This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Resolve EBCDIC inconsistency towards simpler
[perl5.git] / regcomp.c
CommitLineData
a0d0e21e
LW
1/* regcomp.c
2 */
3
4/*
4ac71550
TC
5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
6 *
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
a0d0e21e
LW
8 */
9
61296642
DM
10/* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
166f8a29 12 * a regular expression.
e4a054ea
DM
13 *
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
166f8a29
DM
18 */
19
a687059c
LW
20/* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
22 */
23
24/* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
27 */
28
e50aee73 29/* The names of the functions have been changed from regcomp and
3b753521 30 * regexec to pregcomp and pregexec in order to avoid conflicts
e50aee73
AD
31 * with the POSIX routines of the same names.
32*/
33
b9d5759e 34#ifdef PERL_EXT_RE_BUILD
54df2634 35#include "re_top.h"
b81d288d 36#endif
56953603 37
a687059c 38/*
e50aee73 39 * pregcomp and pregexec -- regsub and regerror are not used in perl
a687059c
LW
40 *
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
43 *
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
47 *
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
50 * from defects in it.
51 *
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
54 *
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
57 *
58 *
59 **** Alterations to Henry's code are...
60 ****
4bb101f2 61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
1129b882
NC
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
a687059c 64 ****
9ef589d8
LW
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
67
a687059c
LW
68 *
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
72 */
73#include "EXTERN.h"
864dbfa3 74#define PERL_IN_REGCOMP_C
a687059c 75#include "perl.h"
d06ea78c 76
acfe0abc 77#ifndef PERL_IN_XSUB_RE
d06ea78c
GS
78# include "INTERN.h"
79#endif
c277df42
IZ
80
81#define REG_COMP_C
54df2634
NC
82#ifdef PERL_IN_XSUB_RE
83# include "re_comp.h"
40b6423c 84extern const struct regexp_engine my_reg_engine;
54df2634
NC
85#else
86# include "regcomp.h"
87#endif
a687059c 88
04e98a4d 89#include "dquote_static.c"
370b8f2f
TC
90#ifndef PERL_IN_XSUB_RE
91# include "charclass_invlists.h"
92#endif
04e98a4d 93
94dc5c2d
KW
94#define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95
d4cce5f1 96#ifdef op
11343788 97#undef op
d4cce5f1 98#endif /* op */
11343788 99
fe14fcc3 100#ifdef MSDOS
7e4e8c89 101# if defined(BUGGY_MSC6)
fe14fcc3 102 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
7e4e8c89 103# pragma optimize("a",off)
fe14fcc3 104 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
7e4e8c89
NC
105# pragma optimize("w",on )
106# endif /* BUGGY_MSC6 */
fe14fcc3
LW
107#endif /* MSDOS */
108
a687059c
LW
109#ifndef STATIC
110#define STATIC static
111#endif
112
b1603ef8 113
830247a4 114typedef struct RExC_state_t {
514a91f1
DM
115 U32 flags; /* RXf_* are we folding, multilining? */
116 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 117 char *precomp; /* uncompiled string. */
288b8c02 118 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf
YO
119 regexp *rx; /* perl core regexp structure */
120 regexp_internal *rxi; /* internal data for regexp object pprivate field */
fac92740 121 char *start; /* Start of input for compile */
830247a4
IZ
122 char *end; /* End of input for compile */
123 char *parse; /* Input-scan pointer. */
124 I32 whilem_seen; /* number of WHILEM in this expr */
fac92740 125 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 126 regnode *emit_bound; /* First regnode outside of the allocated space */
ffc61ed2 127 regnode *emit; /* Code-emit pointer; &regdummy = don't = compiling */
830247a4
IZ
128 I32 naughty; /* How bad is this pattern? */
129 I32 sawback; /* Did we see \1, ...? */
130 U32 seen;
131 I32 size; /* Code size. */
c74340f9
YO
132 I32 npar; /* Capture buffer count, (OPEN). */
133 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 134 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
135 I32 extralen;
136 I32 seen_zerolen;
40d049e4
YO
137 regnode **open_parens; /* pointers to open parens */
138 regnode **close_parens; /* pointers to close parens */
139 regnode *opend; /* END node in program */
02daf0ab
YO
140 I32 utf8; /* whether the pattern is utf8 or not */
141 I32 orig_utf8; /* whether the pattern was originally in utf8 */
142 /* XXX use this for future optimisation of case
143 * where pattern must be upgraded to utf8. */
e40e74fe
KW
144 I32 uni_semantics; /* If a d charset modifier should use unicode
145 rules, even if the pattern is not in
146 utf8 */
81714fb9 147 HV *paren_names; /* Paren names */
1f1031fe 148
40d049e4
YO
149 regnode **recurse; /* Recurse regops */
150 I32 recurse_count; /* Number of recurse regops */
b57e4118 151 I32 in_lookbehind;
4624b182 152 I32 contains_locale;
bb3f3ed2 153 I32 override_recoding;
3d2bd50a 154 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 155 within pattern */
b1603ef8
DM
156 int num_code_blocks; /* size of code_blocks[] */
157 int code_index; /* next code_blocks[] slot */
830247a4
IZ
158#if ADD_TO_REGEXEC
159 char *starttry; /* -Dr: where regtry was called. */
160#define RExC_starttry (pRExC_state->starttry)
161#endif
d24ca0c5 162 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 163#ifdef DEBUGGING
be8e71aa 164 const char *lastparse;
3dab1dad 165 I32 lastnum;
1f1031fe 166 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
167#define RExC_lastparse (pRExC_state->lastparse)
168#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 169#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 170#endif
830247a4
IZ
171} RExC_state_t;
172
e2509266 173#define RExC_flags (pRExC_state->flags)
514a91f1 174#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 175#define RExC_precomp (pRExC_state->precomp)
288b8c02 176#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 177#define RExC_rx (pRExC_state->rx)
f8fc2ecf 178#define RExC_rxi (pRExC_state->rxi)
fac92740 179#define RExC_start (pRExC_state->start)
830247a4
IZ
180#define RExC_end (pRExC_state->end)
181#define RExC_parse (pRExC_state->parse)
182#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
183#ifdef RE_TRACK_PATTERN_OFFSETS
184#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
185#endif
830247a4 186#define RExC_emit (pRExC_state->emit)
fac92740 187#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 188#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
189#define RExC_naughty (pRExC_state->naughty)
190#define RExC_sawback (pRExC_state->sawback)
191#define RExC_seen (pRExC_state->seen)
192#define RExC_size (pRExC_state->size)
193#define RExC_npar (pRExC_state->npar)
e2e6a0f1 194#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
195#define RExC_extralen (pRExC_state->extralen)
196#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 197#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 198#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 199#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
200#define RExC_open_parens (pRExC_state->open_parens)
201#define RExC_close_parens (pRExC_state->close_parens)
202#define RExC_opend (pRExC_state->opend)
81714fb9 203#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
204#define RExC_recurse (pRExC_state->recurse)
205#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 206#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 207#define RExC_contains_locale (pRExC_state->contains_locale)
bb3f3ed2 208#define RExC_override_recoding (pRExC_state->override_recoding)
830247a4 209
cde0cee5 210
a687059c
LW
211#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
212#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
213 ((*s) == '{' && regcurly(s)))
a687059c 214
35c8bce7
LW
215#ifdef SPSTART
216#undef SPSTART /* dratted cpp namespace... */
217#endif
a687059c
LW
218/*
219 * Flags to be passed up and down.
220 */
a687059c 221#define WORST 0 /* Worst case. */
a3b492c3 222#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 223
18890cbd 224/* Simple enough to be STAR/PLUS operand; in an EXACT node must be a single
cf8c372d
KW
225 * character, and if utf8, must be invariant. Note that this is not the same
226 * thing as REGNODE_SIMPLE */
fda99bee 227#define SIMPLE 0x02
a3b492c3
YO
228#define SPSTART 0x04 /* Starts with * or +. */
229#define TRYAGAIN 0x08 /* Weeded out a declaration. */
230#define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
a687059c 231
3dab1dad
YO
232#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
233
07be1b83
YO
234/* whether trie related optimizations are enabled */
235#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
236#define TRIE_STUDY_OPT
786e8c11 237#define FULL_TRIE_STUDY
07be1b83
YO
238#define TRIE_STCLASS
239#endif
1de06328
YO
240
241
40d049e4
YO
242
243#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
244#define PBITVAL(paren) (1 << ((paren) & 7))
245#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
246#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
247#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
248
bbd61b5f
KW
249/* If not already in utf8, do a longjmp back to the beginning */
250#define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
251#define REQUIRE_UTF8 STMT_START { \
252 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
253 } STMT_END
40d049e4 254
1de06328
YO
255/* About scan_data_t.
256
257 During optimisation we recurse through the regexp program performing
258 various inplace (keyhole style) optimisations. In addition study_chunk
259 and scan_commit populate this data structure with information about
260 what strings MUST appear in the pattern. We look for the longest
3b753521 261 string that must appear at a fixed location, and we look for the
1de06328
YO
262 longest string that may appear at a floating location. So for instance
263 in the pattern:
264
265 /FOO[xX]A.*B[xX]BAR/
266
267 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
268 strings (because they follow a .* construct). study_chunk will identify
269 both FOO and BAR as being the longest fixed and floating strings respectively.
270
271 The strings can be composites, for instance
272
273 /(f)(o)(o)/
274
275 will result in a composite fixed substring 'foo'.
276
277 For each string some basic information is maintained:
278
279 - offset or min_offset
280 This is the position the string must appear at, or not before.
281 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
282 characters must match before the string we are searching for.
283 Likewise when combined with minlenp and the length of the string it
1de06328
YO
284 tells us how many characters must appear after the string we have
285 found.
286
287 - max_offset
288 Only used for floating strings. This is the rightmost point that
3b753521 289 the string can appear at. If set to I32 max it indicates that the
1de06328
YO
290 string can occur infinitely far to the right.
291
292 - minlenp
293 A pointer to the minimum length of the pattern that the string
294 was found inside. This is important as in the case of positive
295 lookahead or positive lookbehind we can have multiple patterns
296 involved. Consider
297
298 /(?=FOO).*F/
299
300 The minimum length of the pattern overall is 3, the minimum length
301 of the lookahead part is 3, but the minimum length of the part that
302 will actually match is 1. So 'FOO's minimum length is 3, but the
303 minimum length for the F is 1. This is important as the minimum length
304 is used to determine offsets in front of and behind the string being
305 looked for. Since strings can be composites this is the length of the
486ec47a 306 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
307 the length is calculated by study_chunk, so that the minimum lengths
308 are not known until the full pattern has been compiled, thus the
309 pointer to the value.
310
311 - lookbehind
312
313 In the case of lookbehind the string being searched for can be
314 offset past the start point of the final matching string.
315 If this value was just blithely removed from the min_offset it would
316 invalidate some of the calculations for how many chars must match
317 before or after (as they are derived from min_offset and minlen and
318 the length of the string being searched for).
319 When the final pattern is compiled and the data is moved from the
320 scan_data_t structure into the regexp structure the information
321 about lookbehind is factored in, with the information that would
322 have been lost precalculated in the end_shift field for the
323 associated string.
324
325 The fields pos_min and pos_delta are used to store the minimum offset
326 and the delta to the maximum offset at the current point in the pattern.
327
328*/
2c2d71f5
JH
329
330typedef struct scan_data_t {
1de06328
YO
331 /*I32 len_min; unused */
332 /*I32 len_delta; unused */
2c2d71f5
JH
333 I32 pos_min;
334 I32 pos_delta;
335 SV *last_found;
1de06328 336 I32 last_end; /* min value, <0 unless valid. */
2c2d71f5
JH
337 I32 last_start_min;
338 I32 last_start_max;
1de06328
YO
339 SV **longest; /* Either &l_fixed, or &l_float. */
340 SV *longest_fixed; /* longest fixed string found in pattern */
341 I32 offset_fixed; /* offset where it starts */
486ec47a 342 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
343 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
344 SV *longest_float; /* longest floating string found in pattern */
345 I32 offset_float_min; /* earliest point in string it can appear */
346 I32 offset_float_max; /* latest point in string it can appear */
486ec47a 347 I32 *minlen_float; /* pointer to the minlen relevant to the string */
1de06328 348 I32 lookbehind_float; /* is the position of the string modified by LB */
2c2d71f5
JH
349 I32 flags;
350 I32 whilem_c;
cb434fcc 351 I32 *last_closep;
653099ff 352 struct regnode_charclass_class *start_class;
2c2d71f5
JH
353} scan_data_t;
354
a687059c 355/*
e50aee73 356 * Forward declarations for pregcomp()'s friends.
a687059c 357 */
a0d0e21e 358
27da23d5 359static const scan_data_t zero_scan_data =
1de06328 360 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
361
362#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
363#define SF_BEFORE_SEOL 0x0001
364#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
365#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
366#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
367
09b7f37c
CB
368#ifdef NO_UNARY_PLUS
369# define SF_FIX_SHIFT_EOL (0+2)
370# define SF_FL_SHIFT_EOL (0+4)
371#else
372# define SF_FIX_SHIFT_EOL (+2)
373# define SF_FL_SHIFT_EOL (+4)
374#endif
c277df42
IZ
375
376#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
377#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
378
379#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
380#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
381#define SF_IS_INF 0x0040
382#define SF_HAS_PAR 0x0080
383#define SF_IN_PAR 0x0100
384#define SF_HAS_EVAL 0x0200
385#define SCF_DO_SUBSTR 0x0400
653099ff
GS
386#define SCF_DO_STCLASS_AND 0x0800
387#define SCF_DO_STCLASS_OR 0x1000
388#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 389#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 390
786e8c11 391#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 392#define SCF_SEEN_ACCEPT 0x8000
07be1b83 393
43fead97 394#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
395
396/* The enums for all these are ordered so things work out correctly */
a62b1201 397#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
cfaf538b 398#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
00b27cfc 399#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
400#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
401#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1 402#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a725e29c 403#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 404
43fead97 405#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 406
93733859 407#define OOB_NAMEDCLASS -1
b8c5462f 408
8e661ac5
KW
409/* There is no code point that is out-of-bounds, so this is problematic. But
410 * its only current use is to initialize a variable that is always set before
411 * looked at. */
412#define OOB_UNICODE 0xDEADBEEF
413
a0ed51b3
LW
414#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
415#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
416
8615cb43 417
b45f050a
JF
418/* length of regex to show in messages that don't mark a position within */
419#define RegexLengthToShowInErrorMessages 127
420
421/*
422 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
423 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
424 * op/pragma/warn/regcomp.
425 */
7253e4e3
RK
426#define MARKER1 "<-- HERE" /* marker as it appears in the description */
427#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 428
7253e4e3 429#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
430
431/*
432 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
433 * arg. Show regex, up to a maximum length. If it's too long, chop and add
434 * "...".
435 */
58e23c8d 436#define _FAIL(code) STMT_START { \
bfed75c6 437 const char *ellipses = ""; \
ccb2c380
MP
438 IV len = RExC_end - RExC_precomp; \
439 \
440 if (!SIZE_ONLY) \
288b8c02 441 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
442 if (len > RegexLengthToShowInErrorMessages) { \
443 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
444 len = RegexLengthToShowInErrorMessages - 10; \
445 ellipses = "..."; \
446 } \
58e23c8d 447 code; \
ccb2c380 448} STMT_END
8615cb43 449
58e23c8d
YO
450#define FAIL(msg) _FAIL( \
451 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
452 msg, (int)len, RExC_precomp, ellipses))
453
454#define FAIL2(msg,arg) _FAIL( \
455 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
456 arg, (int)len, RExC_precomp, ellipses))
457
b45f050a 458/*
b45f050a
JF
459 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
460 */
ccb2c380 461#define Simple_vFAIL(m) STMT_START { \
a28509cc 462 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
463 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
464 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
465} STMT_END
b45f050a
JF
466
467/*
468 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
469 */
ccb2c380
MP
470#define vFAIL(m) STMT_START { \
471 if (!SIZE_ONLY) \
288b8c02 472 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
473 Simple_vFAIL(m); \
474} STMT_END
b45f050a
JF
475
476/*
477 * Like Simple_vFAIL(), but accepts two arguments.
478 */
ccb2c380 479#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 480 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
481 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
482 (int)offset, RExC_precomp, RExC_precomp + offset); \
483} STMT_END
b45f050a
JF
484
485/*
486 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
487 */
ccb2c380
MP
488#define vFAIL2(m,a1) STMT_START { \
489 if (!SIZE_ONLY) \
288b8c02 490 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
491 Simple_vFAIL2(m, a1); \
492} STMT_END
b45f050a
JF
493
494
495/*
496 * Like Simple_vFAIL(), but accepts three arguments.
497 */
ccb2c380 498#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 499 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
500 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
501 (int)offset, RExC_precomp, RExC_precomp + offset); \
502} STMT_END
b45f050a
JF
503
504/*
505 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
506 */
ccb2c380
MP
507#define vFAIL3(m,a1,a2) STMT_START { \
508 if (!SIZE_ONLY) \
288b8c02 509 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
510 Simple_vFAIL3(m, a1, a2); \
511} STMT_END
b45f050a
JF
512
513/*
514 * Like Simple_vFAIL(), but accepts four arguments.
515 */
ccb2c380 516#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 517 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
518 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
519 (int)offset, RExC_precomp, RExC_precomp + offset); \
520} STMT_END
b45f050a 521
668c081a 522#define ckWARNreg(loc,m) STMT_START { \
a28509cc 523 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
524 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
525 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
526} STMT_END
527
668c081a 528#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 529 const IV offset = loc - RExC_precomp; \
d1d15184 530 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
531 m REPORT_LOCATION, \
532 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
533} STMT_END
534
2335b3d3
KW
535#define ckWARN2regdep(loc,m, a1) STMT_START { \
536 const IV offset = loc - RExC_precomp; \
537 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
538 m REPORT_LOCATION, \
539 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
540} STMT_END
541
668c081a 542#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 543 const IV offset = loc - RExC_precomp; \
668c081a 544 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
545 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
546} STMT_END
547
548#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 549 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
550 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
551 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
552} STMT_END
553
668c081a
NC
554#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
555 const IV offset = loc - RExC_precomp; \
556 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
557 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
558} STMT_END
559
ccb2c380 560#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 561 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
562 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
563 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
564} STMT_END
565
668c081a
NC
566#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
567 const IV offset = loc - RExC_precomp; \
568 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
569 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
570} STMT_END
571
ccb2c380 572#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 573 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
574 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
575 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
576} STMT_END
9d1d55b5 577
8615cb43 578
cd439c50 579/* Allow for side effects in s */
ccb2c380
MP
580#define REGC(c,s) STMT_START { \
581 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
582} STMT_END
cd439c50 583
fac92740
MJD
584/* Macros for recording node offsets. 20001227 mjd@plover.com
585 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
586 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
587 * Element 0 holds the number n.
07be1b83 588 * Position is 1 indexed.
fac92740 589 */
7122b237
YO
590#ifndef RE_TRACK_PATTERN_OFFSETS
591#define Set_Node_Offset_To_R(node,byte)
592#define Set_Node_Offset(node,byte)
593#define Set_Cur_Node_Offset
594#define Set_Node_Length_To_R(node,len)
595#define Set_Node_Length(node,len)
596#define Set_Node_Cur_Length(node)
597#define Node_Offset(n)
598#define Node_Length(n)
599#define Set_Node_Offset_Length(node,offset,len)
600#define ProgLen(ri) ri->u.proglen
601#define SetProgLen(ri,x) ri->u.proglen = x
602#else
603#define ProgLen(ri) ri->u.offsets[0]
604#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
605#define Set_Node_Offset_To_R(node,byte) STMT_START { \
606 if (! SIZE_ONLY) { \
607 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 608 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 609 if((node) < 0) { \
551405c4 610 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
611 } else { \
612 RExC_offsets[2*(node)-1] = (byte); \
613 } \
614 } \
615} STMT_END
616
617#define Set_Node_Offset(node,byte) \
618 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
619#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
620
621#define Set_Node_Length_To_R(node,len) STMT_START { \
622 if (! SIZE_ONLY) { \
623 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 624 __LINE__, (int)(node), (int)(len))); \
ccb2c380 625 if((node) < 0) { \
551405c4 626 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
627 } else { \
628 RExC_offsets[2*(node)] = (len); \
629 } \
630 } \
631} STMT_END
632
633#define Set_Node_Length(node,len) \
634 Set_Node_Length_To_R((node)-RExC_emit_start, len)
635#define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
636#define Set_Node_Cur_Length(node) \
637 Set_Node_Length(node, RExC_parse - parse_start)
fac92740
MJD
638
639/* Get offsets and lengths */
640#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
641#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
642
07be1b83
YO
643#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
644 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
645 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
646} STMT_END
7122b237 647#endif
07be1b83
YO
648
649#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
650#define EXPERIMENTAL_INPLACESCAN
f427392e 651#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 652
304ee84b
YO
653#define DEBUG_STUDYDATA(str,data,depth) \
654DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 655 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
656 "%*s" str "Pos:%"IVdf"/%"IVdf \
657 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
658 (int)(depth)*2, "", \
659 (IV)((data)->pos_min), \
660 (IV)((data)->pos_delta), \
304ee84b 661 (UV)((data)->flags), \
1de06328 662 (IV)((data)->whilem_c), \
304ee84b
YO
663 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
664 is_inf ? "INF " : "" \
1de06328
YO
665 ); \
666 if ((data)->last_found) \
667 PerlIO_printf(Perl_debug_log, \
668 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
669 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
670 SvPVX_const((data)->last_found), \
671 (IV)((data)->last_end), \
672 (IV)((data)->last_start_min), \
673 (IV)((data)->last_start_max), \
674 ((data)->longest && \
675 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
676 SvPVX_const((data)->longest_fixed), \
677 (IV)((data)->offset_fixed), \
678 ((data)->longest && \
679 (data)->longest==&((data)->longest_float)) ? "*" : "", \
680 SvPVX_const((data)->longest_float), \
681 (IV)((data)->offset_float_min), \
682 (IV)((data)->offset_float_max) \
683 ); \
684 PerlIO_printf(Perl_debug_log,"\n"); \
685});
686
acfe0abc 687static void clear_re(pTHX_ void *r);
4327152a 688
653099ff 689/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 690 Update the longest found anchored substring and the longest found
653099ff
GS
691 floating substrings if needed. */
692
4327152a 693STATIC void
304ee84b 694S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 695{
e1ec3a88
AL
696 const STRLEN l = CHR_SVLEN(data->last_found);
697 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 698 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 699
7918f24d
NC
700 PERL_ARGS_ASSERT_SCAN_COMMIT;
701
c277df42 702 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 703 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
704 if (*data->longest == data->longest_fixed) {
705 data->offset_fixed = l ? data->last_start_min : data->pos_min;
706 if (data->flags & SF_BEFORE_EOL)
b81d288d 707 data->flags
c277df42
IZ
708 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
709 else
710 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 711 data->minlen_fixed=minlenp;
1de06328 712 data->lookbehind_fixed=0;
a0ed51b3 713 }
304ee84b 714 else { /* *data->longest == data->longest_float */
c277df42 715 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
716 data->offset_float_max = (l
717 ? data->last_start_max
c277df42 718 : data->pos_min + data->pos_delta);
304ee84b 719 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 720 data->offset_float_max = I32_MAX;
c277df42 721 if (data->flags & SF_BEFORE_EOL)
b81d288d 722 data->flags
c277df42
IZ
723 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
724 else
725 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
726 data->minlen_float=minlenp;
727 data->lookbehind_float=0;
c277df42
IZ
728 }
729 }
730 SvCUR_set(data->last_found, 0);
0eda9292 731 {
a28509cc 732 SV * const sv = data->last_found;
097eb12c
AL
733 if (SvUTF8(sv) && SvMAGICAL(sv)) {
734 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
735 if (mg)
736 mg->mg_len = 0;
737 }
0eda9292 738 }
c277df42
IZ
739 data->last_end = -1;
740 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 741 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
742}
743
653099ff
GS
744/* Can match anything (initialization) */
745STATIC void
3fffb88a 746S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 747{
7918f24d
NC
748 PERL_ARGS_ASSERT_CL_ANYTHING;
749
f8bef550 750 ANYOF_BITMAP_SETALL(cl);
dd58aee1 751 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
3ad98780 752 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
3fffb88a
KW
753
754 /* If any portion of the regex is to operate under locale rules,
755 * initialization includes it. The reason this isn't done for all regexes
756 * is that the optimizer was written under the assumption that locale was
757 * all-or-nothing. Given the complexity and lack of documentation in the
758 * optimizer, and that there are inadequate test cases for locale, so many
759 * parts of it may not work properly, it is safest to avoid locale unless
760 * necessary. */
761 if (RExC_contains_locale) {
9d7a1e63 762 ANYOF_CLASS_SETALL(cl); /* /l uses class */
3fffb88a
KW
763 cl->flags |= ANYOF_LOCALE;
764 }
9d7a1e63
KW
765 else {
766 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
767 }
653099ff
GS
768}
769
770/* Can match anything (initialization) */
771STATIC int
5f66b61c 772S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
773{
774 int value;
775
7918f24d
NC
776 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
777
aaa51d5e 778 for (value = 0; value <= ANYOF_MAX; value += 2)
653099ff
GS
779 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
780 return 1;
1aa99e6b
IH
781 if (!(cl->flags & ANYOF_UNICODE_ALL))
782 return 0;
10edeb5d 783 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 784 return 0;
653099ff
GS
785 return 1;
786}
787
788/* Can match anything (initialization) */
789STATIC void
e755fd73 790S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 791{
7918f24d
NC
792 PERL_ARGS_ASSERT_CL_INIT;
793
8ecf7187 794 Zero(cl, 1, struct regnode_charclass_class);
653099ff 795 cl->type = ANYOF;
3fffb88a 796 cl_anything(pRExC_state, cl);
1411dba4 797 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
798}
799
1051e1c4
KW
800/* These two functions currently do the exact same thing */
801#define cl_init_zero S_cl_init
653099ff 802
dd58aee1
KW
803/* 'AND' a given class with another one. Can create false positives. 'cl'
804 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
805 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 806STATIC void
5f66b61c 807S_cl_and(struct regnode_charclass_class *cl,
a28509cc 808 const struct regnode_charclass_class *and_with)
653099ff 809{
7918f24d 810 PERL_ARGS_ASSERT_CL_AND;
40d049e4
YO
811
812 assert(and_with->type == ANYOF);
1e6ade67 813
c6b76537 814 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
815 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
816 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 817 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
818 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
819 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
653099ff
GS
820 int i;
821
822 if (and_with->flags & ANYOF_INVERT)
823 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
824 cl->bitmap[i] &= ~and_with->bitmap[i];
825 else
826 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
827 cl->bitmap[i] &= and_with->bitmap[i];
828 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 829
c6b76537 830 if (and_with->flags & ANYOF_INVERT) {
8951c461 831
c6b76537
KW
832 /* Here, the and'ed node is inverted. Get the AND of the flags that
833 * aren't affected by the inversion. Those that are affected are
834 * handled individually below */
835 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
836 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
837 cl->flags |= affected_flags;
838
839 /* We currently don't know how to deal with things that aren't in the
840 * bitmap, but we know that the intersection is no greater than what
841 * is already in cl, so let there be false positives that get sorted
842 * out after the synthetic start class succeeds, and the node is
843 * matched for real. */
844
845 /* The inversion of these two flags indicate that the resulting
846 * intersection doesn't have them */
847 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
848 cl->flags &= ~ANYOF_UNICODE_ALL;
849 }
c6b76537
KW
850 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
851 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 852 }
1aa99e6b 853 }
c6b76537 854 else { /* and'd node is not inverted */
3ad98780
KW
855 U8 outside_bitmap_but_not_utf8; /* Temp variable */
856
137165a6 857 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
858
859 /* Here 'and_with' doesn't match anything outside the bitmap
860 * (except possibly ANYOF_UNICODE_ALL), which means the
861 * intersection can't either, except for ANYOF_UNICODE_ALL, in
862 * which case we don't know what the intersection is, but it's no
863 * greater than what cl already has, so can just leave it alone,
864 * with possible false positives */
865 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
866 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 867 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 868 }
137165a6 869 }
c6b76537
KW
870 else if (! ANYOF_NONBITMAP(cl)) {
871
872 /* Here, 'and_with' does match something outside the bitmap, and cl
873 * doesn't have a list of things to match outside the bitmap. If
874 * cl can match all code points above 255, the intersection will
3ad98780
KW
875 * be those above-255 code points that 'and_with' matches. If cl
876 * can't match all Unicode code points, it means that it can't
877 * match anything outside the bitmap (since the 'if' that got us
878 * into this block tested for that), so we leave the bitmap empty.
879 */
c6b76537
KW
880 if (cl->flags & ANYOF_UNICODE_ALL) {
881 ARG_SET(cl, ARG(and_with));
3ad98780
KW
882
883 /* and_with's ARG may match things that don't require UTF8.
884 * And now cl's will too, in spite of this being an 'and'. See
885 * the comments below about the kludge */
886 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
887 }
888 }
889 else {
890 /* Here, both 'and_with' and cl match something outside the
891 * bitmap. Currently we do not do the intersection, so just match
892 * whatever cl had at the beginning. */
893 }
894
895
3ad98780
KW
896 /* Take the intersection of the two sets of flags. However, the
897 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
898 * kludge around the fact that this flag is not treated like the others
899 * which are initialized in cl_anything(). The way the optimizer works
900 * is that the synthetic start class (SSC) is initialized to match
901 * anything, and then the first time a real node is encountered, its
902 * values are AND'd with the SSC's with the result being the values of
903 * the real node. However, there are paths through the optimizer where
904 * the AND never gets called, so those initialized bits are set
905 * inappropriately, which is not usually a big deal, as they just cause
906 * false positives in the SSC, which will just mean a probably
907 * imperceptible slow down in execution. However this bit has a
908 * higher false positive consequence in that it can cause utf8.pm,
909 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
910 * bigger slowdown and also causes significant extra memory to be used.
911 * In order to prevent this, the code now takes a different tack. The
912 * bit isn't set unless some part of the regular expression needs it,
913 * but once set it won't get cleared. This means that these extra
914 * modules won't get loaded unless there was some path through the
915 * pattern that would have required them anyway, and so any false
916 * positives that occur by not ANDing them out when they could be
917 * aren't as severe as they would be if we treated this bit like all
918 * the others */
919 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
920 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 921 cl->flags &= and_with->flags;
3ad98780 922 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 923 }
653099ff
GS
924}
925
dd58aee1
KW
926/* 'OR' a given class with another one. Can create false positives. 'cl'
927 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
928 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 929STATIC void
3fffb88a 930S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 931{
7918f24d
NC
932 PERL_ARGS_ASSERT_CL_OR;
933
653099ff 934 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
935
936 /* Here, the or'd node is to be inverted. This means we take the
937 * complement of everything not in the bitmap, but currently we don't
938 * know what that is, so give up and match anything */
939 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 940 cl_anything(pRExC_state, cl);
c6b76537 941 }
653099ff
GS
942 /* We do not use
943 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
944 * <= (B1 | !B2) | (CL1 | !CL2)
945 * which is wasteful if CL2 is small, but we ignore CL2:
946 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
947 * XXXX Can we handle case-fold? Unclear:
948 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
949 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
950 */
c6b76537 951 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
952 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
953 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
653099ff
GS
954 int i;
955
956 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
957 cl->bitmap[i] |= ~or_with->bitmap[i];
958 } /* XXXX: logic is complicated otherwise */
959 else {
3fffb88a 960 cl_anything(pRExC_state, cl);
653099ff 961 }
c6b76537
KW
962
963 /* And, we can just take the union of the flags that aren't affected
964 * by the inversion */
965 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
966
967 /* For the remaining flags:
968 ANYOF_UNICODE_ALL and inverted means to not match anything above
969 255, which means that the union with cl should just be
970 what cl has in it, so can ignore this flag
971 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
972 is 127-255 to match them, but then invert that, so the
973 union with cl should just be what cl has in it, so can
974 ignore this flag
975 */
976 } else { /* 'or_with' is not inverted */
653099ff
GS
977 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
978 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
979 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
980 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
653099ff
GS
981 int i;
982
983 /* OR char bitmap and class bitmap separately */
984 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
985 cl->bitmap[i] |= or_with->bitmap[i];
1e6ade67 986 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
653099ff
GS
987 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
988 cl->classflags[i] |= or_with->classflags[i];
989 cl->flags |= ANYOF_CLASS;
990 }
991 }
992 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 993 cl_anything(pRExC_state, cl);
653099ff 994 }
9826f543 995
c6b76537
KW
996 if (ANYOF_NONBITMAP(or_with)) {
997
998 /* Use the added node's outside-the-bit-map match if there isn't a
999 * conflict. If there is a conflict (both nodes match something
1000 * outside the bitmap, but what they match outside is not the same
1001 * pointer, and hence not easily compared until XXX we extend
1002 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
1003 * match everything outside the bitmap. If that stuff is all above
1004 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1005 if (! ANYOF_NONBITMAP(cl)) {
1006 ARG_SET(cl, ARG(or_with));
1007 }
1008 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1009
1010 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1011 cl_anything(pRExC_state, cl);
1012 }
1013 else {
1014 cl->flags |= ANYOF_UNICODE_ALL;
1015 }
c6b76537 1016 }
4c34a693 1017 }
0b9668ee
KW
1018
1019 /* Take the union */
1020 cl->flags |= or_with->flags;
1aa99e6b 1021 }
653099ff
GS
1022}
1023
a3621e74
YO
1024#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1025#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1026#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1027#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1028
3dab1dad
YO
1029
1030#ifdef DEBUGGING
07be1b83 1031/*
2b8b4781
NC
1032 dump_trie(trie,widecharmap,revcharmap)
1033 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1034 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1035
1036 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1037 The _interim_ variants are used for debugging the interim
1038 tables that are used to generate the final compressed
1039 representation which is what dump_trie expects.
1040
486ec47a 1041 Part of the reason for their existence is to provide a form
3dab1dad 1042 of documentation as to how the different representations function.
07be1b83
YO
1043
1044*/
3dab1dad
YO
1045
1046/*
3dab1dad
YO
1047 Dumps the final compressed table form of the trie to Perl_debug_log.
1048 Used for debugging make_trie().
1049*/
b9a59e08 1050
3dab1dad 1051STATIC void
2b8b4781
NC
1052S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1053 AV *revcharmap, U32 depth)
3dab1dad
YO
1054{
1055 U32 state;
ab3bbdeb 1056 SV *sv=sv_newmortal();
55eed653 1057 int colwidth= widecharmap ? 6 : 4;
2e64971a 1058 U16 word;
3dab1dad
YO
1059 GET_RE_DEBUG_FLAGS_DECL;
1060
7918f24d 1061 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1062
3dab1dad
YO
1063 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1064 (int)depth * 2 + 2,"",
1065 "Match","Base","Ofs" );
1066
1067 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1068 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1069 if ( tmp ) {
ab3bbdeb
YO
1070 PerlIO_printf( Perl_debug_log, "%*s",
1071 colwidth,
ddc5bc0f 1072 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1073 PL_colors[0], PL_colors[1],
1074 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1075 PERL_PV_ESCAPE_FIRSTCHAR
1076 )
1077 );
3dab1dad
YO
1078 }
1079 }
1080 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1081 (int)depth * 2 + 2,"");
1082
1083 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1084 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1085 PerlIO_printf( Perl_debug_log, "\n");
1086
1e2e3d02 1087 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1088 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1089
1090 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1091
1092 if ( trie->states[ state ].wordnum ) {
1093 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1094 } else {
1095 PerlIO_printf( Perl_debug_log, "%6s", "" );
1096 }
1097
1098 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1099
1100 if ( base ) {
1101 U32 ofs = 0;
1102
1103 while( ( base + ofs < trie->uniquecharcount ) ||
1104 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1105 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1106 ofs++;
1107
1108 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1109
1110 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1111 if ( ( base + ofs >= trie->uniquecharcount ) &&
1112 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1113 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1114 {
ab3bbdeb
YO
1115 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1116 colwidth,
3dab1dad
YO
1117 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1118 } else {
ab3bbdeb 1119 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1120 }
1121 }
1122
1123 PerlIO_printf( Perl_debug_log, "]");
1124
1125 }
1126 PerlIO_printf( Perl_debug_log, "\n" );
1127 }
2e64971a
DM
1128 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1129 for (word=1; word <= trie->wordcount; word++) {
1130 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1131 (int)word, (int)(trie->wordinfo[word].prev),
1132 (int)(trie->wordinfo[word].len));
1133 }
1134 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1135}
1136/*
3dab1dad
YO
1137 Dumps a fully constructed but uncompressed trie in list form.
1138 List tries normally only are used for construction when the number of
1139 possible chars (trie->uniquecharcount) is very high.
1140 Used for debugging make_trie().
1141*/
1142STATIC void
55eed653 1143S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1144 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1145 U32 depth)
3dab1dad
YO
1146{
1147 U32 state;
ab3bbdeb 1148 SV *sv=sv_newmortal();
55eed653 1149 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1150 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1151
1152 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1153
3dab1dad 1154 /* print out the table precompression. */
ab3bbdeb
YO
1155 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1156 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1157 "------:-----+-----------------\n" );
3dab1dad
YO
1158
1159 for( state=1 ; state < next_alloc ; state ++ ) {
1160 U16 charid;
1161
ab3bbdeb 1162 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1163 (int)depth * 2 + 2,"", (UV)state );
1164 if ( ! trie->states[ state ].wordnum ) {
1165 PerlIO_printf( Perl_debug_log, "%5s| ","");
1166 } else {
1167 PerlIO_printf( Perl_debug_log, "W%4x| ",
1168 trie->states[ state ].wordnum
1169 );
1170 }
1171 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1172 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1173 if ( tmp ) {
1174 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1175 colwidth,
ddc5bc0f 1176 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1177 PL_colors[0], PL_colors[1],
1178 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1179 PERL_PV_ESCAPE_FIRSTCHAR
1180 ) ,
1e2e3d02
YO
1181 TRIE_LIST_ITEM(state,charid).forid,
1182 (UV)TRIE_LIST_ITEM(state,charid).newstate
1183 );
1184 if (!(charid % 10))
664e119d
RGS
1185 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1186 (int)((depth * 2) + 14), "");
1e2e3d02 1187 }
ab3bbdeb
YO
1188 }
1189 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1190 }
1191}
1192
1193/*
3dab1dad
YO
1194 Dumps a fully constructed but uncompressed trie in table form.
1195 This is the normal DFA style state transition table, with a few
1196 twists to facilitate compression later.
1197 Used for debugging make_trie().
1198*/
1199STATIC void
55eed653 1200S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1201 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1202 U32 depth)
3dab1dad
YO
1203{
1204 U32 state;
1205 U16 charid;
ab3bbdeb 1206 SV *sv=sv_newmortal();
55eed653 1207 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1208 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1209
1210 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1211
1212 /*
1213 print out the table precompression so that we can do a visual check
1214 that they are identical.
1215 */
1216
1217 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1218
1219 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1220 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1221 if ( tmp ) {
ab3bbdeb
YO
1222 PerlIO_printf( Perl_debug_log, "%*s",
1223 colwidth,
ddc5bc0f 1224 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1225 PL_colors[0], PL_colors[1],
1226 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1227 PERL_PV_ESCAPE_FIRSTCHAR
1228 )
1229 );
3dab1dad
YO
1230 }
1231 }
1232
1233 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1234
1235 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1236 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1237 }
1238
1239 PerlIO_printf( Perl_debug_log, "\n" );
1240
1241 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1242
1243 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1244 (int)depth * 2 + 2,"",
1245 (UV)TRIE_NODENUM( state ) );
1246
1247 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1248 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1249 if (v)
1250 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1251 else
1252 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1253 }
1254 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1255 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1256 } else {
1257 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1258 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1259 }
1260 }
07be1b83 1261}
3dab1dad
YO
1262
1263#endif
1264
2e64971a 1265
786e8c11
YO
1266/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1267 startbranch: the first branch in the whole branch sequence
1268 first : start branch of sequence of branch-exact nodes.
1269 May be the same as startbranch
1270 last : Thing following the last branch.
1271 May be the same as tail.
1272 tail : item following the branch sequence
1273 count : words in the sequence
1274 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1275 depth : indent depth
3dab1dad 1276
786e8c11 1277Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1278
786e8c11
YO
1279A trie is an N'ary tree where the branches are determined by digital
1280decomposition of the key. IE, at the root node you look up the 1st character and
1281follow that branch repeat until you find the end of the branches. Nodes can be
1282marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1283
786e8c11 1284 /he|she|his|hers/
72f13be8 1285
786e8c11
YO
1286would convert into the following structure. Numbers represent states, letters
1287following numbers represent valid transitions on the letter from that state, if
1288the number is in square brackets it represents an accepting state, otherwise it
1289will be in parenthesis.
07be1b83 1290
786e8c11
YO
1291 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1292 | |
1293 | (2)
1294 | |
1295 (1) +-i->(6)-+-s->[7]
1296 |
1297 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1298
786e8c11
YO
1299 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1300
1301This shows that when matching against the string 'hers' we will begin at state 1
1302read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1303then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1304is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1305single traverse. We store a mapping from accepting to state to which word was
1306matched, and then when we have multiple possibilities we try to complete the
1307rest of the regex in the order in which they occured in the alternation.
1308
1309The only prior NFA like behaviour that would be changed by the TRIE support is
1310the silent ignoring of duplicate alternations which are of the form:
1311
1312 / (DUPE|DUPE) X? (?{ ... }) Y /x
1313
4b714af6 1314Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1315and without the optimisation. With the optimisations dupes will be silently
486ec47a 1316ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1317the following demonstrates:
1318
1319 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1320
1321which prints out 'word' three times, but
1322
1323 'words'=~/(word|word|word)(?{ print $1 })S/
1324
1325which doesnt print it out at all. This is due to other optimisations kicking in.
1326
1327Example of what happens on a structural level:
1328
486ec47a 1329The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1330
1331 1: CURLYM[1] {1,32767}(18)
1332 5: BRANCH(8)
1333 6: EXACT <ac>(16)
1334 8: BRANCH(11)
1335 9: EXACT <ad>(16)
1336 11: BRANCH(14)
1337 12: EXACT <ab>(16)
1338 16: SUCCEED(0)
1339 17: NOTHING(18)
1340 18: END(0)
1341
1342This would be optimizable with startbranch=5, first=5, last=16, tail=16
1343and should turn into:
1344
1345 1: CURLYM[1] {1,32767}(18)
1346 5: TRIE(16)
1347 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1348 <ac>
1349 <ad>
1350 <ab>
1351 16: SUCCEED(0)
1352 17: NOTHING(18)
1353 18: END(0)
1354
1355Cases where tail != last would be like /(?foo|bar)baz/:
1356
1357 1: BRANCH(4)
1358 2: EXACT <foo>(8)
1359 4: BRANCH(7)
1360 5: EXACT <bar>(8)
1361 7: TAIL(8)
1362 8: EXACT <baz>(10)
1363 10: END(0)
1364
1365which would be optimizable with startbranch=1, first=1, last=7, tail=8
1366and would end up looking like:
1367
1368 1: TRIE(8)
1369 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1370 <foo>
1371 <bar>
1372 7: TAIL(8)
1373 8: EXACT <baz>(10)
1374 10: END(0)
1375
1376 d = uvuni_to_utf8_flags(d, uv, 0);
1377
1378is the recommended Unicode-aware way of saying
1379
1380 *(d++) = uv;
1381*/
1382
fab2782b 1383#define TRIE_STORE_REVCHAR(val) \
786e8c11 1384 STMT_START { \
73031816 1385 if (UTF) { \
fab2782b 1386 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1387 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
fab2782b 1388 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
73031816
NC
1389 SvCUR_set(zlopp, kapow - flrbbbbb); \
1390 SvPOK_on(zlopp); \
1391 SvUTF8_on(zlopp); \
1392 av_push(revcharmap, zlopp); \
1393 } else { \
fab2782b 1394 char ooooff = (char)val; \
73031816
NC
1395 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1396 } \
1397 } STMT_END
786e8c11 1398
fab2782b
YO
1399#define TRIE_READ_CHAR STMT_START { \
1400 wordlen++; \
1401 if ( UTF ) { \
1402 /* if it is UTF then it is either already folded, or does not need folding */ \
1403 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1404 } \
1405 else if (folder == PL_fold_latin1) { \
1406 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1407 if ( foldlen > 0 ) { \
1408 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1409 foldlen -= len; \
1410 scan += len; \
1411 len = 0; \
1412 } else { \
1413 len = 1; \
1414 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1415 skiplen = UNISKIP(uvc); \
1416 foldlen -= skiplen; \
1417 scan = foldbuf + skiplen; \
1418 } \
1419 } else { \
1420 /* raw data, will be folded later if needed */ \
1421 uvc = (U32)*uc; \
1422 len = 1; \
1423 } \
786e8c11
YO
1424} STMT_END
1425
1426
1427
1428#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1429 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1430 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1431 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1432 } \
1433 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1434 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1435 TRIE_LIST_CUR( state )++; \
1436} STMT_END
07be1b83 1437
786e8c11
YO
1438#define TRIE_LIST_NEW(state) STMT_START { \
1439 Newxz( trie->states[ state ].trans.list, \
1440 4, reg_trie_trans_le ); \
1441 TRIE_LIST_CUR( state ) = 1; \
1442 TRIE_LIST_LEN( state ) = 4; \
1443} STMT_END
07be1b83 1444
786e8c11
YO
1445#define TRIE_HANDLE_WORD(state) STMT_START { \
1446 U16 dupe= trie->states[ state ].wordnum; \
1447 regnode * const noper_next = regnext( noper ); \
1448 \
786e8c11
YO
1449 DEBUG_r({ \
1450 /* store the word for dumping */ \
1451 SV* tmp; \
1452 if (OP(noper) != NOTHING) \
740cce10 1453 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1454 else \
740cce10 1455 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1456 av_push( trie_words, tmp ); \
786e8c11
YO
1457 }); \
1458 \
1459 curword++; \
2e64971a
DM
1460 trie->wordinfo[curword].prev = 0; \
1461 trie->wordinfo[curword].len = wordlen; \
1462 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1463 \
1464 if ( noper_next < tail ) { \
1465 if (!trie->jump) \
c944940b 1466 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1467 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1468 if (!jumper) \
1469 jumper = noper_next; \
1470 if (!nextbranch) \
1471 nextbranch= regnext(cur); \
1472 } \
1473 \
1474 if ( dupe ) { \
2e64971a
DM
1475 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1476 /* chain, so that when the bits of chain are later */\
1477 /* linked together, the dups appear in the chain */\
1478 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1479 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1480 } else { \
1481 /* we haven't inserted this word yet. */ \
1482 trie->states[ state ].wordnum = curword; \
1483 } \
1484} STMT_END
07be1b83 1485
3dab1dad 1486
786e8c11
YO
1487#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1488 ( ( base + charid >= ucharcount \
1489 && base + charid < ubound \
1490 && state == trie->trans[ base - ucharcount + charid ].check \
1491 && trie->trans[ base - ucharcount + charid ].next ) \
1492 ? trie->trans[ base - ucharcount + charid ].next \
1493 : ( state==1 ? special : 0 ) \
1494 )
3dab1dad 1495
786e8c11
YO
1496#define MADE_TRIE 1
1497#define MADE_JUMP_TRIE 2
1498#define MADE_EXACT_TRIE 4
3dab1dad 1499
a3621e74 1500STATIC I32
786e8c11 1501S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1502{
27da23d5 1503 dVAR;
a3621e74
YO
1504 /* first pass, loop through and scan words */
1505 reg_trie_data *trie;
55eed653 1506 HV *widecharmap = NULL;
2b8b4781 1507 AV *revcharmap = newAV();
a3621e74 1508 regnode *cur;
9f7f3913 1509 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1510 STRLEN len = 0;
1511 UV uvc = 0;
1512 U16 curword = 0;
1513 U32 next_alloc = 0;
786e8c11
YO
1514 regnode *jumper = NULL;
1515 regnode *nextbranch = NULL;
7f69552c 1516 regnode *convert = NULL;
2e64971a 1517 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1518 /* we just use folder as a flag in utf8 */
1e696034 1519 const U8 * folder = NULL;
a3621e74 1520
2b8b4781
NC
1521#ifdef DEBUGGING
1522 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1523 AV *trie_words = NULL;
1524 /* along with revcharmap, this only used during construction but both are
1525 * useful during debugging so we store them in the struct when debugging.
8e11feef 1526 */
2b8b4781
NC
1527#else
1528 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1529 STRLEN trie_charcount=0;
3dab1dad 1530#endif
2b8b4781 1531 SV *re_trie_maxbuff;
a3621e74 1532 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1533
1534 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1535#ifndef DEBUGGING
1536 PERL_UNUSED_ARG(depth);
1537#endif
a3621e74 1538
1e696034 1539 switch (flags) {
c46d03cf 1540 case EXACT: break;
2f7f8cb1 1541 case EXACTFA:
fab2782b
YO
1542 case EXACTFU_SS:
1543 case EXACTFU_TRICKYFOLD:
1e696034
KW
1544 case EXACTFU: folder = PL_fold_latin1; break;
1545 case EXACTF: folder = PL_fold; break;
1546 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1547 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1548 }
1549
c944940b 1550 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1551 trie->refcount = 1;
3dab1dad 1552 trie->startstate = 1;
786e8c11 1553 trie->wordcount = word_count;
f8fc2ecf 1554 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1555 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1556 if (flags == EXACT)
c944940b 1557 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1558 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1559 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1560
a3621e74 1561 DEBUG_r({
2b8b4781 1562 trie_words = newAV();
a3621e74 1563 });
a3621e74 1564
0111c4fd 1565 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1566 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1567 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1568 }
df826430 1569 DEBUG_TRIE_COMPILE_r({
3dab1dad 1570 PerlIO_printf( Perl_debug_log,
786e8c11 1571 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1572 (int)depth * 2 + 2, "",
1573 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1574 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1575 (int)depth);
3dab1dad 1576 });
7f69552c
YO
1577
1578 /* Find the node we are going to overwrite */
1579 if ( first == startbranch && OP( last ) != BRANCH ) {
1580 /* whole branch chain */
1581 convert = first;
1582 } else {
1583 /* branch sub-chain */
1584 convert = NEXTOPER( first );
1585 }
1586
a3621e74
YO
1587 /* -- First loop and Setup --
1588
1589 We first traverse the branches and scan each word to determine if it
1590 contains widechars, and how many unique chars there are, this is
1591 important as we have to build a table with at least as many columns as we
1592 have unique chars.
1593
1594 We use an array of integers to represent the character codes 0..255
38a44b82 1595 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1596 native representation of the character value as the key and IV's for the
1597 coded index.
1598
1599 *TODO* If we keep track of how many times each character is used we can
1600 remap the columns so that the table compression later on is more
3b753521 1601 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1602 middle and the least common are on the outside. IMO this would be better
1603 than a most to least common mapping as theres a decent chance the most
1604 common letter will share a node with the least common, meaning the node
486ec47a 1605 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1606 case is when we have the least common nodes twice.
1607
1608 */
1609
a3621e74 1610 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1611 regnode *noper = NEXTOPER( cur );
e1ec3a88 1612 const U8 *uc = (U8*)STRING( noper );
df826430 1613 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1614 STRLEN foldlen = 0;
1615 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1616 STRLEN skiplen = 0;
2af232bd 1617 const U8 *scan = (U8*)NULL;
07be1b83 1618 U32 wordlen = 0; /* required init */
02daf0ab
YO
1619 STRLEN chars = 0;
1620 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1621
3dab1dad 1622 if (OP(noper) == NOTHING) {
df826430
YO
1623 regnode *noper_next= regnext(noper);
1624 if (noper_next != tail && OP(noper_next) == flags) {
1625 noper = noper_next;
1626 uc= (U8*)STRING(noper);
1627 e= uc + STR_LEN(noper);
1628 trie->minlen= STR_LEN(noper);
1629 } else {
1630 trie->minlen= 0;
1631 continue;
1632 }
3dab1dad 1633 }
df826430 1634
fab2782b 1635 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1636 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1637 regardless of encoding */
fab2782b
YO
1638 if (OP( noper ) == EXACTFU_SS) {
1639 /* false positives are ok, so just set this */
1640 TRIE_BITMAP_SET(trie,0xDF);
1641 }
1642 }
a3621e74 1643 for ( ; uc < e ; uc += len ) {
3dab1dad 1644 TRIE_CHARCOUNT(trie)++;
a3621e74 1645 TRIE_READ_CHAR;
3dab1dad 1646 chars++;
a3621e74 1647 if ( uvc < 256 ) {
fab2782b
YO
1648 if ( folder ) {
1649 U8 folded= folder[ (U8) uvc ];
1650 if ( !trie->charmap[ folded ] ) {
1651 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1652 TRIE_STORE_REVCHAR( folded );
1653 }
1654 }
a3621e74
YO
1655 if ( !trie->charmap[ uvc ] ) {
1656 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1657 TRIE_STORE_REVCHAR( uvc );
a3621e74 1658 }
02daf0ab 1659 if ( set_bit ) {
62012aee
KW
1660 /* store the codepoint in the bitmap, and its folded
1661 * equivalent. */
fab2782b 1662 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1663
1664 /* store the folded codepoint */
fab2782b 1665 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1666
1667 if ( !UTF ) {
1668 /* store first byte of utf8 representation of
acdf4139
KW
1669 variant codepoints */
1670 if (! UNI_IS_INVARIANT(uvc)) {
1671 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1672 }
1673 }
02daf0ab
YO
1674 set_bit = 0; /* We've done our bit :-) */
1675 }
a3621e74
YO
1676 } else {
1677 SV** svpp;
55eed653
NC
1678 if ( !widecharmap )
1679 widecharmap = newHV();
a3621e74 1680
55eed653 1681 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1682
1683 if ( !svpp )
e4584336 1684 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1685
1686 if ( !SvTRUE( *svpp ) ) {
1687 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1688 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1689 }
1690 }
1691 }
3dab1dad 1692 if( cur == first ) {
fab2782b
YO
1693 trie->minlen = chars;
1694 trie->maxlen = chars;
3dab1dad 1695 } else if (chars < trie->minlen) {
fab2782b 1696 trie->minlen = chars;
3dab1dad 1697 } else if (chars > trie->maxlen) {
fab2782b
YO
1698 trie->maxlen = chars;
1699 }
1700 if (OP( noper ) == EXACTFU_SS) {
1701 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1702 if (trie->minlen > 1)
1703 trie->minlen= 1;
1704 }
1705 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1706 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1707 * - We assume that any such sequence might match a 2 byte string */
1708 if (trie->minlen > 2 )
1709 trie->minlen= 2;
3dab1dad
YO
1710 }
1711
a3621e74
YO
1712 } /* end first pass */
1713 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1714 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1715 (int)depth * 2 + 2,"",
55eed653 1716 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1717 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1718 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1719 );
a3621e74
YO
1720
1721 /*
1722 We now know what we are dealing with in terms of unique chars and
1723 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1724 representation using a flat table will take. If it's over a reasonable
1725 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1726 conservative but potentially much slower representation using an array
1727 of lists.
1728
1729 At the end we convert both representations into the same compressed
1730 form that will be used in regexec.c for matching with. The latter
1731 is a form that cannot be used to construct with but has memory
1732 properties similar to the list form and access properties similar
1733 to the table form making it both suitable for fast searches and
1734 small enough that its feasable to store for the duration of a program.
1735
1736 See the comment in the code where the compressed table is produced
1737 inplace from the flat tabe representation for an explanation of how
1738 the compression works.
1739
1740 */
1741
1742
2e64971a
DM
1743 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1744 prev_states[1] = 0;
1745
3dab1dad 1746 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1747 /*
1748 Second Pass -- Array Of Lists Representation
1749
1750 Each state will be represented by a list of charid:state records
1751 (reg_trie_trans_le) the first such element holds the CUR and LEN
1752 points of the allocated array. (See defines above).
1753
1754 We build the initial structure using the lists, and then convert
1755 it into the compressed table form which allows faster lookups
1756 (but cant be modified once converted).
a3621e74
YO
1757 */
1758
a3621e74
YO
1759 STRLEN transcount = 1;
1760
1e2e3d02
YO
1761 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1762 "%*sCompiling trie using list compiler\n",
1763 (int)depth * 2 + 2, ""));
686b73d4 1764
c944940b
JH
1765 trie->states = (reg_trie_state *)
1766 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1767 sizeof(reg_trie_state) );
a3621e74
YO
1768 TRIE_LIST_NEW(1);
1769 next_alloc = 2;
1770
1771 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1772
df826430 1773 regnode *noper = NEXTOPER( cur );
c445ea15 1774 U8 *uc = (U8*)STRING( noper );
df826430 1775 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1776 U32 state = 1; /* required init */
1777 U16 charid = 0; /* sanity init */
1778 U8 *scan = (U8*)NULL; /* sanity init */
1779 STRLEN foldlen = 0; /* required init */
07be1b83 1780 U32 wordlen = 0; /* required init */
c445ea15 1781 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1782 STRLEN skiplen = 0;
c445ea15 1783
df826430
YO
1784 if (OP(noper) == NOTHING) {
1785 regnode *noper_next= regnext(noper);
1786 if (noper_next != tail && OP(noper_next) == flags) {
1787 noper = noper_next;
1788 uc= (U8*)STRING(noper);
1789 e= uc + STR_LEN(noper);
1790 }
1791 }
1792
3dab1dad 1793 if (OP(noper) != NOTHING) {
786e8c11 1794 for ( ; uc < e ; uc += len ) {
c445ea15 1795
786e8c11 1796 TRIE_READ_CHAR;
c445ea15 1797
786e8c11
YO
1798 if ( uvc < 256 ) {
1799 charid = trie->charmap[ uvc ];
c445ea15 1800 } else {
55eed653 1801 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1802 if ( !svpp ) {
1803 charid = 0;
1804 } else {
1805 charid=(U16)SvIV( *svpp );
1806 }
c445ea15 1807 }
786e8c11
YO
1808 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1809 if ( charid ) {
a3621e74 1810
786e8c11
YO
1811 U16 check;
1812 U32 newstate = 0;
a3621e74 1813
786e8c11
YO
1814 charid--;
1815 if ( !trie->states[ state ].trans.list ) {
1816 TRIE_LIST_NEW( state );
c445ea15 1817 }
786e8c11
YO
1818 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1819 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1820 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1821 break;
1822 }
1823 }
1824 if ( ! newstate ) {
1825 newstate = next_alloc++;
2e64971a 1826 prev_states[newstate] = state;
786e8c11
YO
1827 TRIE_LIST_PUSH( state, charid, newstate );
1828 transcount++;
1829 }
1830 state = newstate;
1831 } else {
1832 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1833 }
a28509cc 1834 }
c445ea15 1835 }
3dab1dad 1836 TRIE_HANDLE_WORD(state);
a3621e74
YO
1837
1838 } /* end second pass */
1839
1e2e3d02
YO
1840 /* next alloc is the NEXT state to be allocated */
1841 trie->statecount = next_alloc;
c944940b
JH
1842 trie->states = (reg_trie_state *)
1843 PerlMemShared_realloc( trie->states,
1844 next_alloc
1845 * sizeof(reg_trie_state) );
a3621e74 1846
3dab1dad 1847 /* and now dump it out before we compress it */
2b8b4781
NC
1848 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1849 revcharmap, next_alloc,
1850 depth+1)
1e2e3d02 1851 );
a3621e74 1852
c944940b
JH
1853 trie->trans = (reg_trie_trans *)
1854 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1855 {
1856 U32 state;
a3621e74
YO
1857 U32 tp = 0;
1858 U32 zp = 0;
1859
1860
1861 for( state=1 ; state < next_alloc ; state ++ ) {
1862 U32 base=0;
1863
1864 /*
1865 DEBUG_TRIE_COMPILE_MORE_r(
1866 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1867 );
1868 */
1869
1870 if (trie->states[state].trans.list) {
1871 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1872 U16 maxid=minid;
a28509cc 1873 U16 idx;
a3621e74
YO
1874
1875 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1876 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1877 if ( forid < minid ) {
1878 minid=forid;
1879 } else if ( forid > maxid ) {
1880 maxid=forid;
1881 }
a3621e74
YO
1882 }
1883 if ( transcount < tp + maxid - minid + 1) {
1884 transcount *= 2;
c944940b
JH
1885 trie->trans = (reg_trie_trans *)
1886 PerlMemShared_realloc( trie->trans,
446bd890
NC
1887 transcount
1888 * sizeof(reg_trie_trans) );
a3621e74
YO
1889 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1890 }
1891 base = trie->uniquecharcount + tp - minid;
1892 if ( maxid == minid ) {
1893 U32 set = 0;
1894 for ( ; zp < tp ; zp++ ) {
1895 if ( ! trie->trans[ zp ].next ) {
1896 base = trie->uniquecharcount + zp - minid;
1897 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1898 trie->trans[ zp ].check = state;
1899 set = 1;
1900 break;
1901 }
1902 }
1903 if ( !set ) {
1904 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1905 trie->trans[ tp ].check = state;
1906 tp++;
1907 zp = tp;
1908 }
1909 } else {
1910 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1911 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1912 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1913 trie->trans[ tid ].check = state;
1914 }
1915 tp += ( maxid - minid + 1 );
1916 }
1917 Safefree(trie->states[ state ].trans.list);
1918 }
1919 /*
1920 DEBUG_TRIE_COMPILE_MORE_r(
1921 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1922 );
1923 */
1924 trie->states[ state ].trans.base=base;
1925 }
cc601c31 1926 trie->lasttrans = tp + 1;
a3621e74
YO
1927 }
1928 } else {
1929 /*
1930 Second Pass -- Flat Table Representation.
1931
1932 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1933 We know that we will need Charcount+1 trans at most to store the data
1934 (one row per char at worst case) So we preallocate both structures
1935 assuming worst case.
1936
1937 We then construct the trie using only the .next slots of the entry
1938 structs.
1939
3b753521 1940 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1941 make compression both faster and easier by keeping track of how many non
1942 zero fields are in the node.
1943
1944 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1945 transition.
1946
1947 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1948 number representing the first entry of the node, and state as a
1949 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1950 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1951 are 2 entrys per node. eg:
1952
1953 A B A B
1954 1. 2 4 1. 3 7
1955 2. 0 3 3. 0 5
1956 3. 0 0 5. 0 0
1957 4. 0 0 7. 0 0
1958
1959 The table is internally in the right hand, idx form. However as we also
1960 have to deal with the states array which is indexed by nodenum we have to
1961 use TRIE_NODENUM() to convert.
1962
1963 */
1e2e3d02
YO
1964 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1965 "%*sCompiling trie using table compiler\n",
1966 (int)depth * 2 + 2, ""));
3dab1dad 1967
c944940b
JH
1968 trie->trans = (reg_trie_trans *)
1969 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1970 * trie->uniquecharcount + 1,
1971 sizeof(reg_trie_trans) );
1972 trie->states = (reg_trie_state *)
1973 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1974 sizeof(reg_trie_state) );
a3621e74
YO
1975 next_alloc = trie->uniquecharcount + 1;
1976
3dab1dad 1977
a3621e74
YO
1978 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1979
df826430 1980 regnode *noper = NEXTOPER( cur );
a28509cc 1981 const U8 *uc = (U8*)STRING( noper );
df826430 1982 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1983
1984 U32 state = 1; /* required init */
1985
1986 U16 charid = 0; /* sanity init */
1987 U32 accept_state = 0; /* sanity init */
1988 U8 *scan = (U8*)NULL; /* sanity init */
1989
1990 STRLEN foldlen = 0; /* required init */
07be1b83 1991 U32 wordlen = 0; /* required init */
fab2782b 1992 STRLEN skiplen = 0;
a3621e74
YO
1993 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1994
df826430
YO
1995 if (OP(noper) == NOTHING) {
1996 regnode *noper_next= regnext(noper);
1997 if (noper_next != tail && OP(noper_next) == flags) {
1998 noper = noper_next;
1999 uc= (U8*)STRING(noper);
2000 e= uc + STR_LEN(noper);
2001 }
2002 }
fab2782b 2003
3dab1dad 2004 if ( OP(noper) != NOTHING ) {
786e8c11 2005 for ( ; uc < e ; uc += len ) {
a3621e74 2006
786e8c11 2007 TRIE_READ_CHAR;
a3621e74 2008
786e8c11
YO
2009 if ( uvc < 256 ) {
2010 charid = trie->charmap[ uvc ];
2011 } else {
55eed653 2012 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2013 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2014 }
786e8c11
YO
2015 if ( charid ) {
2016 charid--;
2017 if ( !trie->trans[ state + charid ].next ) {
2018 trie->trans[ state + charid ].next = next_alloc;
2019 trie->trans[ state ].check++;
2e64971a
DM
2020 prev_states[TRIE_NODENUM(next_alloc)]
2021 = TRIE_NODENUM(state);
786e8c11
YO
2022 next_alloc += trie->uniquecharcount;
2023 }
2024 state = trie->trans[ state + charid ].next;
2025 } else {
2026 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2027 }
2028 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2029 }
a3621e74 2030 }
3dab1dad
YO
2031 accept_state = TRIE_NODENUM( state );
2032 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2033
2034 } /* end second pass */
2035
3dab1dad 2036 /* and now dump it out before we compress it */
2b8b4781
NC
2037 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2038 revcharmap,
2039 next_alloc, depth+1));
a3621e74 2040
a3621e74
YO
2041 {
2042 /*
2043 * Inplace compress the table.*
2044
2045 For sparse data sets the table constructed by the trie algorithm will
2046 be mostly 0/FAIL transitions or to put it another way mostly empty.
2047 (Note that leaf nodes will not contain any transitions.)
2048
2049 This algorithm compresses the tables by eliminating most such
2050 transitions, at the cost of a modest bit of extra work during lookup:
2051
2052 - Each states[] entry contains a .base field which indicates the
2053 index in the state[] array wheres its transition data is stored.
2054
3b753521 2055 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2056
2057 - If .base is nonzero then charid is added to it to find an entry in
2058 the trans array.
2059
2060 -If trans[states[state].base+charid].check!=state then the
2061 transition is taken to be a 0/Fail transition. Thus if there are fail
2062 transitions at the front of the node then the .base offset will point
2063 somewhere inside the previous nodes data (or maybe even into a node
2064 even earlier), but the .check field determines if the transition is
2065 valid.
2066
786e8c11 2067 XXX - wrong maybe?
a3621e74 2068 The following process inplace converts the table to the compressed
3b753521 2069 table: We first do not compress the root node 1,and mark all its
a3621e74 2070 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2071 allows us to do a DFA construction from the compressed table later,
2072 and ensures that any .base pointers we calculate later are greater
2073 than 0.
a3621e74
YO
2074
2075 - We set 'pos' to indicate the first entry of the second node.
2076
2077 - We then iterate over the columns of the node, finding the first and
2078 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2079 and set the .check pointers accordingly, and advance pos
2080 appropriately and repreat for the next node. Note that when we copy
2081 the next pointers we have to convert them from the original
2082 NODEIDX form to NODENUM form as the former is not valid post
2083 compression.
2084
2085 - If a node has no transitions used we mark its base as 0 and do not
2086 advance the pos pointer.
2087
2088 - If a node only has one transition we use a second pointer into the
2089 structure to fill in allocated fail transitions from other states.
2090 This pointer is independent of the main pointer and scans forward
2091 looking for null transitions that are allocated to a state. When it
2092 finds one it writes the single transition into the "hole". If the
786e8c11 2093 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2094
2095 - Once compressed we can Renew/realloc the structures to release the
2096 excess space.
2097
2098 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2099 specifically Fig 3.47 and the associated pseudocode.
2100
2101 demq
2102 */
a3b680e6 2103 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2104 U32 state, charid;
a3621e74 2105 U32 pos = 0, zp=0;
1e2e3d02 2106 trie->statecount = laststate;
a3621e74
YO
2107
2108 for ( state = 1 ; state < laststate ; state++ ) {
2109 U8 flag = 0;
a28509cc
AL
2110 const U32 stateidx = TRIE_NODEIDX( state );
2111 const U32 o_used = trie->trans[ stateidx ].check;
2112 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2113 trie->trans[ stateidx ].check = 0;
2114
2115 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2116 if ( flag || trie->trans[ stateidx + charid ].next ) {
2117 if ( trie->trans[ stateidx + charid ].next ) {
2118 if (o_used == 1) {
2119 for ( ; zp < pos ; zp++ ) {
2120 if ( ! trie->trans[ zp ].next ) {
2121 break;
2122 }
2123 }
2124 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2125 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2126 trie->trans[ zp ].check = state;
2127 if ( ++zp > pos ) pos = zp;
2128 break;
2129 }
2130 used--;
2131 }
2132 if ( !flag ) {
2133 flag = 1;
2134 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2135 }
2136 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2137 trie->trans[ pos ].check = state;
2138 pos++;
2139 }
2140 }
2141 }
cc601c31 2142 trie->lasttrans = pos + 1;
c944940b
JH
2143 trie->states = (reg_trie_state *)
2144 PerlMemShared_realloc( trie->states, laststate
2145 * sizeof(reg_trie_state) );
a3621e74 2146 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2147 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2148 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2149 (int)depth * 2 + 2,"",
2150 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2151 (IV)next_alloc,
2152 (IV)pos,
a3621e74
YO
2153 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2154 );
2155
2156 } /* end table compress */
2157 }
1e2e3d02
YO
2158 DEBUG_TRIE_COMPILE_MORE_r(
2159 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2160 (int)depth * 2 + 2, "",
2161 (UV)trie->statecount,
2162 (UV)trie->lasttrans)
2163 );
cc601c31 2164 /* resize the trans array to remove unused space */
c944940b
JH
2165 trie->trans = (reg_trie_trans *)
2166 PerlMemShared_realloc( trie->trans, trie->lasttrans
2167 * sizeof(reg_trie_trans) );
a3621e74 2168
3b753521 2169 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2170 U8 nodetype =(U8)(flags & 0xFF);
2171 char *str=NULL;
786e8c11 2172
07be1b83 2173#ifdef DEBUGGING
e62cc96a 2174 regnode *optimize = NULL;
7122b237
YO
2175#ifdef RE_TRACK_PATTERN_OFFSETS
2176
b57a0404
JH
2177 U32 mjd_offset = 0;
2178 U32 mjd_nodelen = 0;
7122b237
YO
2179#endif /* RE_TRACK_PATTERN_OFFSETS */
2180#endif /* DEBUGGING */
a3621e74 2181 /*
3dab1dad
YO
2182 This means we convert either the first branch or the first Exact,
2183 depending on whether the thing following (in 'last') is a branch
2184 or not and whther first is the startbranch (ie is it a sub part of
2185 the alternation or is it the whole thing.)
3b753521 2186 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2187 the whole branch sequence, including the first.
a3621e74 2188 */
3dab1dad 2189 /* Find the node we are going to overwrite */
7f69552c 2190 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2191 /* branch sub-chain */
3dab1dad 2192 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2193#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2194 DEBUG_r({
2195 mjd_offset= Node_Offset((convert));
2196 mjd_nodelen= Node_Length((convert));
2197 });
7122b237 2198#endif
7f69552c 2199 /* whole branch chain */
7122b237
YO
2200 }
2201#ifdef RE_TRACK_PATTERN_OFFSETS
2202 else {
7f69552c
YO
2203 DEBUG_r({
2204 const regnode *nop = NEXTOPER( convert );
2205 mjd_offset= Node_Offset((nop));
2206 mjd_nodelen= Node_Length((nop));
2207 });
07be1b83
YO
2208 }
2209 DEBUG_OPTIMISE_r(
2210 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2211 (int)depth * 2 + 2, "",
786e8c11 2212 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2213 );
7122b237 2214#endif
3dab1dad
YO
2215 /* But first we check to see if there is a common prefix we can
2216 split out as an EXACT and put in front of the TRIE node. */
2217 trie->startstate= 1;
55eed653 2218 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2219 U32 state;
1e2e3d02 2220 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2221 U32 ofs = 0;
8e11feef
RGS
2222 I32 idx = -1;
2223 U32 count = 0;
2224 const U32 base = trie->states[ state ].trans.base;
a3621e74 2225
3dab1dad 2226 if ( trie->states[state].wordnum )
8e11feef 2227 count = 1;
a3621e74 2228
8e11feef 2229 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2230 if ( ( base + ofs >= trie->uniquecharcount ) &&
2231 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2232 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2233 {
3dab1dad 2234 if ( ++count > 1 ) {
2b8b4781 2235 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2236 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2237 if ( state == 1 ) break;
3dab1dad
YO
2238 if ( count == 2 ) {
2239 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2240 DEBUG_OPTIMISE_r(
8e11feef
RGS
2241 PerlIO_printf(Perl_debug_log,
2242 "%*sNew Start State=%"UVuf" Class: [",
2243 (int)depth * 2 + 2, "",
786e8c11 2244 (UV)state));
be8e71aa 2245 if (idx >= 0) {
2b8b4781 2246 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2247 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2248
3dab1dad 2249 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2250 if ( folder )
2251 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2252 DEBUG_OPTIMISE_r(
f1f66076 2253 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2254 );
8e11feef
RGS
2255 }
2256 }
2257 TRIE_BITMAP_SET(trie,*ch);
2258 if ( folder )
2259 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2260 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2261 }
2262 idx = ofs;
2263 }
3dab1dad
YO
2264 }
2265 if ( count == 1 ) {
2b8b4781 2266 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2267 STRLEN len;
2268 char *ch = SvPV( *tmp, len );
de734bd5
A
2269 DEBUG_OPTIMISE_r({
2270 SV *sv=sv_newmortal();
8e11feef
RGS
2271 PerlIO_printf( Perl_debug_log,
2272 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2273 (int)depth * 2 + 2, "",
de734bd5
A
2274 (UV)state, (UV)idx,
2275 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2276 PL_colors[0], PL_colors[1],
2277 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2278 PERL_PV_ESCAPE_FIRSTCHAR
2279 )
2280 );
2281 });
3dab1dad
YO
2282 if ( state==1 ) {
2283 OP( convert ) = nodetype;
2284 str=STRING(convert);
2285 STR_LEN(convert)=0;
2286 }
c490c714
YO
2287 STR_LEN(convert) += len;
2288 while (len--)
de734bd5 2289 *str++ = *ch++;
8e11feef 2290 } else {
f9049ba1 2291#ifdef DEBUGGING
8e11feef
RGS
2292 if (state>1)
2293 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2294#endif
8e11feef
RGS
2295 break;
2296 }
2297 }
2e64971a 2298 trie->prefixlen = (state-1);
3dab1dad 2299 if (str) {
8e11feef 2300 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2301 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2302 trie->startstate = state;
07be1b83
YO
2303 trie->minlen -= (state - 1);
2304 trie->maxlen -= (state - 1);
33809eae
JH
2305#ifdef DEBUGGING
2306 /* At least the UNICOS C compiler choked on this
2307 * being argument to DEBUG_r(), so let's just have
2308 * it right here. */
2309 if (
2310#ifdef PERL_EXT_RE_BUILD
2311 1
2312#else
2313 DEBUG_r_TEST
2314#endif
2315 ) {
2316 regnode *fix = convert;
2317 U32 word = trie->wordcount;
2318 mjd_nodelen++;
2319 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2320 while( ++fix < n ) {
2321 Set_Node_Offset_Length(fix, 0, 0);
2322 }
2323 while (word--) {
2324 SV ** const tmp = av_fetch( trie_words, word, 0 );
2325 if (tmp) {
2326 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2327 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2328 else
2329 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2330 }
2331 }
2332 }
2333#endif
8e11feef
RGS
2334 if (trie->maxlen) {
2335 convert = n;
2336 } else {
3dab1dad 2337 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2338 DEBUG_r(optimize= n);
3dab1dad
YO
2339 }
2340 }
2341 }
a5ca303d
YO
2342 if (!jumper)
2343 jumper = last;
3dab1dad 2344 if ( trie->maxlen ) {
8e11feef
RGS
2345 NEXT_OFF( convert ) = (U16)(tail - convert);
2346 ARG_SET( convert, data_slot );
786e8c11
YO
2347 /* Store the offset to the first unabsorbed branch in
2348 jump[0], which is otherwise unused by the jump logic.
2349 We use this when dumping a trie and during optimisation. */
2350 if (trie->jump)
7f69552c 2351 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2352
6c48061a
YO
2353 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2354 * and there is a bitmap
2355 * and the first "jump target" node we found leaves enough room
2356 * then convert the TRIE node into a TRIEC node, with the bitmap
2357 * embedded inline in the opcode - this is hypothetically faster.
2358 */
2359 if ( !trie->states[trie->startstate].wordnum
2360 && trie->bitmap
2361 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2362 {
2363 OP( convert ) = TRIEC;
2364 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2365 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2366 trie->bitmap= NULL;
2367 } else
2368 OP( convert ) = TRIE;
a3621e74 2369
3dab1dad
YO
2370 /* store the type in the flags */
2371 convert->flags = nodetype;
a5ca303d
YO
2372 DEBUG_r({
2373 optimize = convert
2374 + NODE_STEP_REGNODE
2375 + regarglen[ OP( convert ) ];
2376 });
2377 /* XXX We really should free up the resource in trie now,
2378 as we won't use them - (which resources?) dmq */
3dab1dad 2379 }
a3621e74 2380 /* needed for dumping*/
e62cc96a 2381 DEBUG_r(if (optimize) {
07be1b83 2382 regnode *opt = convert;
bcdf7404 2383
e62cc96a 2384 while ( ++opt < optimize) {
07be1b83
YO
2385 Set_Node_Offset_Length(opt,0,0);
2386 }
786e8c11
YO
2387 /*
2388 Try to clean up some of the debris left after the
2389 optimisation.
a3621e74 2390 */
786e8c11 2391 while( optimize < jumper ) {
07be1b83 2392 mjd_nodelen += Node_Length((optimize));
a3621e74 2393 OP( optimize ) = OPTIMIZED;
07be1b83 2394 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2395 optimize++;
2396 }
07be1b83 2397 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2398 });
2399 } /* end node insert */
2e64971a
DM
2400
2401 /* Finish populating the prev field of the wordinfo array. Walk back
2402 * from each accept state until we find another accept state, and if
2403 * so, point the first word's .prev field at the second word. If the
2404 * second already has a .prev field set, stop now. This will be the
2405 * case either if we've already processed that word's accept state,
3b753521
FN
2406 * or that state had multiple words, and the overspill words were
2407 * already linked up earlier.
2e64971a
DM
2408 */
2409 {
2410 U16 word;
2411 U32 state;
2412 U16 prev;
2413
2414 for (word=1; word <= trie->wordcount; word++) {
2415 prev = 0;
2416 if (trie->wordinfo[word].prev)
2417 continue;
2418 state = trie->wordinfo[word].accept;
2419 while (state) {
2420 state = prev_states[state];
2421 if (!state)
2422 break;
2423 prev = trie->states[state].wordnum;
2424 if (prev)
2425 break;
2426 }
2427 trie->wordinfo[word].prev = prev;
2428 }
2429 Safefree(prev_states);
2430 }
2431
2432
2433 /* and now dump out the compressed format */
2434 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2435
55eed653 2436 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2437#ifdef DEBUGGING
2438 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2439 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2440#else
2441 SvREFCNT_dec(revcharmap);
07be1b83 2442#endif
786e8c11
YO
2443 return trie->jump
2444 ? MADE_JUMP_TRIE
2445 : trie->startstate>1
2446 ? MADE_EXACT_TRIE
2447 : MADE_TRIE;
2448}
2449
2450STATIC void
2451S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2452{
3b753521 2453/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2454
2455 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2456 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2457 ISBN 0-201-10088-6
2458
2459 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2460 suffix of the current state's 'word' that is also a proper prefix of another word in our
2461 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2462 the DFA not to have to restart after its tried and failed a word at a given point, it
2463 simply continues as though it had been matching the other word in the first place.
2464 Consider
2465 'abcdgu'=~/abcdefg|cdgu/
2466 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2467 fail, which would bring us to the state representing 'd' in the second word where we would
2468 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2469 */
2470 /* add a fail transition */
3251b653
NC
2471 const U32 trie_offset = ARG(source);
2472 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2473 U32 *q;
2474 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2475 const U32 numstates = trie->statecount;
786e8c11
YO
2476 const U32 ubound = trie->lasttrans + ucharcount;
2477 U32 q_read = 0;
2478 U32 q_write = 0;
2479 U32 charid;
2480 U32 base = trie->states[ 1 ].trans.base;
2481 U32 *fail;
2482 reg_ac_data *aho;
2483 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2484 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2485
2486 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2487#ifndef DEBUGGING
2488 PERL_UNUSED_ARG(depth);
2489#endif
2490
2491
2492 ARG_SET( stclass, data_slot );
c944940b 2493 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2494 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2495 aho->trie=trie_offset;
446bd890
NC
2496 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2497 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2498 Newxz( q, numstates, U32);
c944940b 2499 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2500 aho->refcount = 1;
2501 fail = aho->fail;
2502 /* initialize fail[0..1] to be 1 so that we always have
2503 a valid final fail state */
2504 fail[ 0 ] = fail[ 1 ] = 1;
2505
2506 for ( charid = 0; charid < ucharcount ; charid++ ) {
2507 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2508 if ( newstate ) {
2509 q[ q_write ] = newstate;
2510 /* set to point at the root */
2511 fail[ q[ q_write++ ] ]=1;
2512 }
2513 }
2514 while ( q_read < q_write) {
2515 const U32 cur = q[ q_read++ % numstates ];
2516 base = trie->states[ cur ].trans.base;
2517
2518 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2519 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2520 if (ch_state) {
2521 U32 fail_state = cur;
2522 U32 fail_base;
2523 do {
2524 fail_state = fail[ fail_state ];
2525 fail_base = aho->states[ fail_state ].trans.base;
2526 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2527
2528 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2529 fail[ ch_state ] = fail_state;
2530 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2531 {
2532 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2533 }
2534 q[ q_write++ % numstates] = ch_state;
2535 }
2536 }
2537 }
2538 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2539 when we fail in state 1, this allows us to use the
2540 charclass scan to find a valid start char. This is based on the principle
2541 that theres a good chance the string being searched contains lots of stuff
2542 that cant be a start char.
2543 */
2544 fail[ 0 ] = fail[ 1 ] = 0;
2545 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2546 PerlIO_printf(Perl_debug_log,
2547 "%*sStclass Failtable (%"UVuf" states): 0",
2548 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2549 );
786e8c11
YO
2550 for( q_read=1; q_read<numstates; q_read++ ) {
2551 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2552 }
2553 PerlIO_printf(Perl_debug_log, "\n");
2554 });
2555 Safefree(q);
2556 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2557}
2558
786e8c11 2559
a3621e74 2560/*
5d1c421c
JH
2561 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2562 * These need to be revisited when a newer toolchain becomes available.
2563 */
2564#if defined(__sparc64__) && defined(__GNUC__)
2565# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2566# undef SPARC64_GCC_WORKAROUND
2567# define SPARC64_GCC_WORKAROUND 1
2568# endif
2569#endif
2570
07be1b83 2571#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2572 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2573 SV * const mysv=sv_newmortal(); \
2574 regnode *Next = regnext(scan); \
2575 regprop(RExC_rx, mysv, scan); \
7f69552c 2576 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2577 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2578 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2579 }});
07be1b83 2580
1de06328 2581
bb914485
KW
2582/* The below joins as many adjacent EXACTish nodes as possible into a single
2583 * one, and looks for problematic sequences of characters whose folds vs.
2584 * non-folds have sufficiently different lengths, that the optimizer would be
2585 * fooled into rejecting legitimate matches of them, and the trie construction
2586 * code can't cope with them. The joining is only done if:
2587 * 1) there is room in the current conglomerated node to entirely contain the
2588 * next one.
2589 * 2) they are the exact same node type
2590 *
2591 * The adjacent nodes actually may be separated by NOTHING kind nodes, and
2592 * these get optimized out
2593 *
9d071ca8
KW
2594 * If there are problematic code sequences, *min_subtract is set to the delta
2595 * that the minimum size of the node can be less than its actual size. And,
2596 * the node type of the result is changed to reflect that it contains these
bb914485
KW
2597 * sequences.
2598 *
a0c4c608
KW
2599 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2600 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2601 *
bb914485
KW
2602 * This is as good a place as any to discuss the design of handling these
2603 * problematic sequences. It's been wrong in Perl for a very long time. There
2604 * are three code points in Unicode whose folded lengths differ so much from
2605 * the un-folded lengths that it causes problems for the optimizer and trie
2606 * construction. Why only these are problematic, and not others where lengths
2607 * also differ is something I (khw) do not understand. New versions of Unicode
2608 * might add more such code points. Hopefully the logic in fold_grind.t that
287722f3 2609 * figures out what to test (in part by verifying that each size-combination
bb914485 2610 * gets tested) will catch any that do come along, so they can be added to the
287722f3
KW
2611 * special handling below. The chances of new ones are actually rather small,
2612 * as most, if not all, of the world's scripts that have casefolding have
2613 * already been encoded by Unicode. Also, a number of Unicode's decisions were
2614 * made to allow compatibility with pre-existing standards, and almost all of
2615 * those have already been dealt with. These would otherwise be the most
2616 * likely candidates for generating further tricky sequences. In other words,
2617 * Unicode by itself is unlikely to add new ones unless it is for compatibility
a0c4c608 2618 * with pre-existing standards, and there aren't many of those left.
bb914485
KW
2619 *
2620 * The previous designs for dealing with these involved assigning a special
2621 * node for them. This approach doesn't work, as evidenced by this example:
a0c4c608 2622 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
bb914485
KW
2623 * Both these fold to "sss", but if the pattern is parsed to create a node of
2624 * that would match just the \xDF, it won't be able to handle the case where a
2625 * successful match would have to cross the node's boundary. The new approach
2626 * that hopefully generally solves the problem generates an EXACTFU_SS node
2627 * that is "sss".
2628 *
2629 * There are a number of components to the approach (a lot of work for just
2630 * three code points!):
2631 * 1) This routine examines each EXACTFish node that could contain the
9d071ca8
KW
2632 * problematic sequences. It returns in *min_subtract how much to
2633 * subtract from the the actual length of the string to get a real minimum
2634 * for one that could match it. This number is usually 0 except for the
2635 * problematic sequences. This delta is used by the caller to adjust the
2636 * min length of the match, and the delta between min and max, so that the
2637 * optimizer doesn't reject these possibilities based on size constraints.
f7926742
KW
2638 * 2) These sequences require special handling by the trie code, so it
2639 * changes the joined node type to ops for the trie's benefit, those new
2640 * ops being EXACTFU_SS and EXACTFU_TRICKYFOLD.
bb914485
KW
2641 * 3) This is sufficient for the two Greek sequences (described below), but
2642 * the one involving the Sharp s (\xDF) needs more. The node type
2643 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2644 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2645 * case where there is a possible fold length change. That means that a
2646 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2647 * itself with length changes, and so can be processed faster. regexec.c
2648 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2649 * is pre-folded by regcomp.c. This saves effort in regex matching.
2650 * However, probably mostly for historical reasons, the pre-folding isn't
a0c4c608
KW
2651 * done for non-UTF8 patterns (and it can't be for EXACTF and EXACTFL
2652 * nodes, as what they fold to isn't known until runtime.) The fold
2653 * possibilities for the non-UTF8 patterns are quite simple, except for
2654 * the sharp s. All the ones that don't involve a UTF-8 target string
2655 * are members of a fold-pair, and arrays are set up for all of them
2656 * that quickly find the other member of the pair. It might actually
2657 * be faster to pre-fold these, but it isn't currently done, except for
2658 * the sharp s. Code elsewhere in this file makes sure that it gets
2659 * folded to 'ss', even if the pattern isn't UTF-8. This avoids the
2660 * issues described in the next item.
bb914485
KW
2661 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2662 * 'ss' or not is not knowable at compile time. It will match iff the
2663 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2664 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2665 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2666 * described in item 3). An assumption that the optimizer part of
2667 * regexec.c (probably unwittingly) makes is that a character in the
2668 * pattern corresponds to at most a single character in the target string.
2669 * (And I do mean character, and not byte here, unlike other parts of the
2670 * documentation that have never been updated to account for multibyte
2671 * Unicode.) This assumption is wrong only in this case, as all other
2672 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2673 * virtue of having this file pre-fold UTF-8 patterns. I'm
2674 * reluctant to try to change this assumption, so instead the code punts.
9d071ca8
KW
2675 * This routine examines EXACTF nodes for the sharp s, and returns a
2676 * boolean indicating whether or not the node is an EXACTF node that
2677 * contains a sharp s. When it is true, the caller sets a flag that later
2678 * causes the optimizer in this file to not set values for the floating
2679 * and fixed string lengths, and thus avoids the optimizer code in
2680 * regexec.c that makes the invalid assumption. Thus, there is no
2681 * optimization based on string lengths for EXACTF nodes that contain the
2682 * sharp s. This only happens for /id rules (which means the pattern
2683 * isn't in UTF-8).
bb914485 2684 */
1de06328 2685
9d071ca8 2686#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2687 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2688 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2689
be8e71aa 2690STATIC U32
9d071ca8 2691S_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
2692 /* Merge several consecutive EXACTish nodes into one. */
2693 regnode *n = regnext(scan);
2694 U32 stringok = 1;
2695 regnode *next = scan + NODE_SZ_STR(scan);
2696 U32 merged = 0;
2697 U32 stopnow = 0;
2698#ifdef DEBUGGING
2699 regnode *stop = scan;
72f13be8 2700 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2701#else
d47053eb
RGS
2702 PERL_UNUSED_ARG(depth);
2703#endif
7918f24d
NC
2704
2705 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2706#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2707 PERL_UNUSED_ARG(flags);
2708 PERL_UNUSED_ARG(val);
07be1b83 2709#endif
07be1b83 2710 DEBUG_PEEP("join",scan,depth);
bb914485 2711
3f410cf6
KW
2712 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2713 * EXACT ones that are mergeable to the current one. */
2714 while (n
2715 && (PL_regkind[OP(n)] == NOTHING
2716 || (stringok && OP(n) == OP(scan)))
07be1b83 2717 && NEXT_OFF(n)
3f410cf6
KW
2718 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2719 {
07be1b83
YO
2720
2721 if (OP(n) == TAIL || n > next)
2722 stringok = 0;
2723 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2724 DEBUG_PEEP("skip:",n,depth);
2725 NEXT_OFF(scan) += NEXT_OFF(n);
2726 next = n + NODE_STEP_REGNODE;
2727#ifdef DEBUGGING
2728 if (stringok)
2729 stop = n;
2730#endif
2731 n = regnext(n);
2732 }
2733 else if (stringok) {
786e8c11 2734 const unsigned int oldl = STR_LEN(scan);
07be1b83 2735 regnode * const nnext = regnext(n);
b2230d39
KW
2736
2737 if (oldl + STR_LEN(n) > U8_MAX)
2738 break;
07be1b83
YO
2739
2740 DEBUG_PEEP("merg",n,depth);
07be1b83 2741 merged++;
b2230d39 2742
07be1b83
YO
2743 NEXT_OFF(scan) += NEXT_OFF(n);
2744 STR_LEN(scan) += STR_LEN(n);
2745 next = n + NODE_SZ_STR(n);
2746 /* Now we can overwrite *n : */
2747 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2748#ifdef DEBUGGING
2749 stop = next - 1;
2750#endif
2751 n = nnext;
2752 if (stopnow) break;
2753 }
2754
d47053eb
RGS
2755#ifdef EXPERIMENTAL_INPLACESCAN
2756 if (flags && !NEXT_OFF(n)) {
2757 DEBUG_PEEP("atch", val, depth);
2758 if (reg_off_by_arg[OP(n)]) {
2759 ARG_SET(n, val - n);
2760 }
2761 else {
2762 NEXT_OFF(n) = val - n;
2763 }
2764 stopnow = 1;
2765 }
07be1b83
YO
2766#endif
2767 }
2c2b7f86 2768
9d071ca8 2769 *min_subtract = 0;
f758bddf 2770 *has_exactf_sharp_s = FALSE;
f646642f 2771
3f410cf6
KW
2772 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2773 * can now analyze for sequences of problematic code points. (Prior to
2774 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2775 * hence missed). The sequences only happen in folding, hence for any
2776 * non-EXACT EXACTish node */
86d6fcad 2777 if (OP(scan) != EXACT) {
f758bddf
KW
2778 U8 *s;
2779 U8 * s0 = (U8*) STRING(scan);
2780 U8 * const s_end = s0 + STR_LEN(scan);
2781
2782 /* The below is perhaps overboard, but this allows us to save a test
2783 * each time through the loop at the expense of a mask. This is
2784 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2785 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2786 * This uses an exclusive 'or' to find that bit and then inverts it to
2787 * form a mask, with just a single 0, in the bit position where 'S' and
2788 * 's' differ. */
dbeb8947 2789 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
f758bddf
KW
2790 const U8 s_masked = 's' & S_or_s_mask;
2791
2792 /* One pass is made over the node's string looking for all the
2793 * possibilities. to avoid some tests in the loop, there are two main
2794 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2795 * non-UTF-8 */
2796 if (UTF) {
86d6fcad 2797
f758bddf
KW
2798 /* There are two problematic Greek code points in Unicode
2799 * casefolding
86d6fcad
KW
2800 *
2801 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2802 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2803 *
2804 * which casefold to
2805 *
2806 * Unicode UTF-8
2807 *
2808 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2809 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2810 *
2811 * This means that in case-insensitive matching (or "loose
2812 * matching", as Unicode calls it), an EXACTF of length six (the
2813 * UTF-8 encoded byte length of the above casefolded versions) can
2814 * match a target string of length two (the byte length of UTF-8
2815 * encoded U+0390 or U+03B0). This would rather mess up the
2816 * minimum length computation. (there are other code points that
2817 * also fold to these two sequences, but the delta is smaller)
2818 *
f758bddf
KW
2819 * If these sequences are found, the minimum length is decreased by
2820 * four (six minus two).
86d6fcad 2821 *
f758bddf
KW
2822 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2823 * LETTER SHARP S. We decrease the min length by 1 for each
2824 * occurrence of 'ss' found */
3f410cf6 2825
e294cc5d 2826#ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
f758bddf
KW
2827# define U390_first_byte 0xb4
2828 const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
2829# define U3B0_first_byte 0xb5
2830 const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
e294cc5d 2831#else
f758bddf
KW
2832# define U390_first_byte 0xce
2833 const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
2834# define U3B0_first_byte 0xcf
2835 const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
e294cc5d 2836#endif
f758bddf
KW
2837 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2838 yields a net of 0 */
2839 /* Examine the string for one of the problematic sequences */
2840 for (s = s0;
2841 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2842 * sequence we are looking for is 2 */
2843 s += UTF8SKIP(s))
86d6fcad 2844 {
bb914485 2845
f758bddf
KW
2846 /* Look for the first byte in each problematic sequence */
2847 switch (*s) {
2848 /* We don't have to worry about other things that fold to
2849 * 's' (such as the long s, U+017F), as all above-latin1
2850 * code points have been pre-folded */
2851 case 's':
2852 case 'S':
2853
a0c4c608
KW
2854 /* Current character is an 's' or 'S'. If next one is
2855 * as well, we have the dreaded sequence */
f758bddf
KW
2856 if (((*(s+1) & S_or_s_mask) == s_masked)
2857 /* These two node types don't have special handling
2858 * for 'ss' */
2859 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2860 {
9d071ca8 2861 *min_subtract += 1;
f758bddf
KW
2862 OP(scan) = EXACTFU_SS;
2863 s++; /* No need to look at this character again */
2864 }
2865 break;
2866
2867 case U390_first_byte:
2868 if (s_end - s >= len
2869
2870 /* The 1's are because are skipping comparing the
2871 * first byte */
2872 && memEQ(s + 1, U390_tail, len - 1))
2873 {
2874 goto greek_sequence;
2875 }
2876 break;
2877
2878 case U3B0_first_byte:
2879 if (! (s_end - s >= len
2880 && memEQ(s + 1, U3B0_tail, len - 1)))
2881 {
2882 break;
2883 }
2884 greek_sequence:
9d071ca8 2885 *min_subtract += 4;
f758bddf
KW
2886
2887 /* This can't currently be handled by trie's, so change
2888 * the node type to indicate this. If EXACTFA and
2889 * EXACTFL were ever to be handled by trie's, this
2890 * would have to be changed. If this node has already
2891 * been changed to EXACTFU_SS in this loop, leave it as
2892 * is. (I (khw) think it doesn't matter in regexec.c
2893 * for UTF patterns, but no need to change it */
2894 if (OP(scan) == EXACTFU) {
fab2782b 2895 OP(scan) = EXACTFU_TRICKYFOLD;
f758bddf
KW
2896 }
2897 s += 6; /* We already know what this sequence is. Skip
2898 the rest of it */
2899 break;
bb914485
KW
2900 }
2901 }
2902 }
f758bddf 2903 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
bb914485 2904
f758bddf
KW
2905 /* Here, the pattern is not UTF-8. We need to look only for the
2906 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2907 * in the final position. Otherwise we can stop looking 1 byte
2908 * earlier because have to find both the first and second 's' */
2909 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2910
2911 for (s = s0; s < upper; s++) {
2912 switch (*s) {
2913 case 'S':
2914 case 's':
2915 if (s_end - s > 1
2916 && ((*(s+1) & S_or_s_mask) == s_masked))
2917 {
9d071ca8 2918 *min_subtract += 1;
f758bddf
KW
2919
2920 /* EXACTF nodes need to know that the minimum
2921 * length changed so that a sharp s in the string
2922 * can match this ss in the pattern, but they
2923 * remain EXACTF nodes, as they are not trie'able,
2924 * so don't have to invent a new node type to
2925 * exclude them from the trie code */
2926 if (OP(scan) != EXACTF) {
2927 OP(scan) = EXACTFU_SS;
2928 }
2929 s++;
2930 }
2931 break;
2932 case LATIN_SMALL_LETTER_SHARP_S:
2933 if (OP(scan) == EXACTF) {
2934 *has_exactf_sharp_s = TRUE;
2935 }
2936 break;
86d6fcad
KW
2937 }
2938 }
2939 }
07be1b83 2940 }
3f410cf6 2941
07be1b83 2942#ifdef DEBUGGING
bb789b09
DM
2943 /* Allow dumping but overwriting the collection of skipped
2944 * ops and/or strings with fake optimized ops */
07be1b83
YO
2945 n = scan + NODE_SZ_STR(scan);
2946 while (n <= stop) {
bb789b09
DM
2947 OP(n) = OPTIMIZED;
2948 FLAGS(n) = 0;
2949 NEXT_OFF(n) = 0;
07be1b83
YO
2950 n++;
2951 }
2952#endif
2953 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2954 return stopnow;
2955}
2956
486ec47a 2957/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2958 Finds fixed substrings. */
2959
a0288114 2960/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2961 to the position after last scanned or to NULL. */
2962
40d049e4
YO
2963#define INIT_AND_WITHP \
2964 assert(!and_withp); \
2965 Newx(and_withp,1,struct regnode_charclass_class); \
2966 SAVEFREEPV(and_withp)
07be1b83 2967
b515a41d 2968/* this is a chain of data about sub patterns we are processing that
486ec47a 2969 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2970 we can simulate recursion without losing state. */
2971struct scan_frame;
2972typedef struct scan_frame {
2973 regnode *last; /* last node to process in this frame */
2974 regnode *next; /* next node to process when last is reached */
2975 struct scan_frame *prev; /*previous frame*/
2976 I32 stop; /* what stopparen do we use */
2977} scan_frame;
2978
304ee84b
YO
2979
2980#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2981
e1d1eefb
YO
2982#define CASE_SYNST_FNC(nAmE) \
2983case nAmE: \
2984 if (flags & SCF_DO_STCLASS_AND) { \
2985 for (value = 0; value < 256; value++) \
2986 if (!is_ ## nAmE ## _cp(value)) \
2987 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2988 } \
2989 else { \
2990 for (value = 0; value < 256; value++) \
2991 if (is_ ## nAmE ## _cp(value)) \
2992 ANYOF_BITMAP_SET(data->start_class, value); \
2993 } \
2994 break; \
2995case N ## nAmE: \
2996 if (flags & SCF_DO_STCLASS_AND) { \
2997 for (value = 0; value < 256; value++) \
2998 if (is_ ## nAmE ## _cp(value)) \
2999 ANYOF_BITMAP_CLEAR(data->start_class, value); \
3000 } \
3001 else { \
3002 for (value = 0; value < 256; value++) \
3003 if (!is_ ## nAmE ## _cp(value)) \
3004 ANYOF_BITMAP_SET(data->start_class, value); \
3005 } \
3006 break
3007
3008
3009
76e3520e 3010STATIC I32
40d049e4 3011S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 3012 I32 *minlenp, I32 *deltap,
40d049e4
YO
3013 regnode *last,
3014 scan_data_t *data,
3015 I32 stopparen,
3016 U8* recursed,
3017 struct regnode_charclass_class *and_withp,
3018 U32 flags, U32 depth)
c277df42
IZ
3019 /* scanp: Start here (read-write). */
3020 /* deltap: Write maxlen-minlen here. */
3021 /* last: Stop before this one. */
40d049e4
YO
3022 /* data: string data about the pattern */
3023 /* stopparen: treat close N as END */
3024 /* recursed: which subroutines have we recursed into */
3025 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 3026{
97aff369 3027 dVAR;
c277df42
IZ
3028 I32 min = 0, pars = 0, code;
3029 regnode *scan = *scanp, *next;
3030 I32 delta = 0;
3031 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 3032 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
3033 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3034 scan_data_t data_fake;
a3621e74 3035 SV *re_trie_maxbuff = NULL;
786e8c11 3036 regnode *first_non_open = scan;
e2e6a0f1 3037 I32 stopmin = I32_MAX;
8aa23a47 3038 scan_frame *frame = NULL;
a3621e74 3039 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 3040
7918f24d
NC
3041 PERL_ARGS_ASSERT_STUDY_CHUNK;
3042
13a24bad 3043#ifdef DEBUGGING
40d049e4 3044 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 3045#endif
40d049e4 3046
786e8c11 3047 if ( depth == 0 ) {
40d049e4 3048 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
3049 first_non_open=regnext(first_non_open);
3050 }
3051
b81d288d 3052
8aa23a47
YO
3053 fake_study_recurse:
3054 while ( scan && OP(scan) != END && scan < last ){
9d071ca8
KW
3055 UV min_subtract = 0; /* How much to subtract from the minimum node
3056 length to get a real minimum (because the
3057 folded version may be shorter) */
f758bddf 3058 bool has_exactf_sharp_s = FALSE;
8aa23a47 3059 /* Peephole optimizer: */
304ee84b 3060 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3061 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3062
3063 /* Its not clear to khw or hv why this is done here, and not in the
3064 * clauses that deal with EXACT nodes. khw's guess is that it's
3065 * because of a previous design */
9d071ca8 3066 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3067
3068 /* Follow the next-chain of the current node and optimize
3069 away all the NOTHINGs from it. */
3070 if (OP(scan) != CURLYX) {
3071 const int max = (reg_off_by_arg[OP(scan)]
3072 ? I32_MAX
3073 /* I32 may be smaller than U16 on CRAYs! */
3074 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3075 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3076 int noff;
3077 regnode *n = scan;
686b73d4 3078
8aa23a47
YO
3079 /* Skip NOTHING and LONGJMP. */
3080 while ((n = regnext(n))
3081 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3082 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3083 && off + noff < max)
3084 off += noff;
3085 if (reg_off_by_arg[OP(scan)])
3086 ARG(scan) = off;
3087 else
3088 NEXT_OFF(scan) = off;
3089 }
a3621e74 3090
c277df42 3091
8aa23a47
YO
3092
3093 /* The principal pseudo-switch. Cannot be a switch, since we
3094 look into several different things. */
3095 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3096 || OP(scan) == IFTHEN) {
3097 next = regnext(scan);
3098 code = OP(scan);
3099 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3100
8aa23a47
YO
3101 if (OP(next) == code || code == IFTHEN) {
3102 /* NOTE - There is similar code to this block below for handling
3103 TRIE nodes on a re-study. If you change stuff here check there
3104 too. */
3105 I32 max1 = 0, min1 = I32_MAX, num = 0;
3106 struct regnode_charclass_class accum;
3107 regnode * const startbranch=scan;
686b73d4 3108
8aa23a47 3109 if (flags & SCF_DO_SUBSTR)
304ee84b 3110 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3111 if (flags & SCF_DO_STCLASS)
e755fd73 3112 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3113
3114 while (OP(scan) == code) {
3115 I32 deltanext, minnext, f = 0, fake;
3116 struct regnode_charclass_class this_class;
3117
3118 num++;
3119 data_fake.flags = 0;
3120 if (data) {
3121 data_fake.whilem_c = data->whilem_c;
3122 data_fake.last_closep = data->last_closep;
3123 }
3124 else
3125 data_fake.last_closep = &fake;
58e23c8d
YO
3126
3127 data_fake.pos_delta = delta;
8aa23a47
YO
3128 next = regnext(scan);
3129 scan = NEXTOPER(scan);
3130 if (code != BRANCH)
c277df42 3131 scan = NEXTOPER(scan);
8aa23a47 3132 if (flags & SCF_DO_STCLASS) {
e755fd73 3133 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3134 data_fake.start_class = &this_class;
3135 f = SCF_DO_STCLASS_AND;
58e23c8d 3136 }
8aa23a47
YO
3137 if (flags & SCF_WHILEM_VISITED_POS)
3138 f |= SCF_WHILEM_VISITED_POS;
3139
3140 /* we suppose the run is continuous, last=next...*/
3141 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3142 next, &data_fake,
3143 stopparen, recursed, NULL, f,depth+1);
3144 if (min1 > minnext)
3145 min1 = minnext;
3146 if (max1 < minnext + deltanext)
3147 max1 = minnext + deltanext;
3148 if (deltanext == I32_MAX)
3149 is_inf = is_inf_internal = 1;
3150 scan = next;
3151 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3152 pars++;
3153 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3154 if ( stopmin > minnext)
3155 stopmin = min + min1;
3156 flags &= ~SCF_DO_SUBSTR;
3157 if (data)
3158 data->flags |= SCF_SEEN_ACCEPT;
3159 }
3160 if (data) {
3161 if (data_fake.flags & SF_HAS_EVAL)
3162 data->flags |= SF_HAS_EVAL;
3163 data->whilem_c = data_fake.whilem_c;
3dab1dad 3164 }
8aa23a47 3165 if (flags & SCF_DO_STCLASS)
3fffb88a 3166 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3167 }
3168 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3169 min1 = 0;
3170 if (flags & SCF_DO_SUBSTR) {
3171 data->pos_min += min1;
3172 data->pos_delta += max1 - min1;
3173 if (max1 != min1 || is_inf)
3174 data->longest = &(data->longest_float);
3175 }
3176 min += min1;
3177 delta += max1 - min1;
3178 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3179 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3180 if (min1) {
3181 cl_and(data->start_class, and_withp);
3182 flags &= ~SCF_DO_STCLASS;
653099ff 3183 }
8aa23a47
YO
3184 }
3185 else if (flags & SCF_DO_STCLASS_AND) {
3186 if (min1) {
3187 cl_and(data->start_class, &accum);
3188 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3189 }
8aa23a47
YO
3190 else {
3191 /* Switch to OR mode: cache the old value of
3192 * data->start_class */
3193 INIT_AND_WITHP;
3194 StructCopy(data->start_class, and_withp,
3195 struct regnode_charclass_class);
3196 flags &= ~SCF_DO_STCLASS_AND;
3197 StructCopy(&accum, data->start_class,
3198 struct regnode_charclass_class);
3199 flags |= SCF_DO_STCLASS_OR;
3200 data->start_class->flags |= ANYOF_EOS;
de0c8cb8 3201 }
8aa23a47 3202 }
a3621e74 3203
8aa23a47
YO
3204 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3205 /* demq.
a3621e74 3206
8aa23a47
YO
3207 Assuming this was/is a branch we are dealing with: 'scan' now
3208 points at the item that follows the branch sequence, whatever
3209 it is. We now start at the beginning of the sequence and look
3210 for subsequences of
a3621e74 3211
8aa23a47
YO
3212 BRANCH->EXACT=>x1
3213 BRANCH->EXACT=>x2
3214 tail
a3621e74 3215
8aa23a47 3216 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3217
486ec47a 3218 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3219 element into a trie and then add the subsequent branch exact
3220 strings to the trie.
a3621e74 3221
8aa23a47 3222 We have two cases
a3621e74 3223
3b753521 3224 1. patterns where the whole set of branches can be converted.
a3621e74 3225
8aa23a47 3226 2. patterns where only a subset can be converted.
a3621e74 3227
8aa23a47
YO
3228 In case 1 we can replace the whole set with a single regop
3229 for the trie. In case 2 we need to keep the start and end
3b753521 3230 branches so
a3621e74 3231
8aa23a47
YO
3232 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3233 becomes BRANCH TRIE; BRANCH X;
786e8c11 3234
8aa23a47
YO
3235 There is an additional case, that being where there is a
3236 common prefix, which gets split out into an EXACT like node
3237 preceding the TRIE node.
a3621e74 3238
8aa23a47
YO
3239 If x(1..n)==tail then we can do a simple trie, if not we make
3240 a "jump" trie, such that when we match the appropriate word
486ec47a 3241 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3242 a nested if into a case structure of sorts.
b515a41d 3243
8aa23a47 3244 */
686b73d4 3245
8aa23a47
YO
3246 int made=0;
3247 if (!re_trie_maxbuff) {
3248 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3249 if (!SvIOK(re_trie_maxbuff))
3250 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3251 }
3252 if ( SvIV(re_trie_maxbuff)>=0 ) {
3253 regnode *cur;
3254 regnode *first = (regnode *)NULL;
3255 regnode *last = (regnode *)NULL;
3256 regnode *tail = scan;
fab2782b 3257 U8 trietype = 0;
8aa23a47 3258 U32 count=0;
a3621e74
YO
3259
3260#ifdef DEBUGGING
8aa23a47 3261 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3262#endif
8aa23a47
YO
3263 /* var tail is used because there may be a TAIL
3264 regop in the way. Ie, the exacts will point to the
3265 thing following the TAIL, but the last branch will
3266 point at the TAIL. So we advance tail. If we
3267 have nested (?:) we may have to move through several
3268 tails.
3269 */
3270
3271 while ( OP( tail ) == TAIL ) {
3272 /* this is the TAIL generated by (?:) */
3273 tail = regnext( tail );
3274 }
a3621e74 3275
8aa23a47 3276
df826430 3277 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3278 regprop(RExC_rx, mysv, tail );
3279 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3280 (int)depth * 2 + 2, "",
3281 "Looking for TRIE'able sequences. Tail node is: ",
3282 SvPV_nolen_const( mysv )
3283 );
3284 });
3285
3286 /*
3287
fab2782b
YO
3288 Step through the branches
3289 cur represents each branch,
3290 noper is the first thing to be matched as part of that branch
3291 noper_next is the regnext() of that node.
3292
3293 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3294 via a "jump trie" but we also support building with NOJUMPTRIE,
3295 which restricts the trie logic to structures like /FOO|BAR/.
3296
3297 If noper is a trieable nodetype then the branch is a possible optimization
3298 target. If we are building under NOJUMPTRIE then we require that noper_next
3299 is the same as scan (our current position in the regex program).
3300
3301 Once we have two or more consecutive such branches we can create a
3302 trie of the EXACT's contents and stitch it in place into the program.
3303
3304 If the sequence represents all of the branches in the alternation we
3305 replace the entire thing with a single TRIE node.
3306
3307 Otherwise when it is a subsequence we need to stitch it in place and
3308 replace only the relevant branches. This means the first branch has
3309 to remain as it is used by the alternation logic, and its next pointer,
3310 and needs to be repointed at the item on the branch chain following
3311 the last branch we have optimized away.
3312
3313 This could be either a BRANCH, in which case the subsequence is internal,
3314 or it could be the item following the branch sequence in which case the
3315 subsequence is at the end (which does not necessarily mean the first node
3316 is the start of the alternation).
3317
3318 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3319
3320 optype | trietype
3321 ----------------+-----------
3322 NOTHING | NOTHING
3323 EXACT | EXACT
3324 EXACTFU | EXACTFU
3325 EXACTFU_SS | EXACTFU
3326 EXACTFU_TRICKYFOLD | EXACTFU
3327 EXACTFA | 0
3328
8aa23a47
YO
3329
3330 */
fab2782b
YO
3331#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3332 ( EXACT == (X) ) ? EXACT : \
3333 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3334 0 )
8aa23a47
YO
3335
3336 /* dont use tail as the end marker for this traverse */
3337 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3338 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3339 U8 noper_type = OP( noper );
3340 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3341#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3342 regnode * const noper_next = regnext( noper );
df826430
YO
3343 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3344 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3345#endif
3346
df826430 3347 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3348 regprop(RExC_rx, mysv, cur);
3349 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3350 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3351
3352 regprop(RExC_rx, mysv, noper);
3353 PerlIO_printf( Perl_debug_log, " -> %s",
3354 SvPV_nolen_const(mysv));
3355
3356 if ( noper_next ) {
3357 regprop(RExC_rx, mysv, noper_next );
3358 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3359 SvPV_nolen_const(mysv));
3360 }
df826430
YO
3361 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3362 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3363 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3364 );
8aa23a47 3365 });
fab2782b
YO
3366
3367 /* Is noper a trieable nodetype that can be merged with the
3368 * current trie (if there is one)? */
3369 if ( noper_trietype
3370 &&
3371 (
df826430
YO
3372 ( noper_trietype == NOTHING)
3373 || ( trietype == NOTHING )
a40630bf 3374 || ( trietype == noper_trietype )
fab2782b 3375 )
786e8c11 3376#ifdef NOJUMPTRIE
8aa23a47 3377 && noper_next == tail
786e8c11 3378#endif
8aa23a47
YO
3379 && count < U16_MAX)
3380 {
fab2782b
YO
3381 /* Handle mergable triable node
3382 * Either we are the first node in a new trieable sequence,
3383 * in which case we do some bookkeeping, otherwise we update
3384 * the end pointer. */
fab2782b 3385 if ( !first ) {
3b6759a6 3386 first = cur;
df826430
YO
3387 if ( noper_trietype == NOTHING ) {
3388#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3389 regnode * const noper_next = regnext( noper );
3b6759a6 3390 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3391 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3392#endif
3393
190c1910 3394 if ( noper_next_trietype ) {
df826430 3395 trietype = noper_next_trietype;
190c1910
YO
3396 } else if (noper_next_type) {
3397 /* a NOTHING regop is 1 regop wide. We need at least two
3398 * for a trie so we can't merge this in */
3399 first = NULL;
3400 }
3401 } else {
3402 trietype = noper_trietype;
3b6759a6 3403 }
8aa23a47 3404 } else {
fab2782b
YO
3405 if ( trietype == NOTHING )
3406 trietype = noper_trietype;
8aa23a47
YO
3407 last = cur;
3408 }
df826430
YO
3409 if (first)
3410 count++;
fab2782b
YO
3411 } /* end handle mergable triable node */
3412 else {
3413 /* handle unmergable node -
3414 * noper may either be a triable node which can not be tried
3415 * together with the current trie, or a non triable node */
729aaeb5
YO
3416 if ( last ) {
3417 /* If last is set and trietype is not NOTHING then we have found
3418 * at least two triable branch sequences in a row of a similar
3419 * trietype so we can turn them into a trie. If/when we
3420 * allow NOTHING to start a trie sequence this condition will be
3421 * required, and it isn't expensive so we leave it in for now. */
3422 if ( trietype != NOTHING )
3423 make_trie( pRExC_state,
3424 startbranch, first, cur, tail, count,
3425 trietype, depth+1 );
fab2782b 3426 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3427 }
fab2782b 3428 if ( noper_trietype
786e8c11 3429#ifdef NOJUMPTRIE
8aa23a47 3430 && noper_next == tail
786e8c11 3431#endif
8aa23a47 3432 ){
fab2782b 3433 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3434 count = 1;
3435 first = cur;
fab2782b
YO
3436 trietype = noper_trietype;
3437 } else if (first) {
3438 /* if we already saw a first but the current node is not triable then we have
3439 * to reset the first information. */
8aa23a47
YO
3440 count = 0;
3441 first = NULL;
fab2782b 3442 trietype = 0;
8aa23a47 3443 }
fab2782b
YO
3444 } /* end handle unmergable node */
3445 } /* loop over branches */
df826430 3446 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3447 regprop(RExC_rx, mysv, cur);
3448 PerlIO_printf( Perl_debug_log,
3449 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3450 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3451
3452 });
3b6759a6
YO
3453 if ( last ) {
3454 if ( trietype != NOTHING ) {
3455 /* the last branch of the sequence was part of a trie,
3456 * so we have to construct it here outside of the loop
3457 */
3458 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3459#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3460 if ( ((made == MADE_EXACT_TRIE &&
3461 startbranch == first)
3462 || ( first_non_open == first )) &&
3463 depth==0 ) {
3464 flags |= SCF_TRIE_RESTUDY;
3465 if ( startbranch == first
3466 && scan == tail )
3467 {
3468 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3469 }
8aa23a47 3470 }
3dab1dad 3471#endif
3b6759a6
YO
3472 } else {
3473 /* at this point we know whatever we have is a NOTHING sequence/branch
3474 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3475 */
3476 if ( startbranch == first ) {
3477 regnode *opt;
3478 /* the entire thing is a NOTHING sequence, something like this:
3479 * (?:|) So we can turn it into a plain NOTHING op. */
3480 DEBUG_TRIE_COMPILE_r({
3481 regprop(RExC_rx, mysv, cur);
3482 PerlIO_printf( Perl_debug_log,
3483 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3484 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3485
3486 });
3487 OP(startbranch)= NOTHING;
3488 NEXT_OFF(startbranch)= tail - startbranch;
3489 for ( opt= startbranch + 1; opt < tail ; opt++ )
3490 OP(opt)= OPTIMIZED;
3491 }
3492 }
fab2782b
YO
3493 } /* end if ( last) */
3494 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3495
3496 } /* do trie */
3497
653099ff 3498 }
8aa23a47
YO
3499 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3500 scan = NEXTOPER(NEXTOPER(scan));
3501 } else /* single branch is optimized. */
3502 scan = NEXTOPER(scan);
3503 continue;
3504 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3505 scan_frame *newframe = NULL;
3506 I32 paren;
3507 regnode *start;
3508 regnode *end;
3509
3510 if (OP(scan) != SUSPEND) {
3511 /* set the pointer */
3512 if (OP(scan) == GOSUB) {
3513 paren = ARG(scan);
3514 RExC_recurse[ARG2L(scan)] = scan;
3515 start = RExC_open_parens[paren-1];
3516 end = RExC_close_parens[paren-1];
3517 } else {
3518 paren = 0;
f8fc2ecf 3519 start = RExC_rxi->program + 1;
8aa23a47
YO
3520 end = RExC_opend;
3521 }
3522 if (!recursed) {
3523 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3524 SAVEFREEPV(recursed);
3525 }
3526 if (!PAREN_TEST(recursed,paren+1)) {
3527 PAREN_SET(recursed,paren+1);
3528 Newx(newframe,1,scan_frame);
3529 } else {
3530 if (flags & SCF_DO_SUBSTR) {
304ee84b 3531 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3532 data->longest = &(data->longest_float);
3533 }
3534 is_inf = is_inf_internal = 1;
3535 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3536 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3537 flags &= ~SCF_DO_STCLASS;
3538 }
3539 } else {
3540 Newx(newframe,1,scan_frame);
3541 paren = stopparen;
3542 start = scan+2;
3543 end = regnext(scan);
3544 }
3545 if (newframe) {
3546 assert(start);
3547 assert(end);
3548 SAVEFREEPV(newframe);
3549 newframe->next = regnext(scan);
3550 newframe->last = last;
3551 newframe->stop = stopparen;
3552 newframe->prev = frame;
3553
3554 frame = newframe;
3555 scan = start;
3556 stopparen = paren;
3557 last = end;
3558
3559 continue;
3560 }
3561 }
3562 else if (OP(scan) == EXACT) {
3563 I32 l = STR_LEN(scan);
3564 UV uc;
3565 if (UTF) {
3566 const U8 * const s = (U8*)STRING(scan);
4b88fb76 3567 uc = utf8_to_uvchr_buf(s, s + l, NULL);
8aa23a47 3568 l = utf8_length(s, s + l);
8aa23a47
YO
3569 } else {
3570 uc = *((U8*)STRING(scan));
3571 }
3572 min += l;
3573 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3574 /* The code below prefers earlier match for fixed
3575 offset, later match for variable offset. */
3576 if (data->last_end == -1) { /* Update the start info. */
3577 data->last_start_min = data->pos_min;
3578 data->last_start_max = is_inf
3579 ? I32_MAX : data->pos_min + data->pos_delta;
b515a41d 3580 }
8aa23a47
YO
3581 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3582 if (UTF)
3583 SvUTF8_on(data->last_found);
3584 {
3585 SV * const sv = data->last_found;
3586 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3587 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3588 if (mg && mg->mg_len >= 0)
3589 mg->mg_len += utf8_length((U8*)STRING(scan),
3590 (U8*)STRING(scan)+STR_LEN(scan));
b515a41d 3591 }
8aa23a47
YO
3592 data->last_end = data->pos_min + l;
3593 data->pos_min += l; /* As in the first entry. */
3594 data->flags &= ~SF_BEFORE_EOL;
3595 }
3596 if (flags & SCF_DO_STCLASS_AND) {
3597 /* Check whether it is compatible with what we know already! */
3598 int compat = 1;
3599