This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
Add utility and .h for character's UTF-8
[perl5.git] / regcomp.c
CommitLineData
a0d0e21e
LW
1/* regcomp.c
2 */
3
4/*
4ac71550
TC
5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
6 *
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
a0d0e21e
LW
8 */
9
61296642
DM
10/* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
166f8a29 12 * a regular expression.
e4a054ea
DM
13 *
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
166f8a29
DM
18 */
19
a687059c
LW
20/* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
22 */
23
24/* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
27 */
28
e50aee73 29/* The names of the functions have been changed from regcomp and
3b753521 30 * regexec to pregcomp and pregexec in order to avoid conflicts
e50aee73
AD
31 * with the POSIX routines of the same names.
32*/
33
b9d5759e 34#ifdef PERL_EXT_RE_BUILD
54df2634 35#include "re_top.h"
b81d288d 36#endif
56953603 37
a687059c 38/*
e50aee73 39 * pregcomp and pregexec -- regsub and regerror are not used in perl
a687059c
LW
40 *
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
43 *
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
47 *
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
50 * from defects in it.
51 *
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
54 *
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
57 *
58 *
59 **** Alterations to Henry's code are...
60 ****
4bb101f2 61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
1129b882
NC
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
a687059c 64 ****
9ef589d8
LW
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
67
a687059c
LW
68 *
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
72 */
73#include "EXTERN.h"
864dbfa3 74#define PERL_IN_REGCOMP_C
a687059c 75#include "perl.h"
d06ea78c 76
acfe0abc 77#ifndef PERL_IN_XSUB_RE
d06ea78c
GS
78# include "INTERN.h"
79#endif
c277df42
IZ
80
81#define REG_COMP_C
54df2634
NC
82#ifdef PERL_IN_XSUB_RE
83# include "re_comp.h"
40b6423c 84extern const struct regexp_engine my_reg_engine;
54df2634
NC
85#else
86# include "regcomp.h"
87#endif
a687059c 88
04e98a4d 89#include "dquote_static.c"
26faadbd 90#include "charclass_invlists.h"
81e983c1 91#include "inline_invlist.c"
61dad979 92#include "utf8_strings.h"
04e98a4d 93
94dc5c2d 94#define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
26faadbd 95#define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94dc5c2d 96
d4cce5f1 97#ifdef op
11343788 98#undef op
d4cce5f1 99#endif /* op */
11343788 100
fe14fcc3 101#ifdef MSDOS
7e4e8c89 102# if defined(BUGGY_MSC6)
fe14fcc3 103 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
7e4e8c89 104# pragma optimize("a",off)
fe14fcc3 105 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
7e4e8c89
NC
106# pragma optimize("w",on )
107# endif /* BUGGY_MSC6 */
fe14fcc3
LW
108#endif /* MSDOS */
109
a687059c
LW
110#ifndef STATIC
111#define STATIC static
112#endif
113
b1603ef8 114
830247a4 115typedef struct RExC_state_t {
514a91f1
DM
116 U32 flags; /* RXf_* are we folding, multilining? */
117 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 118 char *precomp; /* uncompiled string. */
288b8c02 119 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf
YO
120 regexp *rx; /* perl core regexp structure */
121 regexp_internal *rxi; /* internal data for regexp object pprivate field */
fac92740 122 char *start; /* Start of input for compile */
830247a4
IZ
123 char *end; /* End of input for compile */
124 char *parse; /* Input-scan pointer. */
125 I32 whilem_seen; /* number of WHILEM in this expr */
fac92740 126 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 127 regnode *emit_bound; /* First regnode outside of the allocated space */
ffc61ed2 128 regnode *emit; /* Code-emit pointer; &regdummy = don't = compiling */
830247a4
IZ
129 I32 naughty; /* How bad is this pattern? */
130 I32 sawback; /* Did we see \1, ...? */
131 U32 seen;
132 I32 size; /* Code size. */
c74340f9
YO
133 I32 npar; /* Capture buffer count, (OPEN). */
134 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 135 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
136 I32 extralen;
137 I32 seen_zerolen;
40d049e4
YO
138 regnode **open_parens; /* pointers to open parens */
139 regnode **close_parens; /* pointers to close parens */
140 regnode *opend; /* END node in program */
02daf0ab
YO
141 I32 utf8; /* whether the pattern is utf8 or not */
142 I32 orig_utf8; /* whether the pattern was originally in utf8 */
143 /* XXX use this for future optimisation of case
144 * where pattern must be upgraded to utf8. */
e40e74fe
KW
145 I32 uni_semantics; /* If a d charset modifier should use unicode
146 rules, even if the pattern is not in
147 utf8 */
81714fb9 148 HV *paren_names; /* Paren names */
1f1031fe 149
40d049e4
YO
150 regnode **recurse; /* Recurse regops */
151 I32 recurse_count; /* Number of recurse regops */
b57e4118 152 I32 in_lookbehind;
4624b182 153 I32 contains_locale;
bb3f3ed2 154 I32 override_recoding;
3d2bd50a 155 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 156 within pattern */
b1603ef8
DM
157 int num_code_blocks; /* size of code_blocks[] */
158 int code_index; /* next code_blocks[] slot */
830247a4
IZ
159#if ADD_TO_REGEXEC
160 char *starttry; /* -Dr: where regtry was called. */
161#define RExC_starttry (pRExC_state->starttry)
162#endif
d24ca0c5 163 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 164#ifdef DEBUGGING
be8e71aa 165 const char *lastparse;
3dab1dad 166 I32 lastnum;
1f1031fe 167 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
168#define RExC_lastparse (pRExC_state->lastparse)
169#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 170#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 171#endif
830247a4
IZ
172} RExC_state_t;
173
e2509266 174#define RExC_flags (pRExC_state->flags)
514a91f1 175#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 176#define RExC_precomp (pRExC_state->precomp)
288b8c02 177#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 178#define RExC_rx (pRExC_state->rx)
f8fc2ecf 179#define RExC_rxi (pRExC_state->rxi)
fac92740 180#define RExC_start (pRExC_state->start)
830247a4
IZ
181#define RExC_end (pRExC_state->end)
182#define RExC_parse (pRExC_state->parse)
183#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
184#ifdef RE_TRACK_PATTERN_OFFSETS
185#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
186#endif
830247a4 187#define RExC_emit (pRExC_state->emit)
fac92740 188#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 189#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
190#define RExC_naughty (pRExC_state->naughty)
191#define RExC_sawback (pRExC_state->sawback)
192#define RExC_seen (pRExC_state->seen)
193#define RExC_size (pRExC_state->size)
194#define RExC_npar (pRExC_state->npar)
e2e6a0f1 195#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
196#define RExC_extralen (pRExC_state->extralen)
197#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 198#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 199#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 200#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
201#define RExC_open_parens (pRExC_state->open_parens)
202#define RExC_close_parens (pRExC_state->close_parens)
203#define RExC_opend (pRExC_state->opend)
81714fb9 204#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
205#define RExC_recurse (pRExC_state->recurse)
206#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 207#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 208#define RExC_contains_locale (pRExC_state->contains_locale)
bb3f3ed2 209#define RExC_override_recoding (pRExC_state->override_recoding)
830247a4 210
cde0cee5 211
a687059c
LW
212#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
213#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
214 ((*s) == '{' && regcurly(s)))
a687059c 215
35c8bce7
LW
216#ifdef SPSTART
217#undef SPSTART /* dratted cpp namespace... */
218#endif
a687059c
LW
219/*
220 * Flags to be passed up and down.
221 */
a687059c 222#define WORST 0 /* Worst case. */
a3b492c3 223#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 224
18890cbd 225/* Simple enough to be STAR/PLUS operand; in an EXACT node must be a single
cf8c372d
KW
226 * character, and if utf8, must be invariant. Note that this is not the same
227 * thing as REGNODE_SIMPLE */
fda99bee 228#define SIMPLE 0x02
a3b492c3
YO
229#define SPSTART 0x04 /* Starts with * or +. */
230#define TRYAGAIN 0x08 /* Weeded out a declaration. */
231#define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
a687059c 232
3dab1dad
YO
233#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
234
07be1b83
YO
235/* whether trie related optimizations are enabled */
236#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
237#define TRIE_STUDY_OPT
786e8c11 238#define FULL_TRIE_STUDY
07be1b83
YO
239#define TRIE_STCLASS
240#endif
1de06328
YO
241
242
40d049e4
YO
243
244#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
245#define PBITVAL(paren) (1 << ((paren) & 7))
246#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
247#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
248#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
249
bbd61b5f
KW
250/* If not already in utf8, do a longjmp back to the beginning */
251#define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
252#define REQUIRE_UTF8 STMT_START { \
253 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
254 } STMT_END
40d049e4 255
1de06328
YO
256/* About scan_data_t.
257
258 During optimisation we recurse through the regexp program performing
259 various inplace (keyhole style) optimisations. In addition study_chunk
260 and scan_commit populate this data structure with information about
261 what strings MUST appear in the pattern. We look for the longest
3b753521 262 string that must appear at a fixed location, and we look for the
1de06328
YO
263 longest string that may appear at a floating location. So for instance
264 in the pattern:
265
266 /FOO[xX]A.*B[xX]BAR/
267
268 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
269 strings (because they follow a .* construct). study_chunk will identify
270 both FOO and BAR as being the longest fixed and floating strings respectively.
271
272 The strings can be composites, for instance
273
274 /(f)(o)(o)/
275
276 will result in a composite fixed substring 'foo'.
277
278 For each string some basic information is maintained:
279
280 - offset or min_offset
281 This is the position the string must appear at, or not before.
282 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
283 characters must match before the string we are searching for.
284 Likewise when combined with minlenp and the length of the string it
1de06328
YO
285 tells us how many characters must appear after the string we have
286 found.
287
288 - max_offset
289 Only used for floating strings. This is the rightmost point that
3b753521 290 the string can appear at. If set to I32 max it indicates that the
1de06328
YO
291 string can occur infinitely far to the right.
292
293 - minlenp
294 A pointer to the minimum length of the pattern that the string
295 was found inside. This is important as in the case of positive
296 lookahead or positive lookbehind we can have multiple patterns
297 involved. Consider
298
299 /(?=FOO).*F/
300
301 The minimum length of the pattern overall is 3, the minimum length
302 of the lookahead part is 3, but the minimum length of the part that
303 will actually match is 1. So 'FOO's minimum length is 3, but the
304 minimum length for the F is 1. This is important as the minimum length
305 is used to determine offsets in front of and behind the string being
306 looked for. Since strings can be composites this is the length of the
486ec47a 307 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
308 the length is calculated by study_chunk, so that the minimum lengths
309 are not known until the full pattern has been compiled, thus the
310 pointer to the value.
311
312 - lookbehind
313
314 In the case of lookbehind the string being searched for can be
315 offset past the start point of the final matching string.
316 If this value was just blithely removed from the min_offset it would
317 invalidate some of the calculations for how many chars must match
318 before or after (as they are derived from min_offset and minlen and
319 the length of the string being searched for).
320 When the final pattern is compiled and the data is moved from the
321 scan_data_t structure into the regexp structure the information
322 about lookbehind is factored in, with the information that would
323 have been lost precalculated in the end_shift field for the
324 associated string.
325
326 The fields pos_min and pos_delta are used to store the minimum offset
327 and the delta to the maximum offset at the current point in the pattern.
328
329*/
2c2d71f5
JH
330
331typedef struct scan_data_t {
1de06328
YO
332 /*I32 len_min; unused */
333 /*I32 len_delta; unused */
2c2d71f5
JH
334 I32 pos_min;
335 I32 pos_delta;
336 SV *last_found;
1de06328 337 I32 last_end; /* min value, <0 unless valid. */
2c2d71f5
JH
338 I32 last_start_min;
339 I32 last_start_max;
1de06328
YO
340 SV **longest; /* Either &l_fixed, or &l_float. */
341 SV *longest_fixed; /* longest fixed string found in pattern */
342 I32 offset_fixed; /* offset where it starts */
486ec47a 343 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
344 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
345 SV *longest_float; /* longest floating string found in pattern */
346 I32 offset_float_min; /* earliest point in string it can appear */
347 I32 offset_float_max; /* latest point in string it can appear */
486ec47a 348 I32 *minlen_float; /* pointer to the minlen relevant to the string */
1de06328 349 I32 lookbehind_float; /* is the position of the string modified by LB */
2c2d71f5
JH
350 I32 flags;
351 I32 whilem_c;
cb434fcc 352 I32 *last_closep;
653099ff 353 struct regnode_charclass_class *start_class;
2c2d71f5
JH
354} scan_data_t;
355
a687059c 356/*
e50aee73 357 * Forward declarations for pregcomp()'s friends.
a687059c 358 */
a0d0e21e 359
27da23d5 360static const scan_data_t zero_scan_data =
1de06328 361 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
362
363#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
364#define SF_BEFORE_SEOL 0x0001
365#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
366#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
367#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
368
09b7f37c
CB
369#ifdef NO_UNARY_PLUS
370# define SF_FIX_SHIFT_EOL (0+2)
371# define SF_FL_SHIFT_EOL (0+4)
372#else
373# define SF_FIX_SHIFT_EOL (+2)
374# define SF_FL_SHIFT_EOL (+4)
375#endif
c277df42
IZ
376
377#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
378#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
379
380#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
381#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
382#define SF_IS_INF 0x0040
383#define SF_HAS_PAR 0x0080
384#define SF_IN_PAR 0x0100
385#define SF_HAS_EVAL 0x0200
386#define SCF_DO_SUBSTR 0x0400
653099ff
GS
387#define SCF_DO_STCLASS_AND 0x0800
388#define SCF_DO_STCLASS_OR 0x1000
389#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 390#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 391
786e8c11 392#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 393#define SCF_SEEN_ACCEPT 0x8000
07be1b83 394
43fead97 395#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
396
397/* The enums for all these are ordered so things work out correctly */
a62b1201 398#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
cfaf538b 399#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
00b27cfc 400#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
401#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
402#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1 403#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a725e29c 404#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 405
43fead97 406#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 407
93733859 408#define OOB_NAMEDCLASS -1
b8c5462f 409
8e661ac5
KW
410/* There is no code point that is out-of-bounds, so this is problematic. But
411 * its only current use is to initialize a variable that is always set before
412 * looked at. */
413#define OOB_UNICODE 0xDEADBEEF
414
a0ed51b3
LW
415#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
416#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
417
8615cb43 418
b45f050a
JF
419/* length of regex to show in messages that don't mark a position within */
420#define RegexLengthToShowInErrorMessages 127
421
422/*
423 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
424 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
425 * op/pragma/warn/regcomp.
426 */
7253e4e3
RK
427#define MARKER1 "<-- HERE" /* marker as it appears in the description */
428#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 429
7253e4e3 430#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
431
432/*
433 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
434 * arg. Show regex, up to a maximum length. If it's too long, chop and add
435 * "...".
436 */
58e23c8d 437#define _FAIL(code) STMT_START { \
bfed75c6 438 const char *ellipses = ""; \
ccb2c380
MP
439 IV len = RExC_end - RExC_precomp; \
440 \
441 if (!SIZE_ONLY) \
288b8c02 442 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
443 if (len > RegexLengthToShowInErrorMessages) { \
444 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
445 len = RegexLengthToShowInErrorMessages - 10; \
446 ellipses = "..."; \
447 } \
58e23c8d 448 code; \
ccb2c380 449} STMT_END
8615cb43 450
58e23c8d
YO
451#define FAIL(msg) _FAIL( \
452 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
453 msg, (int)len, RExC_precomp, ellipses))
454
455#define FAIL2(msg,arg) _FAIL( \
456 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
457 arg, (int)len, RExC_precomp, ellipses))
458
b45f050a 459/*
b45f050a
JF
460 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
461 */
ccb2c380 462#define Simple_vFAIL(m) STMT_START { \
a28509cc 463 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
464 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
465 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
466} STMT_END
b45f050a
JF
467
468/*
469 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
470 */
ccb2c380
MP
471#define vFAIL(m) STMT_START { \
472 if (!SIZE_ONLY) \
288b8c02 473 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
474 Simple_vFAIL(m); \
475} STMT_END
b45f050a
JF
476
477/*
478 * Like Simple_vFAIL(), but accepts two arguments.
479 */
ccb2c380 480#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 481 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
482 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
483 (int)offset, RExC_precomp, RExC_precomp + offset); \
484} STMT_END
b45f050a
JF
485
486/*
487 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
488 */
ccb2c380
MP
489#define vFAIL2(m,a1) STMT_START { \
490 if (!SIZE_ONLY) \
288b8c02 491 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
492 Simple_vFAIL2(m, a1); \
493} STMT_END
b45f050a
JF
494
495
496/*
497 * Like Simple_vFAIL(), but accepts three arguments.
498 */
ccb2c380 499#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 500 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
501 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
502 (int)offset, RExC_precomp, RExC_precomp + offset); \
503} STMT_END
b45f050a
JF
504
505/*
506 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
507 */
ccb2c380
MP
508#define vFAIL3(m,a1,a2) STMT_START { \
509 if (!SIZE_ONLY) \
288b8c02 510 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
511 Simple_vFAIL3(m, a1, a2); \
512} STMT_END
b45f050a
JF
513
514/*
515 * Like Simple_vFAIL(), but accepts four arguments.
516 */
ccb2c380 517#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 518 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
519 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
520 (int)offset, RExC_precomp, RExC_precomp + offset); \
521} STMT_END
b45f050a 522
668c081a 523#define ckWARNreg(loc,m) STMT_START { \
a28509cc 524 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
525 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
526 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
527} STMT_END
528
668c081a 529#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 530 const IV offset = loc - RExC_precomp; \
d1d15184 531 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
532 m REPORT_LOCATION, \
533 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
534} STMT_END
535
2335b3d3
KW
536#define ckWARN2regdep(loc,m, a1) STMT_START { \
537 const IV offset = loc - RExC_precomp; \
538 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
539 m REPORT_LOCATION, \
540 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
541} STMT_END
542
668c081a 543#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 544 const IV offset = loc - RExC_precomp; \
668c081a 545 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
546 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
547} STMT_END
548
549#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 550 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
551 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
552 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
553} STMT_END
554
668c081a
NC
555#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
556 const IV offset = loc - RExC_precomp; \
557 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
558 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
559} STMT_END
560
ccb2c380 561#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 562 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
563 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
564 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
565} STMT_END
566
668c081a
NC
567#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
568 const IV offset = loc - RExC_precomp; \
569 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
570 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
571} STMT_END
572
ccb2c380 573#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 574 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
575 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
576 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
577} STMT_END
9d1d55b5 578
8615cb43 579
cd439c50 580/* Allow for side effects in s */
ccb2c380
MP
581#define REGC(c,s) STMT_START { \
582 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
583} STMT_END
cd439c50 584
fac92740
MJD
585/* Macros for recording node offsets. 20001227 mjd@plover.com
586 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
587 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
588 * Element 0 holds the number n.
07be1b83 589 * Position is 1 indexed.
fac92740 590 */
7122b237
YO
591#ifndef RE_TRACK_PATTERN_OFFSETS
592#define Set_Node_Offset_To_R(node,byte)
593#define Set_Node_Offset(node,byte)
594#define Set_Cur_Node_Offset
595#define Set_Node_Length_To_R(node,len)
596#define Set_Node_Length(node,len)
597#define Set_Node_Cur_Length(node)
598#define Node_Offset(n)
599#define Node_Length(n)
600#define Set_Node_Offset_Length(node,offset,len)
601#define ProgLen(ri) ri->u.proglen
602#define SetProgLen(ri,x) ri->u.proglen = x
603#else
604#define ProgLen(ri) ri->u.offsets[0]
605#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
606#define Set_Node_Offset_To_R(node,byte) STMT_START { \
607 if (! SIZE_ONLY) { \
608 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 609 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 610 if((node) < 0) { \
551405c4 611 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
612 } else { \
613 RExC_offsets[2*(node)-1] = (byte); \
614 } \
615 } \
616} STMT_END
617
618#define Set_Node_Offset(node,byte) \
619 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
620#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
621
622#define Set_Node_Length_To_R(node,len) STMT_START { \
623 if (! SIZE_ONLY) { \
624 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 625 __LINE__, (int)(node), (int)(len))); \
ccb2c380 626 if((node) < 0) { \
551405c4 627 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
628 } else { \
629 RExC_offsets[2*(node)] = (len); \
630 } \
631 } \
632} STMT_END
633
634#define Set_Node_Length(node,len) \
635 Set_Node_Length_To_R((node)-RExC_emit_start, len)
636#define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
637#define Set_Node_Cur_Length(node) \
638 Set_Node_Length(node, RExC_parse - parse_start)
fac92740
MJD
639
640/* Get offsets and lengths */
641#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
642#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
643
07be1b83
YO
644#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
645 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
646 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
647} STMT_END
7122b237 648#endif
07be1b83
YO
649
650#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
651#define EXPERIMENTAL_INPLACESCAN
f427392e 652#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 653
304ee84b
YO
654#define DEBUG_STUDYDATA(str,data,depth) \
655DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 656 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
657 "%*s" str "Pos:%"IVdf"/%"IVdf \
658 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
659 (int)(depth)*2, "", \
660 (IV)((data)->pos_min), \
661 (IV)((data)->pos_delta), \
304ee84b 662 (UV)((data)->flags), \
1de06328 663 (IV)((data)->whilem_c), \
304ee84b
YO
664 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
665 is_inf ? "INF " : "" \
1de06328
YO
666 ); \
667 if ((data)->last_found) \
668 PerlIO_printf(Perl_debug_log, \
669 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
670 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
671 SvPVX_const((data)->last_found), \
672 (IV)((data)->last_end), \
673 (IV)((data)->last_start_min), \
674 (IV)((data)->last_start_max), \
675 ((data)->longest && \
676 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
677 SvPVX_const((data)->longest_fixed), \
678 (IV)((data)->offset_fixed), \
679 ((data)->longest && \
680 (data)->longest==&((data)->longest_float)) ? "*" : "", \
681 SvPVX_const((data)->longest_float), \
682 (IV)((data)->offset_float_min), \
683 (IV)((data)->offset_float_max) \
684 ); \
685 PerlIO_printf(Perl_debug_log,"\n"); \
686});
687
acfe0abc 688static void clear_re(pTHX_ void *r);
4327152a 689
653099ff 690/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 691 Update the longest found anchored substring and the longest found
653099ff
GS
692 floating substrings if needed. */
693
4327152a 694STATIC void
304ee84b 695S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 696{
e1ec3a88
AL
697 const STRLEN l = CHR_SVLEN(data->last_found);
698 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 699 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 700
7918f24d
NC
701 PERL_ARGS_ASSERT_SCAN_COMMIT;
702
c277df42 703 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 704 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
705 if (*data->longest == data->longest_fixed) {
706 data->offset_fixed = l ? data->last_start_min : data->pos_min;
707 if (data->flags & SF_BEFORE_EOL)
b81d288d 708 data->flags
c277df42
IZ
709 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
710 else
711 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 712 data->minlen_fixed=minlenp;
1de06328 713 data->lookbehind_fixed=0;
a0ed51b3 714 }
304ee84b 715 else { /* *data->longest == data->longest_float */
c277df42 716 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
717 data->offset_float_max = (l
718 ? data->last_start_max
c277df42 719 : data->pos_min + data->pos_delta);
304ee84b 720 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 721 data->offset_float_max = I32_MAX;
c277df42 722 if (data->flags & SF_BEFORE_EOL)
b81d288d 723 data->flags
c277df42
IZ
724 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
725 else
726 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
727 data->minlen_float=minlenp;
728 data->lookbehind_float=0;
c277df42
IZ
729 }
730 }
731 SvCUR_set(data->last_found, 0);
0eda9292 732 {
a28509cc 733 SV * const sv = data->last_found;
097eb12c
AL
734 if (SvUTF8(sv) && SvMAGICAL(sv)) {
735 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
736 if (mg)
737 mg->mg_len = 0;
738 }
0eda9292 739 }
c277df42
IZ
740 data->last_end = -1;
741 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 742 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
743}
744
653099ff
GS
745/* Can match anything (initialization) */
746STATIC void
3fffb88a 747S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 748{
7918f24d
NC
749 PERL_ARGS_ASSERT_CL_ANYTHING;
750
f8bef550 751 ANYOF_BITMAP_SETALL(cl);
dd58aee1 752 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
3ad98780 753 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
3fffb88a
KW
754
755 /* If any portion of the regex is to operate under locale rules,
756 * initialization includes it. The reason this isn't done for all regexes
757 * is that the optimizer was written under the assumption that locale was
758 * all-or-nothing. Given the complexity and lack of documentation in the
759 * optimizer, and that there are inadequate test cases for locale, so many
760 * parts of it may not work properly, it is safest to avoid locale unless
761 * necessary. */
762 if (RExC_contains_locale) {
9d7a1e63 763 ANYOF_CLASS_SETALL(cl); /* /l uses class */
3fffb88a
KW
764 cl->flags |= ANYOF_LOCALE;
765 }
9d7a1e63
KW
766 else {
767 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
768 }
653099ff
GS
769}
770
771/* Can match anything (initialization) */
772STATIC int
5f66b61c 773S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
774{
775 int value;
776
7918f24d
NC
777 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
778
aaa51d5e 779 for (value = 0; value <= ANYOF_MAX; value += 2)
653099ff
GS
780 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
781 return 1;
1aa99e6b
IH
782 if (!(cl->flags & ANYOF_UNICODE_ALL))
783 return 0;
10edeb5d 784 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 785 return 0;
653099ff
GS
786 return 1;
787}
788
789/* Can match anything (initialization) */
790STATIC void
e755fd73 791S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 792{
7918f24d
NC
793 PERL_ARGS_ASSERT_CL_INIT;
794
8ecf7187 795 Zero(cl, 1, struct regnode_charclass_class);
653099ff 796 cl->type = ANYOF;
3fffb88a 797 cl_anything(pRExC_state, cl);
1411dba4 798 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
799}
800
1051e1c4
KW
801/* These two functions currently do the exact same thing */
802#define cl_init_zero S_cl_init
653099ff 803
dd58aee1
KW
804/* 'AND' a given class with another one. Can create false positives. 'cl'
805 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
806 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 807STATIC void
5f66b61c 808S_cl_and(struct regnode_charclass_class *cl,
a28509cc 809 const struct regnode_charclass_class *and_with)
653099ff 810{
7918f24d 811 PERL_ARGS_ASSERT_CL_AND;
40d049e4
YO
812
813 assert(and_with->type == ANYOF);
1e6ade67 814
c6b76537 815 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
816 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
817 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 818 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
819 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
820 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
653099ff
GS
821 int i;
822
823 if (and_with->flags & ANYOF_INVERT)
824 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
825 cl->bitmap[i] &= ~and_with->bitmap[i];
826 else
827 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
828 cl->bitmap[i] &= and_with->bitmap[i];
829 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 830
c6b76537 831 if (and_with->flags & ANYOF_INVERT) {
8951c461 832
c6b76537
KW
833 /* Here, the and'ed node is inverted. Get the AND of the flags that
834 * aren't affected by the inversion. Those that are affected are
835 * handled individually below */
836 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
837 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
838 cl->flags |= affected_flags;
839
840 /* We currently don't know how to deal with things that aren't in the
841 * bitmap, but we know that the intersection is no greater than what
842 * is already in cl, so let there be false positives that get sorted
843 * out after the synthetic start class succeeds, and the node is
844 * matched for real. */
845
846 /* The inversion of these two flags indicate that the resulting
847 * intersection doesn't have them */
848 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
849 cl->flags &= ~ANYOF_UNICODE_ALL;
850 }
c6b76537
KW
851 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
852 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 853 }
1aa99e6b 854 }
c6b76537 855 else { /* and'd node is not inverted */
3ad98780
KW
856 U8 outside_bitmap_but_not_utf8; /* Temp variable */
857
137165a6 858 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
859
860 /* Here 'and_with' doesn't match anything outside the bitmap
861 * (except possibly ANYOF_UNICODE_ALL), which means the
862 * intersection can't either, except for ANYOF_UNICODE_ALL, in
863 * which case we don't know what the intersection is, but it's no
864 * greater than what cl already has, so can just leave it alone,
865 * with possible false positives */
866 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
867 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 868 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 869 }
137165a6 870 }
c6b76537
KW
871 else if (! ANYOF_NONBITMAP(cl)) {
872
873 /* Here, 'and_with' does match something outside the bitmap, and cl
874 * doesn't have a list of things to match outside the bitmap. If
875 * cl can match all code points above 255, the intersection will
3ad98780
KW
876 * be those above-255 code points that 'and_with' matches. If cl
877 * can't match all Unicode code points, it means that it can't
878 * match anything outside the bitmap (since the 'if' that got us
879 * into this block tested for that), so we leave the bitmap empty.
880 */
c6b76537
KW
881 if (cl->flags & ANYOF_UNICODE_ALL) {
882 ARG_SET(cl, ARG(and_with));
3ad98780
KW
883
884 /* and_with's ARG may match things that don't require UTF8.
885 * And now cl's will too, in spite of this being an 'and'. See
886 * the comments below about the kludge */
887 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
888 }
889 }
890 else {
891 /* Here, both 'and_with' and cl match something outside the
892 * bitmap. Currently we do not do the intersection, so just match
893 * whatever cl had at the beginning. */
894 }
895
896
3ad98780
KW
897 /* Take the intersection of the two sets of flags. However, the
898 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
899 * kludge around the fact that this flag is not treated like the others
900 * which are initialized in cl_anything(). The way the optimizer works
901 * is that the synthetic start class (SSC) is initialized to match
902 * anything, and then the first time a real node is encountered, its
903 * values are AND'd with the SSC's with the result being the values of
904 * the real node. However, there are paths through the optimizer where
905 * the AND never gets called, so those initialized bits are set
906 * inappropriately, which is not usually a big deal, as they just cause
907 * false positives in the SSC, which will just mean a probably
908 * imperceptible slow down in execution. However this bit has a
909 * higher false positive consequence in that it can cause utf8.pm,
910 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
911 * bigger slowdown and also causes significant extra memory to be used.
912 * In order to prevent this, the code now takes a different tack. The
913 * bit isn't set unless some part of the regular expression needs it,
914 * but once set it won't get cleared. This means that these extra
915 * modules won't get loaded unless there was some path through the
916 * pattern that would have required them anyway, and so any false
917 * positives that occur by not ANDing them out when they could be
918 * aren't as severe as they would be if we treated this bit like all
919 * the others */
920 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
921 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 922 cl->flags &= and_with->flags;
3ad98780 923 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 924 }
653099ff
GS
925}
926
dd58aee1
KW
927/* 'OR' a given class with another one. Can create false positives. 'cl'
928 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
929 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 930STATIC void
3fffb88a 931S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 932{
7918f24d
NC
933 PERL_ARGS_ASSERT_CL_OR;
934
653099ff 935 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
936
937 /* Here, the or'd node is to be inverted. This means we take the
938 * complement of everything not in the bitmap, but currently we don't
939 * know what that is, so give up and match anything */
940 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 941 cl_anything(pRExC_state, cl);
c6b76537 942 }
653099ff
GS
943 /* We do not use
944 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
945 * <= (B1 | !B2) | (CL1 | !CL2)
946 * which is wasteful if CL2 is small, but we ignore CL2:
947 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
948 * XXXX Can we handle case-fold? Unclear:
949 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
950 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
951 */
c6b76537 952 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
953 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
954 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
653099ff
GS
955 int i;
956
957 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
958 cl->bitmap[i] |= ~or_with->bitmap[i];
959 } /* XXXX: logic is complicated otherwise */
960 else {
3fffb88a 961 cl_anything(pRExC_state, cl);
653099ff 962 }
c6b76537
KW
963
964 /* And, we can just take the union of the flags that aren't affected
965 * by the inversion */
966 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
967
968 /* For the remaining flags:
969 ANYOF_UNICODE_ALL and inverted means to not match anything above
970 255, which means that the union with cl should just be
971 what cl has in it, so can ignore this flag
972 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
973 is 127-255 to match them, but then invert that, so the
974 union with cl should just be what cl has in it, so can
975 ignore this flag
976 */
977 } else { /* 'or_with' is not inverted */
653099ff
GS
978 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
979 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
980 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
981 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
653099ff
GS
982 int i;
983
984 /* OR char bitmap and class bitmap separately */
985 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
986 cl->bitmap[i] |= or_with->bitmap[i];
1e6ade67 987 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
653099ff
GS
988 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
989 cl->classflags[i] |= or_with->classflags[i];
990 cl->flags |= ANYOF_CLASS;
991 }
992 }
993 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 994 cl_anything(pRExC_state, cl);
653099ff 995 }
9826f543 996
c6b76537
KW
997 if (ANYOF_NONBITMAP(or_with)) {
998
999 /* Use the added node's outside-the-bit-map match if there isn't a
1000 * conflict. If there is a conflict (both nodes match something
1001 * outside the bitmap, but what they match outside is not the same
1002 * pointer, and hence not easily compared until XXX we extend
1003 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
1004 * match everything outside the bitmap. If that stuff is all above
1005 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1006 if (! ANYOF_NONBITMAP(cl)) {
1007 ARG_SET(cl, ARG(or_with));
1008 }
1009 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1010
1011 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1012 cl_anything(pRExC_state, cl);
1013 }
1014 else {
1015 cl->flags |= ANYOF_UNICODE_ALL;
1016 }
c6b76537 1017 }
4c34a693 1018 }
0b9668ee
KW
1019
1020 /* Take the union */
1021 cl->flags |= or_with->flags;
1aa99e6b 1022 }
653099ff
GS
1023}
1024
a3621e74
YO
1025#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1026#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1027#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1028#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1029
3dab1dad
YO
1030
1031#ifdef DEBUGGING
07be1b83 1032/*
2b8b4781
NC
1033 dump_trie(trie,widecharmap,revcharmap)
1034 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1035 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1036
1037 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1038 The _interim_ variants are used for debugging the interim
1039 tables that are used to generate the final compressed
1040 representation which is what dump_trie expects.
1041
486ec47a 1042 Part of the reason for their existence is to provide a form
3dab1dad 1043 of documentation as to how the different representations function.
07be1b83
YO
1044
1045*/
3dab1dad
YO
1046
1047/*
3dab1dad
YO
1048 Dumps the final compressed table form of the trie to Perl_debug_log.
1049 Used for debugging make_trie().
1050*/
b9a59e08 1051
3dab1dad 1052STATIC void
2b8b4781
NC
1053S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1054 AV *revcharmap, U32 depth)
3dab1dad
YO
1055{
1056 U32 state;
ab3bbdeb 1057 SV *sv=sv_newmortal();
55eed653 1058 int colwidth= widecharmap ? 6 : 4;
2e64971a 1059 U16 word;
3dab1dad
YO
1060 GET_RE_DEBUG_FLAGS_DECL;
1061
7918f24d 1062 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1063
3dab1dad
YO
1064 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1065 (int)depth * 2 + 2,"",
1066 "Match","Base","Ofs" );
1067
1068 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1069 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1070 if ( tmp ) {
ab3bbdeb
YO
1071 PerlIO_printf( Perl_debug_log, "%*s",
1072 colwidth,
ddc5bc0f 1073 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1074 PL_colors[0], PL_colors[1],
1075 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1076 PERL_PV_ESCAPE_FIRSTCHAR
1077 )
1078 );
3dab1dad
YO
1079 }
1080 }
1081 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1082 (int)depth * 2 + 2,"");
1083
1084 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1085 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1086 PerlIO_printf( Perl_debug_log, "\n");
1087
1e2e3d02 1088 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1089 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1090
1091 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1092
1093 if ( trie->states[ state ].wordnum ) {
1094 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1095 } else {
1096 PerlIO_printf( Perl_debug_log, "%6s", "" );
1097 }
1098
1099 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1100
1101 if ( base ) {
1102 U32 ofs = 0;
1103
1104 while( ( base + ofs < trie->uniquecharcount ) ||
1105 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1106 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1107 ofs++;
1108
1109 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1110
1111 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1112 if ( ( base + ofs >= trie->uniquecharcount ) &&
1113 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1114 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1115 {
ab3bbdeb
YO
1116 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1117 colwidth,
3dab1dad
YO
1118 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1119 } else {
ab3bbdeb 1120 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1121 }
1122 }
1123
1124 PerlIO_printf( Perl_debug_log, "]");
1125
1126 }
1127 PerlIO_printf( Perl_debug_log, "\n" );
1128 }
2e64971a
DM
1129 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1130 for (word=1; word <= trie->wordcount; word++) {
1131 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1132 (int)word, (int)(trie->wordinfo[word].prev),
1133 (int)(trie->wordinfo[word].len));
1134 }
1135 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1136}
1137/*
3dab1dad
YO
1138 Dumps a fully constructed but uncompressed trie in list form.
1139 List tries normally only are used for construction when the number of
1140 possible chars (trie->uniquecharcount) is very high.
1141 Used for debugging make_trie().
1142*/
1143STATIC void
55eed653 1144S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1145 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1146 U32 depth)
3dab1dad
YO
1147{
1148 U32 state;
ab3bbdeb 1149 SV *sv=sv_newmortal();
55eed653 1150 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1151 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1152
1153 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1154
3dab1dad 1155 /* print out the table precompression. */
ab3bbdeb
YO
1156 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1157 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1158 "------:-----+-----------------\n" );
3dab1dad
YO
1159
1160 for( state=1 ; state < next_alloc ; state ++ ) {
1161 U16 charid;
1162
ab3bbdeb 1163 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1164 (int)depth * 2 + 2,"", (UV)state );
1165 if ( ! trie->states[ state ].wordnum ) {
1166 PerlIO_printf( Perl_debug_log, "%5s| ","");
1167 } else {
1168 PerlIO_printf( Perl_debug_log, "W%4x| ",
1169 trie->states[ state ].wordnum
1170 );
1171 }
1172 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1173 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1174 if ( tmp ) {
1175 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1176 colwidth,
ddc5bc0f 1177 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1178 PL_colors[0], PL_colors[1],
1179 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1180 PERL_PV_ESCAPE_FIRSTCHAR
1181 ) ,
1e2e3d02
YO
1182 TRIE_LIST_ITEM(state,charid).forid,
1183 (UV)TRIE_LIST_ITEM(state,charid).newstate
1184 );
1185 if (!(charid % 10))
664e119d
RGS
1186 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1187 (int)((depth * 2) + 14), "");
1e2e3d02 1188 }
ab3bbdeb
YO
1189 }
1190 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1191 }
1192}
1193
1194/*
3dab1dad
YO
1195 Dumps a fully constructed but uncompressed trie in table form.
1196 This is the normal DFA style state transition table, with a few
1197 twists to facilitate compression later.
1198 Used for debugging make_trie().
1199*/
1200STATIC void
55eed653 1201S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1202 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1203 U32 depth)
3dab1dad
YO
1204{
1205 U32 state;
1206 U16 charid;
ab3bbdeb 1207 SV *sv=sv_newmortal();
55eed653 1208 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1209 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1210
1211 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1212
1213 /*
1214 print out the table precompression so that we can do a visual check
1215 that they are identical.
1216 */
1217
1218 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1219
1220 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1221 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1222 if ( tmp ) {
ab3bbdeb
YO
1223 PerlIO_printf( Perl_debug_log, "%*s",
1224 colwidth,
ddc5bc0f 1225 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1226 PL_colors[0], PL_colors[1],
1227 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1228 PERL_PV_ESCAPE_FIRSTCHAR
1229 )
1230 );
3dab1dad
YO
1231 }
1232 }
1233
1234 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1235
1236 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1237 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1238 }
1239
1240 PerlIO_printf( Perl_debug_log, "\n" );
1241
1242 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1243
1244 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1245 (int)depth * 2 + 2,"",
1246 (UV)TRIE_NODENUM( state ) );
1247
1248 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1249 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1250 if (v)
1251 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1252 else
1253 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1254 }
1255 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1256 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1257 } else {
1258 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1259 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1260 }
1261 }
07be1b83 1262}
3dab1dad
YO
1263
1264#endif
1265
2e64971a 1266
786e8c11
YO
1267/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1268 startbranch: the first branch in the whole branch sequence
1269 first : start branch of sequence of branch-exact nodes.
1270 May be the same as startbranch
1271 last : Thing following the last branch.
1272 May be the same as tail.
1273 tail : item following the branch sequence
1274 count : words in the sequence
1275 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1276 depth : indent depth
3dab1dad 1277
786e8c11 1278Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1279
786e8c11
YO
1280A trie is an N'ary tree where the branches are determined by digital
1281decomposition of the key. IE, at the root node you look up the 1st character and
1282follow that branch repeat until you find the end of the branches. Nodes can be
1283marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1284
786e8c11 1285 /he|she|his|hers/
72f13be8 1286
786e8c11
YO
1287would convert into the following structure. Numbers represent states, letters
1288following numbers represent valid transitions on the letter from that state, if
1289the number is in square brackets it represents an accepting state, otherwise it
1290will be in parenthesis.
07be1b83 1291
786e8c11
YO
1292 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1293 | |
1294 | (2)
1295 | |
1296 (1) +-i->(6)-+-s->[7]
1297 |
1298 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1299
786e8c11
YO
1300 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1301
1302This shows that when matching against the string 'hers' we will begin at state 1
1303read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1304then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1305is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1306single traverse. We store a mapping from accepting to state to which word was
1307matched, and then when we have multiple possibilities we try to complete the
1308rest of the regex in the order in which they occured in the alternation.
1309
1310The only prior NFA like behaviour that would be changed by the TRIE support is
1311the silent ignoring of duplicate alternations which are of the form:
1312
1313 / (DUPE|DUPE) X? (?{ ... }) Y /x
1314
4b714af6 1315Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1316and without the optimisation. With the optimisations dupes will be silently
486ec47a 1317ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1318the following demonstrates:
1319
1320 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1321
1322which prints out 'word' three times, but
1323
1324 'words'=~/(word|word|word)(?{ print $1 })S/
1325
1326which doesnt print it out at all. This is due to other optimisations kicking in.
1327
1328Example of what happens on a structural level:
1329
486ec47a 1330The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1331
1332 1: CURLYM[1] {1,32767}(18)
1333 5: BRANCH(8)
1334 6: EXACT <ac>(16)
1335 8: BRANCH(11)
1336 9: EXACT <ad>(16)
1337 11: BRANCH(14)
1338 12: EXACT <ab>(16)
1339 16: SUCCEED(0)
1340 17: NOTHING(18)
1341 18: END(0)
1342
1343This would be optimizable with startbranch=5, first=5, last=16, tail=16
1344and should turn into:
1345
1346 1: CURLYM[1] {1,32767}(18)
1347 5: TRIE(16)
1348 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1349 <ac>
1350 <ad>
1351 <ab>
1352 16: SUCCEED(0)
1353 17: NOTHING(18)
1354 18: END(0)
1355
1356Cases where tail != last would be like /(?foo|bar)baz/:
1357
1358 1: BRANCH(4)
1359 2: EXACT <foo>(8)
1360 4: BRANCH(7)
1361 5: EXACT <bar>(8)
1362 7: TAIL(8)
1363 8: EXACT <baz>(10)
1364 10: END(0)
1365
1366which would be optimizable with startbranch=1, first=1, last=7, tail=8
1367and would end up looking like:
1368
1369 1: TRIE(8)
1370 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1371 <foo>
1372 <bar>
1373 7: TAIL(8)
1374 8: EXACT <baz>(10)
1375 10: END(0)
1376
1377 d = uvuni_to_utf8_flags(d, uv, 0);
1378
1379is the recommended Unicode-aware way of saying
1380
1381 *(d++) = uv;
1382*/
1383
fab2782b 1384#define TRIE_STORE_REVCHAR(val) \
786e8c11 1385 STMT_START { \
73031816 1386 if (UTF) { \
fab2782b 1387 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1388 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
fab2782b 1389 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
73031816
NC
1390 SvCUR_set(zlopp, kapow - flrbbbbb); \
1391 SvPOK_on(zlopp); \
1392 SvUTF8_on(zlopp); \
1393 av_push(revcharmap, zlopp); \
1394 } else { \
fab2782b 1395 char ooooff = (char)val; \
73031816
NC
1396 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1397 } \
1398 } STMT_END
786e8c11 1399
fab2782b
YO
1400#define TRIE_READ_CHAR STMT_START { \
1401 wordlen++; \
1402 if ( UTF ) { \
1403 /* if it is UTF then it is either already folded, or does not need folding */ \
1404 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1405 } \
1406 else if (folder == PL_fold_latin1) { \
1407 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1408 if ( foldlen > 0 ) { \
1409 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1410 foldlen -= len; \
1411 scan += len; \
1412 len = 0; \
1413 } else { \
1414 len = 1; \
1415 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1416 skiplen = UNISKIP(uvc); \
1417 foldlen -= skiplen; \
1418 scan = foldbuf + skiplen; \
1419 } \
1420 } else { \
1421 /* raw data, will be folded later if needed */ \
1422 uvc = (U32)*uc; \
1423 len = 1; \
1424 } \
786e8c11
YO
1425} STMT_END
1426
1427
1428
1429#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1430 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1431 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1432 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1433 } \
1434 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1435 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1436 TRIE_LIST_CUR( state )++; \
1437} STMT_END
07be1b83 1438
786e8c11
YO
1439#define TRIE_LIST_NEW(state) STMT_START { \
1440 Newxz( trie->states[ state ].trans.list, \
1441 4, reg_trie_trans_le ); \
1442 TRIE_LIST_CUR( state ) = 1; \
1443 TRIE_LIST_LEN( state ) = 4; \
1444} STMT_END
07be1b83 1445
786e8c11
YO
1446#define TRIE_HANDLE_WORD(state) STMT_START { \
1447 U16 dupe= trie->states[ state ].wordnum; \
1448 regnode * const noper_next = regnext( noper ); \
1449 \
786e8c11
YO
1450 DEBUG_r({ \
1451 /* store the word for dumping */ \
1452 SV* tmp; \
1453 if (OP(noper) != NOTHING) \
740cce10 1454 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1455 else \
740cce10 1456 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1457 av_push( trie_words, tmp ); \
786e8c11
YO
1458 }); \
1459 \
1460 curword++; \
2e64971a
DM
1461 trie->wordinfo[curword].prev = 0; \
1462 trie->wordinfo[curword].len = wordlen; \
1463 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1464 \
1465 if ( noper_next < tail ) { \
1466 if (!trie->jump) \
c944940b 1467 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1468 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1469 if (!jumper) \
1470 jumper = noper_next; \
1471 if (!nextbranch) \
1472 nextbranch= regnext(cur); \
1473 } \
1474 \
1475 if ( dupe ) { \
2e64971a
DM
1476 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1477 /* chain, so that when the bits of chain are later */\
1478 /* linked together, the dups appear in the chain */\
1479 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1480 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1481 } else { \
1482 /* we haven't inserted this word yet. */ \
1483 trie->states[ state ].wordnum = curword; \
1484 } \
1485} STMT_END
07be1b83 1486
3dab1dad 1487
786e8c11
YO
1488#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1489 ( ( base + charid >= ucharcount \
1490 && base + charid < ubound \
1491 && state == trie->trans[ base - ucharcount + charid ].check \
1492 && trie->trans[ base - ucharcount + charid ].next ) \
1493 ? trie->trans[ base - ucharcount + charid ].next \
1494 : ( state==1 ? special : 0 ) \
1495 )
3dab1dad 1496
786e8c11
YO
1497#define MADE_TRIE 1
1498#define MADE_JUMP_TRIE 2
1499#define MADE_EXACT_TRIE 4
3dab1dad 1500
a3621e74 1501STATIC I32
786e8c11 1502S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1503{
27da23d5 1504 dVAR;
a3621e74
YO
1505 /* first pass, loop through and scan words */
1506 reg_trie_data *trie;
55eed653 1507 HV *widecharmap = NULL;
2b8b4781 1508 AV *revcharmap = newAV();
a3621e74 1509 regnode *cur;
9f7f3913 1510 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1511 STRLEN len = 0;
1512 UV uvc = 0;
1513 U16 curword = 0;
1514 U32 next_alloc = 0;
786e8c11
YO
1515 regnode *jumper = NULL;
1516 regnode *nextbranch = NULL;
7f69552c 1517 regnode *convert = NULL;
2e64971a 1518 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1519 /* we just use folder as a flag in utf8 */
1e696034 1520 const U8 * folder = NULL;
a3621e74 1521
2b8b4781
NC
1522#ifdef DEBUGGING
1523 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1524 AV *trie_words = NULL;
1525 /* along with revcharmap, this only used during construction but both are
1526 * useful during debugging so we store them in the struct when debugging.
8e11feef 1527 */
2b8b4781
NC
1528#else
1529 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1530 STRLEN trie_charcount=0;
3dab1dad 1531#endif
2b8b4781 1532 SV *re_trie_maxbuff;
a3621e74 1533 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1534
1535 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1536#ifndef DEBUGGING
1537 PERL_UNUSED_ARG(depth);
1538#endif
a3621e74 1539
1e696034 1540 switch (flags) {
c46d03cf 1541 case EXACT: break;
2f7f8cb1 1542 case EXACTFA:
fab2782b
YO
1543 case EXACTFU_SS:
1544 case EXACTFU_TRICKYFOLD:
1e696034
KW
1545 case EXACTFU: folder = PL_fold_latin1; break;
1546 case EXACTF: folder = PL_fold; break;
1547 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1548 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1549 }
1550
c944940b 1551 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1552 trie->refcount = 1;
3dab1dad 1553 trie->startstate = 1;
786e8c11 1554 trie->wordcount = word_count;
f8fc2ecf 1555 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1556 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1557 if (flags == EXACT)
c944940b 1558 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1559 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1560 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1561
a3621e74 1562 DEBUG_r({
2b8b4781 1563 trie_words = newAV();
a3621e74 1564 });
a3621e74 1565
0111c4fd 1566 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1567 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1568 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1569 }
df826430 1570 DEBUG_TRIE_COMPILE_r({
3dab1dad 1571 PerlIO_printf( Perl_debug_log,
786e8c11 1572 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1573 (int)depth * 2 + 2, "",
1574 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1575 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1576 (int)depth);
3dab1dad 1577 });
7f69552c
YO
1578
1579 /* Find the node we are going to overwrite */
1580 if ( first == startbranch && OP( last ) != BRANCH ) {
1581 /* whole branch chain */
1582 convert = first;
1583 } else {
1584 /* branch sub-chain */
1585 convert = NEXTOPER( first );
1586 }
1587
a3621e74
YO
1588 /* -- First loop and Setup --
1589
1590 We first traverse the branches and scan each word to determine if it
1591 contains widechars, and how many unique chars there are, this is
1592 important as we have to build a table with at least as many columns as we
1593 have unique chars.
1594
1595 We use an array of integers to represent the character codes 0..255
38a44b82 1596 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1597 native representation of the character value as the key and IV's for the
1598 coded index.
1599
1600 *TODO* If we keep track of how many times each character is used we can
1601 remap the columns so that the table compression later on is more
3b753521 1602 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1603 middle and the least common are on the outside. IMO this would be better
1604 than a most to least common mapping as theres a decent chance the most
1605 common letter will share a node with the least common, meaning the node
486ec47a 1606 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1607 case is when we have the least common nodes twice.
1608
1609 */
1610
a3621e74 1611 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1612 regnode *noper = NEXTOPER( cur );
e1ec3a88 1613 const U8 *uc = (U8*)STRING( noper );
df826430 1614 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1615 STRLEN foldlen = 0;
1616 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1617 STRLEN skiplen = 0;
2af232bd 1618 const U8 *scan = (U8*)NULL;
07be1b83 1619 U32 wordlen = 0; /* required init */
02daf0ab
YO
1620 STRLEN chars = 0;
1621 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1622
3dab1dad 1623 if (OP(noper) == NOTHING) {
df826430
YO
1624 regnode *noper_next= regnext(noper);
1625 if (noper_next != tail && OP(noper_next) == flags) {
1626 noper = noper_next;
1627 uc= (U8*)STRING(noper);
1628 e= uc + STR_LEN(noper);
1629 trie->minlen= STR_LEN(noper);
1630 } else {
1631 trie->minlen= 0;
1632 continue;
1633 }
3dab1dad 1634 }
df826430 1635
fab2782b 1636 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1637 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1638 regardless of encoding */
fab2782b
YO
1639 if (OP( noper ) == EXACTFU_SS) {
1640 /* false positives are ok, so just set this */
1641 TRIE_BITMAP_SET(trie,0xDF);
1642 }
1643 }
a3621e74 1644 for ( ; uc < e ; uc += len ) {
3dab1dad 1645 TRIE_CHARCOUNT(trie)++;
a3621e74 1646 TRIE_READ_CHAR;
3dab1dad 1647 chars++;
a3621e74 1648 if ( uvc < 256 ) {
fab2782b
YO
1649 if ( folder ) {
1650 U8 folded= folder[ (U8) uvc ];
1651 if ( !trie->charmap[ folded ] ) {
1652 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1653 TRIE_STORE_REVCHAR( folded );
1654 }
1655 }
a3621e74
YO
1656 if ( !trie->charmap[ uvc ] ) {
1657 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1658 TRIE_STORE_REVCHAR( uvc );
a3621e74 1659 }
02daf0ab 1660 if ( set_bit ) {
62012aee
KW
1661 /* store the codepoint in the bitmap, and its folded
1662 * equivalent. */
fab2782b 1663 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1664
1665 /* store the folded codepoint */
fab2782b 1666 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1667
1668 if ( !UTF ) {
1669 /* store first byte of utf8 representation of
acdf4139
KW
1670 variant codepoints */
1671 if (! UNI_IS_INVARIANT(uvc)) {
1672 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1673 }
1674 }
02daf0ab
YO
1675 set_bit = 0; /* We've done our bit :-) */
1676 }
a3621e74
YO
1677 } else {
1678 SV** svpp;
55eed653
NC
1679 if ( !widecharmap )
1680 widecharmap = newHV();
a3621e74 1681
55eed653 1682 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1683
1684 if ( !svpp )
e4584336 1685 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1686
1687 if ( !SvTRUE( *svpp ) ) {
1688 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1689 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1690 }
1691 }
1692 }
3dab1dad 1693 if( cur == first ) {
fab2782b
YO
1694 trie->minlen = chars;
1695 trie->maxlen = chars;
3dab1dad 1696 } else if (chars < trie->minlen) {
fab2782b 1697 trie->minlen = chars;
3dab1dad 1698 } else if (chars > trie->maxlen) {
fab2782b
YO
1699 trie->maxlen = chars;
1700 }
1701 if (OP( noper ) == EXACTFU_SS) {
1702 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1703 if (trie->minlen > 1)
1704 trie->minlen= 1;
1705 }
1706 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1707 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1708 * - We assume that any such sequence might match a 2 byte string */
1709 if (trie->minlen > 2 )
1710 trie->minlen= 2;
3dab1dad
YO
1711 }
1712
a3621e74
YO
1713 } /* end first pass */
1714 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1715 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1716 (int)depth * 2 + 2,"",
55eed653 1717 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1718 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1719 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1720 );
a3621e74
YO
1721
1722 /*
1723 We now know what we are dealing with in terms of unique chars and
1724 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1725 representation using a flat table will take. If it's over a reasonable
1726 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1727 conservative but potentially much slower representation using an array
1728 of lists.
1729
1730 At the end we convert both representations into the same compressed
1731 form that will be used in regexec.c for matching with. The latter
1732 is a form that cannot be used to construct with but has memory
1733 properties similar to the list form and access properties similar
1734 to the table form making it both suitable for fast searches and
1735 small enough that its feasable to store for the duration of a program.
1736
1737 See the comment in the code where the compressed table is produced
1738 inplace from the flat tabe representation for an explanation of how
1739 the compression works.
1740
1741 */
1742
1743
2e64971a
DM
1744 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1745 prev_states[1] = 0;
1746
3dab1dad 1747 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1748 /*
1749 Second Pass -- Array Of Lists Representation
1750
1751 Each state will be represented by a list of charid:state records
1752 (reg_trie_trans_le) the first such element holds the CUR and LEN
1753 points of the allocated array. (See defines above).
1754
1755 We build the initial structure using the lists, and then convert
1756 it into the compressed table form which allows faster lookups
1757 (but cant be modified once converted).
a3621e74
YO
1758 */
1759
a3621e74
YO
1760 STRLEN transcount = 1;
1761
1e2e3d02
YO
1762 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1763 "%*sCompiling trie using list compiler\n",
1764 (int)depth * 2 + 2, ""));
686b73d4 1765
c944940b
JH
1766 trie->states = (reg_trie_state *)
1767 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1768 sizeof(reg_trie_state) );
a3621e74
YO
1769 TRIE_LIST_NEW(1);
1770 next_alloc = 2;
1771
1772 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1773
df826430 1774 regnode *noper = NEXTOPER( cur );
c445ea15 1775 U8 *uc = (U8*)STRING( noper );
df826430 1776 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1777 U32 state = 1; /* required init */
1778 U16 charid = 0; /* sanity init */
1779 U8 *scan = (U8*)NULL; /* sanity init */
1780 STRLEN foldlen = 0; /* required init */
07be1b83 1781 U32 wordlen = 0; /* required init */
c445ea15 1782 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1783 STRLEN skiplen = 0;
c445ea15 1784
df826430
YO
1785 if (OP(noper) == NOTHING) {
1786 regnode *noper_next= regnext(noper);
1787 if (noper_next != tail && OP(noper_next) == flags) {
1788 noper = noper_next;
1789 uc= (U8*)STRING(noper);
1790 e= uc + STR_LEN(noper);
1791 }
1792 }
1793
3dab1dad 1794 if (OP(noper) != NOTHING) {
786e8c11 1795 for ( ; uc < e ; uc += len ) {
c445ea15 1796
786e8c11 1797 TRIE_READ_CHAR;
c445ea15 1798
786e8c11
YO
1799 if ( uvc < 256 ) {
1800 charid = trie->charmap[ uvc ];
c445ea15 1801 } else {
55eed653 1802 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1803 if ( !svpp ) {
1804 charid = 0;
1805 } else {
1806 charid=(U16)SvIV( *svpp );
1807 }
c445ea15 1808 }
786e8c11
YO
1809 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1810 if ( charid ) {
a3621e74 1811
786e8c11
YO
1812 U16 check;
1813 U32 newstate = 0;
a3621e74 1814
786e8c11
YO
1815 charid--;
1816 if ( !trie->states[ state ].trans.list ) {
1817 TRIE_LIST_NEW( state );
c445ea15 1818 }
786e8c11
YO
1819 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1820 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1821 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1822 break;
1823 }
1824 }
1825 if ( ! newstate ) {
1826 newstate = next_alloc++;
2e64971a 1827 prev_states[newstate] = state;
786e8c11
YO
1828 TRIE_LIST_PUSH( state, charid, newstate );
1829 transcount++;
1830 }
1831 state = newstate;
1832 } else {
1833 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1834 }
a28509cc 1835 }
c445ea15 1836 }
3dab1dad 1837 TRIE_HANDLE_WORD(state);
a3621e74
YO
1838
1839 } /* end second pass */
1840
1e2e3d02
YO
1841 /* next alloc is the NEXT state to be allocated */
1842 trie->statecount = next_alloc;
c944940b
JH
1843 trie->states = (reg_trie_state *)
1844 PerlMemShared_realloc( trie->states,
1845 next_alloc
1846 * sizeof(reg_trie_state) );
a3621e74 1847
3dab1dad 1848 /* and now dump it out before we compress it */
2b8b4781
NC
1849 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1850 revcharmap, next_alloc,
1851 depth+1)
1e2e3d02 1852 );
a3621e74 1853
c944940b
JH
1854 trie->trans = (reg_trie_trans *)
1855 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1856 {
1857 U32 state;
a3621e74
YO
1858 U32 tp = 0;
1859 U32 zp = 0;
1860
1861
1862 for( state=1 ; state < next_alloc ; state ++ ) {
1863 U32 base=0;
1864
1865 /*
1866 DEBUG_TRIE_COMPILE_MORE_r(
1867 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1868 );
1869 */
1870
1871 if (trie->states[state].trans.list) {
1872 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1873 U16 maxid=minid;
a28509cc 1874 U16 idx;
a3621e74
YO
1875
1876 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1877 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1878 if ( forid < minid ) {
1879 minid=forid;
1880 } else if ( forid > maxid ) {
1881 maxid=forid;
1882 }
a3621e74
YO
1883 }
1884 if ( transcount < tp + maxid - minid + 1) {
1885 transcount *= 2;
c944940b
JH
1886 trie->trans = (reg_trie_trans *)
1887 PerlMemShared_realloc( trie->trans,
446bd890
NC
1888 transcount
1889 * sizeof(reg_trie_trans) );
a3621e74
YO
1890 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1891 }
1892 base = trie->uniquecharcount + tp - minid;
1893 if ( maxid == minid ) {
1894 U32 set = 0;
1895 for ( ; zp < tp ; zp++ ) {
1896 if ( ! trie->trans[ zp ].next ) {
1897 base = trie->uniquecharcount + zp - minid;
1898 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1899 trie->trans[ zp ].check = state;
1900 set = 1;
1901 break;
1902 }
1903 }
1904 if ( !set ) {
1905 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1906 trie->trans[ tp ].check = state;
1907 tp++;
1908 zp = tp;
1909 }
1910 } else {
1911 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1912 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1913 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1914 trie->trans[ tid ].check = state;
1915 }
1916 tp += ( maxid - minid + 1 );
1917 }
1918 Safefree(trie->states[ state ].trans.list);
1919 }
1920 /*
1921 DEBUG_TRIE_COMPILE_MORE_r(
1922 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1923 );
1924 */
1925 trie->states[ state ].trans.base=base;
1926 }
cc601c31 1927 trie->lasttrans = tp + 1;
a3621e74
YO
1928 }
1929 } else {
1930 /*
1931 Second Pass -- Flat Table Representation.
1932
1933 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1934 We know that we will need Charcount+1 trans at most to store the data
1935 (one row per char at worst case) So we preallocate both structures
1936 assuming worst case.
1937
1938 We then construct the trie using only the .next slots of the entry
1939 structs.
1940
3b753521 1941 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1942 make compression both faster and easier by keeping track of how many non
1943 zero fields are in the node.
1944
1945 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1946 transition.
1947
1948 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1949 number representing the first entry of the node, and state as a
1950 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1951 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1952 are 2 entrys per node. eg:
1953
1954 A B A B
1955 1. 2 4 1. 3 7
1956 2. 0 3 3. 0 5
1957 3. 0 0 5. 0 0
1958 4. 0 0 7. 0 0
1959
1960 The table is internally in the right hand, idx form. However as we also
1961 have to deal with the states array which is indexed by nodenum we have to
1962 use TRIE_NODENUM() to convert.
1963
1964 */
1e2e3d02
YO
1965 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1966 "%*sCompiling trie using table compiler\n",
1967 (int)depth * 2 + 2, ""));
3dab1dad 1968
c944940b
JH
1969 trie->trans = (reg_trie_trans *)
1970 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1971 * trie->uniquecharcount + 1,
1972 sizeof(reg_trie_trans) );
1973 trie->states = (reg_trie_state *)
1974 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1975 sizeof(reg_trie_state) );
a3621e74
YO
1976 next_alloc = trie->uniquecharcount + 1;
1977
3dab1dad 1978
a3621e74
YO
1979 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1980
df826430 1981 regnode *noper = NEXTOPER( cur );
a28509cc 1982 const U8 *uc = (U8*)STRING( noper );
df826430 1983 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1984
1985 U32 state = 1; /* required init */
1986
1987 U16 charid = 0; /* sanity init */
1988 U32 accept_state = 0; /* sanity init */
1989 U8 *scan = (U8*)NULL; /* sanity init */
1990
1991 STRLEN foldlen = 0; /* required init */
07be1b83 1992 U32 wordlen = 0; /* required init */
fab2782b 1993 STRLEN skiplen = 0;
a3621e74
YO
1994 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1995
df826430
YO
1996 if (OP(noper) == NOTHING) {
1997 regnode *noper_next= regnext(noper);
1998 if (noper_next != tail && OP(noper_next) == flags) {
1999 noper = noper_next;
2000 uc= (U8*)STRING(noper);
2001 e= uc + STR_LEN(noper);
2002 }
2003 }
fab2782b 2004
3dab1dad 2005 if ( OP(noper) != NOTHING ) {
786e8c11 2006 for ( ; uc < e ; uc += len ) {
a3621e74 2007
786e8c11 2008 TRIE_READ_CHAR;
a3621e74 2009
786e8c11
YO
2010 if ( uvc < 256 ) {
2011 charid = trie->charmap[ uvc ];
2012 } else {
55eed653 2013 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2014 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2015 }
786e8c11
YO
2016 if ( charid ) {
2017 charid--;
2018 if ( !trie->trans[ state + charid ].next ) {
2019 trie->trans[ state + charid ].next = next_alloc;
2020 trie->trans[ state ].check++;
2e64971a
DM
2021 prev_states[TRIE_NODENUM(next_alloc)]
2022 = TRIE_NODENUM(state);
786e8c11
YO
2023 next_alloc += trie->uniquecharcount;
2024 }
2025 state = trie->trans[ state + charid ].next;
2026 } else {
2027 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2028 }
2029 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2030 }
a3621e74 2031 }
3dab1dad
YO
2032 accept_state = TRIE_NODENUM( state );
2033 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2034
2035 } /* end second pass */
2036
3dab1dad 2037 /* and now dump it out before we compress it */
2b8b4781
NC
2038 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2039 revcharmap,
2040 next_alloc, depth+1));
a3621e74 2041
a3621e74
YO
2042 {
2043 /*
2044 * Inplace compress the table.*
2045
2046 For sparse data sets the table constructed by the trie algorithm will
2047 be mostly 0/FAIL transitions or to put it another way mostly empty.
2048 (Note that leaf nodes will not contain any transitions.)
2049
2050 This algorithm compresses the tables by eliminating most such
2051 transitions, at the cost of a modest bit of extra work during lookup:
2052
2053 - Each states[] entry contains a .base field which indicates the
2054 index in the state[] array wheres its transition data is stored.
2055
3b753521 2056 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2057
2058 - If .base is nonzero then charid is added to it to find an entry in
2059 the trans array.
2060
2061 -If trans[states[state].base+charid].check!=state then the
2062 transition is taken to be a 0/Fail transition. Thus if there are fail
2063 transitions at the front of the node then the .base offset will point
2064 somewhere inside the previous nodes data (or maybe even into a node
2065 even earlier), but the .check field determines if the transition is
2066 valid.
2067
786e8c11 2068 XXX - wrong maybe?
a3621e74 2069 The following process inplace converts the table to the compressed
3b753521 2070 table: We first do not compress the root node 1,and mark all its
a3621e74 2071 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2072 allows us to do a DFA construction from the compressed table later,
2073 and ensures that any .base pointers we calculate later are greater
2074 than 0.
a3621e74
YO
2075
2076 - We set 'pos' to indicate the first entry of the second node.
2077
2078 - We then iterate over the columns of the node, finding the first and
2079 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2080 and set the .check pointers accordingly, and advance pos
2081 appropriately and repreat for the next node. Note that when we copy
2082 the next pointers we have to convert them from the original
2083 NODEIDX form to NODENUM form as the former is not valid post
2084 compression.
2085
2086 - If a node has no transitions used we mark its base as 0 and do not
2087 advance the pos pointer.
2088
2089 - If a node only has one transition we use a second pointer into the
2090 structure to fill in allocated fail transitions from other states.
2091 This pointer is independent of the main pointer and scans forward
2092 looking for null transitions that are allocated to a state. When it
2093 finds one it writes the single transition into the "hole". If the
786e8c11 2094 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2095
2096 - Once compressed we can Renew/realloc the structures to release the
2097 excess space.
2098
2099 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2100 specifically Fig 3.47 and the associated pseudocode.
2101
2102 demq
2103 */
a3b680e6 2104 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2105 U32 state, charid;
a3621e74 2106 U32 pos = 0, zp=0;
1e2e3d02 2107 trie->statecount = laststate;
a3621e74
YO
2108
2109 for ( state = 1 ; state < laststate ; state++ ) {
2110 U8 flag = 0;
a28509cc
AL
2111 const U32 stateidx = TRIE_NODEIDX( state );
2112 const U32 o_used = trie->trans[ stateidx ].check;
2113 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2114 trie->trans[ stateidx ].check = 0;
2115
2116 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2117 if ( flag || trie->trans[ stateidx + charid ].next ) {
2118 if ( trie->trans[ stateidx + charid ].next ) {
2119 if (o_used == 1) {
2120 for ( ; zp < pos ; zp++ ) {
2121 if ( ! trie->trans[ zp ].next ) {
2122 break;
2123 }
2124 }
2125 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2126 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2127 trie->trans[ zp ].check = state;
2128 if ( ++zp > pos ) pos = zp;
2129 break;
2130 }
2131 used--;
2132 }
2133 if ( !flag ) {
2134 flag = 1;
2135 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2136 }
2137 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2138 trie->trans[ pos ].check = state;
2139 pos++;
2140 }
2141 }
2142 }
cc601c31 2143 trie->lasttrans = pos + 1;
c944940b
JH
2144 trie->states = (reg_trie_state *)
2145 PerlMemShared_realloc( trie->states, laststate
2146 * sizeof(reg_trie_state) );
a3621e74 2147 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2148 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2149 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2150 (int)depth * 2 + 2,"",
2151 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2152 (IV)next_alloc,
2153 (IV)pos,
a3621e74
YO
2154 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2155 );
2156
2157 } /* end table compress */
2158 }
1e2e3d02
YO
2159 DEBUG_TRIE_COMPILE_MORE_r(
2160 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2161 (int)depth * 2 + 2, "",
2162 (UV)trie->statecount,
2163 (UV)trie->lasttrans)
2164 );
cc601c31 2165 /* resize the trans array to remove unused space */
c944940b
JH
2166 trie->trans = (reg_trie_trans *)
2167 PerlMemShared_realloc( trie->trans, trie->lasttrans
2168 * sizeof(reg_trie_trans) );
a3621e74 2169
3b753521 2170 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2171 U8 nodetype =(U8)(flags & 0xFF);
2172 char *str=NULL;
786e8c11 2173
07be1b83 2174#ifdef DEBUGGING
e62cc96a 2175 regnode *optimize = NULL;
7122b237
YO
2176#ifdef RE_TRACK_PATTERN_OFFSETS
2177
b57a0404
JH
2178 U32 mjd_offset = 0;
2179 U32 mjd_nodelen = 0;
7122b237
YO
2180#endif /* RE_TRACK_PATTERN_OFFSETS */
2181#endif /* DEBUGGING */
a3621e74 2182 /*
3dab1dad
YO
2183 This means we convert either the first branch or the first Exact,
2184 depending on whether the thing following (in 'last') is a branch
2185 or not and whther first is the startbranch (ie is it a sub part of
2186 the alternation or is it the whole thing.)
3b753521 2187 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2188 the whole branch sequence, including the first.
a3621e74 2189 */
3dab1dad 2190 /* Find the node we are going to overwrite */
7f69552c 2191 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2192 /* branch sub-chain */
3dab1dad 2193 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2194#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2195 DEBUG_r({
2196 mjd_offset= Node_Offset((convert));
2197 mjd_nodelen= Node_Length((convert));
2198 });
7122b237 2199#endif
7f69552c 2200 /* whole branch chain */
7122b237
YO
2201 }
2202#ifdef RE_TRACK_PATTERN_OFFSETS
2203 else {
7f69552c
YO
2204 DEBUG_r({
2205 const regnode *nop = NEXTOPER( convert );
2206 mjd_offset= Node_Offset((nop));
2207 mjd_nodelen= Node_Length((nop));
2208 });
07be1b83
YO
2209 }
2210 DEBUG_OPTIMISE_r(
2211 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2212 (int)depth * 2 + 2, "",
786e8c11 2213 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2214 );
7122b237 2215#endif
3dab1dad
YO
2216 /* But first we check to see if there is a common prefix we can
2217 split out as an EXACT and put in front of the TRIE node. */
2218 trie->startstate= 1;
55eed653 2219 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2220 U32 state;
1e2e3d02 2221 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2222 U32 ofs = 0;
8e11feef
RGS
2223 I32 idx = -1;
2224 U32 count = 0;
2225 const U32 base = trie->states[ state ].trans.base;
a3621e74 2226
3dab1dad 2227 if ( trie->states[state].wordnum )
8e11feef 2228 count = 1;
a3621e74 2229
8e11feef 2230 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2231 if ( ( base + ofs >= trie->uniquecharcount ) &&
2232 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2233 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2234 {
3dab1dad 2235 if ( ++count > 1 ) {
2b8b4781 2236 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2237 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2238 if ( state == 1 ) break;
3dab1dad
YO
2239 if ( count == 2 ) {
2240 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2241 DEBUG_OPTIMISE_r(
8e11feef
RGS
2242 PerlIO_printf(Perl_debug_log,
2243 "%*sNew Start State=%"UVuf" Class: [",
2244 (int)depth * 2 + 2, "",
786e8c11 2245 (UV)state));
be8e71aa 2246 if (idx >= 0) {
2b8b4781 2247 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2248 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2249
3dab1dad 2250 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2251 if ( folder )
2252 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2253 DEBUG_OPTIMISE_r(
f1f66076 2254 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2255 );
8e11feef
RGS
2256 }
2257 }
2258 TRIE_BITMAP_SET(trie,*ch);
2259 if ( folder )
2260 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2261 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2262 }
2263 idx = ofs;
2264 }
3dab1dad
YO
2265 }
2266 if ( count == 1 ) {
2b8b4781 2267 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2268 STRLEN len;
2269 char *ch = SvPV( *tmp, len );
de734bd5
A
2270 DEBUG_OPTIMISE_r({
2271 SV *sv=sv_newmortal();
8e11feef
RGS
2272 PerlIO_printf( Perl_debug_log,
2273 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2274 (int)depth * 2 + 2, "",
de734bd5
A
2275 (UV)state, (UV)idx,
2276 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2277 PL_colors[0], PL_colors[1],
2278 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2279 PERL_PV_ESCAPE_FIRSTCHAR
2280 )
2281 );
2282 });
3dab1dad
YO
2283 if ( state==1 ) {
2284 OP( convert ) = nodetype;
2285 str=STRING(convert);
2286 STR_LEN(convert)=0;
2287 }
c490c714
YO
2288 STR_LEN(convert) += len;
2289 while (len--)
de734bd5 2290 *str++ = *ch++;
8e11feef 2291 } else {
f9049ba1 2292#ifdef DEBUGGING
8e11feef
RGS
2293 if (state>1)
2294 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2295#endif
8e11feef
RGS
2296 break;
2297 }
2298 }
2e64971a 2299 trie->prefixlen = (state-1);
3dab1dad 2300 if (str) {
8e11feef 2301 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2302 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2303 trie->startstate = state;
07be1b83
YO
2304 trie->minlen -= (state - 1);
2305 trie->maxlen -= (state - 1);
33809eae
JH
2306#ifdef DEBUGGING
2307 /* At least the UNICOS C compiler choked on this
2308 * being argument to DEBUG_r(), so let's just have
2309 * it right here. */
2310 if (
2311#ifdef PERL_EXT_RE_BUILD
2312 1
2313#else
2314 DEBUG_r_TEST
2315#endif
2316 ) {
2317 regnode *fix = convert;
2318 U32 word = trie->wordcount;
2319 mjd_nodelen++;
2320 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2321 while( ++fix < n ) {
2322 Set_Node_Offset_Length(fix, 0, 0);
2323 }
2324 while (word--) {
2325 SV ** const tmp = av_fetch( trie_words, word, 0 );
2326 if (tmp) {
2327 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2328 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2329 else
2330 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2331 }
2332 }
2333 }
2334#endif
8e11feef
RGS
2335 if (trie->maxlen) {
2336 convert = n;
2337 } else {
3dab1dad 2338 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2339 DEBUG_r(optimize= n);
3dab1dad
YO
2340 }
2341 }
2342 }
a5ca303d
YO
2343 if (!jumper)
2344 jumper = last;
3dab1dad 2345 if ( trie->maxlen ) {
8e11feef
RGS
2346 NEXT_OFF( convert ) = (U16)(tail - convert);
2347 ARG_SET( convert, data_slot );
786e8c11
YO
2348 /* Store the offset to the first unabsorbed branch in
2349 jump[0], which is otherwise unused by the jump logic.
2350 We use this when dumping a trie and during optimisation. */
2351 if (trie->jump)
7f69552c 2352 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2353
6c48061a
YO
2354 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2355 * and there is a bitmap
2356 * and the first "jump target" node we found leaves enough room
2357 * then convert the TRIE node into a TRIEC node, with the bitmap
2358 * embedded inline in the opcode - this is hypothetically faster.
2359 */
2360 if ( !trie->states[trie->startstate].wordnum
2361 && trie->bitmap
2362 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2363 {
2364 OP( convert ) = TRIEC;
2365 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2366 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2367 trie->bitmap= NULL;
2368 } else
2369 OP( convert ) = TRIE;
a3621e74 2370
3dab1dad
YO
2371 /* store the type in the flags */
2372 convert->flags = nodetype;
a5ca303d
YO
2373 DEBUG_r({
2374 optimize = convert
2375 + NODE_STEP_REGNODE
2376 + regarglen[ OP( convert ) ];
2377 });
2378 /* XXX We really should free up the resource in trie now,
2379 as we won't use them - (which resources?) dmq */
3dab1dad 2380 }
a3621e74 2381 /* needed for dumping*/
e62cc96a 2382 DEBUG_r(if (optimize) {
07be1b83 2383 regnode *opt = convert;
bcdf7404 2384
e62cc96a 2385 while ( ++opt < optimize) {
07be1b83
YO
2386 Set_Node_Offset_Length(opt,0,0);
2387 }
786e8c11
YO
2388 /*
2389 Try to clean up some of the debris left after the
2390 optimisation.
a3621e74 2391 */
786e8c11 2392 while( optimize < jumper ) {
07be1b83 2393 mjd_nodelen += Node_Length((optimize));
a3621e74 2394 OP( optimize ) = OPTIMIZED;
07be1b83 2395 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2396 optimize++;
2397 }
07be1b83 2398 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2399 });
2400 } /* end node insert */
2e64971a
DM
2401
2402 /* Finish populating the prev field of the wordinfo array. Walk back
2403 * from each accept state until we find another accept state, and if
2404 * so, point the first word's .prev field at the second word. If the
2405 * second already has a .prev field set, stop now. This will be the
2406 * case either if we've already processed that word's accept state,
3b753521
FN
2407 * or that state had multiple words, and the overspill words were
2408 * already linked up earlier.
2e64971a
DM
2409 */
2410 {
2411 U16 word;
2412 U32 state;
2413 U16 prev;
2414
2415 for (word=1; word <= trie->wordcount; word++) {
2416 prev = 0;
2417 if (trie->wordinfo[word].prev)
2418 continue;
2419 state = trie->wordinfo[word].accept;
2420 while (state) {
2421 state = prev_states[state];
2422 if (!state)
2423 break;
2424 prev = trie->states[state].wordnum;
2425 if (prev)
2426 break;
2427 }
2428 trie->wordinfo[word].prev = prev;
2429 }
2430 Safefree(prev_states);
2431 }
2432
2433
2434 /* and now dump out the compressed format */
2435 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2436
55eed653 2437 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2438#ifdef DEBUGGING
2439 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2440 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2441#else
2442 SvREFCNT_dec(revcharmap);
07be1b83 2443#endif
786e8c11
YO
2444 return trie->jump
2445 ? MADE_JUMP_TRIE
2446 : trie->startstate>1
2447 ? MADE_EXACT_TRIE
2448 : MADE_TRIE;
2449}
2450
2451STATIC void
2452S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2453{
3b753521 2454/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2455
2456 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2457 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2458 ISBN 0-201-10088-6
2459
2460 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2461 suffix of the current state's 'word' that is also a proper prefix of another word in our
2462 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2463 the DFA not to have to restart after its tried and failed a word at a given point, it
2464 simply continues as though it had been matching the other word in the first place.
2465 Consider
2466 'abcdgu'=~/abcdefg|cdgu/
2467 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2468 fail, which would bring us to the state representing 'd' in the second word where we would
2469 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2470 */
2471 /* add a fail transition */
3251b653
NC
2472 const U32 trie_offset = ARG(source);
2473 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2474 U32 *q;
2475 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2476 const U32 numstates = trie->statecount;
786e8c11
YO
2477 const U32 ubound = trie->lasttrans + ucharcount;
2478 U32 q_read = 0;
2479 U32 q_write = 0;
2480 U32 charid;
2481 U32 base = trie->states[ 1 ].trans.base;
2482 U32 *fail;
2483 reg_ac_data *aho;
2484 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2485 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2486
2487 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2488#ifndef DEBUGGING
2489 PERL_UNUSED_ARG(depth);
2490#endif
2491
2492
2493 ARG_SET( stclass, data_slot );
c944940b 2494 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2495 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2496 aho->trie=trie_offset;
446bd890
NC
2497 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2498 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2499 Newxz( q, numstates, U32);
c944940b 2500 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2501 aho->refcount = 1;
2502 fail = aho->fail;
2503 /* initialize fail[0..1] to be 1 so that we always have
2504 a valid final fail state */
2505 fail[ 0 ] = fail[ 1 ] = 1;
2506
2507 for ( charid = 0; charid < ucharcount ; charid++ ) {
2508 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2509 if ( newstate ) {
2510 q[ q_write ] = newstate;
2511 /* set to point at the root */
2512 fail[ q[ q_write++ ] ]=1;
2513 }
2514 }
2515 while ( q_read < q_write) {
2516 const U32 cur = q[ q_read++ % numstates ];
2517 base = trie->states[ cur ].trans.base;
2518
2519 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2520 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2521 if (ch_state) {
2522 U32 fail_state = cur;
2523 U32 fail_base;
2524 do {
2525 fail_state = fail[ fail_state ];
2526 fail_base = aho->states[ fail_state ].trans.base;
2527 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2528
2529 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2530 fail[ ch_state ] = fail_state;
2531 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2532 {
2533 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2534 }
2535 q[ q_write++ % numstates] = ch_state;
2536 }
2537 }
2538 }
2539 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2540 when we fail in state 1, this allows us to use the
2541 charclass scan to find a valid start char. This is based on the principle
2542 that theres a good chance the string being searched contains lots of stuff
2543 that cant be a start char.
2544 */
2545 fail[ 0 ] = fail[ 1 ] = 0;
2546 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2547 PerlIO_printf(Perl_debug_log,
2548 "%*sStclass Failtable (%"UVuf" states): 0",
2549 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2550 );
786e8c11
YO
2551 for( q_read=1; q_read<numstates; q_read++ ) {
2552 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2553 }
2554 PerlIO_printf(Perl_debug_log, "\n");
2555 });
2556 Safefree(q);
2557 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2558}
2559
786e8c11 2560
a3621e74 2561/*
5d1c421c
JH
2562 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2563 * These need to be revisited when a newer toolchain becomes available.
2564 */
2565#if defined(__sparc64__) && defined(__GNUC__)
2566# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2567# undef SPARC64_GCC_WORKAROUND
2568# define SPARC64_GCC_WORKAROUND 1
2569# endif
2570#endif
2571
07be1b83 2572#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2573 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2574 SV * const mysv=sv_newmortal(); \
2575 regnode *Next = regnext(scan); \
2576 regprop(RExC_rx, mysv, scan); \
7f69552c 2577 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2578 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2579 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2580 }});
07be1b83 2581
1de06328 2582
bb914485
KW
2583/* The below joins as many adjacent EXACTish nodes as possible into a single
2584 * one, and looks for problematic sequences of characters whose folds vs.
2585 * non-folds have sufficiently different lengths, that the optimizer would be
2586 * fooled into rejecting legitimate matches of them, and the trie construction
87b8b349 2587 * code needs to handle specially. The joining is only done if:
bb914485
KW
2588 * 1) there is room in the current conglomerated node to entirely contain the
2589 * next one.
2590 * 2) they are the exact same node type
2591 *
87b8b349 2592 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
2593 * these get optimized out
2594 *
9d071ca8
KW
2595 * If there are problematic code sequences, *min_subtract is set to the delta
2596 * that the minimum size of the node can be less than its actual size. And,
2597 * the node type of the result is changed to reflect that it contains these
bb914485
KW
2598 * sequences.
2599 *
a0c4c608
KW
2600 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2601 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2602 *
bb914485
KW
2603 * This is as good a place as any to discuss the design of handling these
2604 * problematic sequences. It's been wrong in Perl for a very long time. There
87b8b349
KW
2605 * are three code points currently in Unicode whose folded lengths differ so
2606 * much from the un-folded lengths that it causes problems for the optimizer
2607 * and trie construction. Why only these are problematic, and not others where
2608 * lengths also differ is something I (khw) do not understand. New versions of
2609 * Unicode might add more such code points. Hopefully the logic in
2610 * fold_grind.t that figures out what to test (in part by verifying that each
2611 * size-combination gets tested) will catch any that do come along, so they can
2612 * be added to the special handling below. The chances of new ones are
2613 * actually rather small, as most, if not all, of the world's scripts that have
2614 * casefolding have already been encoded by Unicode. Also, a number of
2615 * Unicode's decisions were made to allow compatibility with pre-existing
2616 * standards, and almost all of those have already been dealt with. These
2617 * would otherwise be the most likely candidates for generating further tricky
2618 * sequences. In other words, Unicode by itself is unlikely to add new ones
2619 * unless it is for compatibility with pre-existing standards, and there aren't
2620 * many of those left.
bb914485
KW
2621 *
2622 * The previous designs for dealing with these involved assigning a special
2623 * node for them. This approach doesn't work, as evidenced by this example:
a0c4c608 2624 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
87b8b349 2625 * Both these fold to "sss", but if the pattern is parsed to create a node
bb914485
KW
2626 * that would match just the \xDF, it won't be able to handle the case where a
2627 * successful match would have to cross the node's boundary. The new approach
2628 * that hopefully generally solves the problem generates an EXACTFU_SS node
2629 * that is "sss".
2630 *
2631 * There are a number of components to the approach (a lot of work for just
2632 * three code points!):
2633 * 1) This routine examines each EXACTFish node that could contain the
9d071ca8
KW
2634 * problematic sequences. It returns in *min_subtract how much to
2635 * subtract from the the actual length of the string to get a real minimum
2636 * for one that could match it. This number is usually 0 except for the
2637 * problematic sequences. This delta is used by the caller to adjust the
2638 * min length of the match, and the delta between min and max, so that the
2639 * optimizer doesn't reject these possibilities based on size constraints.
87b8b349
KW
2640 * 2) These sequences require special handling by the trie code, so this code
2641 * changes the joined node type to special ops: EXACTFU_TRICKYFOLD and
2642 * EXACTFU_SS.
bb914485
KW
2643 * 3) This is sufficient for the two Greek sequences (described below), but
2644 * the one involving the Sharp s (\xDF) needs more. The node type
2645 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2646 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2647 * case where there is a possible fold length change. That means that a
2648 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2649 * itself with length changes, and so can be processed faster. regexec.c
2650 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2651 * is pre-folded by regcomp.c. This saves effort in regex matching.
87b8b349
KW
2652 * However, the pre-folding isn't done for non-UTF8 patterns because the
2653 * fold of the MICRO SIGN requires UTF-8. Also what EXACTF and EXACTFL
2654 * nodes fold to isn't known until runtime. The fold possibilities for
2655 * the non-UTF8 patterns are quite simple, except for the sharp s. All
2656 * the ones that don't involve a UTF-8 target string are members of a
2657 * fold-pair, and arrays are set up for all of them so that the other
2658 * member of the pair can be found quickly. Code elsewhere in this file
2659 * makes sure that in EXACTFU nodes, the sharp s gets folded to 'ss', even
2660 * if the pattern isn't UTF-8. This avoids the issues described in the
2661 * next item.
bb914485
KW
2662 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2663 * 'ss' or not is not knowable at compile time. It will match iff the
2664 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2665 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
6012a526 2666 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
bb914485
KW
2667 * described in item 3). An assumption that the optimizer part of
2668 * regexec.c (probably unwittingly) makes is that a character in the
2669 * pattern corresponds to at most a single character in the target string.
2670 * (And I do mean character, and not byte here, unlike other parts of the
2671 * documentation that have never been updated to account for multibyte
2672 * Unicode.) This assumption is wrong only in this case, as all other
2673 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2674 * virtue of having this file pre-fold UTF-8 patterns. I'm
2675 * reluctant to try to change this assumption, so instead the code punts.
9d071ca8
KW
2676 * This routine examines EXACTF nodes for the sharp s, and returns a
2677 * boolean indicating whether or not the node is an EXACTF node that
2678 * contains a sharp s. When it is true, the caller sets a flag that later
2679 * causes the optimizer in this file to not set values for the floating
2680 * and fixed string lengths, and thus avoids the optimizer code in
2681 * regexec.c that makes the invalid assumption. Thus, there is no
2682 * optimization based on string lengths for EXACTF nodes that contain the
2683 * sharp s. This only happens for /id rules (which means the pattern
2684 * isn't in UTF-8).
bb914485 2685 */
1de06328 2686
9d071ca8 2687#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2688 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2689 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2690
be8e71aa 2691STATIC U32
9d071ca8 2692S_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
2693 /* Merge several consecutive EXACTish nodes into one. */
2694 regnode *n = regnext(scan);
2695 U32 stringok = 1;
2696 regnode *next = scan + NODE_SZ_STR(scan);
2697 U32 merged = 0;
2698 U32 stopnow = 0;
2699#ifdef DEBUGGING
2700 regnode *stop = scan;
72f13be8 2701 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2702#else
d47053eb
RGS
2703 PERL_UNUSED_ARG(depth);
2704#endif
7918f24d
NC
2705
2706 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2707#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2708 PERL_UNUSED_ARG(flags);
2709 PERL_UNUSED_ARG(val);
07be1b83 2710#endif
07be1b83 2711 DEBUG_PEEP("join",scan,depth);
bb914485 2712
3f410cf6
KW
2713 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2714 * EXACT ones that are mergeable to the current one. */
2715 while (n
2716 && (PL_regkind[OP(n)] == NOTHING
2717 || (stringok && OP(n) == OP(scan)))
07be1b83 2718 && NEXT_OFF(n)
3f410cf6
KW
2719 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2720 {
07be1b83
YO
2721
2722 if (OP(n) == TAIL || n > next)
2723 stringok = 0;
2724 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2725 DEBUG_PEEP("skip:",n,depth);
2726 NEXT_OFF(scan) += NEXT_OFF(n);
2727 next = n + NODE_STEP_REGNODE;
2728#ifdef DEBUGGING
2729 if (stringok)
2730 stop = n;
2731#endif
2732 n = regnext(n);
2733 }
2734 else if (stringok) {
786e8c11 2735 const unsigned int oldl = STR_LEN(scan);
07be1b83 2736 regnode * const nnext = regnext(n);
b2230d39 2737
87b8b349
KW
2738 /* XXX I (khw) kind of doubt that this works on platforms where
2739 * U8_MAX is above 255 because of lots of other assumptions */
b2230d39
KW
2740 if (oldl + STR_LEN(n) > U8_MAX)
2741 break;
07be1b83
YO
2742
2743 DEBUG_PEEP("merg",n,depth);
07be1b83 2744 merged++;
b2230d39 2745
07be1b83
YO
2746 NEXT_OFF(scan) += NEXT_OFF(n);
2747 STR_LEN(scan) += STR_LEN(n);
2748 next = n + NODE_SZ_STR(n);
2749 /* Now we can overwrite *n : */
2750 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2751#ifdef DEBUGGING
2752 stop = next - 1;
2753#endif
2754 n = nnext;
2755 if (stopnow) break;
2756 }
2757
d47053eb
RGS
2758#ifdef EXPERIMENTAL_INPLACESCAN
2759 if (flags && !NEXT_OFF(n)) {
2760 DEBUG_PEEP("atch", val, depth);
2761 if (reg_off_by_arg[OP(n)]) {
2762 ARG_SET(n, val - n);
2763 }
2764 else {
2765 NEXT_OFF(n) = val - n;
2766 }
2767 stopnow = 1;
2768 }
07be1b83
YO
2769#endif
2770 }
2c2b7f86 2771
9d071ca8 2772 *min_subtract = 0;
f758bddf 2773 *has_exactf_sharp_s = FALSE;
f646642f 2774
3f410cf6
KW
2775 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2776 * can now analyze for sequences of problematic code points. (Prior to
2777 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2778 * hence missed). The sequences only happen in folding, hence for any
2779 * non-EXACT EXACTish node */
86d6fcad 2780 if (OP(scan) != EXACT) {
f758bddf
KW
2781 U8 *s;
2782 U8 * s0 = (U8*) STRING(scan);
2783 U8 * const s_end = s0 + STR_LEN(scan);
2784
2785 /* The below is perhaps overboard, but this allows us to save a test
2786 * each time through the loop at the expense of a mask. This is
2787 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2788 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2789 * This uses an exclusive 'or' to find that bit and then inverts it to
2790 * form a mask, with just a single 0, in the bit position where 'S' and
2791 * 's' differ. */
dbeb8947 2792 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
f758bddf
KW
2793 const U8 s_masked = 's' & S_or_s_mask;
2794
2795 /* One pass is made over the node's string looking for all the
2796 * possibilities. to avoid some tests in the loop, there are two main
2797 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2798 * non-UTF-8 */
2799 if (UTF) {
86d6fcad 2800
f758bddf
KW
2801 /* There are two problematic Greek code points in Unicode
2802 * casefolding
86d6fcad
KW
2803 *
2804 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2805 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2806 *
2807 * which casefold to
2808 *
2809 * Unicode UTF-8
2810 *
2811 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2812 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2813 *
2814 * This means that in case-insensitive matching (or "loose
2815 * matching", as Unicode calls it), an EXACTF of length six (the
2816 * UTF-8 encoded byte length of the above casefolded versions) can
2817 * match a target string of length two (the byte length of UTF-8
2818 * encoded U+0390 or U+03B0). This would rather mess up the
2819 * minimum length computation. (there are other code points that
2820 * also fold to these two sequences, but the delta is smaller)
2821 *
f758bddf
KW
2822 * If these sequences are found, the minimum length is decreased by
2823 * four (six minus two).
86d6fcad 2824 *
f758bddf
KW
2825 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2826 * LETTER SHARP S. We decrease the min length by 1 for each
2827 * occurrence of 'ss' found */
3f410cf6 2828
61dad979
KW
2829#define U390_FIRST_BYTE GREEK_SMALL_LETTER_IOTA_UTF8_FIRST_BYTE
2830#define U3B0_FIRST_BYTE GREEK_SMALL_LETTER_UPSILON_UTF8_FIRST_BYTE
2831 const U8 U390_tail[] = GREEK_SMALL_LETTER_IOTA_UTF8_TAIL
2832 COMBINING_DIAERESIS_UTF8
2833 COMBINING_ACUTE_ACCENT_UTF8;
2834 const U8 U3B0_tail[] = GREEK_SMALL_LETTER_UPSILON_UTF8_TAIL
2835 COMBINING_DIAERESIS_UTF8
2836 COMBINING_ACUTE_ACCENT_UTF8;
2837 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
f758bddf
KW
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
61dad979 2867 case U390_FIRST_BYTE:
f758bddf
KW
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
61dad979 2878 case U3B0_FIRST_BYTE:
f758bddf
KW
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 2886
87b8b349 2887 /* This requires special handling by trie's, so change
f758bddf
KW
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
87b8b349
KW
2923 * remain EXACTF nodes, as they won't match this
2924 * unless the target string is is UTF-8, which we
2925 * don't know until runtime */
f758bddf
KW
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<