE_r(
1910 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1911 );
1912 */
1913 trie->states[ state ].trans.base=base;
1914 }
cc601c31 1915 trie->lasttrans = tp + 1;
a3621e74
YO
1916 }
1917 } else {
1918 /*
1919 Second Pass -- Flat Table Representation.
1920
1921 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1922 We know that we will need Charcount+1 trans at most to store the data
1923 (one row per char at worst case) So we preallocate both structures
1924 assuming worst case.
1925
1926 We then construct the trie using only the .next slots of the entry
1927 structs.
1928
3b753521 1929 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1930 make compression both faster and easier by keeping track of how many non
1931 zero fields are in the node.
1932
1933 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1934 transition.
1935
1936 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1937 number representing the first entry of the node, and state as a
1938 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1939 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1940 are 2 entrys per node. eg:
1941
1942 A B A B
1943 1. 2 4 1. 3 7
1944 2. 0 3 3. 0 5
1945 3. 0 0 5. 0 0
1946 4. 0 0 7. 0 0
1947
1948 The table is internally in the right hand, idx form. However as we also
1949 have to deal with the states array which is indexed by nodenum we have to
1950 use TRIE_NODENUM() to convert.
1951
1952 */
1e2e3d02
YO
1953 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1954 "%*sCompiling trie using table compiler\n",
1955 (int)depth * 2 + 2, ""));
3dab1dad 1956
c944940b
JH
1957 trie->trans = (reg_trie_trans *)
1958 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1959 * trie->uniquecharcount + 1,
1960 sizeof(reg_trie_trans) );
1961 trie->states = (reg_trie_state *)
1962 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1963 sizeof(reg_trie_state) );
a3621e74
YO
1964 next_alloc = trie->uniquecharcount + 1;
1965
3dab1dad 1966
a3621e74
YO
1967 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1968
df826430 1969 regnode *noper = NEXTOPER( cur );
a28509cc 1970 const U8 *uc = (U8*)STRING( noper );
df826430 1971 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1972
1973 U32 state = 1; /* required init */
1974
1975 U16 charid = 0; /* sanity init */
1976 U32 accept_state = 0; /* sanity init */
1977 U8 *scan = (U8*)NULL; /* sanity init */
1978
1979 STRLEN foldlen = 0; /* required init */
07be1b83 1980 U32 wordlen = 0; /* required init */
fab2782b 1981 STRLEN skiplen = 0;
a3621e74
YO
1982 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1983
df826430
YO
1984 if (OP(noper) == NOTHING) {
1985 regnode *noper_next= regnext(noper);
1986 if (noper_next != tail && OP(noper_next) == flags) {
1987 noper = noper_next;
1988 uc= (U8*)STRING(noper);
1989 e= uc + STR_LEN(noper);
1990 }
1991 }
fab2782b 1992
3dab1dad 1993 if ( OP(noper) != NOTHING ) {
786e8c11 1994 for ( ; uc < e ; uc += len ) {
a3621e74 1995
786e8c11 1996 TRIE_READ_CHAR;
a3621e74 1997
786e8c11
YO
1998 if ( uvc < 256 ) {
1999 charid = trie->charmap[ uvc ];
2000 } else {
55eed653 2001 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2002 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2003 }
786e8c11
YO
2004 if ( charid ) {
2005 charid--;
2006 if ( !trie->trans[ state + charid ].next ) {
2007 trie->trans[ state + charid ].next = next_alloc;
2008 trie->trans[ state ].check++;
2e64971a
DM
2009 prev_states[TRIE_NODENUM(next_alloc)]
2010 = TRIE_NODENUM(state);
786e8c11
YO
2011 next_alloc += trie->uniquecharcount;
2012 }
2013 state = trie->trans[ state + charid ].next;
2014 } else {
2015 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2016 }
2017 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2018 }
a3621e74 2019 }
3dab1dad
YO
2020 accept_state = TRIE_NODENUM( state );
2021 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2022
2023 } /* end second pass */
2024
3dab1dad 2025 /* and now dump it out before we compress it */
2b8b4781
NC
2026 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2027 revcharmap,
2028 next_alloc, depth+1));
a3621e74 2029
a3621e74
YO
2030 {
2031 /*
2032 * Inplace compress the table.*
2033
2034 For sparse data sets the table constructed by the trie algorithm will
2035 be mostly 0/FAIL transitions or to put it another way mostly empty.
2036 (Note that leaf nodes will not contain any transitions.)
2037
2038 This algorithm compresses the tables by eliminating most such
2039 transitions, at the cost of a modest bit of extra work during lookup:
2040
2041 - Each states[] entry contains a .base field which indicates the
2042 index in the state[] array wheres its transition data is stored.
2043
3b753521 2044 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2045
2046 - If .base is nonzero then charid is added to it to find an entry in
2047 the trans array.
2048
2049 -If trans[states[state].base+charid].check!=state then the
2050 transition is taken to be a 0/Fail transition. Thus if there are fail
2051 transitions at the front of the node then the .base offset will point
2052 somewhere inside the previous nodes data (or maybe even into a node
2053 even earlier), but the .check field determines if the transition is
2054 valid.
2055
786e8c11 2056 XXX - wrong maybe?
a3621e74 2057 The following process inplace converts the table to the compressed
3b753521 2058 table: We first do not compress the root node 1,and mark all its
a3621e74 2059 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2060 allows us to do a DFA construction from the compressed table later,
2061 and ensures that any .base pointers we calculate later are greater
2062 than 0.
a3621e74
YO
2063
2064 - We set 'pos' to indicate the first entry of the second node.
2065
2066 - We then iterate over the columns of the node, finding the first and
2067 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2068 and set the .check pointers accordingly, and advance pos
2069 appropriately and repreat for the next node. Note that when we copy
2070 the next pointers we have to convert them from the original
2071 NODEIDX form to NODENUM form as the former is not valid post
2072 compression.
2073
2074 - If a node has no transitions used we mark its base as 0 and do not
2075 advance the pos pointer.
2076
2077 - If a node only has one transition we use a second pointer into the
2078 structure to fill in allocated fail transitions from other states.
2079 This pointer is independent of the main pointer and scans forward
2080 looking for null transitions that are allocated to a state. When it
2081 finds one it writes the single transition into the "hole". If the
786e8c11 2082 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2083
2084 - Once compressed we can Renew/realloc the structures to release the
2085 excess space.
2086
2087 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2088 specifically Fig 3.47 and the associated pseudocode.
2089
2090 demq
2091 */
a3b680e6 2092 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2093 U32 state, charid;
a3621e74 2094 U32 pos = 0, zp=0;
1e2e3d02 2095 trie->statecount = laststate;
a3621e74
YO
2096
2097 for ( state = 1 ; state < laststate ; state++ ) {
2098 U8 flag = 0;
a28509cc
AL
2099 const U32 stateidx = TRIE_NODEIDX( state );
2100 const U32 o_used = trie->trans[ stateidx ].check;
2101 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2102 trie->trans[ stateidx ].check = 0;
2103
2104 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2105 if ( flag || trie->trans[ stateidx + charid ].next ) {
2106 if ( trie->trans[ stateidx + charid ].next ) {
2107 if (o_used == 1) {
2108 for ( ; zp < pos ; zp++ ) {
2109 if ( ! trie->trans[ zp ].next ) {
2110 break;
2111 }
2112 }
2113 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2114 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2115 trie->trans[ zp ].check = state;
2116 if ( ++zp > pos ) pos = zp;
2117 break;
2118 }
2119 used--;
2120 }
2121 if ( !flag ) {
2122 flag = 1;
2123 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2124 }
2125 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2126 trie->trans[ pos ].check = state;
2127 pos++;
2128 }
2129 }
2130 }
cc601c31 2131 trie->lasttrans = pos + 1;
c944940b
JH
2132 trie->states = (reg_trie_state *)
2133 PerlMemShared_realloc( trie->states, laststate
2134 * sizeof(reg_trie_state) );
a3621e74 2135 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2136 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2137 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2138 (int)depth * 2 + 2,"",
2139 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2140 (IV)next_alloc,
2141 (IV)pos,
a3621e74
YO
2142 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2143 );
2144
2145 } /* end table compress */
2146 }
1e2e3d02
YO
2147 DEBUG_TRIE_COMPILE_MORE_r(
2148 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2149 (int)depth * 2 + 2, "",
2150 (UV)trie->statecount,
2151 (UV)trie->lasttrans)
2152 );
cc601c31 2153 /* resize the trans array to remove unused space */
c944940b
JH
2154 trie->trans = (reg_trie_trans *)
2155 PerlMemShared_realloc( trie->trans, trie->lasttrans
2156 * sizeof(reg_trie_trans) );
a3621e74 2157
3b753521 2158 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2159 U8 nodetype =(U8)(flags & 0xFF);
2160 char *str=NULL;
786e8c11 2161
07be1b83 2162#ifdef DEBUGGING
e62cc96a 2163 regnode *optimize = NULL;
7122b237
YO
2164#ifdef RE_TRACK_PATTERN_OFFSETS
2165
b57a0404
JH
2166 U32 mjd_offset = 0;
2167 U32 mjd_nodelen = 0;
7122b237
YO
2168#endif /* RE_TRACK_PATTERN_OFFSETS */
2169#endif /* DEBUGGING */
a3621e74 2170 /*
3dab1dad
YO
2171 This means we convert either the first branch or the first Exact,
2172 depending on whether the thing following (in 'last') is a branch
2173 or not and whther first is the startbranch (ie is it a sub part of
2174 the alternation or is it the whole thing.)
3b753521 2175 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2176 the whole branch sequence, including the first.
a3621e74 2177 */
3dab1dad 2178 /* Find the node we are going to overwrite */
7f69552c 2179 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2180 /* branch sub-chain */
3dab1dad 2181 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2182#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2183 DEBUG_r({
2184 mjd_offset= Node_Offset((convert));
2185 mjd_nodelen= Node_Length((convert));
2186 });
7122b237 2187#endif
7f69552c 2188 /* whole branch chain */
7122b237
YO
2189 }
2190#ifdef RE_TRACK_PATTERN_OFFSETS
2191 else {
7f69552c
YO
2192 DEBUG_r({
2193 const regnode *nop = NEXTOPER( convert );
2194 mjd_offset= Node_Offset((nop));
2195 mjd_nodelen= Node_Length((nop));
2196 });
07be1b83
YO
2197 }
2198 DEBUG_OPTIMISE_r(
2199 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2200 (int)depth * 2 + 2, "",
786e8c11 2201 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2202 );
7122b237 2203#endif
3dab1dad
YO
2204 /* But first we check to see if there is a common prefix we can
2205 split out as an EXACT and put in front of the TRIE node. */
2206 trie->startstate= 1;
55eed653 2207 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2208 U32 state;
1e2e3d02 2209 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2210 U32 ofs = 0;
8e11feef
RGS
2211 I32 idx = -1;
2212 U32 count = 0;
2213 const U32 base = trie->states[ state ].trans.base;
a3621e74 2214
3dab1dad 2215 if ( trie->states[state].wordnum )
8e11feef 2216 count = 1;
a3621e74 2217
8e11feef 2218 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2219 if ( ( base + ofs >= trie->uniquecharcount ) &&
2220 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2221 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2222 {
3dab1dad 2223 if ( ++count > 1 ) {
2b8b4781 2224 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2225 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2226 if ( state == 1 ) break;
3dab1dad
YO
2227 if ( count == 2 ) {
2228 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2229 DEBUG_OPTIMISE_r(
8e11feef
RGS
2230 PerlIO_printf(Perl_debug_log,
2231 "%*sNew Start State=%"UVuf" Class: [",
2232 (int)depth * 2 + 2, "",
786e8c11 2233 (UV)state));
be8e71aa 2234 if (idx >= 0) {
2b8b4781 2235 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2236 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2237
3dab1dad 2238 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2239 if ( folder )
2240 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2241 DEBUG_OPTIMISE_r(
f1f66076 2242 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2243 );
8e11feef
RGS
2244 }
2245 }
2246 TRIE_BITMAP_SET(trie,*ch);
2247 if ( folder )
2248 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2249 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2250 }
2251 idx = ofs;
2252 }
3dab1dad
YO
2253 }
2254 if ( count == 1 ) {
2b8b4781 2255 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2256 STRLEN len;
2257 char *ch = SvPV( *tmp, len );
de734bd5
A
2258 DEBUG_OPTIMISE_r({
2259 SV *sv=sv_newmortal();
8e11feef
RGS
2260 PerlIO_printf( Perl_debug_log,
2261 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2262 (int)depth * 2 + 2, "",
de734bd5
A
2263 (UV)state, (UV)idx,
2264 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2265 PL_colors[0], PL_colors[1],
2266 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2267 PERL_PV_ESCAPE_FIRSTCHAR
2268 )
2269 );
2270 });
3dab1dad
YO
2271 if ( state==1 ) {
2272 OP( convert ) = nodetype;
2273 str=STRING(convert);
2274 STR_LEN(convert)=0;
2275 }
c490c714
YO
2276 STR_LEN(convert) += len;
2277 while (len--)
de734bd5 2278 *str++ = *ch++;
8e11feef 2279 } else {
f9049ba1 2280#ifdef DEBUGGING
8e11feef
RGS
2281 if (state>1)
2282 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2283#endif
8e11feef
RGS
2284 break;
2285 }
2286 }
2e64971a 2287 trie->prefixlen = (state-1);
3dab1dad 2288 if (str) {
8e11feef 2289 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2290 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2291 trie->startstate = state;
07be1b83
YO
2292 trie->minlen -= (state - 1);
2293 trie->maxlen -= (state - 1);
33809eae
JH
2294#ifdef DEBUGGING
2295 /* At least the UNICOS C compiler choked on this
2296 * being argument to DEBUG_r(), so let's just have
2297 * it right here. */
2298 if (
2299#ifdef PERL_EXT_RE_BUILD
2300 1
2301#else
2302 DEBUG_r_TEST
2303#endif
2304 ) {
2305 regnode *fix = convert;
2306 U32 word = trie->wordcount;
2307 mjd_nodelen++;
2308 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2309 while( ++fix < n ) {
2310 Set_Node_Offset_Length(fix, 0, 0);
2311 }
2312 while (word--) {
2313 SV ** const tmp = av_fetch( trie_words, word, 0 );
2314 if (tmp) {
2315 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2316 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2317 else
2318 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2319 }
2320 }
2321 }
2322#endif
8e11feef
RGS
2323 if (trie->maxlen) {
2324 convert = n;
2325 } else {
3dab1dad 2326 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2327 DEBUG_r(optimize= n);
3dab1dad
YO
2328 }
2329 }
2330 }
a5ca303d
YO
2331 if (!jumper)
2332 jumper = last;
3dab1dad 2333 if ( trie->maxlen ) {
8e11feef
RGS
2334 NEXT_OFF( convert ) = (U16)(tail - convert);
2335 ARG_SET( convert, data_slot );
786e8c11
YO
2336 /* Store the offset to the first unabsorbed branch in
2337 jump[0], which is otherwise unused by the jump logic.
2338 We use this when dumping a trie and during optimisation. */
2339 if (trie->jump)
7f69552c 2340 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2341
6c48061a
YO
2342 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2343 * and there is a bitmap
2344 * and the first "jump target" node we found leaves enough room
2345 * then convert the TRIE node into a TRIEC node, with the bitmap
2346 * embedded inline in the opcode - this is hypothetically faster.
2347 */
2348 if ( !trie->states[trie->startstate].wordnum
2349 && trie->bitmap
2350 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2351 {
2352 OP( convert ) = TRIEC;
2353 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2354 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2355 trie->bitmap= NULL;
2356 } else
2357 OP( convert ) = TRIE;
a3621e74 2358
3dab1dad
YO
2359 /* store the type in the flags */
2360 convert->flags = nodetype;
a5ca303d
YO
2361 DEBUG_r({
2362 optimize = convert
2363 + NODE_STEP_REGNODE
2364 + regarglen[ OP( convert ) ];
2365 });
2366 /* XXX We really should free up the resource in trie now,
2367 as we won't use them - (which resources?) dmq */
3dab1dad 2368 }
a3621e74 2369 /* needed for dumping*/
e62cc96a 2370 DEBUG_r(if (optimize) {
07be1b83 2371 regnode *opt = convert;
bcdf7404 2372
e62cc96a 2373 while ( ++opt < optimize) {
07be1b83
YO
2374 Set_Node_Offset_Length(opt,0,0);
2375 }
786e8c11
YO
2376 /*
2377 Try to clean up some of the debris left after the
2378 optimisation.
a3621e74 2379 */
786e8c11 2380 while( optimize < jumper ) {
07be1b83 2381 mjd_nodelen += Node_Length((optimize));
a3621e74 2382 OP( optimize ) = OPTIMIZED;
07be1b83 2383 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2384 optimize++;
2385 }
07be1b83 2386 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2387 });
2388 } /* end node insert */
2e64971a
DM
2389
2390 /* Finish populating the prev field of the wordinfo array. Walk back
2391 * from each accept state until we find another accept state, and if
2392 * so, point the first word's .prev field at the second word. If the
2393 * second already has a .prev field set, stop now. This will be the
2394 * case either if we've already processed that word's accept state,
3b753521
FN
2395 * or that state had multiple words, and the overspill words were
2396 * already linked up earlier.
2e64971a
DM
2397 */
2398 {
2399 U16 word;
2400 U32 state;
2401 U16 prev;
2402
2403 for (word=1; word <= trie->wordcount; word++) {
2404 prev = 0;
2405 if (trie->wordinfo[word].prev)
2406 continue;
2407 state = trie->wordinfo[word].accept;
2408 while (state) {
2409 state = prev_states[state];
2410 if (!state)
2411 break;
2412 prev = trie->states[state].wordnum;
2413 if (prev)
2414 break;
2415 }
2416 trie->wordinfo[word].prev = prev;
2417 }
2418 Safefree(prev_states);
2419 }
2420
2421
2422 /* and now dump out the compressed format */
2423 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2424
55eed653 2425 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2426#ifdef DEBUGGING
2427 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2428 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2429#else
2430 SvREFCNT_dec(revcharmap);
07be1b83 2431#endif
786e8c11
YO
2432 return trie->jump
2433 ? MADE_JUMP_TRIE
2434 : trie->startstate>1
2435 ? MADE_EXACT_TRIE
2436 : MADE_TRIE;
2437}
2438
2439STATIC void
2440S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2441{
3b753521 2442/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2443
2444 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2445 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2446 ISBN 0-201-10088-6
2447
2448 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2449 suffix of the current state's 'word' that is also a proper prefix of another word in our
2450 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2451 the DFA not to have to restart after its tried and failed a word at a given point, it
2452 simply continues as though it had been matching the other word in the first place.
2453 Consider
2454 'abcdgu'=~/abcdefg|cdgu/
2455 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2456 fail, which would bring us to the state representing 'd' in the second word where we would
2457 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2458 */
2459 /* add a fail transition */
3251b653
NC
2460 const U32 trie_offset = ARG(source);
2461 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2462 U32 *q;
2463 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2464 const U32 numstates = trie->statecount;
786e8c11
YO
2465 const U32 ubound = trie->lasttrans + ucharcount;
2466 U32 q_read = 0;
2467 U32 q_write = 0;
2468 U32 charid;
2469 U32 base = trie->states[ 1 ].trans.base;
2470 U32 *fail;
2471 reg_ac_data *aho;
2472 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2473 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2474
2475 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2476#ifndef DEBUGGING
2477 PERL_UNUSED_ARG(depth);
2478#endif
2479
2480
2481 ARG_SET( stclass, data_slot );
c944940b 2482 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2483 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2484 aho->trie=trie_offset;
446bd890
NC
2485 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2486 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2487 Newxz( q, numstates, U32);
c944940b 2488 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2489 aho->refcount = 1;
2490 fail = aho->fail;
2491 /* initialize fail[0..1] to be 1 so that we always have
2492 a valid final fail state */
2493 fail[ 0 ] = fail[ 1 ] = 1;
2494
2495 for ( charid = 0; charid < ucharcount ; charid++ ) {
2496 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2497 if ( newstate ) {
2498 q[ q_write ] = newstate;
2499 /* set to point at the root */
2500 fail[ q[ q_write++ ] ]=1;
2501 }
2502 }
2503 while ( q_read < q_write) {
2504 const U32 cur = q[ q_read++ % numstates ];
2505 base = trie->states[ cur ].trans.base;
2506
2507 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2508 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2509 if (ch_state) {
2510 U32 fail_state = cur;
2511 U32 fail_base;
2512 do {
2513 fail_state = fail[ fail_state ];
2514 fail_base = aho->states[ fail_state ].trans.base;
2515 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2516
2517 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2518 fail[ ch_state ] = fail_state;
2519 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2520 {
2521 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2522 }
2523 q[ q_write++ % numstates] = ch_state;
2524 }
2525 }
2526 }
2527 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2528 when we fail in state 1, this allows us to use the
2529 charclass scan to find a valid start char. This is based on the principle
2530 that theres a good chance the string being searched contains lots of stuff
2531 that cant be a start char.
2532 */
2533 fail[ 0 ] = fail[ 1 ] = 0;
2534 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2535 PerlIO_printf(Perl_debug_log,
2536 "%*sStclass Failtable (%"UVuf" states): 0",
2537 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2538 );
786e8c11
YO
2539 for( q_read=1; q_read<numstates; q_read++ ) {
2540 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2541 }
2542 PerlIO_printf(Perl_debug_log, "\n");
2543 });
2544 Safefree(q);
2545 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2546}
2547
786e8c11 2548
a3621e74 2549/*
5d1c421c
JH
2550 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2551 * These need to be revisited when a newer toolchain becomes available.
2552 */
2553#if defined(__sparc64__) && defined(__GNUC__)
2554# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2555# undef SPARC64_GCC_WORKAROUND
2556# define SPARC64_GCC_WORKAROUND 1
2557# endif
2558#endif
2559
07be1b83 2560#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2561 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2562 SV * const mysv=sv_newmortal(); \
2563 regnode *Next = regnext(scan); \
2564 regprop(RExC_rx, mysv, scan); \
7f69552c 2565 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2566 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2567 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2568 }});
07be1b83 2569
1de06328 2570
bb914485
KW
2571/* The below joins as many adjacent EXACTish nodes as possible into a single
2572 * one, and looks for problematic sequences of characters whose folds vs.
2573 * non-folds have sufficiently different lengths, that the optimizer would be
2574 * fooled into rejecting legitimate matches of them, and the trie construction
2575 * code can't cope with them. The joining is only done if:
2576 * 1) there is room in the current conglomerated node to entirely contain the
2577 * next one.
2578 * 2) they are the exact same node type
2579 *
2580 * The adjacent nodes actually may be separated by NOTHING kind nodes, and
2581 * these get optimized out
2582 *
9d071ca8
KW
2583 * If there are problematic code sequences, *min_subtract is set to the delta
2584 * that the minimum size of the node can be less than its actual size. And,
2585 * the node type of the result is changed to reflect that it contains these
bb914485
KW
2586 * sequences.
2587 *
a0c4c608
KW
2588 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2589 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2590 *
bb914485
KW
2591 * This is as good a place as any to discuss the design of handling these
2592 * problematic sequences. It's been wrong in Perl for a very long time. There
2593 * are three code points in Unicode whose folded lengths differ so much from
2594 * the un-folded lengths that it causes problems for the optimizer and trie
2595 * construction. Why only these are problematic, and not others where lengths
2596 * also differ is something I (khw) do not understand. New versions of Unicode
2597 * might add more such code points. Hopefully the logic in fold_grind.t that
287722f3 2598 * figures out what to test (in part by verifying that each size-combination
bb914485 2599 * gets tested) will catch any that do come along, so they can be added to the
287722f3
KW
2600 * special handling below. The chances of new ones are actually rather small,
2601 * as most, if not all, of the world's scripts that have casefolding have
2602 * already been encoded by Unicode. Also, a number of Unicode's decisions were
2603 * made to allow compatibility with pre-existing standards, and almost all of
2604 * those have already been dealt with. These would otherwise be the most
2605 * likely candidates for generating further tricky sequences. In other words,
2606 * Unicode by itself is unlikely to add new ones unless it is for compatibility
a0c4c608 2607 * with pre-existing standards, and there aren't many of those left.
bb914485
KW
2608 *
2609 * The previous designs for dealing with these involved assigning a special
2610 * node for them. This approach doesn't work, as evidenced by this example:
a0c4c608 2611 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
bb914485
KW
2612 * Both these fold to "sss", but if the pattern is parsed to create a node of
2613 * that would match just the \xDF, it won't be able to handle the case where a
2614 * successful match would have to cross the node's boundary. The new approach
2615 * that hopefully generally solves the problem generates an EXACTFU_SS node
2616 * that is "sss".
2617 *
2618 * There are a number of components to the approach (a lot of work for just
2619 * three code points!):
2620 * 1) This routine examines each EXACTFish node that could contain the
9d071ca8
KW
2621 * problematic sequences. It returns in *min_subtract how much to
2622 * subtract from the the actual length of the string to get a real minimum
2623 * for one that could match it. This number is usually 0 except for the
2624 * problematic sequences. This delta is used by the caller to adjust the
2625 * min length of the match, and the delta between min and max, so that the
2626 * optimizer doesn't reject these possibilities based on size constraints.
bb914485
KW
2627 * 2) These sequences are not currently correctly handled by the trie code
2628 * either, so it changes the joined node type to ops that are not handled
fab2782b 2629 * by trie's, those new ops being EXACTFU_SS and EXACTFU_TRICKYFOLD.
bb914485
KW
2630 * 3) This is sufficient for the two Greek sequences (described below), but
2631 * the one involving the Sharp s (\xDF) needs more. The node type
2632 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2633 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2634 * case where there is a possible fold length change. That means that a
2635 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2636 * itself with length changes, and so can be processed faster. regexec.c
2637 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2638 * is pre-folded by regcomp.c. This saves effort in regex matching.
2639 * However, probably mostly for historical reasons, the pre-folding isn't
a0c4c608
KW
2640 * done for non-UTF8 patterns (and it can't be for EXACTF and EXACTFL
2641 * nodes, as what they fold to isn't known until runtime.) The fold
2642 * possibilities for the non-UTF8 patterns are quite simple, except for
2643 * the sharp s. All the ones that don't involve a UTF-8 target string
2644 * are members of a fold-pair, and arrays are set up for all of them
2645 * that quickly find the other member of the pair. It might actually
2646 * be faster to pre-fold these, but it isn't currently done, except for
2647 * the sharp s. Code elsewhere in this file makes sure that it gets
2648 * folded to 'ss', even if the pattern isn't UTF-8. This avoids the
2649 * issues described in the next item.
bb914485
KW
2650 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2651 * 'ss' or not is not knowable at compile time. It will match iff the
2652 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2653 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2654 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2655 * described in item 3). An assumption that the optimizer part of
2656 * regexec.c (probably unwittingly) makes is that a character in the
2657 * pattern corresponds to at most a single character in the target string.
2658 * (And I do mean character, and not byte here, unlike other parts of the
2659 * documentation that have never been updated to account for multibyte
2660 * Unicode.) This assumption is wrong only in this case, as all other
2661 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2662 * virtue of having this file pre-fold UTF-8 patterns. I'm
2663 * reluctant to try to change this assumption, so instead the code punts.
9d071ca8
KW
2664 * This routine examines EXACTF nodes for the sharp s, and returns a
2665 * boolean indicating whether or not the node is an EXACTF node that
2666 * contains a sharp s. When it is true, the caller sets a flag that later
2667 * causes the optimizer in this file to not set values for the floating
2668 * and fixed string lengths, and thus avoids the optimizer code in
2669 * regexec.c that makes the invalid assumption. Thus, there is no
2670 * optimization based on string lengths for EXACTF nodes that contain the
2671 * sharp s. This only happens for /id rules (which means the pattern
2672 * isn't in UTF-8).
bb914485 2673 */
1de06328 2674
9d071ca8 2675#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2676 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2677 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2678
be8e71aa 2679STATIC U32
9d071ca8 2680S_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
2681 /* Merge several consecutive EXACTish nodes into one. */
2682 regnode *n = regnext(scan);
2683 U32 stringok = 1;
2684 regnode *next = scan + NODE_SZ_STR(scan);
2685 U32 merged = 0;
2686 U32 stopnow = 0;
2687#ifdef DEBUGGING
2688 regnode *stop = scan;
72f13be8 2689 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2690#else
d47053eb
RGS
2691 PERL_UNUSED_ARG(depth);
2692#endif
7918f24d
NC
2693
2694 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2695#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2696 PERL_UNUSED_ARG(flags);
2697 PERL_UNUSED_ARG(val);
07be1b83 2698#endif
07be1b83 2699 DEBUG_PEEP("join",scan,depth);
bb914485 2700
3f410cf6
KW
2701 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2702 * EXACT ones that are mergeable to the current one. */
2703 while (n
2704 && (PL_regkind[OP(n)] == NOTHING
2705 || (stringok && OP(n) == OP(scan)))
07be1b83 2706 && NEXT_OFF(n)
3f410cf6
KW
2707 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2708 {
07be1b83
YO
2709
2710 if (OP(n) == TAIL || n > next)
2711 stringok = 0;
2712 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2713 DEBUG_PEEP("skip:",n,depth);
2714 NEXT_OFF(scan) += NEXT_OFF(n);
2715 next = n + NODE_STEP_REGNODE;
2716#ifdef DEBUGGING
2717 if (stringok)
2718 stop = n;
2719#endif
2720 n = regnext(n);
2721 }
2722 else if (stringok) {
786e8c11 2723 const unsigned int oldl = STR_LEN(scan);
07be1b83 2724 regnode * const nnext = regnext(n);
b2230d39
KW
2725
2726 if (oldl + STR_LEN(n) > U8_MAX)
2727 break;
07be1b83
YO
2728
2729 DEBUG_PEEP("merg",n,depth);
07be1b83 2730 merged++;
b2230d39 2731
07be1b83
YO
2732 NEXT_OFF(scan) += NEXT_OFF(n);
2733 STR_LEN(scan) += STR_LEN(n);
2734 next = n + NODE_SZ_STR(n);
2735 /* Now we can overwrite *n : */
2736 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2737#ifdef DEBUGGING
2738 stop = next - 1;
2739#endif
2740 n = nnext;
2741 if (stopnow) break;
2742 }
2743
d47053eb
RGS
2744#ifdef EXPERIMENTAL_INPLACESCAN
2745 if (flags && !NEXT_OFF(n)) {
2746 DEBUG_PEEP("atch", val, depth);
2747 if (reg_off_by_arg[OP(n)]) {
2748 ARG_SET(n, val - n);
2749 }
2750 else {
2751 NEXT_OFF(n) = val - n;
2752 }
2753 stopnow = 1;
2754 }
07be1b83
YO
2755#endif
2756 }
2c2b7f86 2757
9d071ca8 2758 *min_subtract = 0;
f758bddf 2759 *has_exactf_sharp_s = FALSE;
f646642f 2760
3f410cf6
KW
2761 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2762 * can now analyze for sequences of problematic code points. (Prior to
2763 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2764 * hence missed). The sequences only happen in folding, hence for any
2765 * non-EXACT EXACTish node */
86d6fcad 2766 if (OP(scan) != EXACT) {
f758bddf
KW
2767 U8 *s;
2768 U8 * s0 = (U8*) STRING(scan);
2769 U8 * const s_end = s0 + STR_LEN(scan);
2770
2771 /* The below is perhaps overboard, but this allows us to save a test
2772 * each time through the loop at the expense of a mask. This is
2773 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2774 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2775 * This uses an exclusive 'or' to find that bit and then inverts it to
2776 * form a mask, with just a single 0, in the bit position where 'S' and
2777 * 's' differ. */
dbeb8947 2778 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
f758bddf
KW
2779 const U8 s_masked = 's' & S_or_s_mask;
2780
2781 /* One pass is made over the node's string looking for all the
2782 * possibilities. to avoid some tests in the loop, there are two main
2783 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2784 * non-UTF-8 */
2785 if (UTF) {
86d6fcad 2786
f758bddf
KW
2787 /* There are two problematic Greek code points in Unicode
2788 * casefolding
86d6fcad
KW
2789 *
2790 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2791 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2792 *
2793 * which casefold to
2794 *
2795 * Unicode UTF-8
2796 *
2797 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2798 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2799 *
2800 * This means that in case-insensitive matching (or "loose
2801 * matching", as Unicode calls it), an EXACTF of length six (the
2802 * UTF-8 encoded byte length of the above casefolded versions) can
2803 * match a target string of length two (the byte length of UTF-8
2804 * encoded U+0390 or U+03B0). This would rather mess up the
2805 * minimum length computation. (there are other code points that
2806 * also fold to these two sequences, but the delta is smaller)
2807 *
f758bddf
KW
2808 * If these sequences are found, the minimum length is decreased by
2809 * four (six minus two).
86d6fcad 2810 *
f758bddf
KW
2811 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2812 * LETTER SHARP S. We decrease the min length by 1 for each
2813 * occurrence of 'ss' found */
3f410cf6 2814
e294cc5d 2815#ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
f758bddf
KW
2816# define U390_first_byte 0xb4
2817 const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
2818# define U3B0_first_byte 0xb5
2819 const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
e294cc5d 2820#else
f758bddf
KW
2821# define U390_first_byte 0xce
2822 const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
2823# define U3B0_first_byte 0xcf
2824 const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
e294cc5d 2825#endif
f758bddf
KW
2826 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2827 yields a net of 0 */
2828 /* Examine the string for one of the problematic sequences */
2829 for (s = s0;
2830 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2831 * sequence we are looking for is 2 */
2832 s += UTF8SKIP(s))
86d6fcad 2833 {
bb914485 2834
f758bddf
KW
2835 /* Look for the first byte in each problematic sequence */
2836 switch (*s) {
2837 /* We don't have to worry about other things that fold to
2838 * 's' (such as the long s, U+017F), as all above-latin1
2839 * code points have been pre-folded */
2840 case 's':
2841 case 'S':
2842
a0c4c608
KW
2843 /* Current character is an 's' or 'S'. If next one is
2844 * as well, we have the dreaded sequence */
f758bddf
KW
2845 if (((*(s+1) & S_or_s_mask) == s_masked)
2846 /* These two node types don't have special handling
2847 * for 'ss' */
2848 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2849 {
9d071ca8 2850 *min_subtract += 1;
f758bddf
KW
2851 OP(scan) = EXACTFU_SS;
2852 s++; /* No need to look at this character again */
2853 }
2854 break;
2855
2856 case U390_first_byte:
2857 if (s_end - s >= len
2858
2859 /* The 1's are because are skipping comparing the
2860 * first byte */
2861 && memEQ(s + 1, U390_tail, len - 1))
2862 {
2863 goto greek_sequence;
2864 }
2865 break;
2866
2867 case U3B0_first_byte:
2868 if (! (s_end - s >= len
2869 && memEQ(s + 1, U3B0_tail, len - 1)))
2870 {
2871 break;
2872 }
2873 greek_sequence:
9d071ca8 2874 *min_subtract += 4;
f758bddf
KW
2875
2876 /* This can't currently be handled by trie's, so change
2877 * the node type to indicate this. If EXACTFA and
2878 * EXACTFL were ever to be handled by trie's, this
2879 * would have to be changed. If this node has already
2880 * been changed to EXACTFU_SS in this loop, leave it as
2881 * is. (I (khw) think it doesn't matter in regexec.c
2882 * for UTF patterns, but no need to change it */
2883 if (OP(scan) == EXACTFU) {
fab2782b 2884 OP(scan) = EXACTFU_TRICKYFOLD;
f758bddf
KW
2885 }
2886 s += 6; /* We already know what this sequence is. Skip
2887 the rest of it */
2888 break;
bb914485
KW
2889 }
2890 }
2891 }
f758bddf 2892 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
bb914485 2893
f758bddf
KW
2894 /* Here, the pattern is not UTF-8. We need to look only for the
2895 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2896 * in the final position. Otherwise we can stop looking 1 byte
2897 * earlier because have to find both the first and second 's' */
2898 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2899
2900 for (s = s0; s < upper; s++) {
2901 switch (*s) {
2902 case 'S':
2903 case 's':
2904 if (s_end - s > 1
2905 && ((*(s+1) & S_or_s_mask) == s_masked))
2906 {
9d071ca8 2907 *min_subtract += 1;
f758bddf
KW
2908
2909 /* EXACTF nodes need to know that the minimum
2910 * length changed so that a sharp s in the string
2911 * can match this ss in the pattern, but they
2912 * remain EXACTF nodes, as they are not trie'able,
2913 * so don't have to invent a new node type to
2914 * exclude them from the trie code */
2915 if (OP(scan) != EXACTF) {
2916 OP(scan) = EXACTFU_SS;
2917 }
2918 s++;
2919 }
2920 break;
2921 case LATIN_SMALL_LETTER_SHARP_S:
2922 if (OP(scan) == EXACTF) {
2923 *has_exactf_sharp_s = TRUE;
2924 }
2925 break;
86d6fcad
KW
2926 }
2927 }
2928 }
07be1b83 2929 }
3f410cf6 2930
07be1b83 2931#ifdef DEBUGGING
bb789b09
DM
2932 /* Allow dumping but overwriting the collection of skipped
2933 * ops and/or strings with fake optimized ops */
07be1b83
YO
2934 n = scan + NODE_SZ_STR(scan);
2935 while (n <= stop) {
bb789b09
DM
2936 OP(n) = OPTIMIZED;
2937 FLAGS(n) = 0;
2938 NEXT_OFF(n) = 0;
07be1b83
YO
2939 n++;
2940 }
2941#endif
2942 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2943 return stopnow;
2944}
2945
486ec47a 2946/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2947 Finds fixed substrings. */
2948
a0288114 2949/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2950 to the position after last scanned or to NULL. */
2951
40d049e4
YO
2952#define INIT_AND_WITHP \
2953 assert(!and_withp); \
2954 Newx(and_withp,1,struct regnode_charclass_class); \
2955 SAVEFREEPV(and_withp)
07be1b83 2956
b515a41d 2957/* this is a chain of data about sub patterns we are processing that
486ec47a 2958 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2959 we can simulate recursion without losing state. */
2960struct scan_frame;
2961typedef struct scan_frame {
2962 regnode *last; /* last node to process in this frame */
2963 regnode *next; /* next node to process when last is reached */
2964 struct scan_frame *prev; /*previous frame*/
2965 I32 stop; /* what stopparen do we use */
2966} scan_frame;
2967
304ee84b
YO
2968
2969#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2970
e1d1eefb
YO
2971#define CASE_SYNST_FNC(nAmE) \
2972case nAmE: \
2973 if (flags & SCF_DO_STCLASS_AND) { \
2974 for (value = 0; value < 256; value++) \
2975 if (!is_ ## nAmE ## _cp(value)) \
2976 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2977 } \
2978 else { \
2979 for (value = 0; value < 256; value++) \
2980 if (is_ ## nAmE ## _cp(value)) \
2981 ANYOF_BITMAP_SET(data->start_class, value); \
2982 } \
2983 break; \
2984case N ## nAmE: \
2985 if (flags & SCF_DO_STCLASS_AND) { \
2986 for (value = 0; value < 256; value++) \
2987 if (is_ ## nAmE ## _cp(value)) \
2988 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2989 } \
2990 else { \
2991 for (value = 0; value < 256; value++) \
2992 if (!is_ ## nAmE ## _cp(value)) \
2993 ANYOF_BITMAP_SET(data->start_class, value); \
2994 } \
2995 break
2996
2997
2998
76e3520e 2999STATIC I32
40d049e4 3000S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 3001 I32 *minlenp, I32 *deltap,
40d049e4
YO
3002 regnode *last,
3003 scan_data_t *data,
3004 I32 stopparen,
3005 U8* recursed,
3006 struct regnode_charclass_class *and_withp,
3007 U32 flags, U32 depth)
c277df42
IZ
3008 /* scanp: Start here (read-write). */
3009 /* deltap: Write maxlen-minlen here. */
3010 /* last: Stop before this one. */
40d049e4
YO
3011 /* data: string data about the pattern */
3012 /* stopparen: treat close N as END */
3013 /* recursed: which subroutines have we recursed into */
3014 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 3015{
97aff369 3016 dVAR;
c277df42
IZ
3017 I32 min = 0, pars = 0, code;
3018 regnode *scan = *scanp, *next;
3019 I32 delta = 0;
3020 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 3021 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
3022 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3023 scan_data_t data_fake;
a3621e74 3024 SV *re_trie_maxbuff = NULL;
786e8c11 3025 regnode *first_non_open = scan;
e2e6a0f1 3026 I32 stopmin = I32_MAX;
8aa23a47 3027 scan_frame *frame = NULL;
a3621e74 3028 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 3029
7918f24d
NC
3030 PERL_ARGS_ASSERT_STUDY_CHUNK;
3031
13a24bad 3032#ifdef DEBUGGING
40d049e4 3033 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 3034#endif
40d049e4 3035
786e8c11 3036 if ( depth == 0 ) {
40d049e4 3037 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
3038 first_non_open=regnext(first_non_open);
3039 }
3040
b81d288d 3041
8aa23a47
YO
3042 fake_study_recurse:
3043 while ( scan && OP(scan) != END && scan < last ){
9d071ca8
KW
3044 UV min_subtract = 0; /* How much to subtract from the minimum node
3045 length to get a real minimum (because the
3046 folded version may be shorter) */
f758bddf 3047 bool has_exactf_sharp_s = FALSE;
8aa23a47 3048 /* Peephole optimizer: */
304ee84b 3049 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3050 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3051
3052 /* Its not clear to khw or hv why this is done here, and not in the
3053 * clauses that deal with EXACT nodes. khw's guess is that it's
3054 * because of a previous design */
9d071ca8 3055 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3056
3057 /* Follow the next-chain of the current node and optimize
3058 away all the NOTHINGs from it. */
3059 if (OP(scan) != CURLYX) {
3060 const int max = (reg_off_by_arg[OP(scan)]
3061 ? I32_MAX
3062 /* I32 may be smaller than U16 on CRAYs! */
3063 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3064 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3065 int noff;
3066 regnode *n = scan;
686b73d4 3067
8aa23a47
YO
3068 /* Skip NOTHING and LONGJMP. */
3069 while ((n = regnext(n))
3070 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3071 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3072 && off + noff < max)
3073 off += noff;
3074 if (reg_off_by_arg[OP(scan)])
3075 ARG(scan) = off;
3076 else
3077 NEXT_OFF(scan) = off;
3078 }
a3621e74 3079
c277df42 3080
8aa23a47
YO
3081
3082 /* The principal pseudo-switch. Cannot be a switch, since we
3083 look into several different things. */
3084 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3085 || OP(scan) == IFTHEN) {
3086 next = regnext(scan);
3087 code = OP(scan);
3088 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3089
8aa23a47
YO
3090 if (OP(next) == code || code == IFTHEN) {
3091 /* NOTE - There is similar code to this block below for handling
3092 TRIE nodes on a re-study. If you change stuff here check there
3093 too. */
3094 I32 max1 = 0, min1 = I32_MAX, num = 0;
3095 struct regnode_charclass_class accum;
3096 regnode * const startbranch=scan;
686b73d4 3097
8aa23a47 3098 if (flags & SCF_DO_SUBSTR)
304ee84b 3099 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3100 if (flags & SCF_DO_STCLASS)
e755fd73 3101 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3102
3103 while (OP(scan) == code) {
3104 I32 deltanext, minnext, f = 0, fake;
3105 struct regnode_charclass_class this_class;
3106
3107 num++;
3108 data_fake.flags = 0;
3109 if (data) {
3110 data_fake.whilem_c = data->whilem_c;
3111 data_fake.last_closep = data->last_closep;
3112 }
3113 else
3114 data_fake.last_closep = &fake;
58e23c8d
YO
3115
3116 data_fake.pos_delta = delta;
8aa23a47
YO
3117 next = regnext(scan);
3118 scan = NEXTOPER(scan);
3119 if (code != BRANCH)
c277df42 3120 scan = NEXTOPER(scan);
8aa23a47 3121 if (flags & SCF_DO_STCLASS) {
e755fd73 3122 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3123 data_fake.start_class = &this_class;
3124 f = SCF_DO_STCLASS_AND;
58e23c8d 3125 }
8aa23a47
YO
3126 if (flags & SCF_WHILEM_VISITED_POS)
3127 f |= SCF_WHILEM_VISITED_POS;
3128
3129 /* we suppose the run is continuous, last=next...*/
3130 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3131 next, &data_fake,
3132 stopparen, recursed, NULL, f,depth+1);
3133 if (min1 > minnext)
3134 min1 = minnext;
3135 if (max1 < minnext + deltanext)
3136 max1 = minnext + deltanext;
3137 if (deltanext == I32_MAX)
3138 is_inf = is_inf_internal = 1;
3139 scan = next;
3140 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3141 pars++;
3142 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3143 if ( stopmin > minnext)
3144 stopmin = min + min1;
3145 flags &= ~SCF_DO_SUBSTR;
3146 if (data)
3147 data->flags |= SCF_SEEN_ACCEPT;
3148 }
3149 if (data) {
3150 if (data_fake.flags & SF_HAS_EVAL)
3151 data->flags |= SF_HAS_EVAL;
3152 data->whilem_c = data_fake.whilem_c;
3dab1dad 3153 }
8aa23a47 3154 if (flags & SCF_DO_STCLASS)
3fffb88a 3155 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3156 }
3157 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3158 min1 = 0;
3159 if (flags & SCF_DO_SUBSTR) {
3160 data->pos_min += min1;
3161 data->pos_delta += max1 - min1;
3162 if (max1 != min1 || is_inf)
3163 data->longest = &(data->longest_float);
3164 }
3165 min += min1;
3166 delta += max1 - min1;
3167 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3168 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3169 if (min1) {
3170 cl_and(data->start_class, and_withp);
3171 flags &= ~SCF_DO_STCLASS;
653099ff 3172 }
8aa23a47
YO
3173 }
3174 else if (flags & SCF_DO_STCLASS_AND) {
3175 if (min1) {
3176 cl_and(data->start_class, &accum);
3177 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3178 }
8aa23a47
YO
3179 else {
3180 /* Switch to OR mode: cache the old value of
3181 * data->start_class */
3182 INIT_AND_WITHP;
3183 StructCopy(data->start_class, and_withp,
3184 struct regnode_charclass_class);
3185 flags &= ~SCF_DO_STCLASS_AND;
3186 StructCopy(&accum, data->start_class,
3187 struct regnode_charclass_class);
3188 flags |= SCF_DO_STCLASS_OR;
3189 data->start_class->flags |= ANYOF_EOS;
de0c8cb8 3190 }
8aa23a47 3191 }
a3621e74 3192
8aa23a47
YO
3193 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3194 /* demq.
a3621e74 3195
8aa23a47
YO
3196 Assuming this was/is a branch we are dealing with: 'scan' now
3197 points at the item that follows the branch sequence, whatever
3198 it is. We now start at the beginning of the sequence and look
3199 for subsequences of
a3621e74 3200
8aa23a47
YO
3201 BRANCH->EXACT=>x1
3202 BRANCH->EXACT=>x2
3203 tail
a3621e74 3204
8aa23a47 3205 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3206
486ec47a 3207 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3208 element into a trie and then add the subsequent branch exact
3209 strings to the trie.
a3621e74 3210
8aa23a47 3211 We have two cases
a3621e74 3212
3b753521 3213 1. patterns where the whole set of branches can be converted.
a3621e74 3214
8aa23a47 3215 2. patterns where only a subset can be converted.
a3621e74 3216
8aa23a47
YO
3217 In case 1 we can replace the whole set with a single regop
3218 for the trie. In case 2 we need to keep the start and end
3b753521 3219 branches so
a3621e74 3220
8aa23a47
YO
3221 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3222 becomes BRANCH TRIE; BRANCH X;
786e8c11 3223
8aa23a47
YO
3224 There is an additional case, that being where there is a
3225 common prefix, which gets split out into an EXACT like node
3226 preceding the TRIE node.
a3621e74 3227
8aa23a47
YO
3228 If x(1..n)==tail then we can do a simple trie, if not we make
3229 a "jump" trie, such that when we match the appropriate word
486ec47a 3230 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3231 a nested if into a case structure of sorts.
b515a41d 3232
8aa23a47 3233 */
686b73d4 3234
8aa23a47
YO
3235 int made=0;
3236 if (!re_trie_maxbuff) {
3237 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3238 if (!SvIOK(re_trie_maxbuff))
3239 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3240 }
3241 if ( SvIV(re_trie_maxbuff)>=0 ) {
3242 regnode *cur;
3243 regnode *first = (regnode *)NULL;
3244 regnode *last = (regnode *)NULL;
3245 regnode *tail = scan;
fab2782b 3246 U8 trietype = 0;
8aa23a47 3247 U32 count=0;
a3621e74
YO
3248
3249#ifdef DEBUGGING
8aa23a47 3250 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3251#endif
8aa23a47
YO
3252 /* var tail is used because there may be a TAIL
3253 regop in the way. Ie, the exacts will point to the
3254 thing following the TAIL, but the last branch will
3255 point at the TAIL. So we advance tail. If we
3256 have nested (?:) we may have to move through several
3257 tails.
3258 */
3259
3260 while ( OP( tail ) == TAIL ) {
3261 /* this is the TAIL generated by (?:) */
3262 tail = regnext( tail );
3263 }
a3621e74 3264
8aa23a47 3265
df826430 3266 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3267 regprop(RExC_rx, mysv, tail );
3268 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3269 (int)depth * 2 + 2, "",
3270 "Looking for TRIE'able sequences. Tail node is: ",
3271 SvPV_nolen_const( mysv )
3272 );
3273 });
3274
3275 /*
3276
fab2782b
YO
3277 Step through the branches
3278 cur represents each branch,
3279 noper is the first thing to be matched as part of that branch
3280 noper_next is the regnext() of that node.
3281
3282 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3283 via a "jump trie" but we also support building with NOJUMPTRIE,
3284 which restricts the trie logic to structures like /FOO|BAR/.
3285
3286 If noper is a trieable nodetype then the branch is a possible optimization
3287 target. If we are building under NOJUMPTRIE then we require that noper_next
3288 is the same as scan (our current position in the regex program).
3289
3290 Once we have two or more consecutive such branches we can create a
3291 trie of the EXACT's contents and stitch it in place into the program.
3292
3293 If the sequence represents all of the branches in the alternation we
3294 replace the entire thing with a single TRIE node.
3295
3296 Otherwise when it is a subsequence we need to stitch it in place and
3297 replace only the relevant branches. This means the first branch has
3298 to remain as it is used by the alternation logic, and its next pointer,
3299 and needs to be repointed at the item on the branch chain following
3300 the last branch we have optimized away.
3301
3302 This could be either a BRANCH, in which case the subsequence is internal,
3303 or it could be the item following the branch sequence in which case the
3304 subsequence is at the end (which does not necessarily mean the first node
3305 is the start of the alternation).
3306
3307 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3308
3309 optype | trietype
3310 ----------------+-----------
3311 NOTHING | NOTHING
3312 EXACT | EXACT
3313 EXACTFU | EXACTFU
3314 EXACTFU_SS | EXACTFU
3315 EXACTFU_TRICKYFOLD | EXACTFU
3316 EXACTFA | 0
3317
8aa23a47
YO
3318
3319 */
fab2782b
YO
3320#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3321 ( EXACT == (X) ) ? EXACT : \
3322 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3323 0 )
8aa23a47
YO
3324
3325 /* dont use tail as the end marker for this traverse */
3326 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3327 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3328 U8 noper_type = OP( noper );
3329 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3330#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3331 regnode * const noper_next = regnext( noper );
df826430
YO
3332 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3333 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3334#endif
3335
df826430 3336 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3337 regprop(RExC_rx, mysv, cur);
3338 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3339 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3340
3341 regprop(RExC_rx, mysv, noper);
3342 PerlIO_printf( Perl_debug_log, " -> %s",
3343 SvPV_nolen_const(mysv));
3344
3345 if ( noper_next ) {
3346 regprop(RExC_rx, mysv, noper_next );
3347 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3348 SvPV_nolen_const(mysv));
3349 }
df826430
YO
3350 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3351 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3352 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3353 );
8aa23a47 3354 });
fab2782b
YO
3355
3356 /* Is noper a trieable nodetype that can be merged with the
3357 * current trie (if there is one)? */
3358 if ( noper_trietype
3359 &&
3360 (
df826430
YO
3361 ( noper_trietype == NOTHING)
3362 || ( trietype == NOTHING )
a40630bf 3363 || ( trietype == noper_trietype )
fab2782b 3364 )
786e8c11 3365#ifdef NOJUMPTRIE
8aa23a47 3366 && noper_next == tail
786e8c11 3367#endif
8aa23a47
YO
3368 && count < U16_MAX)
3369 {
fab2782b
YO
3370 /* Handle mergable triable node
3371 * Either we are the first node in a new trieable sequence,
3372 * in which case we do some bookkeeping, otherwise we update
3373 * the end pointer. */
fab2782b 3374 if ( !first ) {
3b6759a6
YO
3375 first = cur;
3376 trietype = noper_trietype;
df826430
YO
3377 if ( noper_trietype == NOTHING ) {
3378#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3379 regnode * const noper_next = regnext( noper );
3b6759a6 3380 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3381 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3382#endif
3383
3b6759a6 3384 if ( noper_next_trietype )
df826430 3385 trietype = noper_next_trietype;
3b6759a6 3386 }
8aa23a47 3387 } else {
fab2782b
YO
3388 if ( trietype == NOTHING )
3389 trietype = noper_trietype;
8aa23a47
YO
3390 last = cur;
3391 }
df826430
YO
3392 if (first)
3393 count++;
fab2782b
YO
3394 } /* end handle mergable triable node */
3395 else {
3396 /* handle unmergable node -
3397 * noper may either be a triable node which can not be tried
3398 * together with the current trie, or a non triable node */
729aaeb5
YO
3399 if ( last ) {
3400 /* If last is set and trietype is not NOTHING then we have found
3401 * at least two triable branch sequences in a row of a similar
3402 * trietype so we can turn them into a trie. If/when we
3403 * allow NOTHING to start a trie sequence this condition will be
3404 * required, and it isn't expensive so we leave it in for now. */
3405 if ( trietype != NOTHING )
3406 make_trie( pRExC_state,
3407 startbranch, first, cur, tail, count,
3408 trietype, depth+1 );
fab2782b 3409 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3410 }
fab2782b 3411 if ( noper_trietype
786e8c11 3412#ifdef NOJUMPTRIE
8aa23a47 3413 && noper_next == tail
786e8c11 3414#endif
8aa23a47 3415 ){
fab2782b 3416 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3417 count = 1;
3418 first = cur;
fab2782b
YO
3419 trietype = noper_trietype;
3420 } else if (first) {
3421 /* if we already saw a first but the current node is not triable then we have
3422 * to reset the first information. */
8aa23a47
YO
3423 count = 0;
3424 first = NULL;
fab2782b 3425 trietype = 0;
8aa23a47 3426 }
fab2782b
YO
3427 } /* end handle unmergable node */
3428 } /* loop over branches */
df826430 3429 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3430 regprop(RExC_rx, mysv, cur);
3431 PerlIO_printf( Perl_debug_log,
3432 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3433 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3434
3435 });
3b6759a6
YO
3436 if ( last ) {
3437 if ( trietype != NOTHING ) {
3438 /* the last branch of the sequence was part of a trie,
3439 * so we have to construct it here outside of the loop
3440 */
3441 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3442#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3443 if ( ((made == MADE_EXACT_TRIE &&
3444 startbranch == first)
3445 || ( first_non_open == first )) &&
3446 depth==0 ) {
3447 flags |= SCF_TRIE_RESTUDY;
3448 if ( startbranch == first
3449 && scan == tail )
3450 {
3451 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3452 }
8aa23a47 3453 }
3dab1dad 3454#endif
3b6759a6
YO
3455 } else {
3456 /* at this point we know whatever we have is a NOTHING sequence/branch
3457 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3458 */
3459 if ( startbranch == first ) {
3460 regnode *opt;
3461 /* the entire thing is a NOTHING sequence, something like this:
3462 * (?:|) So we can turn it into a plain NOTHING op. */
3463 DEBUG_TRIE_COMPILE_r({
3464 regprop(RExC_rx, mysv, cur);
3465 PerlIO_printf( Perl_debug_log,
3466 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3467 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3468
3469 });
3470 OP(startbranch)= NOTHING;
3471 NEXT_OFF(startbranch)= tail - startbranch;
3472 for ( opt= startbranch + 1; opt < tail ; opt++ )
3473 OP(opt)= OPTIMIZED;
3474 }
3475 }
fab2782b
YO
3476 } /* end if ( last) */
3477 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3478
3479 } /* do trie */
3480
653099ff 3481 }
8aa23a47
YO
3482 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3483 scan = NEXTOPER(NEXTOPER(scan));
3484 } else /* single branch is optimized. */
3485 scan = NEXTOPER(scan);
3486 continue;
3487 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3488 scan_frame *newframe = NULL;
3489 I32 paren;
3490 regnode *start;
3491 regnode *end;
3492
3493 if (OP(scan) != SUSPEND) {
3494 /* set the pointer */
3495 if (OP(scan) == GOSUB) {
3496 paren = ARG(scan);
3497 RExC_recurse[ARG2L(scan)] = scan;
3498 start = RExC_open_parens[paren-1];
3499 end = RExC_close_parens[paren-1];
3500 } else {
3501 paren = 0;
f8fc2ecf 3502 start = RExC_rxi->program + 1;
8aa23a47
YO
3503 end = RExC_opend;
3504 }
3505 if (!recursed) {
3506 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3507 SAVEFREEPV(recursed);
3508 }
3509 if (!PAREN_TEST(recursed,paren+1)) {
3510 PAREN_SET(recursed,paren+1);
3511 Newx(newframe,1,scan_frame);
3512 } else {
3513 if (flags & SCF_DO_SUBSTR) {
304ee84b 3514 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3515 data->longest = &(data->longest_float);
3516 }
3517 is_inf = is_inf_internal = 1;
3518 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3519 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3520 flags &= ~SCF_DO_STCLASS;
3521 }
3522 } else {
3523 Newx(newframe,1,scan_frame);
3524 paren = stopparen;
3525 start = scan+2;
3526 end = regnext(scan);
3527 }
3528 if (newframe) {
3529 assert(start);
3530 assert(end);
3531 SAVEFREEPV(newframe);
3532 newframe->next = regnext(scan);
3533 newframe->last = last;
3534 newframe->stop = stopparen;
3535 newframe->prev = frame;
3536
3537 frame = newframe;
3538 scan = start;
3539 stopparen = paren;
3540 last = end;
3541
3542 continue;
3543 }
3544 }
3545 else if (OP(scan) == EXACT) {
3546 I32 l = STR_LEN(scan);
3547 UV uc;
3548 if (UTF) {
3549 const U8 * const s = (U8*)STRING(scan);
4b88fb76 3550 uc = utf8_to_uvchr_buf(s, s + l, NULL);
8aa23a47 3551 l = utf8_length(s, s + l);
8aa23a47
YO
3552 } else {
3553 uc = *((U8*)STRING(scan));
3554 }
3555 min += l;
3556 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3557 /* The code below prefers earlier match for fixed
3558 offset, later match for variable offset. */
3559 if (data->last_end == -1) { /* Update the start info. */
3560 data->last_start_min = data->pos_min;
3561 data->last_start_max = is_inf
3562 ? I32_MAX : data->pos_min + data->pos_delta;
b515a41d 3563 }
8aa23a47
YO
3564 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3565 if (UTF)
3566 SvUTF8_on(data->last_found);
3567 {
3568 SV * const sv = data->last_found;
3569 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3570 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3571 if (mg && mg->mg_len >= 0)
3572 mg->mg_len += utf8_length((U8*)STRING(scan),
3573 (U8*)STRING(scan)+STR_LEN(scan));
b515a41d 3574 }
8aa23a47
YO
3575 data->last_end = data->pos_min + l;
3576 data->pos_min += l; /* As in the first entry. */
3577 data->flags &= ~SF_BEFORE_EOL;
3578 }
3579 if (flags & SCF_DO_STCLASS_AND) {
3580 /* Check whether it is compatible with what we know already! */
3581 int compat = 1;
3582
54251c2e 3583
486ec47a 3584 /* If compatible, we or it in below. It is compatible if is
54251c2e
KW
3585 * in the bitmp and either 1) its bit or its fold is set, or 2)
3586 * it's for a locale. Even if there isn't unicode semantics
3587 * here, at runtime there may be because of matching against a
3588 * utf8 string, so accept a possible false positive for
3589 * latin1-range folds */
8aa23a47
YO
3590 if (uc >= 0x100 ||
3591 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3592 && !ANYOF_BITMAP_TEST(data->start_class, uc)
39065660 3593 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
54251c2e 3594 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
8aa23a47 3595 )
d18bf9dc 3596 {
8aa23a47 3597 compat = 0;
d18bf9dc 3598 }
8aa23a47
YO
3599 ANYOF_CLASS_ZERO(data->start_class);