This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Make sure optimizer handles node; Note for 3018b823
[perl5.git] / regcomp.c
CommitLineData
a0d0e21e
LW
1/* regcomp.c
2 */
3
4/*
4ac71550
TC
5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
6 *
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
a0d0e21e
LW
8 */
9
61296642
DM
10/* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
166f8a29 12 * a regular expression.
e4a054ea
DM
13 *
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
166f8a29
DM
18 */
19
a687059c
LW
20/* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
22 */
23
24/* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
27 */
28
e50aee73 29/* The names of the functions have been changed from regcomp and
3b753521 30 * regexec to pregcomp and pregexec in order to avoid conflicts
e50aee73
AD
31 * with the POSIX routines of the same names.
32*/
33
b9d5759e 34#ifdef PERL_EXT_RE_BUILD
54df2634 35#include "re_top.h"
b81d288d 36#endif
56953603 37
a687059c 38/*
e50aee73 39 * pregcomp and pregexec -- regsub and regerror are not used in perl
a687059c
LW
40 *
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
43 *
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
47 *
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
50 * from defects in it.
51 *
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
54 *
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
57 *
58 *
59 **** Alterations to Henry's code are...
60 ****
4bb101f2 61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
1129b882
NC
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
a687059c 64 ****
9ef589d8
LW
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
67
a687059c
LW
68 *
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
72 */
73#include "EXTERN.h"
864dbfa3 74#define PERL_IN_REGCOMP_C
a687059c 75#include "perl.h"
d06ea78c 76
acfe0abc 77#ifndef PERL_IN_XSUB_RE
d06ea78c
GS
78# include "INTERN.h"
79#endif
c277df42
IZ
80
81#define REG_COMP_C
54df2634
NC
82#ifdef PERL_IN_XSUB_RE
83# include "re_comp.h"
40b6423c 84extern const struct regexp_engine my_reg_engine;
54df2634
NC
85#else
86# include "regcomp.h"
87#endif
a687059c 88
04e98a4d 89#include "dquote_static.c"
26faadbd 90#include "charclass_invlists.h"
81e983c1 91#include "inline_invlist.c"
1b0f46bf 92#include "unicode_constants.h"
04e98a4d 93
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)
2c61f163 96#define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94dc5c2d 97
d4cce5f1 98#ifdef op
11343788 99#undef op
d4cce5f1 100#endif /* op */
11343788 101
fe14fcc3 102#ifdef MSDOS
7e4e8c89 103# if defined(BUGGY_MSC6)
fe14fcc3 104 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
7e4e8c89 105# pragma optimize("a",off)
fe14fcc3 106 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
7e4e8c89
NC
107# pragma optimize("w",on )
108# endif /* BUGGY_MSC6 */
fe14fcc3
LW
109#endif /* MSDOS */
110
a687059c
LW
111#ifndef STATIC
112#define STATIC static
113#endif
114
b1603ef8 115
830247a4 116typedef struct RExC_state_t {
514a91f1
DM
117 U32 flags; /* RXf_* are we folding, multilining? */
118 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 119 char *precomp; /* uncompiled string. */
288b8c02 120 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf
YO
121 regexp *rx; /* perl core regexp structure */
122 regexp_internal *rxi; /* internal data for regexp object pprivate field */
fac92740 123 char *start; /* Start of input for compile */
830247a4
IZ
124 char *end; /* End of input for compile */
125 char *parse; /* Input-scan pointer. */
126 I32 whilem_seen; /* number of WHILEM in this expr */
fac92740 127 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 128 regnode *emit_bound; /* First regnode outside of the allocated space */
ffc61ed2 129 regnode *emit; /* Code-emit pointer; &regdummy = don't = compiling */
830247a4
IZ
130 I32 naughty; /* How bad is this pattern? */
131 I32 sawback; /* Did we see \1, ...? */
132 U32 seen;
133 I32 size; /* Code size. */
c74340f9
YO
134 I32 npar; /* Capture buffer count, (OPEN). */
135 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 136 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
137 I32 extralen;
138 I32 seen_zerolen;
40d049e4
YO
139 regnode **open_parens; /* pointers to open parens */
140 regnode **close_parens; /* pointers to close parens */
141 regnode *opend; /* END node in program */
02daf0ab
YO
142 I32 utf8; /* whether the pattern is utf8 or not */
143 I32 orig_utf8; /* whether the pattern was originally in utf8 */
144 /* XXX use this for future optimisation of case
145 * where pattern must be upgraded to utf8. */
e40e74fe
KW
146 I32 uni_semantics; /* If a d charset modifier should use unicode
147 rules, even if the pattern is not in
148 utf8 */
81714fb9 149 HV *paren_names; /* Paren names */
1f1031fe 150
40d049e4
YO
151 regnode **recurse; /* Recurse regops */
152 I32 recurse_count; /* Number of recurse regops */
b57e4118 153 I32 in_lookbehind;
4624b182 154 I32 contains_locale;
bb3f3ed2 155 I32 override_recoding;
9d53c457 156 I32 in_multi_char_class;
3d2bd50a 157 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 158 within pattern */
b1603ef8
DM
159 int num_code_blocks; /* size of code_blocks[] */
160 int code_index; /* next code_blocks[] slot */
830247a4
IZ
161#if ADD_TO_REGEXEC
162 char *starttry; /* -Dr: where regtry was called. */
163#define RExC_starttry (pRExC_state->starttry)
164#endif
d24ca0c5 165 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 166#ifdef DEBUGGING
be8e71aa 167 const char *lastparse;
3dab1dad 168 I32 lastnum;
1f1031fe 169 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
170#define RExC_lastparse (pRExC_state->lastparse)
171#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 172#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 173#endif
830247a4
IZ
174} RExC_state_t;
175
e2509266 176#define RExC_flags (pRExC_state->flags)
514a91f1 177#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 178#define RExC_precomp (pRExC_state->precomp)
288b8c02 179#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 180#define RExC_rx (pRExC_state->rx)
f8fc2ecf 181#define RExC_rxi (pRExC_state->rxi)
fac92740 182#define RExC_start (pRExC_state->start)
830247a4
IZ
183#define RExC_end (pRExC_state->end)
184#define RExC_parse (pRExC_state->parse)
185#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
186#ifdef RE_TRACK_PATTERN_OFFSETS
187#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
188#endif
830247a4 189#define RExC_emit (pRExC_state->emit)
fac92740 190#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 191#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
192#define RExC_naughty (pRExC_state->naughty)
193#define RExC_sawback (pRExC_state->sawback)
194#define RExC_seen (pRExC_state->seen)
195#define RExC_size (pRExC_state->size)
196#define RExC_npar (pRExC_state->npar)
e2e6a0f1 197#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
198#define RExC_extralen (pRExC_state->extralen)
199#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 200#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 201#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 202#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
203#define RExC_open_parens (pRExC_state->open_parens)
204#define RExC_close_parens (pRExC_state->close_parens)
205#define RExC_opend (pRExC_state->opend)
81714fb9 206#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
207#define RExC_recurse (pRExC_state->recurse)
208#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 209#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 210#define RExC_contains_locale (pRExC_state->contains_locale)
9d53c457
KW
211#define RExC_override_recoding (pRExC_state->override_recoding)
212#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
830247a4 213
cde0cee5 214
a687059c
LW
215#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
216#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
217 ((*s) == '{' && regcurly(s)))
a687059c 218
35c8bce7
LW
219#ifdef SPSTART
220#undef SPSTART /* dratted cpp namespace... */
221#endif
a687059c
LW
222/*
223 * Flags to be passed up and down.
224 */
a687059c 225#define WORST 0 /* Worst case. */
a3b492c3 226#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 227
e64f369d 228/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
229 * character. (There needs to be a case: in the switch statement in regexec.c
230 * for any node marked SIMPLE.) Note that this is not the same thing as
231 * REGNODE_SIMPLE */
fda99bee 232#define SIMPLE 0x02
e64f369d 233#define SPSTART 0x04 /* Starts with * or + */
a3b492c3
YO
234#define TRYAGAIN 0x08 /* Weeded out a declaration. */
235#define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
a687059c 236
3dab1dad
YO
237#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
238
07be1b83
YO
239/* whether trie related optimizations are enabled */
240#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
241#define TRIE_STUDY_OPT
786e8c11 242#define FULL_TRIE_STUDY
07be1b83
YO
243#define TRIE_STCLASS
244#endif
1de06328
YO
245
246
40d049e4
YO
247
248#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
249#define PBITVAL(paren) (1 << ((paren) & 7))
250#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
251#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
252#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
253
bbd61b5f
KW
254/* If not already in utf8, do a longjmp back to the beginning */
255#define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
256#define REQUIRE_UTF8 STMT_START { \
257 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
258 } STMT_END
40d049e4 259
f19b1a63
KW
260/* This converts the named class defined in regcomp.h to its equivalent class
261 * number defined in handy.h. */
262#define namedclass_to_classnum(class) ((int) ((class) / 2))
263#define classnum_to_namedclass(classnum) ((classnum) * 2)
264
1de06328
YO
265/* About scan_data_t.
266
267 During optimisation we recurse through the regexp program performing
268 various inplace (keyhole style) optimisations. In addition study_chunk
269 and scan_commit populate this data structure with information about
270 what strings MUST appear in the pattern. We look for the longest
3b753521 271 string that must appear at a fixed location, and we look for the
1de06328
YO
272 longest string that may appear at a floating location. So for instance
273 in the pattern:
274
275 /FOO[xX]A.*B[xX]BAR/
276
277 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
278 strings (because they follow a .* construct). study_chunk will identify
279 both FOO and BAR as being the longest fixed and floating strings respectively.
280
281 The strings can be composites, for instance
282
283 /(f)(o)(o)/
284
285 will result in a composite fixed substring 'foo'.
286
287 For each string some basic information is maintained:
288
289 - offset or min_offset
290 This is the position the string must appear at, or not before.
291 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
292 characters must match before the string we are searching for.
293 Likewise when combined with minlenp and the length of the string it
1de06328
YO
294 tells us how many characters must appear after the string we have
295 found.
296
297 - max_offset
298 Only used for floating strings. This is the rightmost point that
3b753521 299 the string can appear at. If set to I32 max it indicates that the
1de06328
YO
300 string can occur infinitely far to the right.
301
302 - minlenp
2d608413
KW
303 A pointer to the minimum number of characters of the pattern that the
304 string was found inside. This is important as in the case of positive
1de06328
YO
305 lookahead or positive lookbehind we can have multiple patterns
306 involved. Consider
307
308 /(?=FOO).*F/
309
310 The minimum length of the pattern overall is 3, the minimum length
311 of the lookahead part is 3, but the minimum length of the part that
312 will actually match is 1. So 'FOO's minimum length is 3, but the
313 minimum length for the F is 1. This is important as the minimum length
314 is used to determine offsets in front of and behind the string being
315 looked for. Since strings can be composites this is the length of the
486ec47a 316 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
317 the length is calculated by study_chunk, so that the minimum lengths
318 are not known until the full pattern has been compiled, thus the
319 pointer to the value.
320
321 - lookbehind
322
323 In the case of lookbehind the string being searched for can be
324 offset past the start point of the final matching string.
325 If this value was just blithely removed from the min_offset it would
326 invalidate some of the calculations for how many chars must match
327 before or after (as they are derived from min_offset and minlen and
328 the length of the string being searched for).
329 When the final pattern is compiled and the data is moved from the
330 scan_data_t structure into the regexp structure the information
331 about lookbehind is factored in, with the information that would
332 have been lost precalculated in the end_shift field for the
333 associated string.
334
335 The fields pos_min and pos_delta are used to store the minimum offset
336 and the delta to the maximum offset at the current point in the pattern.
337
338*/
2c2d71f5
JH
339
340typedef struct scan_data_t {
1de06328
YO
341 /*I32 len_min; unused */
342 /*I32 len_delta; unused */
2c2d71f5
JH
343 I32 pos_min;
344 I32 pos_delta;
345 SV *last_found;
1de06328 346 I32 last_end; /* min value, <0 unless valid. */
2c2d71f5
JH
347 I32 last_start_min;
348 I32 last_start_max;
1de06328
YO
349 SV **longest; /* Either &l_fixed, or &l_float. */
350 SV *longest_fixed; /* longest fixed string found in pattern */
351 I32 offset_fixed; /* offset where it starts */
486ec47a 352 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
353 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
354 SV *longest_float; /* longest floating string found in pattern */
355 I32 offset_float_min; /* earliest point in string it can appear */
356 I32 offset_float_max; /* latest point in string it can appear */
486ec47a 357 I32 *minlen_float; /* pointer to the minlen relevant to the string */
1de06328 358 I32 lookbehind_float; /* is the position of the string modified by LB */
2c2d71f5
JH
359 I32 flags;
360 I32 whilem_c;
cb434fcc 361 I32 *last_closep;
653099ff 362 struct regnode_charclass_class *start_class;
2c2d71f5
JH
363} scan_data_t;
364
a687059c 365/*
e50aee73 366 * Forward declarations for pregcomp()'s friends.
a687059c 367 */
a0d0e21e 368
27da23d5 369static const scan_data_t zero_scan_data =
1de06328 370 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
371
372#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
373#define SF_BEFORE_SEOL 0x0001
374#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
375#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
376#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
377
09b7f37c
CB
378#ifdef NO_UNARY_PLUS
379# define SF_FIX_SHIFT_EOL (0+2)
380# define SF_FL_SHIFT_EOL (0+4)
381#else
382# define SF_FIX_SHIFT_EOL (+2)
383# define SF_FL_SHIFT_EOL (+4)
384#endif
c277df42
IZ
385
386#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
387#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
388
389#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
390#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
391#define SF_IS_INF 0x0040
392#define SF_HAS_PAR 0x0080
393#define SF_IN_PAR 0x0100
394#define SF_HAS_EVAL 0x0200
395#define SCF_DO_SUBSTR 0x0400
653099ff
GS
396#define SCF_DO_STCLASS_AND 0x0800
397#define SCF_DO_STCLASS_OR 0x1000
398#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 399#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 400
786e8c11 401#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 402#define SCF_SEEN_ACCEPT 0x8000
07be1b83 403
43fead97 404#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
405
406/* The enums for all these are ordered so things work out correctly */
a62b1201 407#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
cfaf538b 408#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
00b27cfc 409#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
410#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
411#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1 412#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a725e29c 413#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 414
43fead97 415#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 416
93733859 417#define OOB_NAMEDCLASS -1
b8c5462f 418
8e661ac5
KW
419/* There is no code point that is out-of-bounds, so this is problematic. But
420 * its only current use is to initialize a variable that is always set before
421 * looked at. */
422#define OOB_UNICODE 0xDEADBEEF
423
a0ed51b3
LW
424#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
425#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
426
8615cb43 427
b45f050a
JF
428/* length of regex to show in messages that don't mark a position within */
429#define RegexLengthToShowInErrorMessages 127
430
431/*
432 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
433 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
434 * op/pragma/warn/regcomp.
435 */
7253e4e3
RK
436#define MARKER1 "<-- HERE" /* marker as it appears in the description */
437#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 438
7253e4e3 439#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
440
441/*
442 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
443 * arg. Show regex, up to a maximum length. If it's too long, chop and add
444 * "...".
445 */
58e23c8d 446#define _FAIL(code) STMT_START { \
bfed75c6 447 const char *ellipses = ""; \
ccb2c380
MP
448 IV len = RExC_end - RExC_precomp; \
449 \
450 if (!SIZE_ONLY) \
a5e7bc51 451 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
452 if (len > RegexLengthToShowInErrorMessages) { \
453 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
454 len = RegexLengthToShowInErrorMessages - 10; \
455 ellipses = "..."; \
456 } \
58e23c8d 457 code; \
ccb2c380 458} STMT_END
8615cb43 459
58e23c8d
YO
460#define FAIL(msg) _FAIL( \
461 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
462 msg, (int)len, RExC_precomp, ellipses))
463
464#define FAIL2(msg,arg) _FAIL( \
465 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
466 arg, (int)len, RExC_precomp, ellipses))
467
b45f050a 468/*
b45f050a
JF
469 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
470 */
ccb2c380 471#define Simple_vFAIL(m) STMT_START { \
a28509cc 472 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
473 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
474 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
475} STMT_END
b45f050a
JF
476
477/*
478 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
479 */
ccb2c380
MP
480#define vFAIL(m) STMT_START { \
481 if (!SIZE_ONLY) \
a5e7bc51 482 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
483 Simple_vFAIL(m); \
484} STMT_END
b45f050a
JF
485
486/*
487 * Like Simple_vFAIL(), but accepts two arguments.
488 */
ccb2c380 489#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 490 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
491 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
492 (int)offset, RExC_precomp, RExC_precomp + offset); \
493} STMT_END
b45f050a
JF
494
495/*
496 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
497 */
ccb2c380
MP
498#define vFAIL2(m,a1) STMT_START { \
499 if (!SIZE_ONLY) \
a5e7bc51 500 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
501 Simple_vFAIL2(m, a1); \
502} STMT_END
b45f050a
JF
503
504
505/*
506 * Like Simple_vFAIL(), but accepts three arguments.
507 */
ccb2c380 508#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 509 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
510 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
511 (int)offset, RExC_precomp, RExC_precomp + offset); \
512} STMT_END
b45f050a
JF
513
514/*
515 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
516 */
ccb2c380
MP
517#define vFAIL3(m,a1,a2) STMT_START { \
518 if (!SIZE_ONLY) \
a5e7bc51 519 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
520 Simple_vFAIL3(m, a1, a2); \
521} STMT_END
b45f050a
JF
522
523/*
524 * Like Simple_vFAIL(), but accepts four arguments.
525 */
ccb2c380 526#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 527 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
528 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
529 (int)offset, RExC_precomp, RExC_precomp + offset); \
530} STMT_END
b45f050a 531
668c081a 532#define ckWARNreg(loc,m) STMT_START { \
a28509cc 533 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
534 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
535 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
536} STMT_END
537
668c081a 538#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 539 const IV offset = loc - RExC_precomp; \
d1d15184 540 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
541 m REPORT_LOCATION, \
542 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
543} STMT_END
544
2335b3d3
KW
545#define ckWARN2regdep(loc,m, a1) STMT_START { \
546 const IV offset = loc - RExC_precomp; \
547 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
548 m REPORT_LOCATION, \
549 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
550} STMT_END
551
668c081a 552#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 553 const IV offset = loc - RExC_precomp; \
668c081a 554 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
555 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
556} STMT_END
557
558#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 559 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
560 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
561 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
562} STMT_END
563
668c081a
NC
564#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
565 const IV offset = loc - RExC_precomp; \
566 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
567 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
568} STMT_END
569
ccb2c380 570#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 571 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
572 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
573 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
574} STMT_END
575
668c081a
NC
576#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
577 const IV offset = loc - RExC_precomp; \
578 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
579 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
580} STMT_END
581
ccb2c380 582#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 583 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
584 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
585 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
586} STMT_END
9d1d55b5 587
8615cb43 588
cd439c50 589/* Allow for side effects in s */
ccb2c380
MP
590#define REGC(c,s) STMT_START { \
591 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
592} STMT_END
cd439c50 593
fac92740
MJD
594/* Macros for recording node offsets. 20001227 mjd@plover.com
595 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
596 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
597 * Element 0 holds the number n.
07be1b83 598 * Position is 1 indexed.
fac92740 599 */
7122b237
YO
600#ifndef RE_TRACK_PATTERN_OFFSETS
601#define Set_Node_Offset_To_R(node,byte)
602#define Set_Node_Offset(node,byte)
603#define Set_Cur_Node_Offset
604#define Set_Node_Length_To_R(node,len)
605#define Set_Node_Length(node,len)
606#define Set_Node_Cur_Length(node)
607#define Node_Offset(n)
608#define Node_Length(n)
609#define Set_Node_Offset_Length(node,offset,len)
610#define ProgLen(ri) ri->u.proglen
611#define SetProgLen(ri,x) ri->u.proglen = x
612#else
613#define ProgLen(ri) ri->u.offsets[0]
614#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
615#define Set_Node_Offset_To_R(node,byte) STMT_START { \
616 if (! SIZE_ONLY) { \
617 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 618 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 619 if((node) < 0) { \
551405c4 620 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
621 } else { \
622 RExC_offsets[2*(node)-1] = (byte); \
623 } \
624 } \
625} STMT_END
626
627#define Set_Node_Offset(node,byte) \
628 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
629#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
630
631#define Set_Node_Length_To_R(node,len) STMT_START { \
632 if (! SIZE_ONLY) { \
633 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 634 __LINE__, (int)(node), (int)(len))); \
ccb2c380 635 if((node) < 0) { \
551405c4 636 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
637 } else { \
638 RExC_offsets[2*(node)] = (len); \
639 } \
640 } \
641} STMT_END
642
643#define Set_Node_Length(node,len) \
644 Set_Node_Length_To_R((node)-RExC_emit_start, len)
645#define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
646#define Set_Node_Cur_Length(node) \
647 Set_Node_Length(node, RExC_parse - parse_start)
fac92740
MJD
648
649/* Get offsets and lengths */
650#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
651#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
652
07be1b83
YO
653#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
654 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
655 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
656} STMT_END
7122b237 657#endif
07be1b83
YO
658
659#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
660#define EXPERIMENTAL_INPLACESCAN
f427392e 661#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 662
304ee84b
YO
663#define DEBUG_STUDYDATA(str,data,depth) \
664DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 665 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
666 "%*s" str "Pos:%"IVdf"/%"IVdf \
667 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
668 (int)(depth)*2, "", \
669 (IV)((data)->pos_min), \
670 (IV)((data)->pos_delta), \
304ee84b 671 (UV)((data)->flags), \
1de06328 672 (IV)((data)->whilem_c), \
304ee84b
YO
673 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
674 is_inf ? "INF " : "" \
1de06328
YO
675 ); \
676 if ((data)->last_found) \
677 PerlIO_printf(Perl_debug_log, \
678 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
679 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
680 SvPVX_const((data)->last_found), \
681 (IV)((data)->last_end), \
682 (IV)((data)->last_start_min), \
683 (IV)((data)->last_start_max), \
684 ((data)->longest && \
685 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
686 SvPVX_const((data)->longest_fixed), \
687 (IV)((data)->offset_fixed), \
688 ((data)->longest && \
689 (data)->longest==&((data)->longest_float)) ? "*" : "", \
690 SvPVX_const((data)->longest_float), \
691 (IV)((data)->offset_float_min), \
692 (IV)((data)->offset_float_max) \
693 ); \
694 PerlIO_printf(Perl_debug_log,"\n"); \
695});
696
653099ff 697/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 698 Update the longest found anchored substring and the longest found
653099ff
GS
699 floating substrings if needed. */
700
4327152a 701STATIC void
304ee84b 702S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 703{
e1ec3a88
AL
704 const STRLEN l = CHR_SVLEN(data->last_found);
705 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 706 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 707
7918f24d
NC
708 PERL_ARGS_ASSERT_SCAN_COMMIT;
709
c277df42 710 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 711 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
712 if (*data->longest == data->longest_fixed) {
713 data->offset_fixed = l ? data->last_start_min : data->pos_min;
714 if (data->flags & SF_BEFORE_EOL)
b81d288d 715 data->flags
c277df42
IZ
716 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
717 else
718 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 719 data->minlen_fixed=minlenp;
1de06328 720 data->lookbehind_fixed=0;
a0ed51b3 721 }
304ee84b 722 else { /* *data->longest == data->longest_float */
c277df42 723 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
724 data->offset_float_max = (l
725 ? data->last_start_max
c277df42 726 : data->pos_min + data->pos_delta);
304ee84b 727 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 728 data->offset_float_max = I32_MAX;
c277df42 729 if (data->flags & SF_BEFORE_EOL)
b81d288d 730 data->flags
c277df42
IZ
731 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
732 else
733 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
734 data->minlen_float=minlenp;
735 data->lookbehind_float=0;
c277df42
IZ
736 }
737 }
738 SvCUR_set(data->last_found, 0);
0eda9292 739 {
a28509cc 740 SV * const sv = data->last_found;
097eb12c
AL
741 if (SvUTF8(sv) && SvMAGICAL(sv)) {
742 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
743 if (mg)
744 mg->mg_len = 0;
745 }
0eda9292 746 }
c277df42
IZ
747 data->last_end = -1;
748 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 749 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
750}
751
899d20b9
KW
752/* These macros set, clear and test whether the synthetic start class ('ssc',
753 * given by the parameter) matches an empty string (EOS). This uses the
754 * 'next_off' field in the node, to save a bit in the flags field. The ssc
755 * stands alone, so there is never a next_off, so this field is otherwise
756 * unused. The EOS information is used only for compilation, but theoretically
757 * it could be passed on to the execution code. This could be used to store
758 * more than one bit of information, but only this one is currently used. */
759#define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
760#define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
761#define TEST_SSC_EOS(node) cBOOL((node)->next_off)
762
653099ff
GS
763/* Can match anything (initialization) */
764STATIC void
3fffb88a 765S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 766{
7918f24d
NC
767 PERL_ARGS_ASSERT_CL_ANYTHING;
768
f8bef550 769 ANYOF_BITMAP_SETALL(cl);
899d20b9
KW
770 cl->flags = ANYOF_UNICODE_ALL;
771 SET_SSC_EOS(cl);
3fffb88a
KW
772
773 /* If any portion of the regex is to operate under locale rules,
774 * initialization includes it. The reason this isn't done for all regexes
775 * is that the optimizer was written under the assumption that locale was
776 * all-or-nothing. Given the complexity and lack of documentation in the
777 * optimizer, and that there are inadequate test cases for locale, so many
778 * parts of it may not work properly, it is safest to avoid locale unless
779 * necessary. */
780 if (RExC_contains_locale) {
9d7a1e63 781 ANYOF_CLASS_SETALL(cl); /* /l uses class */
b25a1036 782 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
3fffb88a 783 }
9d7a1e63
KW
784 else {
785 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
786 }
653099ff
GS
787}
788
789/* Can match anything (initialization) */
790STATIC int
5f66b61c 791S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
792{
793 int value;
794
7918f24d
NC
795 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
796
aaa51d5e 797 for (value = 0; value <= ANYOF_MAX; value += 2)
653099ff
GS
798 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
799 return 1;
1aa99e6b
IH
800 if (!(cl->flags & ANYOF_UNICODE_ALL))
801 return 0;
10edeb5d 802 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 803 return 0;
653099ff
GS
804 return 1;
805}
806
807/* Can match anything (initialization) */
808STATIC void
e755fd73 809S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 810{
7918f24d
NC
811 PERL_ARGS_ASSERT_CL_INIT;
812
8ecf7187 813 Zero(cl, 1, struct regnode_charclass_class);
653099ff 814 cl->type = ANYOF;
3fffb88a 815 cl_anything(pRExC_state, cl);
1411dba4 816 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
817}
818
1051e1c4
KW
819/* These two functions currently do the exact same thing */
820#define cl_init_zero S_cl_init
653099ff 821
dd58aee1
KW
822/* 'AND' a given class with another one. Can create false positives. 'cl'
823 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
824 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 825STATIC void
5f66b61c 826S_cl_and(struct regnode_charclass_class *cl,
a28509cc 827 const struct regnode_charclass_class *and_with)
653099ff 828{
7918f24d 829 PERL_ARGS_ASSERT_CL_AND;
40d049e4 830
954a2af6 831 assert(PL_regkind[and_with->type] == ANYOF);
1e6ade67 832
c6b76537 833 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
834 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
835 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 836 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
837 && !(and_with->flags & ANYOF_LOC_FOLD)
838 && !(cl->flags & ANYOF_LOC_FOLD)) {
653099ff
GS
839 int i;
840
841 if (and_with->flags & ANYOF_INVERT)
842 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
843 cl->bitmap[i] &= ~and_with->bitmap[i];
844 else
845 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
846 cl->bitmap[i] &= and_with->bitmap[i];
847 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 848
c6b76537 849 if (and_with->flags & ANYOF_INVERT) {
8951c461 850
c6b76537
KW
851 /* Here, the and'ed node is inverted. Get the AND of the flags that
852 * aren't affected by the inversion. Those that are affected are
853 * handled individually below */
854 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
855 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
856 cl->flags |= affected_flags;
857
858 /* We currently don't know how to deal with things that aren't in the
859 * bitmap, but we know that the intersection is no greater than what
860 * is already in cl, so let there be false positives that get sorted
861 * out after the synthetic start class succeeds, and the node is
862 * matched for real. */
863
864 /* The inversion of these two flags indicate that the resulting
865 * intersection doesn't have them */
866 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
867 cl->flags &= ~ANYOF_UNICODE_ALL;
868 }
c6b76537
KW
869 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
870 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 871 }
1aa99e6b 872 }
c6b76537 873 else { /* and'd node is not inverted */
3ad98780
KW
874 U8 outside_bitmap_but_not_utf8; /* Temp variable */
875
137165a6 876 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
877
878 /* Here 'and_with' doesn't match anything outside the bitmap
879 * (except possibly ANYOF_UNICODE_ALL), which means the
880 * intersection can't either, except for ANYOF_UNICODE_ALL, in
881 * which case we don't know what the intersection is, but it's no
882 * greater than what cl already has, so can just leave it alone,
883 * with possible false positives */
884 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
885 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 886 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 887 }
137165a6 888 }
c6b76537
KW
889 else if (! ANYOF_NONBITMAP(cl)) {
890
891 /* Here, 'and_with' does match something outside the bitmap, and cl
892 * doesn't have a list of things to match outside the bitmap. If
893 * cl can match all code points above 255, the intersection will
3ad98780
KW
894 * be those above-255 code points that 'and_with' matches. If cl
895 * can't match all Unicode code points, it means that it can't
896 * match anything outside the bitmap (since the 'if' that got us
897 * into this block tested for that), so we leave the bitmap empty.
898 */
c6b76537
KW
899 if (cl->flags & ANYOF_UNICODE_ALL) {
900 ARG_SET(cl, ARG(and_with));
3ad98780
KW
901
902 /* and_with's ARG may match things that don't require UTF8.
903 * And now cl's will too, in spite of this being an 'and'. See
904 * the comments below about the kludge */
905 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
906 }
907 }
908 else {
909 /* Here, both 'and_with' and cl match something outside the
910 * bitmap. Currently we do not do the intersection, so just match
911 * whatever cl had at the beginning. */
912 }
913
914
3ad98780
KW
915 /* Take the intersection of the two sets of flags. However, the
916 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
917 * kludge around the fact that this flag is not treated like the others
918 * which are initialized in cl_anything(). The way the optimizer works
919 * is that the synthetic start class (SSC) is initialized to match
920 * anything, and then the first time a real node is encountered, its
921 * values are AND'd with the SSC's with the result being the values of
922 * the real node. However, there are paths through the optimizer where
923 * the AND never gets called, so those initialized bits are set
924 * inappropriately, which is not usually a big deal, as they just cause
925 * false positives in the SSC, which will just mean a probably
926 * imperceptible slow down in execution. However this bit has a
927 * higher false positive consequence in that it can cause utf8.pm,
928 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
929 * bigger slowdown and also causes significant extra memory to be used.
930 * In order to prevent this, the code now takes a different tack. The
931 * bit isn't set unless some part of the regular expression needs it,
932 * but once set it won't get cleared. This means that these extra
933 * modules won't get loaded unless there was some path through the
934 * pattern that would have required them anyway, and so any false
935 * positives that occur by not ANDing them out when they could be
936 * aren't as severe as they would be if we treated this bit like all
937 * the others */
938 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
939 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 940 cl->flags &= and_with->flags;
3ad98780 941 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 942 }
653099ff
GS
943}
944
dd58aee1
KW
945/* 'OR' a given class with another one. Can create false positives. 'cl'
946 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
947 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 948STATIC void
3fffb88a 949S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 950{
7918f24d
NC
951 PERL_ARGS_ASSERT_CL_OR;
952
653099ff 953 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
954
955 /* Here, the or'd node is to be inverted. This means we take the
956 * complement of everything not in the bitmap, but currently we don't
957 * know what that is, so give up and match anything */
958 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 959 cl_anything(pRExC_state, cl);
c6b76537 960 }
653099ff
GS
961 /* We do not use
962 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
963 * <= (B1 | !B2) | (CL1 | !CL2)
964 * which is wasteful if CL2 is small, but we ignore CL2:
965 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
966 * XXXX Can we handle case-fold? Unclear:
967 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
968 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
969 */
c6b76537 970 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
971 && !(or_with->flags & ANYOF_LOC_FOLD)
972 && !(cl->flags & ANYOF_LOC_FOLD) ) {
653099ff
GS
973 int i;
974
975 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
976 cl->bitmap[i] |= ~or_with->bitmap[i];
977 } /* XXXX: logic is complicated otherwise */
978 else {
3fffb88a 979 cl_anything(pRExC_state, cl);
653099ff 980 }
c6b76537
KW
981
982 /* And, we can just take the union of the flags that aren't affected
983 * by the inversion */
984 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
985
986 /* For the remaining flags:
987 ANYOF_UNICODE_ALL and inverted means to not match anything above
988 255, which means that the union with cl should just be
989 what cl has in it, so can ignore this flag
990 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
991 is 127-255 to match them, but then invert that, so the
992 union with cl should just be what cl has in it, so can
993 ignore this flag
994 */
995 } else { /* 'or_with' is not inverted */
653099ff
GS
996 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
997 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
998 && (!(or_with->flags & ANYOF_LOC_FOLD)
999 || (cl->flags & ANYOF_LOC_FOLD)) ) {
653099ff
GS
1000 int i;
1001
1002 /* OR char bitmap and class bitmap separately */
1003 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1004 cl->bitmap[i] |= or_with->bitmap[i];
9cf78fa0 1005 ANYOF_CLASS_OR(or_with, cl);
653099ff
GS
1006 }
1007 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 1008 cl_anything(pRExC_state, cl);
653099ff 1009 }
9826f543 1010
c6b76537
KW
1011 if (ANYOF_NONBITMAP(or_with)) {
1012
1013 /* Use the added node's outside-the-bit-map match if there isn't a
1014 * conflict. If there is a conflict (both nodes match something
1015 * outside the bitmap, but what they match outside is not the same
1016 * pointer, and hence not easily compared until XXX we extend
1017 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
1018 * match everything outside the bitmap. If that stuff is all above
1019 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1020 if (! ANYOF_NONBITMAP(cl)) {
1021 ARG_SET(cl, ARG(or_with));
1022 }
1023 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1024
1025 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1026 cl_anything(pRExC_state, cl);
1027 }
1028 else {
1029 cl->flags |= ANYOF_UNICODE_ALL;
1030 }
c6b76537 1031 }
4c34a693 1032 }
0b9668ee
KW
1033
1034 /* Take the union */
1035 cl->flags |= or_with->flags;
1aa99e6b 1036 }
653099ff
GS
1037}
1038
a3621e74
YO
1039#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1040#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1041#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1042#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1043
3dab1dad
YO
1044
1045#ifdef DEBUGGING
07be1b83 1046/*
2b8b4781
NC
1047 dump_trie(trie,widecharmap,revcharmap)
1048 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1049 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1050
1051 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1052 The _interim_ variants are used for debugging the interim
1053 tables that are used to generate the final compressed
1054 representation which is what dump_trie expects.
1055
486ec47a 1056 Part of the reason for their existence is to provide a form
3dab1dad 1057 of documentation as to how the different representations function.
07be1b83
YO
1058
1059*/
3dab1dad
YO
1060
1061/*
3dab1dad
YO
1062 Dumps the final compressed table form of the trie to Perl_debug_log.
1063 Used for debugging make_trie().
1064*/
b9a59e08 1065
3dab1dad 1066STATIC void
2b8b4781
NC
1067S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1068 AV *revcharmap, U32 depth)
3dab1dad
YO
1069{
1070 U32 state;
ab3bbdeb 1071 SV *sv=sv_newmortal();
55eed653 1072 int colwidth= widecharmap ? 6 : 4;
2e64971a 1073 U16 word;
3dab1dad
YO
1074 GET_RE_DEBUG_FLAGS_DECL;
1075
7918f24d 1076 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1077
3dab1dad
YO
1078 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1079 (int)depth * 2 + 2,"",
1080 "Match","Base","Ofs" );
1081
1082 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1083 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1084 if ( tmp ) {
ab3bbdeb
YO
1085 PerlIO_printf( Perl_debug_log, "%*s",
1086 colwidth,
ddc5bc0f 1087 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1088 PL_colors[0], PL_colors[1],
1089 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1090 PERL_PV_ESCAPE_FIRSTCHAR
1091 )
1092 );
3dab1dad
YO
1093 }
1094 }
1095 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1096 (int)depth * 2 + 2,"");
1097
1098 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1099 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1100 PerlIO_printf( Perl_debug_log, "\n");
1101
1e2e3d02 1102 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1103 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1104
1105 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1106
1107 if ( trie->states[ state ].wordnum ) {
1108 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1109 } else {
1110 PerlIO_printf( Perl_debug_log, "%6s", "" );
1111 }
1112
1113 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1114
1115 if ( base ) {
1116 U32 ofs = 0;
1117
1118 while( ( base + ofs < trie->uniquecharcount ) ||
1119 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1120 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1121 ofs++;
1122
1123 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1124
1125 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1126 if ( ( base + ofs >= trie->uniquecharcount ) &&
1127 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1128 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1129 {
ab3bbdeb
YO
1130 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1131 colwidth,
3dab1dad
YO
1132 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1133 } else {
ab3bbdeb 1134 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1135 }
1136 }
1137
1138 PerlIO_printf( Perl_debug_log, "]");
1139
1140 }
1141 PerlIO_printf( Perl_debug_log, "\n" );
1142 }
2e64971a
DM
1143 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1144 for (word=1; word <= trie->wordcount; word++) {
1145 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1146 (int)word, (int)(trie->wordinfo[word].prev),
1147 (int)(trie->wordinfo[word].len));
1148 }
1149 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1150}
1151/*
3dab1dad
YO
1152 Dumps a fully constructed but uncompressed trie in list form.
1153 List tries normally only are used for construction when the number of
1154 possible chars (trie->uniquecharcount) is very high.
1155 Used for debugging make_trie().
1156*/
1157STATIC void
55eed653 1158S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1159 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1160 U32 depth)
3dab1dad
YO
1161{
1162 U32 state;
ab3bbdeb 1163 SV *sv=sv_newmortal();
55eed653 1164 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1165 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1166
1167 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1168
3dab1dad 1169 /* print out the table precompression. */
ab3bbdeb
YO
1170 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1171 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1172 "------:-----+-----------------\n" );
3dab1dad
YO
1173
1174 for( state=1 ; state < next_alloc ; state ++ ) {
1175 U16 charid;
1176
ab3bbdeb 1177 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1178 (int)depth * 2 + 2,"", (UV)state );
1179 if ( ! trie->states[ state ].wordnum ) {
1180 PerlIO_printf( Perl_debug_log, "%5s| ","");
1181 } else {
1182 PerlIO_printf( Perl_debug_log, "W%4x| ",
1183 trie->states[ state ].wordnum
1184 );
1185 }
1186 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1187 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1188 if ( tmp ) {
1189 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1190 colwidth,
ddc5bc0f 1191 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1192 PL_colors[0], PL_colors[1],
1193 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1194 PERL_PV_ESCAPE_FIRSTCHAR
1195 ) ,
1e2e3d02
YO
1196 TRIE_LIST_ITEM(state,charid).forid,
1197 (UV)TRIE_LIST_ITEM(state,charid).newstate
1198 );
1199 if (!(charid % 10))
664e119d
RGS
1200 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1201 (int)((depth * 2) + 14), "");
1e2e3d02 1202 }
ab3bbdeb
YO
1203 }
1204 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1205 }
1206}
1207
1208/*
3dab1dad
YO
1209 Dumps a fully constructed but uncompressed trie in table form.
1210 This is the normal DFA style state transition table, with a few
1211 twists to facilitate compression later.
1212 Used for debugging make_trie().
1213*/
1214STATIC void
55eed653 1215S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1216 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1217 U32 depth)
3dab1dad
YO
1218{
1219 U32 state;
1220 U16 charid;
ab3bbdeb 1221 SV *sv=sv_newmortal();
55eed653 1222 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1223 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1224
1225 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1226
1227 /*
1228 print out the table precompression so that we can do a visual check
1229 that they are identical.
1230 */
1231
1232 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1233
1234 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1235 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1236 if ( tmp ) {
ab3bbdeb
YO
1237 PerlIO_printf( Perl_debug_log, "%*s",
1238 colwidth,
ddc5bc0f 1239 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1240 PL_colors[0], PL_colors[1],
1241 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1242 PERL_PV_ESCAPE_FIRSTCHAR
1243 )
1244 );
3dab1dad
YO
1245 }
1246 }
1247
1248 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1249
1250 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1251 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1252 }
1253
1254 PerlIO_printf( Perl_debug_log, "\n" );
1255
1256 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1257
1258 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1259 (int)depth * 2 + 2,"",
1260 (UV)TRIE_NODENUM( state ) );
1261
1262 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1263 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1264 if (v)
1265 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1266 else
1267 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1268 }
1269 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1270 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1271 } else {
1272 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1273 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1274 }
1275 }
07be1b83 1276}
3dab1dad
YO
1277
1278#endif
1279
2e64971a 1280
786e8c11
YO
1281/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1282 startbranch: the first branch in the whole branch sequence
1283 first : start branch of sequence of branch-exact nodes.
1284 May be the same as startbranch
1285 last : Thing following the last branch.
1286 May be the same as tail.
1287 tail : item following the branch sequence
1288 count : words in the sequence
1289 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1290 depth : indent depth
3dab1dad 1291
786e8c11 1292Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1293
786e8c11
YO
1294A trie is an N'ary tree where the branches are determined by digital
1295decomposition of the key. IE, at the root node you look up the 1st character and
1296follow that branch repeat until you find the end of the branches. Nodes can be
1297marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1298
786e8c11 1299 /he|she|his|hers/
72f13be8 1300
786e8c11
YO
1301would convert into the following structure. Numbers represent states, letters
1302following numbers represent valid transitions on the letter from that state, if
1303the number is in square brackets it represents an accepting state, otherwise it
1304will be in parenthesis.
07be1b83 1305
786e8c11
YO
1306 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1307 | |
1308 | (2)
1309 | |
1310 (1) +-i->(6)-+-s->[7]
1311 |
1312 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1313
786e8c11
YO
1314 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1315
1316This shows that when matching against the string 'hers' we will begin at state 1
1317read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1318then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1319is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1320single traverse. We store a mapping from accepting to state to which word was
1321matched, and then when we have multiple possibilities we try to complete the
1322rest of the regex in the order in which they occured in the alternation.
1323
1324The only prior NFA like behaviour that would be changed by the TRIE support is
1325the silent ignoring of duplicate alternations which are of the form:
1326
1327 / (DUPE|DUPE) X? (?{ ... }) Y /x
1328
4b714af6 1329Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1330and without the optimisation. With the optimisations dupes will be silently
486ec47a 1331ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1332the following demonstrates:
1333
1334 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1335
1336which prints out 'word' three times, but
1337
1338 'words'=~/(word|word|word)(?{ print $1 })S/
1339
1340which doesnt print it out at all. This is due to other optimisations kicking in.
1341
1342Example of what happens on a structural level:
1343
486ec47a 1344The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1345
1346 1: CURLYM[1] {1,32767}(18)
1347 5: BRANCH(8)
1348 6: EXACT <ac>(16)
1349 8: BRANCH(11)
1350 9: EXACT <ad>(16)
1351 11: BRANCH(14)
1352 12: EXACT <ab>(16)
1353 16: SUCCEED(0)
1354 17: NOTHING(18)
1355 18: END(0)
1356
1357This would be optimizable with startbranch=5, first=5, last=16, tail=16
1358and should turn into:
1359
1360 1: CURLYM[1] {1,32767}(18)
1361 5: TRIE(16)
1362 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1363 <ac>
1364 <ad>
1365 <ab>
1366 16: SUCCEED(0)
1367 17: NOTHING(18)
1368 18: END(0)
1369
1370Cases where tail != last would be like /(?foo|bar)baz/:
1371
1372 1: BRANCH(4)
1373 2: EXACT <foo>(8)
1374 4: BRANCH(7)
1375 5: EXACT <bar>(8)
1376 7: TAIL(8)
1377 8: EXACT <baz>(10)
1378 10: END(0)
1379
1380which would be optimizable with startbranch=1, first=1, last=7, tail=8
1381and would end up looking like:
1382
1383 1: TRIE(8)
1384 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1385 <foo>
1386 <bar>
1387 7: TAIL(8)
1388 8: EXACT <baz>(10)
1389 10: END(0)
1390
1391 d = uvuni_to_utf8_flags(d, uv, 0);
1392
1393is the recommended Unicode-aware way of saying
1394
1395 *(d++) = uv;
1396*/
1397
fab2782b 1398#define TRIE_STORE_REVCHAR(val) \
786e8c11 1399 STMT_START { \
73031816 1400 if (UTF) { \
fab2782b 1401 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1402 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
fab2782b 1403 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
73031816
NC
1404 SvCUR_set(zlopp, kapow - flrbbbbb); \
1405 SvPOK_on(zlopp); \
1406 SvUTF8_on(zlopp); \
1407 av_push(revcharmap, zlopp); \
1408 } else { \
fab2782b 1409 char ooooff = (char)val; \
73031816
NC
1410 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1411 } \
1412 } STMT_END
786e8c11 1413
fab2782b
YO
1414#define TRIE_READ_CHAR STMT_START { \
1415 wordlen++; \
1416 if ( UTF ) { \
1417 /* if it is UTF then it is either already folded, or does not need folding */ \
1418 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1419 } \
1420 else if (folder == PL_fold_latin1) { \
1421 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1422 if ( foldlen > 0 ) { \
1423 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1424 foldlen -= len; \
1425 scan += len; \
1426 len = 0; \
1427 } else { \
1428 len = 1; \
1429 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1430 skiplen = UNISKIP(uvc); \
1431 foldlen -= skiplen; \
1432 scan = foldbuf + skiplen; \
1433 } \
1434 } else { \
1435 /* raw data, will be folded later if needed */ \
1436 uvc = (U32)*uc; \
1437 len = 1; \
1438 } \
786e8c11
YO
1439} STMT_END
1440
1441
1442
1443#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1444 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1445 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1446 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1447 } \
1448 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1449 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1450 TRIE_LIST_CUR( state )++; \
1451} STMT_END
07be1b83 1452
786e8c11
YO
1453#define TRIE_LIST_NEW(state) STMT_START { \
1454 Newxz( trie->states[ state ].trans.list, \
1455 4, reg_trie_trans_le ); \
1456 TRIE_LIST_CUR( state ) = 1; \
1457 TRIE_LIST_LEN( state ) = 4; \
1458} STMT_END
07be1b83 1459
786e8c11
YO
1460#define TRIE_HANDLE_WORD(state) STMT_START { \
1461 U16 dupe= trie->states[ state ].wordnum; \
1462 regnode * const noper_next = regnext( noper ); \
1463 \
786e8c11
YO
1464 DEBUG_r({ \
1465 /* store the word for dumping */ \
1466 SV* tmp; \
1467 if (OP(noper) != NOTHING) \
740cce10 1468 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1469 else \
740cce10 1470 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1471 av_push( trie_words, tmp ); \
786e8c11
YO
1472 }); \
1473 \
1474 curword++; \
2e64971a
DM
1475 trie->wordinfo[curword].prev = 0; \
1476 trie->wordinfo[curword].len = wordlen; \
1477 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1478 \
1479 if ( noper_next < tail ) { \
1480 if (!trie->jump) \
c944940b 1481 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1482 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1483 if (!jumper) \
1484 jumper = noper_next; \
1485 if (!nextbranch) \
1486 nextbranch= regnext(cur); \
1487 } \
1488 \
1489 if ( dupe ) { \
2e64971a
DM
1490 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1491 /* chain, so that when the bits of chain are later */\
1492 /* linked together, the dups appear in the chain */\
1493 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1494 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1495 } else { \
1496 /* we haven't inserted this word yet. */ \
1497 trie->states[ state ].wordnum = curword; \
1498 } \
1499} STMT_END
07be1b83 1500
3dab1dad 1501
786e8c11
YO
1502#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1503 ( ( base + charid >= ucharcount \
1504 && base + charid < ubound \
1505 && state == trie->trans[ base - ucharcount + charid ].check \
1506 && trie->trans[ base - ucharcount + charid ].next ) \
1507 ? trie->trans[ base - ucharcount + charid ].next \
1508 : ( state==1 ? special : 0 ) \
1509 )
3dab1dad 1510
786e8c11
YO
1511#define MADE_TRIE 1
1512#define MADE_JUMP_TRIE 2
1513#define MADE_EXACT_TRIE 4
3dab1dad 1514
a3621e74 1515STATIC I32
786e8c11 1516S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1517{
27da23d5 1518 dVAR;
a3621e74
YO
1519 /* first pass, loop through and scan words */
1520 reg_trie_data *trie;
55eed653 1521 HV *widecharmap = NULL;
2b8b4781 1522 AV *revcharmap = newAV();
a3621e74 1523 regnode *cur;
9f7f3913 1524 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1525 STRLEN len = 0;
1526 UV uvc = 0;
1527 U16 curword = 0;
1528 U32 next_alloc = 0;
786e8c11
YO
1529 regnode *jumper = NULL;
1530 regnode *nextbranch = NULL;
7f69552c 1531 regnode *convert = NULL;
2e64971a 1532 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1533 /* we just use folder as a flag in utf8 */
1e696034 1534 const U8 * folder = NULL;
a3621e74 1535
2b8b4781
NC
1536#ifdef DEBUGGING
1537 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1538 AV *trie_words = NULL;
1539 /* along with revcharmap, this only used during construction but both are
1540 * useful during debugging so we store them in the struct when debugging.
8e11feef 1541 */
2b8b4781
NC
1542#else
1543 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1544 STRLEN trie_charcount=0;
3dab1dad 1545#endif
2b8b4781 1546 SV *re_trie_maxbuff;
a3621e74 1547 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1548
1549 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1550#ifndef DEBUGGING
1551 PERL_UNUSED_ARG(depth);
1552#endif
a3621e74 1553
1e696034 1554 switch (flags) {
79a81a6e 1555 case EXACT: break;
2f7f8cb1 1556 case EXACTFA:
fab2782b
YO
1557 case EXACTFU_SS:
1558 case EXACTFU_TRICKYFOLD:
1e696034
KW
1559 case EXACTFU: folder = PL_fold_latin1; break;
1560 case EXACTF: folder = PL_fold; break;
1561 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1562 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1563 }
1564
c944940b 1565 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1566 trie->refcount = 1;
3dab1dad 1567 trie->startstate = 1;
786e8c11 1568 trie->wordcount = word_count;
f8fc2ecf 1569 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1570 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1571 if (flags == EXACT)
c944940b 1572 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1573 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1574 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1575
a3621e74 1576 DEBUG_r({
2b8b4781 1577 trie_words = newAV();
a3621e74 1578 });
a3621e74 1579
0111c4fd 1580 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1581 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1582 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1583 }
df826430 1584 DEBUG_TRIE_COMPILE_r({
3dab1dad 1585 PerlIO_printf( Perl_debug_log,
786e8c11 1586 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1587 (int)depth * 2 + 2, "",
1588 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1589 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1590 (int)depth);
3dab1dad 1591 });
7f69552c
YO
1592
1593 /* Find the node we are going to overwrite */
1594 if ( first == startbranch && OP( last ) != BRANCH ) {
1595 /* whole branch chain */
1596 convert = first;
1597 } else {
1598 /* branch sub-chain */
1599 convert = NEXTOPER( first );
1600 }
1601
a3621e74
YO
1602 /* -- First loop and Setup --
1603
1604 We first traverse the branches and scan each word to determine if it
1605 contains widechars, and how many unique chars there are, this is
1606 important as we have to build a table with at least as many columns as we
1607 have unique chars.
1608
1609 We use an array of integers to represent the character codes 0..255
38a44b82 1610 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1611 native representation of the character value as the key and IV's for the
1612 coded index.
1613
1614 *TODO* If we keep track of how many times each character is used we can
1615 remap the columns so that the table compression later on is more
3b753521 1616 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1617 middle and the least common are on the outside. IMO this would be better
1618 than a most to least common mapping as theres a decent chance the most
1619 common letter will share a node with the least common, meaning the node
486ec47a 1620 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1621 case is when we have the least common nodes twice.
1622
1623 */
1624
a3621e74 1625 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1626 regnode *noper = NEXTOPER( cur );
e1ec3a88 1627 const U8 *uc = (U8*)STRING( noper );
df826430 1628 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1629 STRLEN foldlen = 0;
1630 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1631 STRLEN skiplen = 0;
2af232bd 1632 const U8 *scan = (U8*)NULL;
07be1b83 1633 U32 wordlen = 0; /* required init */
02daf0ab
YO
1634 STRLEN chars = 0;
1635 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1636
3dab1dad 1637 if (OP(noper) == NOTHING) {
df826430
YO
1638 regnode *noper_next= regnext(noper);
1639 if (noper_next != tail && OP(noper_next) == flags) {
1640 noper = noper_next;
1641 uc= (U8*)STRING(noper);
1642 e= uc + STR_LEN(noper);
1643 trie->minlen= STR_LEN(noper);
1644 } else {
1645 trie->minlen= 0;
1646 continue;
1647 }
3dab1dad 1648 }
df826430 1649
fab2782b 1650 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1651 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1652 regardless of encoding */
fab2782b
YO
1653 if (OP( noper ) == EXACTFU_SS) {
1654 /* false positives are ok, so just set this */
1655 TRIE_BITMAP_SET(trie,0xDF);
1656 }
1657 }
a3621e74 1658 for ( ; uc < e ; uc += len ) {
3dab1dad 1659 TRIE_CHARCOUNT(trie)++;
a3621e74 1660 TRIE_READ_CHAR;
3dab1dad 1661 chars++;
a3621e74 1662 if ( uvc < 256 ) {
fab2782b
YO
1663 if ( folder ) {
1664 U8 folded= folder[ (U8) uvc ];
1665 if ( !trie->charmap[ folded ] ) {
1666 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1667 TRIE_STORE_REVCHAR( folded );
1668 }
1669 }
a3621e74
YO
1670 if ( !trie->charmap[ uvc ] ) {
1671 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1672 TRIE_STORE_REVCHAR( uvc );
a3621e74 1673 }
02daf0ab 1674 if ( set_bit ) {
62012aee
KW
1675 /* store the codepoint in the bitmap, and its folded
1676 * equivalent. */
fab2782b 1677 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1678
1679 /* store the folded codepoint */
fab2782b 1680 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1681
1682 if ( !UTF ) {
1683 /* store first byte of utf8 representation of
acdf4139
KW
1684 variant codepoints */
1685 if (! UNI_IS_INVARIANT(uvc)) {
1686 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1687 }
1688 }
02daf0ab
YO
1689 set_bit = 0; /* We've done our bit :-) */
1690 }
a3621e74
YO
1691 } else {
1692 SV** svpp;
55eed653
NC
1693 if ( !widecharmap )
1694 widecharmap = newHV();
a3621e74 1695
55eed653 1696 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1697
1698 if ( !svpp )
e4584336 1699 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1700
1701 if ( !SvTRUE( *svpp ) ) {
1702 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1703 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1704 }
1705 }
1706 }
3dab1dad 1707 if( cur == first ) {
fab2782b
YO
1708 trie->minlen = chars;
1709 trie->maxlen = chars;
3dab1dad 1710 } else if (chars < trie->minlen) {
fab2782b 1711 trie->minlen = chars;
3dab1dad 1712 } else if (chars > trie->maxlen) {
fab2782b
YO
1713 trie->maxlen = chars;
1714 }
1715 if (OP( noper ) == EXACTFU_SS) {
1716 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1717 if (trie->minlen > 1)
1718 trie->minlen= 1;
1719 }
1720 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1721 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1722 * - We assume that any such sequence might match a 2 byte string */
1723 if (trie->minlen > 2 )
1724 trie->minlen= 2;
3dab1dad
YO
1725 }
1726
a3621e74
YO
1727 } /* end first pass */
1728 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1729 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1730 (int)depth * 2 + 2,"",
55eed653 1731 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1732 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1733 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1734 );
a3621e74
YO
1735
1736 /*
1737 We now know what we are dealing with in terms of unique chars and
1738 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1739 representation using a flat table will take. If it's over a reasonable
1740 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1741 conservative but potentially much slower representation using an array
1742 of lists.
1743
1744 At the end we convert both representations into the same compressed
1745 form that will be used in regexec.c for matching with. The latter
1746 is a form that cannot be used to construct with but has memory
1747 properties similar to the list form and access properties similar
1748 to the table form making it both suitable for fast searches and
1749 small enough that its feasable to store for the duration of a program.
1750
1751 See the comment in the code where the compressed table is produced
1752 inplace from the flat tabe representation for an explanation of how
1753 the compression works.
1754
1755 */
1756
1757
2e64971a
DM
1758 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1759 prev_states[1] = 0;
1760
3dab1dad 1761 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1762 /*
1763 Second Pass -- Array Of Lists Representation
1764
1765 Each state will be represented by a list of charid:state records
1766 (reg_trie_trans_le) the first such element holds the CUR and LEN
1767 points of the allocated array. (See defines above).
1768
1769 We build the initial structure using the lists, and then convert
1770 it into the compressed table form which allows faster lookups
1771 (but cant be modified once converted).
a3621e74
YO
1772 */
1773
a3621e74
YO
1774 STRLEN transcount = 1;
1775
1e2e3d02
YO
1776 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1777 "%*sCompiling trie using list compiler\n",
1778 (int)depth * 2 + 2, ""));
686b73d4 1779
c944940b
JH
1780 trie->states = (reg_trie_state *)
1781 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1782 sizeof(reg_trie_state) );
a3621e74
YO
1783 TRIE_LIST_NEW(1);
1784 next_alloc = 2;
1785
1786 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1787
df826430 1788 regnode *noper = NEXTOPER( cur );
c445ea15 1789 U8 *uc = (U8*)STRING( noper );
df826430 1790 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1791 U32 state = 1; /* required init */
1792 U16 charid = 0; /* sanity init */
1793 U8 *scan = (U8*)NULL; /* sanity init */
1794 STRLEN foldlen = 0; /* required init */
07be1b83 1795 U32 wordlen = 0; /* required init */
c445ea15 1796 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1797 STRLEN skiplen = 0;
c445ea15 1798
df826430
YO
1799 if (OP(noper) == NOTHING) {
1800 regnode *noper_next= regnext(noper);
1801 if (noper_next != tail && OP(noper_next) == flags) {
1802 noper = noper_next;
1803 uc= (U8*)STRING(noper);
1804 e= uc + STR_LEN(noper);
1805 }
1806 }
1807
3dab1dad 1808 if (OP(noper) != NOTHING) {
786e8c11 1809 for ( ; uc < e ; uc += len ) {
c445ea15 1810
786e8c11 1811 TRIE_READ_CHAR;
c445ea15 1812
786e8c11
YO
1813 if ( uvc < 256 ) {
1814 charid = trie->charmap[ uvc ];
c445ea15 1815 } else {
55eed653 1816 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1817 if ( !svpp ) {
1818 charid = 0;
1819 } else {
1820 charid=(U16)SvIV( *svpp );
1821 }
c445ea15 1822 }
786e8c11
YO
1823 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1824 if ( charid ) {
a3621e74 1825
786e8c11
YO
1826 U16 check;
1827 U32 newstate = 0;
a3621e74 1828
786e8c11
YO
1829 charid--;
1830 if ( !trie->states[ state ].trans.list ) {
1831 TRIE_LIST_NEW( state );
c445ea15 1832 }
786e8c11
YO
1833 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1834 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1835 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1836 break;
1837 }
1838 }
1839 if ( ! newstate ) {
1840 newstate = next_alloc++;
2e64971a 1841 prev_states[newstate] = state;
786e8c11
YO
1842 TRIE_LIST_PUSH( state, charid, newstate );
1843 transcount++;
1844 }
1845 state = newstate;
1846 } else {
1847 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1848 }
a28509cc 1849 }
c445ea15 1850 }
3dab1dad 1851 TRIE_HANDLE_WORD(state);
a3621e74
YO
1852
1853 } /* end second pass */
1854
1e2e3d02
YO
1855 /* next alloc is the NEXT state to be allocated */
1856 trie->statecount = next_alloc;
c944940b
JH
1857 trie->states = (reg_trie_state *)
1858 PerlMemShared_realloc( trie->states,
1859 next_alloc
1860 * sizeof(reg_trie_state) );
a3621e74 1861
3dab1dad 1862 /* and now dump it out before we compress it */
2b8b4781
NC
1863 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1864 revcharmap, next_alloc,
1865 depth+1)
1e2e3d02 1866 );
a3621e74 1867
c944940b
JH
1868 trie->trans = (reg_trie_trans *)
1869 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1870 {
1871 U32 state;
a3621e74
YO
1872 U32 tp = 0;
1873 U32 zp = 0;
1874
1875
1876 for( state=1 ; state < next_alloc ; state ++ ) {
1877 U32 base=0;
1878
1879 /*
1880 DEBUG_TRIE_COMPILE_MORE_r(
1881 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1882 );
1883 */
1884
1885 if (trie->states[state].trans.list) {
1886 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1887 U16 maxid=minid;
a28509cc 1888 U16 idx;
a3621e74
YO
1889
1890 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1891 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1892 if ( forid < minid ) {
1893 minid=forid;
1894 } else if ( forid > maxid ) {
1895 maxid=forid;
1896 }
a3621e74
YO
1897 }
1898 if ( transcount < tp + maxid - minid + 1) {
1899 transcount *= 2;
c944940b
JH
1900 trie->trans = (reg_trie_trans *)
1901 PerlMemShared_realloc( trie->trans,
446bd890
NC
1902 transcount
1903 * sizeof(reg_trie_trans) );
a3621e74
YO
1904 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1905 }
1906 base = trie->uniquecharcount + tp - minid;
1907 if ( maxid == minid ) {
1908 U32 set = 0;
1909 for ( ; zp < tp ; zp++ ) {
1910 if ( ! trie->trans[ zp ].next ) {
1911 base = trie->uniquecharcount + zp - minid;
1912 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1913 trie->trans[ zp ].check = state;
1914 set = 1;
1915 break;
1916 }
1917 }
1918 if ( !set ) {
1919 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1920 trie->trans[ tp ].check = state;
1921 tp++;
1922 zp = tp;
1923 }
1924 } else {
1925 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1926 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1927 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1928 trie->trans[ tid ].check = state;
1929 }
1930 tp += ( maxid - minid + 1 );
1931 }
1932 Safefree(trie->states[ state ].trans.list);
1933 }
1934 /*
1935 DEBUG_TRIE_COMPILE_MORE_r(
1936 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1937 );
1938 */
1939 trie->states[ state ].trans.base=base;
1940 }
cc601c31 1941 trie->lasttrans = tp + 1;
a3621e74
YO
1942 }
1943 } else {
1944 /*
1945 Second Pass -- Flat Table Representation.
1946
1947 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1948 We know that we will need Charcount+1 trans at most to store the data
1949 (one row per char at worst case) So we preallocate both structures
1950 assuming worst case.
1951
1952 We then construct the trie using only the .next slots of the entry
1953 structs.
1954
3b753521 1955 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1956 make compression both faster and easier by keeping track of how many non
1957 zero fields are in the node.
1958
1959 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1960 transition.
1961
1962 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1963 number representing the first entry of the node, and state as a
1964 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1965 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1966 are 2 entrys per node. eg:
1967
1968 A B A B
1969 1. 2 4 1. 3 7
1970 2. 0 3 3. 0 5
1971 3. 0 0 5. 0 0
1972 4. 0 0 7. 0 0
1973
1974 The table is internally in the right hand, idx form. However as we also
1975 have to deal with the states array which is indexed by nodenum we have to
1976 use TRIE_NODENUM() to convert.
1977
1978 */
1e2e3d02
YO
1979 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1980 "%*sCompiling trie using table compiler\n",
1981 (int)depth * 2 + 2, ""));
3dab1dad 1982
c944940b
JH
1983 trie->trans = (reg_trie_trans *)
1984 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1985 * trie->uniquecharcount + 1,
1986 sizeof(reg_trie_trans) );
1987 trie->states = (reg_trie_state *)
1988 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1989 sizeof(reg_trie_state) );
a3621e74
YO
1990 next_alloc = trie->uniquecharcount + 1;
1991
3dab1dad 1992
a3621e74
YO
1993 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1994
df826430 1995 regnode *noper = NEXTOPER( cur );
a28509cc 1996 const U8 *uc = (U8*)STRING( noper );
df826430 1997 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1998
1999 U32 state = 1; /* required init */
2000
2001 U16 charid = 0; /* sanity init */
2002 U32 accept_state = 0; /* sanity init */
2003 U8 *scan = (U8*)NULL; /* sanity init */
2004
2005 STRLEN foldlen = 0; /* required init */
07be1b83 2006 U32 wordlen = 0; /* required init */
fab2782b 2007 STRLEN skiplen = 0;
a3621e74
YO
2008 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2009
df826430
YO
2010 if (OP(noper) == NOTHING) {
2011 regnode *noper_next= regnext(noper);
2012 if (noper_next != tail && OP(noper_next) == flags) {
2013 noper = noper_next;
2014 uc= (U8*)STRING(noper);
2015 e= uc + STR_LEN(noper);
2016 }
2017 }
fab2782b 2018
3dab1dad 2019 if ( OP(noper) != NOTHING ) {
786e8c11 2020 for ( ; uc < e ; uc += len ) {
a3621e74 2021
786e8c11 2022 TRIE_READ_CHAR;
a3621e74 2023
786e8c11
YO
2024 if ( uvc < 256 ) {
2025 charid = trie->charmap[ uvc ];
2026 } else {
55eed653 2027 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2028 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2029 }
786e8c11
YO
2030 if ( charid ) {
2031 charid--;
2032 if ( !trie->trans[ state + charid ].next ) {
2033 trie->trans[ state + charid ].next = next_alloc;
2034 trie->trans[ state ].check++;
2e64971a
DM
2035 prev_states[TRIE_NODENUM(next_alloc)]
2036 = TRIE_NODENUM(state);
786e8c11
YO
2037 next_alloc += trie->uniquecharcount;
2038 }
2039 state = trie->trans[ state + charid ].next;
2040 } else {
2041 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2042 }
2043 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2044 }
a3621e74 2045 }
3dab1dad
YO
2046 accept_state = TRIE_NODENUM( state );
2047 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2048
2049 } /* end second pass */
2050
3dab1dad 2051 /* and now dump it out before we compress it */
2b8b4781
NC
2052 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2053 revcharmap,
2054 next_alloc, depth+1));
a3621e74 2055
a3621e74
YO
2056 {
2057 /*
2058 * Inplace compress the table.*
2059
2060 For sparse data sets the table constructed by the trie algorithm will
2061 be mostly 0/FAIL transitions or to put it another way mostly empty.
2062 (Note that leaf nodes will not contain any transitions.)
2063
2064 This algorithm compresses the tables by eliminating most such
2065 transitions, at the cost of a modest bit of extra work during lookup:
2066
2067 - Each states[] entry contains a .base field which indicates the
2068 index in the state[] array wheres its transition data is stored.
2069
3b753521 2070 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2071
2072 - If .base is nonzero then charid is added to it to find an entry in
2073 the trans array.
2074
2075 -If trans[states[state].base+charid].check!=state then the
2076 transition is taken to be a 0/Fail transition. Thus if there are fail
2077 transitions at the front of the node then the .base offset will point
2078 somewhere inside the previous nodes data (or maybe even into a node
2079 even earlier), but the .check field determines if the transition is
2080 valid.
2081
786e8c11 2082 XXX - wrong maybe?
a3621e74 2083 The following process inplace converts the table to the compressed
3b753521 2084 table: We first do not compress the root node 1,and mark all its
a3621e74 2085 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2086 allows us to do a DFA construction from the compressed table later,
2087 and ensures that any .base pointers we calculate later are greater
2088 than 0.
a3621e74
YO
2089
2090 - We set 'pos' to indicate the first entry of the second node.
2091
2092 - We then iterate over the columns of the node, finding the first and
2093 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2094 and set the .check pointers accordingly, and advance pos
2095 appropriately and repreat for the next node. Note that when we copy
2096 the next pointers we have to convert them from the original
2097 NODEIDX form to NODENUM form as the former is not valid post
2098 compression.
2099
2100 - If a node has no transitions used we mark its base as 0 and do not
2101 advance the pos pointer.
2102
2103 - If a node only has one transition we use a second pointer into the
2104 structure to fill in allocated fail transitions from other states.
2105 This pointer is independent of the main pointer and scans forward
2106 looking for null transitions that are allocated to a state. When it
2107 finds one it writes the single transition into the "hole". If the
786e8c11 2108 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2109
2110 - Once compressed we can Renew/realloc the structures to release the
2111 excess space.
2112
2113 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2114 specifically Fig 3.47 and the associated pseudocode.
2115
2116 demq
2117 */
a3b680e6 2118 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2119 U32 state, charid;
a3621e74 2120 U32 pos = 0, zp=0;
1e2e3d02 2121 trie->statecount = laststate;
a3621e74
YO
2122
2123 for ( state = 1 ; state < laststate ; state++ ) {
2124 U8 flag = 0;
a28509cc
AL
2125 const U32 stateidx = TRIE_NODEIDX( state );
2126 const U32 o_used = trie->trans[ stateidx ].check;
2127 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2128 trie->trans[ stateidx ].check = 0;
2129
2130 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2131 if ( flag || trie->trans[ stateidx + charid ].next ) {
2132 if ( trie->trans[ stateidx + charid ].next ) {
2133 if (o_used == 1) {
2134 for ( ; zp < pos ; zp++ ) {
2135 if ( ! trie->trans[ zp ].next ) {
2136 break;
2137 }
2138 }
2139 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2140 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2141 trie->trans[ zp ].check = state;
2142 if ( ++zp > pos ) pos = zp;
2143 break;
2144 }
2145 used--;
2146 }
2147 if ( !flag ) {
2148 flag = 1;
2149 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2150 }
2151 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2152 trie->trans[ pos ].check = state;
2153 pos++;
2154 }
2155 }
2156 }
cc601c31 2157 trie->lasttrans = pos + 1;
c944940b
JH
2158 trie->states = (reg_trie_state *)
2159 PerlMemShared_realloc( trie->states, laststate
2160 * sizeof(reg_trie_state) );
a3621e74 2161 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2162 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2163 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2164 (int)depth * 2 + 2,"",
2165 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2166 (IV)next_alloc,
2167 (IV)pos,
a3621e74
YO
2168 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2169 );
2170
2171 } /* end table compress */
2172 }
1e2e3d02
YO
2173 DEBUG_TRIE_COMPILE_MORE_r(
2174 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2175 (int)depth * 2 + 2, "",
2176 (UV)trie->statecount,
2177 (UV)trie->lasttrans)
2178 );
cc601c31 2179 /* resize the trans array to remove unused space */
c944940b
JH
2180 trie->trans = (reg_trie_trans *)
2181 PerlMemShared_realloc( trie->trans, trie->lasttrans
2182 * sizeof(reg_trie_trans) );
a3621e74 2183
3b753521 2184 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2185 U8 nodetype =(U8)(flags & 0xFF);
2186 char *str=NULL;
786e8c11 2187
07be1b83 2188#ifdef DEBUGGING
e62cc96a 2189 regnode *optimize = NULL;
7122b237
YO
2190#ifdef RE_TRACK_PATTERN_OFFSETS
2191
b57a0404
JH
2192 U32 mjd_offset = 0;
2193 U32 mjd_nodelen = 0;
7122b237
YO
2194#endif /* RE_TRACK_PATTERN_OFFSETS */
2195#endif /* DEBUGGING */
a3621e74 2196 /*
3dab1dad
YO
2197 This means we convert either the first branch or the first Exact,
2198 depending on whether the thing following (in 'last') is a branch
2199 or not and whther first is the startbranch (ie is it a sub part of
2200 the alternation or is it the whole thing.)
3b753521 2201 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2202 the whole branch sequence, including the first.
a3621e74 2203 */
3dab1dad 2204 /* Find the node we are going to overwrite */
7f69552c 2205 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2206 /* branch sub-chain */
3dab1dad 2207 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2208#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2209 DEBUG_r({
2210 mjd_offset= Node_Offset((convert));
2211 mjd_nodelen= Node_Length((convert));
2212 });
7122b237 2213#endif
7f69552c 2214 /* whole branch chain */
7122b237
YO
2215 }
2216#ifdef RE_TRACK_PATTERN_OFFSETS
2217 else {
7f69552c
YO
2218 DEBUG_r({
2219 const regnode *nop = NEXTOPER( convert );
2220 mjd_offset= Node_Offset((nop));
2221 mjd_nodelen= Node_Length((nop));
2222 });
07be1b83
YO
2223 }
2224 DEBUG_OPTIMISE_r(
2225 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2226 (int)depth * 2 + 2, "",
786e8c11 2227 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2228 );
7122b237 2229#endif
3dab1dad
YO
2230 /* But first we check to see if there is a common prefix we can
2231 split out as an EXACT and put in front of the TRIE node. */
2232 trie->startstate= 1;
55eed653 2233 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2234 U32 state;
1e2e3d02 2235 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2236 U32 ofs = 0;
8e11feef
RGS
2237 I32 idx = -1;
2238 U32 count = 0;
2239 const U32 base = trie->states[ state ].trans.base;
a3621e74 2240
3dab1dad 2241 if ( trie->states[state].wordnum )
8e11feef 2242 count = 1;
a3621e74 2243
8e11feef 2244 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2245 if ( ( base + ofs >= trie->uniquecharcount ) &&
2246 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2247 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2248 {
3dab1dad 2249 if ( ++count > 1 ) {
2b8b4781 2250 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2251 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2252 if ( state == 1 ) break;
3dab1dad
YO
2253 if ( count == 2 ) {
2254 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2255 DEBUG_OPTIMISE_r(
8e11feef
RGS
2256 PerlIO_printf(Perl_debug_log,
2257 "%*sNew Start State=%"UVuf" Class: [",
2258 (int)depth * 2 + 2, "",
786e8c11 2259 (UV)state));
be8e71aa 2260 if (idx >= 0) {
2b8b4781 2261 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2262 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2263
3dab1dad 2264 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2265 if ( folder )
2266 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2267 DEBUG_OPTIMISE_r(
f1f66076 2268 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2269 );
8e11feef
RGS
2270 }
2271 }
2272 TRIE_BITMAP_SET(trie,*ch);
2273 if ( folder )
2274 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2275 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2276 }
2277 idx = ofs;
2278 }
3dab1dad
YO
2279 }
2280 if ( count == 1 ) {
2b8b4781 2281 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2282 STRLEN len;
2283 char *ch = SvPV( *tmp, len );
de734bd5
A
2284 DEBUG_OPTIMISE_r({
2285 SV *sv=sv_newmortal();
8e11feef
RGS
2286 PerlIO_printf( Perl_debug_log,
2287 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2288 (int)depth * 2 + 2, "",
de734bd5
A
2289 (UV)state, (UV)idx,
2290 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2291 PL_colors[0], PL_colors[1],
2292 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2293 PERL_PV_ESCAPE_FIRSTCHAR
2294 )
2295 );
2296 });
3dab1dad
YO
2297 if ( state==1 ) {
2298 OP( convert ) = nodetype;
2299 str=STRING(convert);
2300 STR_LEN(convert)=0;
2301 }
c490c714
YO
2302 STR_LEN(convert) += len;
2303 while (len--)
de734bd5 2304 *str++ = *ch++;
8e11feef 2305 } else {
f9049ba1 2306#ifdef DEBUGGING
8e11feef
RGS
2307 if (state>1)
2308 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2309#endif
8e11feef
RGS
2310 break;
2311 }
2312 }
2e64971a 2313 trie->prefixlen = (state-1);
3dab1dad 2314 if (str) {
8e11feef 2315 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2316 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2317 trie->startstate = state;
07be1b83
YO
2318 trie->minlen -= (state - 1);
2319 trie->maxlen -= (state - 1);
33809eae
JH
2320#ifdef DEBUGGING
2321 /* At least the UNICOS C compiler choked on this
2322 * being argument to DEBUG_r(), so let's just have
2323 * it right here. */
2324 if (
2325#ifdef PERL_EXT_RE_BUILD
2326 1
2327#else
2328 DEBUG_r_TEST
2329#endif
2330 ) {
2331 regnode *fix = convert;
2332 U32 word = trie->wordcount;
2333 mjd_nodelen++;
2334 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2335 while( ++fix < n ) {
2336 Set_Node_Offset_Length(fix, 0, 0);
2337 }
2338 while (word--) {
2339 SV ** const tmp = av_fetch( trie_words, word, 0 );
2340 if (tmp) {
2341 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2342 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2343 else
2344 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2345 }
2346 }
2347 }
2348#endif
8e11feef
RGS
2349 if (trie->maxlen) {
2350 convert = n;
2351 } else {
3dab1dad 2352 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2353 DEBUG_r(optimize= n);
3dab1dad
YO
2354 }
2355 }
2356 }
a5ca303d
YO
2357 if (!jumper)
2358 jumper = last;
3dab1dad 2359 if ( trie->maxlen ) {
8e11feef
RGS
2360 NEXT_OFF( convert ) = (U16)(tail - convert);
2361 ARG_SET( convert, data_slot );
786e8c11
YO
2362 /* Store the offset to the first unabsorbed branch in
2363 jump[0], which is otherwise unused by the jump logic.
2364 We use this when dumping a trie and during optimisation. */
2365 if (trie->jump)
7f69552c 2366 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2367
6c48061a
YO
2368 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2369 * and there is a bitmap
2370 * and the first "jump target" node we found leaves enough room
2371 * then convert the TRIE node into a TRIEC node, with the bitmap
2372 * embedded inline in the opcode - this is hypothetically faster.
2373 */
2374 if ( !trie->states[trie->startstate].wordnum
2375 && trie->bitmap
2376 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2377 {
2378 OP( convert ) = TRIEC;
2379 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2380 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2381 trie->bitmap= NULL;
2382 } else
2383 OP( convert ) = TRIE;
a3621e74 2384
3dab1dad
YO
2385 /* store the type in the flags */
2386 convert->flags = nodetype;
a5ca303d
YO
2387 DEBUG_r({
2388 optimize = convert
2389 + NODE_STEP_REGNODE
2390 + regarglen[ OP( convert ) ];
2391 });
2392 /* XXX We really should free up the resource in trie now,
2393 as we won't use them - (which resources?) dmq */
3dab1dad 2394 }
a3621e74 2395 /* needed for dumping*/
e62cc96a 2396 DEBUG_r(if (optimize) {
07be1b83 2397 regnode *opt = convert;
bcdf7404 2398
e62cc96a 2399 while ( ++opt < optimize) {
07be1b83
YO
2400 Set_Node_Offset_Length(opt,0,0);
2401 }
786e8c11
YO
2402 /*
2403 Try to clean up some of the debris left after the
2404 optimisation.
a3621e74 2405 */
786e8c11 2406 while( optimize < jumper ) {
07be1b83 2407 mjd_nodelen += Node_Length((optimize));
a3621e74 2408 OP( optimize ) = OPTIMIZED;
07be1b83 2409 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2410 optimize++;
2411 }
07be1b83 2412 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2413 });
2414 } /* end node insert */
2e64971a
DM
2415
2416 /* Finish populating the prev field of the wordinfo array. Walk back
2417 * from each accept state until we find another accept state, and if
2418 * so, point the first word's .prev field at the second word. If the
2419 * second already has a .prev field set, stop now. This will be the
2420 * case either if we've already processed that word's accept state,
3b753521
FN
2421 * or that state had multiple words, and the overspill words were
2422 * already linked up earlier.
2e64971a
DM
2423 */
2424 {
2425 U16 word;
2426 U32 state;
2427 U16 prev;
2428
2429 for (word=1; word <= trie->wordcount; word++) {
2430 prev = 0;
2431 if (trie->wordinfo[word].prev)
2432 continue;
2433 state = trie->wordinfo[word].accept;
2434 while (state) {
2435 state = prev_states[state];
2436 if (!state)
2437 break;
2438 prev = trie->states[state].wordnum;
2439 if (prev)
2440 break;
2441 }
2442 trie->wordinfo[word].prev = prev;
2443 }
2444 Safefree(prev_states);
2445 }
2446
2447
2448 /* and now dump out the compressed format */
2449 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2450
55eed653 2451 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2452#ifdef DEBUGGING
2453 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2454 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2455#else
03e70be4 2456 SvREFCNT_dec_NN(revcharmap);
07be1b83 2457#endif
786e8c11
YO
2458 return trie->jump
2459 ? MADE_JUMP_TRIE
2460 : trie->startstate>1
2461 ? MADE_EXACT_TRIE
2462 : MADE_TRIE;
2463}
2464
2465STATIC void
2466S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2467{
3b753521 2468/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2469
2470 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2471 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2472 ISBN 0-201-10088-6
2473
2474 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2475 suffix of the current state's 'word' that is also a proper prefix of another word in our
2476 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2477 the DFA not to have to restart after its tried and failed a word at a given point, it
2478 simply continues as though it had been matching the other word in the first place.
2479 Consider
2480 'abcdgu'=~/abcdefg|cdgu/
2481 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2482 fail, which would bring us to the state representing 'd' in the second word where we would
2483 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2484 */
2485 /* add a fail transition */
3251b653
NC
2486 const U32 trie_offset = ARG(source);
2487 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2488 U32 *q;
2489 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2490 const U32 numstates = trie->statecount;
786e8c11
YO
2491 const U32 ubound = trie->lasttrans + ucharcount;
2492 U32 q_read = 0;
2493 U32 q_write = 0;
2494 U32 charid;
2495 U32 base = trie->states[ 1 ].trans.base;
2496 U32 *fail;
2497 reg_ac_data *aho;
2498 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2499 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2500
2501 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2502#ifndef DEBUGGING
2503 PERL_UNUSED_ARG(depth);
2504#endif
2505
2506
2507 ARG_SET( stclass, data_slot );
c944940b 2508 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2509 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2510 aho->trie=trie_offset;
446bd890
NC
2511 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2512 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2513 Newxz( q, numstates, U32);
c944940b 2514 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2515 aho->refcount = 1;
2516 fail = aho->fail;
2517 /* initialize fail[0..1] to be 1 so that we always have
2518 a valid final fail state */
2519 fail[ 0 ] = fail[ 1 ] = 1;
2520
2521 for ( charid = 0; charid < ucharcount ; charid++ ) {
2522 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2523 if ( newstate ) {
2524 q[ q_write ] = newstate;
2525 /* set to point at the root */
2526 fail[ q[ q_write++ ] ]=1;
2527 }
2528 }
2529 while ( q_read < q_write) {
2530 const U32 cur = q[ q_read++ % numstates ];
2531 base = trie->states[ cur ].trans.base;
2532
2533 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2534 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2535 if (ch_state) {
2536 U32 fail_state = cur;
2537 U32 fail_base;
2538 do {
2539 fail_state = fail[ fail_state ];
2540 fail_base = aho->states[ fail_state ].trans.base;
2541 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2542
2543 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2544 fail[ ch_state ] = fail_state;
2545 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2546 {
2547 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2548 }
2549 q[ q_write++ % numstates] = ch_state;
2550 }
2551 }
2552 }
2553 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2554 when we fail in state 1, this allows us to use the
2555 charclass scan to find a valid start char. This is based on the principle
2556 that theres a good chance the string being searched contains lots of stuff
2557 that cant be a start char.
2558 */
2559 fail[ 0 ] = fail[ 1 ] = 0;
2560 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2561 PerlIO_printf(Perl_debug_log,
2562 "%*sStclass Failtable (%"UVuf" states): 0",
2563 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2564 );
786e8c11
YO
2565 for( q_read=1; q_read<numstates; q_read++ ) {
2566 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2567 }
2568 PerlIO_printf(Perl_debug_log, "\n");
2569 });
2570 Safefree(q);
2571 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2572}
2573
786e8c11 2574
a3621e74 2575/*
5d1c421c
JH
2576 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2577 * These need to be revisited when a newer toolchain becomes available.
2578 */
2579#if defined(__sparc64__) && defined(__GNUC__)
2580# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2581# undef SPARC64_GCC_WORKAROUND
2582# define SPARC64_GCC_WORKAROUND 1
2583# endif
2584#endif
2585
07be1b83 2586#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2587 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2588 SV * const mysv=sv_newmortal(); \
2589 regnode *Next = regnext(scan); \
2590 regprop(RExC_rx, mysv, scan); \
7f69552c 2591 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2592 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2593 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2594 }});
07be1b83 2595
1de06328 2596
bb914485 2597/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
2598 * one. The regop may be changed if the node(s) contain certain sequences that
2599 * require special handling. The joining is only done if:
bb914485
KW
2600 * 1) there is room in the current conglomerated node to entirely contain the
2601 * next one.
2602 * 2) they are the exact same node type
2603 *
87b8b349 2604 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
2605 * these get optimized out
2606 *
0a982f06
KW
2607 * If a node is to match under /i (folded), the number of characters it matches
2608 * can be different than its character length if it contains a multi-character
2609 * fold. *min_subtract is set to the total delta of the input nodes.
bb914485 2610 *
a0c4c608
KW
2611 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2612 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2613 *
bb914485 2614 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
2615 * multi-character fold sequences. It's been wrong in Perl for a very long
2616 * time. There are three code points in Unicode whose multi-character folds
2617 * were long ago discovered to mess things up. The previous designs for
2618 * dealing with these involved assigning a special node for them. This
2619 * approach doesn't work, as evidenced by this example:
a0c4c608 2620 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
0a982f06
KW
2621 * Both these fold to "sss", but if the pattern is parsed to create a node that
2622 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
2623 * successful match would have to cross the node's boundary. The new approach
2624 * that hopefully generally solves the problem generates an EXACTFU_SS node
2625 * that is "sss".
2626 *
0a982f06
KW
2627 * It turns out that there are problems with all multi-character folds, and not
2628 * just these three. Now the code is general, for all such cases, but the
2629 * three still have some special handling. The approach taken is:
2630 * 1) This routine examines each EXACTFish node that could contain multi-
2631 * character fold sequences. It returns in *min_subtract how much to
9d071ca8 2632 * subtract from the the actual length of the string to get a real minimum
0a982f06
KW
2633 * match length; it is 0 if there are no multi-char folds. This delta is
2634 * used by the caller to adjust the min length of the match, and the delta
2635 * between min and max, so that the optimizer doesn't reject these
2636 * possibilities based on size constraints.
2637 * 2) Certain of these sequences require special handling by the trie code,
2638 * so, if found, this code changes the joined node type to special ops:
2639 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2640 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2641 * is used for an EXACTFU node that contains at least one "ss" sequence in
2642 * it. For non-UTF-8 patterns and strings, this is the only case where
2643 * there is a possible fold length change. That means that a regular
2644 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2645 * with length changes, and so can be processed faster. regexec.c takes
2646 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2647 * pre-folded by regcomp.c. This saves effort in regex matching.
87b8b349 2648 * However, the pre-folding isn't done for non-UTF8 patterns because the
0a982f06
KW
2649 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2650 * down by forcing the pattern into UTF8 unless necessary. Also what
2651 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2652 * possibilities for the non-UTF8 patterns are quite simple, except for
2653 * the sharp s. All the ones that don't involve a UTF-8 target string are
2654 * members of a fold-pair, and arrays are set up for all of them so that
2655 * the other member of the pair can be found quickly. Code elsewhere in
2656 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2657 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2658 * described in the next item.
bb914485
KW
2659 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2660 * 'ss' or not is not knowable at compile time. It will match iff the
2661 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2662 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
6012a526 2663 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
bb914485
KW
2664 * described in item 3). An assumption that the optimizer part of
2665 * regexec.c (probably unwittingly) makes is that a character in the
2666 * pattern corresponds to at most a single character in the target string.
2667 * (And I do mean character, and not byte here, unlike other parts of the
2668 * documentation that have never been updated to account for multibyte
2669 * Unicode.) This assumption is wrong only in this case, as all other
2670 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2671 * virtue of having this file pre-fold UTF-8 patterns. I'm
2672 * reluctant to try to change this assumption, so instead the code punts.
9d071ca8
KW
2673 * This routine examines EXACTF nodes for the sharp s, and returns a
2674 * boolean indicating whether or not the node is an EXACTF node that
2675 * contains a sharp s. When it is true, the caller sets a flag that later
2676 * causes the optimizer in this file to not set values for the floating
2677 * and fixed string lengths, and thus avoids the optimizer code in
2678 * regexec.c that makes the invalid assumption. Thus, there is no
2679 * optimization based on string lengths for EXACTF nodes that contain the
2680 * sharp s. This only happens for /id rules (which means the pattern
2681 * isn't in UTF-8).
bb914485 2682 */
1de06328 2683
9d071ca8 2684#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2685 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2686 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2687
be8e71aa 2688STATIC U32
9d071ca8 2689S_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
2690 /* Merge several consecutive EXACTish nodes into one. */
2691 regnode *n = regnext(scan);
2692 U32 stringok = 1;
2693 regnode *next = scan + NODE_SZ_STR(scan);
2694 U32 merged = 0;
2695 U32 stopnow = 0;
2696#ifdef DEBUGGING
2697 regnode *stop = scan;
72f13be8 2698 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2699#else
d47053eb
RGS
2700 PERL_UNUSED_ARG(depth);
2701#endif
7918f24d
NC
2702
2703 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2704#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2705 PERL_UNUSED_ARG(flags);
2706 PERL_UNUSED_ARG(val);
07be1b83 2707#endif
07be1b83 2708 DEBUG_PEEP("join",scan,depth);
bb914485 2709
3f410cf6
KW
2710 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2711 * EXACT ones that are mergeable to the current one. */
2712 while (n
2713 && (PL_regkind[OP(n)] == NOTHING
2714 || (stringok && OP(n) == OP(scan)))
07be1b83 2715 && NEXT_OFF(n)
3f410cf6
KW
2716 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2717 {
07be1b83
YO
2718
2719 if (OP(n) == TAIL || n > next)
2720 stringok = 0;
2721 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2722 DEBUG_PEEP("skip:",n,depth);
2723 NEXT_OFF(scan) += NEXT_OFF(n);
2724 next = n + NODE_STEP_REGNODE;
2725#ifdef DEBUGGING
2726 if (stringok)
2727 stop = n;
2728#endif
2729 n = regnext(n);
2730 }
2731 else if (stringok) {
786e8c11 2732 const unsigned int oldl = STR_LEN(scan);
07be1b83 2733 regnode * const nnext = regnext(n);
b2230d39 2734
87b8b349
KW
2735 /* XXX I (khw) kind of doubt that this works on platforms where
2736 * U8_MAX is above 255 because of lots of other assumptions */
79a81a6e 2737 /* Don't join if the sum can't fit into a single node */
b2230d39
KW
2738 if (oldl + STR_LEN(n) > U8_MAX)
2739 break;
07be1b83
YO
2740
2741 DEBUG_PEEP("merg",n,depth);
07be1b83 2742 merged++;
b2230d39 2743
07be1b83
YO
2744 NEXT_OFF(scan) += NEXT_OFF(n);
2745 STR_LEN(scan) += STR_LEN(n);
2746 next = n + NODE_SZ_STR(n);
2747 /* Now we can overwrite *n : */
2748 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2749#ifdef DEBUGGING
2750 stop = next - 1;
2751#endif
2752 n = nnext;
2753 if (stopnow) break;
2754 }
2755
d47053eb
RGS
2756#ifdef EXPERIMENTAL_INPLACESCAN
2757 if (flags && !NEXT_OFF(n)) {
2758 DEBUG_PEEP("atch", val, depth);
2759 if (reg_off_by_arg[OP(n)]) {
2760 ARG_SET(n, val - n);
2761 }
2762 else {
2763 NEXT_OFF(n) = val - n;
2764 }
2765 stopnow = 1;
2766 }
07be1b83
YO
2767#endif
2768 }
2c2b7f86 2769
9d071ca8 2770 *min_subtract = 0;
f758bddf 2771 *has_exactf_sharp_s = FALSE;
f646642f 2772
3f410cf6
KW
2773 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2774 * can now analyze for sequences of problematic code points. (Prior to
2775 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2776 * hence missed). The sequences only happen in folding, hence for any
2777 * non-EXACT EXACTish node */
86d6fcad 2778 if (OP(scan) != EXACT) {
0a982f06
KW
2779 const U8 * const s0 = (U8*) STRING(scan);
2780 const U8 * s = s0;
2781 const U8 * const s_end = s0 + STR_LEN(scan);
f758bddf
KW
2782
2783 /* One pass is made over the node's string looking for all the
2784 * possibilities. to avoid some tests in the loop, there are two main
2785 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2786 * non-UTF-8 */
2787 if (UTF) {
86d6fcad 2788
0a982f06
KW
2789 /* Examine the string for a multi-character fold sequence. UTF-8
2790 * patterns have all characters pre-folded by the time this code is
2791 * executed */
2792 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2793 length sequence we are looking for is 2 */
86d6fcad 2794 {
0a982f06
KW
2795 int count = 0;
2796 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2797 if (! len) { /* Not a multi-char fold: get next char */
2798 s += UTF8SKIP(s);
2799 continue;
2800 }
bb914485 2801
0a982f06
KW
2802 /* Nodes with 'ss' require special handling, except for EXACTFL
2803 * and EXACTFA for which there is no multi-char fold to this */
2804 if (len == 2 && *s == 's' && *(s+1) == 's'
2805 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2806 {
2807 count = 2;
2808 OP(scan) = EXACTFU_SS;
2809 s += 2;
2810 }
2811 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2812 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2813 COMBINING_DIAERESIS_UTF8
2814 COMBINING_ACUTE_ACCENT_UTF8,
2815 6)
2816 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2817 COMBINING_DIAERESIS_UTF8
2818 COMBINING_ACUTE_ACCENT_UTF8,
2819 6)))
2820 {
2821 count = 3;
2822
2823 /* These two folds require special handling by trie's, so
2824 * change the node type to indicate this. If EXACTFA and
2825 * EXACTFL were ever to be handled by trie's, this would
2826 * have to be changed. If this node has already been
2827 * changed to EXACTFU_SS in this loop, leave it as is. (I
2828 * (khw) think it doesn't matter in regexec.c for UTF
2829 * patterns, but no need to change it */
2830 if (OP(scan) == EXACTFU) {
2831 OP(scan) = EXACTFU_TRICKYFOLD;
2832 }
2833 s += 6;
2834 }
2835 else { /* Here is a generic multi-char fold. */
2836 const U8* multi_end = s + len;
2837
2838 /* Count how many characters in it. In the case of /l and
2839 * /aa, no folds which contain ASCII code points are
2840 * allowed, so check for those, and skip if found. (In
2841 * EXACTFL, no folds are allowed to any Latin1 code point,
2842 * not just ASCII. But there aren't any of these
2843 * currently, nor ever likely, so don't take the time to
2844 * test for them. The code that generates the
2845 * is_MULTI_foo() macros croaks should one actually get put
2846 * into Unicode .) */
2847 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2848 count = utf8_length(s, multi_end);
2849 s = multi_end;
2850 }
2851 else {
2852 while (s < multi_end) {
2853 if (isASCII(*s)) {
2854 s++;
2855 goto next_iteration;
2856 }
2857 else {
2858 s += UTF8SKIP(s);
2859 }
2860 count++;
2861 }
2862 }
2863 }
f758bddf 2864
0a982f06
KW
2865 /* The delta is how long the sequence is minus 1 (1 is how long
2866 * the character that folds to the sequence is) */
2867 *min_subtract += count - 1;
2868 next_iteration: ;
bb914485
KW
2869 }
2870 }
f758bddf 2871 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
bb914485 2872
0a982f06
KW
2873 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2874 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2875 * nodes can't have multi-char folds to this range (and there are
2db97d41 2876 * no existing ones in the upper latin1 range). In the EXACTF
0a982f06
KW
2877 * case we look also for the sharp s, which can be in the final
2878 * position. Otherwise we can stop looking 1 byte earlier because
2879 * have to find at least two characters for a multi-fold */
f758bddf
KW
2880 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2881
0a982f06
KW
2882 /* The below is perhaps overboard, but this allows us to save a
2883 * test each time through the loop at the expense of a mask. This
2884 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2885 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2886 * are 64. This uses an exclusive 'or' to find that bit and then
2887 * inverts it to form a mask, with just a single 0, in the bit
2888 * position where 'S' and 's' differ. */
2889 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2890 const U8 s_masked = 's' & S_or_s_mask;
2891
2892 while (s < upper) {
40b1ba4f 2893 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
0a982f06
KW
2894 if (! len) { /* Not a multi-char fold. */
2895 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2896 {
2897 *has_exactf_sharp_s = TRUE;
2898 }
2899 s++;
2900 continue;
2901 }
2902
2903 if (len == 2
2904 && ((*s & S_or_s_mask) == s_masked)
2905 && ((*(s+1) & S_or_s_mask) == s_masked))
2906 {
2907
2908 /* EXACTF nodes need to know that the minimum length
2909 * changed so that a sharp s in the string can match this
2910 * ss in the pattern, but they remain EXACTF nodes, as they
2911 * won't match this unless the target string is is UTF-8,
2912 * which we don't know until runtime */
2913 if (OP(scan) != EXACTF) {
2914 OP(scan) = EXACTFU_SS;
2915 }
86d6fcad 2916 }
0a982f06
KW
2917
2918 *min_subtract += len - 1;
2919 s += len;
86d6fcad
KW
2920 }
2921 }
07be1b83 2922 }
3f410cf6 2923
07be1b83 2924#ifdef DEBUGGING
bb789b09
DM
2925 /* Allow dumping but overwriting the collection of skipped
2926 * ops and/or strings with fake optimized ops */
07be1b83
YO
2927 n = scan + NODE_SZ_STR(scan);
2928 while (n <= stop) {
bb789b09
DM
2929 OP(n) = OPTIMIZED;
2930 FLAGS(n) = 0;
2931 NEXT_OFF(n) = 0;
07be1b83
YO
2932 n++;
2933 }
2934#endif
2935 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2936 return stopnow;
2937}
2938
486ec47a 2939/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2940 Finds fixed substrings. */
2941
a0288114 2942/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2943 to the position after last scanned or to NULL. */
2944
40d049e4
YO
2945#define INIT_AND_WITHP \
2946 assert(!and_withp); \
2947 Newx(and_withp,1,struct regnode_charclass_class); \
2948 SAVEFREEPV(and_withp)
07be1b83 2949
b515a41d 2950/* this is a chain of data about sub patterns we are processing that
486ec47a 2951 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2952 we can simulate recursion without losing state. */
2953struct scan_frame;
2954typedef struct scan_frame {
2955 regnode *last; /* last node to process in this frame */
2956 regnode *next; /* next node to process when last is reached */
2957 struct scan_frame *prev; /*previous frame*/
2958 I32 stop; /* what stopparen do we use */
2959} scan_frame;
2960
304ee84b
YO
2961
2962#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2963
76e3520e 2964STATIC I32
40d049e4 2965S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 2966 I32 *minlenp, I32 *deltap,
40d049e4
YO
2967 regnode *last,
2968 scan_data_t *data,
2969 I32 stopparen,
2970 U8* recursed,
2971 struct regnode_charclass_class *and_withp,
2972 U32 flags, U32 depth)
c277df42
IZ
2973 /* scanp: Start here (read-write). */
2974 /* deltap: Write maxlen-minlen here. */
2975 /* last: Stop before this one. */
40d049e4
YO
2976 /* data: string data about the pattern */
2977 /* stopparen: treat close N as END */
2978 /* recursed: which subroutines have we recursed into */
2979 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 2980{
97aff369 2981 dVAR;
2d608413
KW
2982 I32 min = 0; /* There must be at least this number of characters to match */
2983 I32 pars = 0, code;
c277df42
IZ
2984 regnode *scan = *scanp, *next;
2985 I32 delta = 0;
2986 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 2987 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
2988 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2989 scan_data_t data_fake;
a3621e74 2990 SV *re_trie_maxbuff = NULL;
786e8c11 2991 regnode *first_non_open = scan;
e2e6a0f1 2992 I32 stopmin = I32_MAX;
8aa23a47 2993 scan_frame *frame = NULL;
a3621e74 2994 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 2995
7918f24d
NC
2996 PERL_ARGS_ASSERT_STUDY_CHUNK;
2997
13a24bad 2998#ifdef DEBUGGING
40d049e4 2999 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 3000#endif
40d049e4 3001
786e8c11 3002 if ( depth == 0 ) {
40d049e4 3003 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
3004 first_non_open=regnext(first_non_open);
3005 }
3006
b81d288d 3007
8aa23a47
YO
3008 fake_study_recurse:
3009 while ( scan && OP(scan) != END && scan < last ){
2d608413
KW
3010 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3011 node length to get a real minimum (because
3012 the folded version may be shorter) */
f758bddf 3013 bool has_exactf_sharp_s = FALSE;
8aa23a47 3014 /* Peephole optimizer: */
304ee84b 3015 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3016 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3017
3018 /* Its not clear to khw or hv why this is done here, and not in the
3019 * clauses that deal with EXACT nodes. khw's guess is that it's
3020 * because of a previous design */
9d071ca8 3021 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3022
3023 /* Follow the next-chain of the current node and optimize
3024 away all the NOTHINGs from it. */
3025 if (OP(scan) != CURLYX) {
3026 const int max = (reg_off_by_arg[OP(scan)]
3027 ? I32_MAX
3028 /* I32 may be smaller than U16 on CRAYs! */
3029 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3030 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3031 int noff;
3032 regnode *n = scan;
686b73d4 3033
8aa23a47
YO
3034 /* Skip NOTHING and LONGJMP. */
3035 while ((n = regnext(n))
3036 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3037 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3038 && off + noff < max)
3039 off += noff;
3040 if (reg_off_by_arg[OP(scan)])
3041 ARG(scan) = off;
3042 else
3043 NEXT_OFF(scan) = off;
3044 }
a3621e74 3045
c277df42 3046
8aa23a47
YO
3047
3048 /* The principal pseudo-switch. Cannot be a switch, since we
3049 look into several different things. */
3050 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3051 || OP(scan) == IFTHEN) {
3052 next = regnext(scan);
3053 code = OP(scan);
3054 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3055
8aa23a47
YO
3056 if (OP(next) == code || code == IFTHEN) {
3057 /* NOTE - There is similar code to this block below for handling
3058 TRIE nodes on a re-study. If you change stuff here check there
3059 too. */
3060 I32 max1 = 0, min1 = I32_MAX, num = 0;
3061 struct regnode_charclass_class accum;
3062 regnode * const startbranch=scan;
686b73d4 3063
8aa23a47 3064 if (flags & SCF_DO_SUBSTR)
304ee84b 3065 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3066 if (flags & SCF_DO_STCLASS)
e755fd73 3067 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3068
3069 while (OP(scan) == code) {
3070 I32 deltanext, minnext, f = 0, fake;
3071 struct regnode_charclass_class this_class;
3072
3073 num++;
3074 data_fake.flags = 0;
3075 if (data) {
3076 data_fake.whilem_c = data->whilem_c;
3077 data_fake.last_closep = data->last_closep;
3078 }
3079 else
3080 data_fake.last_closep = &fake;
58e23c8d
YO
3081
3082 data_fake.pos_delta = delta;
8aa23a47
YO
3083 next = regnext(scan);
3084 scan = NEXTOPER(scan);
3085 if (code != BRANCH)
c277df42 3086 scan = NEXTOPER(scan);
8aa23a47 3087 if (flags & SCF_DO_STCLASS) {
e755fd73 3088 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3089 data_fake.start_class = &this_class;
3090 f = SCF_DO_STCLASS_AND;
58e23c8d 3091 }
8aa23a47
YO
3092 if (flags & SCF_WHILEM_VISITED_POS)
3093 f |= SCF_WHILEM_VISITED_POS;
3094
3095 /* we suppose the run is continuous, last=next...*/
3096 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3097 next, &data_fake,
3098 stopparen, recursed, NULL, f,depth+1);
3099 if (min1 > minnext)
3100 min1 = minnext;
3101 if (max1 < minnext + deltanext)
3102 max1 = minnext + deltanext;
3103 if (deltanext == I32_MAX)
3104 is_inf = is_inf_internal = 1;
3105 scan = next;
3106 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3107 pars++;
3108 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3109 if ( stopmin > minnext)
3110 stopmin = min + min1;
3111 flags &= ~SCF_DO_SUBSTR;
3112 if (data)
3113 data->flags |= SCF_SEEN_ACCEPT;
3114 }
3115 if (data) {
3116 if (data_fake.flags & SF_HAS_EVAL)
3117 data->flags |= SF_HAS_EVAL;
3118 data->whilem_c = data_fake.whilem_c;
3dab1dad 3119 }
8aa23a47 3120 if (flags & SCF_DO_STCLASS)
3fffb88a 3121 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3122 }
3123 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3124 min1 = 0;
3125 if (flags & SCF_DO_SUBSTR) {
3126 data->pos_min += min1;
3127 data->pos_delta += max1 - min1;
3128 if (max1 != min1 || is_inf)
3129 data->longest = &(data->longest_float);
3130 }
3131 min += min1;
3132 delta += max1 - min1;
3133 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3134 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3135 if (min1) {
3136 cl_and(data->start_class, and_withp);
3137 flags &= ~SCF_DO_STCLASS;
653099ff 3138 }
8aa23a47
YO
3139 }
3140 else if (flags & SCF_DO_STCLASS_AND) {
3141 if (min1) {
3142 cl_and(data->start_class, &accum);
3143 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3144 }
8aa23a47
YO
3145 else {
3146 /* Switch to OR mode: cache the old value of
3147 * data->start_class */
3148 INIT_AND_WITHP;
3149 StructCopy(data->start_class, and_withp,
3150 struct regnode_charclass_class);
3151 flags &= ~SCF_DO_STCLASS_AND;
3152 StructCopy(&accum, data->start_class,
3153 struct regnode_charclass_class);
3154 flags |= SCF_DO_STCLASS_OR;
899d20b9 3155 SET_SSC_EOS(data->start_class);
de0c8cb8 3156 }
8aa23a47 3157 }
a3621e74 3158
8aa23a47
YO
3159 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3160 /* demq.
a3621e74 3161
8aa23a47
YO
3162 Assuming this was/is a branch we are dealing with: 'scan' now
3163 points at the item that follows the branch sequence, whatever
3164 it is. We now start at the beginning of the sequence and look
3165 for subsequences of
a3621e74 3166
8aa23a47
YO
3167 BRANCH->EXACT=>x1
3168 BRANCH->EXACT=>x2
3169 tail
a3621e74 3170
8aa23a47 3171 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3172
486ec47a 3173 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3174 element into a trie and then add the subsequent branch exact
3175 strings to the trie.
a3621e74 3176
8aa23a47 3177 We have two cases
a3621e74 3178
3b753521 3179 1. patterns where the whole set of branches can be converted.
a3621e74 3180
8aa23a47 3181 2. patterns where only a subset can be converted.
a3621e74 3182
8aa23a47
YO
3183 In case 1 we can replace the whole set with a single regop
3184 for the trie. In case 2 we need to keep the start and end
3b753521 3185 branches so
a3621e74 3186
8aa23a47
YO
3187 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3188 becomes BRANCH TRIE; BRANCH X;
786e8c11 3189
8aa23a47
YO
3190 There is an additional case, that being where there is a
3191 common prefix, which gets split out into an EXACT like node
3192 preceding the TRIE node.
a3621e74 3193
8aa23a47
YO
3194 If x(1..n)==tail then we can do a simple trie, if not we make
3195 a "jump" trie, such that when we match the appropriate word
486ec47a 3196 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3197 a nested if into a case structure of sorts.
b515a41d 3198
8aa23a47 3199 */
686b73d4 3200
8aa23a47
YO
3201 int made=0;
3202 if (!re_trie_maxbuff) {
3203 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3204 if (!SvIOK(re_trie_maxbuff))
3205 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3206 }
3207 if ( SvIV(re_trie_maxbuff)>=0 ) {
3208 regnode *cur;
3209 regnode *first = (regnode *)NULL;
3210 regnode *last = (regnode *)NULL;
3211 regnode *tail = scan;
fab2782b 3212 U8 trietype = 0;
8aa23a47 3213 U32 count=0;
a3621e74
YO
3214
3215#ifdef DEBUGGING
8aa23a47 3216 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3217#endif
8aa23a47
YO
3218 /* var tail is used because there may be a TAIL
3219 regop in the way. Ie, the exacts will point to the
3220 thing following the TAIL, but the last branch will
3221 point at the TAIL. So we advance tail. If we
3222 have nested (?:) we may have to move through several
3223 tails.
3224 */
3225
3226 while ( OP( tail ) == TAIL ) {
3227 /* this is the TAIL generated by (?:) */
3228 tail = regnext( tail );
3229 }
a3621e74 3230
8aa23a47 3231
df826430 3232 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3233 regprop(RExC_rx, mysv, tail );
3234 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3235 (int)depth * 2 + 2, "",
3236 "Looking for TRIE'able sequences. Tail node is: ",
3237 SvPV_nolen_const( mysv )
3238 );
3239 });
3240
3241 /*
3242
fab2782b
YO
3243 Step through the branches
3244 cur represents each branch,
3245 noper is the first thing to be matched as part of that branch
3246 noper_next is the regnext() of that node.
3247
3248 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3249 via a "jump trie" but we also support building with NOJUMPTRIE,
3250 which restricts the trie logic to structures like /FOO|BAR/.
3251
3252 If noper is a trieable nodetype then the branch is a possible optimization
3253 target. If we are building under NOJUMPTRIE then we require that noper_next
3254 is the same as scan (our current position in the regex program).
3255
3256 Once we have two or more consecutive such branches we can create a
3257 trie of the EXACT's contents and stitch it in place into the program.
3258
3259 If the sequence represents all of the branches in the alternation we
3260 replace the entire thing with a single TRIE node.
3261
3262 Otherwise when it is a subsequence we need to stitch it in place and
3263 replace only the relevant branches. This means the first branch has
3264 to remain as it is used by the alternation logic, and its next pointer,
3265 and needs to be repointed at the item on the branch chain following
3266 the last branch we have optimized away.
3267
3268 This could be either a BRANCH, in which case the subsequence is internal,
3269 or it could be the item following the branch sequence in which case the
3270 subsequence is at the end (which does not necessarily mean the first node
3271 is the start of the alternation).
3272
3273 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3274
3275 optype | trietype
3276 ----------------+-----------
3277 NOTHING | NOTHING
3278 EXACT | EXACT
3279 EXACTFU | EXACTFU
3280 EXACTFU_SS | EXACTFU
3281 EXACTFU_TRICKYFOLD | EXACTFU
3282 EXACTFA | 0
3283
8aa23a47
YO
3284
3285 */
fab2782b
YO
3286#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3287 ( EXACT == (X) ) ? EXACT : \
3288 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3289 0 )
8aa23a47
YO
3290
3291 /* dont use tail as the end marker for this traverse */
3292 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3293 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3294 U8 noper_type = OP( noper );
3295 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3296#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3297 regnode * const noper_next = regnext( noper );
df826430
YO
3298 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3299 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3300#endif
3301
df826430 3302 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3303 regprop(RExC_rx, mysv, cur);
3304 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3305 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3306
3307 regprop(RExC_rx, mysv, noper);
3308 PerlIO_printf( Perl_debug_log, " -> %s",
3309 SvPV_nolen_const(mysv));
3310
3311 if ( noper_next ) {
3312 regprop(RExC_rx, mysv, noper_next );
3313 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3314 SvPV_nolen_const(mysv));
3315 }
df826430
YO
3316 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3317 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3318 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3319 );
8aa23a47 3320 });
fab2782b
YO
3321
3322 /* Is noper a trieable nodetype that can be merged with the
3323 * current trie (if there is one)? */
3324 if ( noper_trietype
3325 &&
3326 (
df826430
YO
3327 ( noper_trietype == NOTHING)
3328 || ( trietype == NOTHING )
a40630bf 3329 || ( trietype == noper_trietype )
fab2782b 3330 )
786e8c11 3331#ifdef NOJUMPTRIE
8aa23a47 3332 && noper_next == tail
786e8c11 3333#endif
8aa23a47
YO
3334 && count < U16_MAX)
3335 {
fab2782b
YO
3336 /* Handle mergable triable node
3337 * Either we are the first node in a new trieable sequence,
3338 * in which case we do some bookkeeping, otherwise we update
3339 * the end pointer. */
fab2782b 3340 if ( !first ) {
3b6759a6 3341 first = cur;
df826430
YO
3342 if ( noper_trietype == NOTHING ) {
3343#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3344 regnode * const noper_next = regnext( noper );
3b6759a6 3345 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3346 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3347#endif
3348
190c1910 3349 if ( noper_next_trietype ) {
df826430 3350 trietype = noper_next_trietype;
190c1910
YO
3351 } else if (noper_next_type) {
3352 /* a NOTHING regop is 1 regop wide. We need at least two
3353 * for a trie so we can't merge this in */
3354 first = NULL;
3355 }
3356 } else {
3357 trietype = noper_trietype;
3b6759a6 3358 }
8aa23a47 3359 } else {
fab2782b
YO
3360 if ( trietype == NOTHING )
3361 trietype = noper_trietype;
8aa23a47
YO
3362 last = cur;
3363 }
df826430
YO
3364 if (first)
3365 count++;
fab2782b
YO
3366 } /* end handle mergable triable node */
3367 else {
3368 /* handle unmergable node -
3369 * noper may either be a triable node which can not be tried
3370 * together with the current trie, or a non triable node */
729aaeb5
YO
3371 if ( last ) {
3372 /* If last is set and trietype is not NOTHING then we have found
3373 * at least two triable branch sequences in a row of a similar
3374 * trietype so we can turn them into a trie. If/when we
3375 * allow NOTHING to start a trie sequence this condition will be
3376 * required, and it isn't expensive so we leave it in for now. */
e6351b37 3377 if ( trietype && trietype != NOTHING )
729aaeb5
YO
3378 make_trie( pRExC_state,
3379 startbranch, first, cur, tail, count,
3380 trietype, depth+1 );
fab2782b 3381 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3382 }
fab2782b 3383 if ( noper_trietype
786e8c11 3384#ifdef NOJUMPTRIE
8aa23a47 3385 && noper_next == tail
786e8c11 3386#endif
8aa23a47 3387 ){
fab2782b 3388 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3389 count = 1;
3390 first = cur;
fab2782b
YO
3391 trietype = noper_trietype;
3392 } else if (first) {
3393 /* if we already saw a first but the current node is not triable then we have
3394 * to reset the first information. */
8aa23a47
YO
3395 count = 0;
3396 first = NULL;
fab2782b 3397 trietype = 0;
8aa23a47 3398 }
fab2782b
YO
3399 } /* end handle unmergable node */
3400 } /* loop over branches */
df826430 3401 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3402 regprop(RExC_rx, mysv, cur);
3403 PerlIO_printf( Perl_debug_log,
3404 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3405 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3406
3407 });
e6351b37 3408 if ( last && trietype ) {
3b6759a6
YO
3409 if ( trietype != NOTHING ) {
3410 /* the last branch of the sequence was part of a trie,
3411 * so we have to construct it here outside of the loop
3412 */
3413 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3414#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3415 if ( ((made == MADE_EXACT_TRIE &&
3416 startbranch == first)
3417 || ( first_non_open == first )) &&
3418 depth==0 ) {
3419 flags |= SCF_TRIE_RESTUDY;
3420 if ( startbranch == first
3421 && scan == tail )
3422 {
3423 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3424 }
8aa23a47 3425 }
3dab1dad 3426#endif
3b6759a6
YO
3427 } else {
3428 /* at this point we know whatever we have is a NOTHING sequence/branch
3429 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3430 */
3431 if ( startbranch == first ) {
3432 regnode *opt;
3433 /* the entire thing is a NOTHING sequence, something like this:
3434 * (?:|) So we can turn it into a plain NOTHING op. */
3435 DEBUG_TRIE_COMPILE_r({
3436 regprop(RExC_rx, mysv, cur);
3437 PerlIO_printf( Perl_debug_log,
3438 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3439 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3440
3441 });
3442 OP(startbranch)= NOTHING;
3443 NEXT_OFF(startbranch)= tail - startbranch;
3444 for ( opt= startbranch + 1; opt < tail ; opt++ )
3445 OP(opt)= OPTIMIZED;
3446 }
3447 }
fab2782b
YO
3448 } /* end if ( last) */
3449 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3450
3451 } /* do trie */
3452
653099ff 3453 }
8aa23a47
YO
3454 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3455 scan = NEXTOPER(NEXTOPER(scan));
3456 } else /* single branch is optimized. */
3457 scan = NEXTOPER(scan);
3458 continue;
3459 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3460 scan_frame *newframe = NULL;
3461 I32 paren;
3462 regnode *start;
3463 regnode *end;
3464
3465 if (OP(scan) != SUSPEND) {
3466 /* set the pointer */
3467 if (OP(scan) == GOSUB) {
3468 paren = ARG(scan);
3469 RExC_recurse[ARG2L(scan)] = scan;
3470 start = RExC_open_parens[paren-1];
3471 end = RExC_close_parens[paren-1];
3472 } else {
3473 paren = 0;
f8fc2ecf 3474 start = RExC_rxi->program + 1;
8aa23a47
YO
3475 end = RExC_opend;
3476 }
3477 if (!recursed) {
3478 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3479 SAVEFREEPV(recursed);
3480 }
3481 if (!PAREN_TEST(recursed,paren+1)) {
3482 PAREN_SET(recursed,paren+1);
3483 Newx(newframe,1,scan_frame);
3484 } else {
3485 if (flags & SCF_DO_SUBSTR) {
304ee84b 3486 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3487 data->longest = &(data->longest_float);
3488 }
3489 is_inf = is_inf_internal = 1;
3490 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3491 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3492 flags &= ~SCF_DO_STCLASS;
3493 }
3494 } else {
3495 Newx(newframe,1,scan_frame);
3496 paren = stopparen;
3497 start = scan+2;
3498 end = regnext(scan);
3499 }
3500 if (newframe) {
3501 assert(start);
3502 assert(end);
3503 SAVEFREEPV(newframe);
3504 newframe->next = regnext(scan);
3505 newframe->last = last;
3506 newframe->stop = stopparen;
3507 newframe->prev = frame;
3508
3509 frame = newframe;
3510 scan = start;
3511 stopparen = paren;
3512 last = end;
3513
3514 continue;
3515 }
3516 }
3517 else if (OP(scan) == EXACT) {
3518 I32 l = STR_LEN(scan);
3519 UV uc;
3520 if (UTF) {
3521 const U8 * const s = (U8*)STRING(scan);
4b88fb76 3522 uc = utf8_to_uvchr_buf(s, s + l, NULL);
8aa23a47 3523 l = utf8_length(s, s + l);
8aa23a47
YO
3524 } else {
3525 uc = *((U8*)STRING(scan));
3526 }
3527 min += l;
3528 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3529 /* The code below prefers earlier match for fixed
3530 offset, later match for variable offset. */
3531 if (data->last_end == -1) { /* Update the start info. */
3532 data->last_start_min = data->pos_min;
3533 data->last_start_max = is_inf
3534 ? I32_MAX : data->pos_min + data->pos_delta;
b515a41d 3535 }
8aa23a47
YO
3536 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3537 if (UTF)
3538 SvUTF8_on(data->last_found);
3539 {
3540 SV * const sv = data->last_found;
3541 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3542 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3543 if (mg && mg->mg_len >= 0)
3544 mg->mg_len += utf8_length((U8*)STRING(scan),
3545 (U8*)STRING(scan)+STR_LEN(scan));
b515a41d 3546 }
8aa23a47
YO
3547 data->last_end = data->pos_min + l;
3548 data->pos_min += l; /* As in the first entry. */
3549 data->flags &= ~SF_BEFORE_EOL;
3550 }
3551 if (flags & SCF_DO_STCLASS_AND) {
3552 /* Check whether it is compatible with what we know already! */
3553 int compat = 1;
3554
54251c2e 3555
486ec47a 3556 /* If compatible, we or it in below. It is compatible if is
54251c2e
KW
3557 * in the bitmp and either 1) its bit or its fold is set, or 2)
3558 * it's for a locale. Even if there isn't unicode semantics
3559 * here, at runtime there may be because of matching against a
3560 * utf8 string, so accept a possible false positive for
3561 * latin1-range folds */
8aa23a47 3562 if (uc >= 0x100 ||
b25a1036 3563 (!(data->start_class->flags & ANYOF_LOCALE)
8aa23a47 3564 && !ANYOF_BITMAP_TEST(data->start_class, uc)
538b546e 3565 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
54251c2e 3566 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
8aa23a47 3567 )
d18bf9dc 3568 {
8aa23a47 3569 compat = 0;
d18bf9dc 3570 }
8aa23a47
YO
3571 ANYOF_CLASS_ZERO(data->start_class);
3572 ANYOF_BITMAP_ZERO(data->start_class);
3573 if (compat)
3574 ANYOF_BITMAP_SET(data->start_class, uc);
d18bf9dc<