This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
ensure regex evals report the right location
[perl5.git] / regcomp.c
CommitLineData
a0d0e21e
LW
1/* regcomp.c
2 */
3
4/*
4ac71550
TC
5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
6 *
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
a0d0e21e
LW
8 */
9
61296642
DM
10/* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
166f8a29 12 * a regular expression.
e4a054ea
DM
13 *
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
166f8a29
DM
18 */
19
a687059c
LW
20/* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
22 */
23
24/* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
27 */
28
e50aee73 29/* The names of the functions have been changed from regcomp and
3b753521 30 * regexec to pregcomp and pregexec in order to avoid conflicts
e50aee73
AD
31 * with the POSIX routines of the same names.
32*/
33
b9d5759e 34#ifdef PERL_EXT_RE_BUILD
54df2634 35#include "re_top.h"
b81d288d 36#endif
56953603 37
a687059c 38/*
e50aee73 39 * pregcomp and pregexec -- regsub and regerror are not used in perl
a687059c
LW
40 *
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
43 *
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
47 *
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
50 * from defects in it.
51 *
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
54 *
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
57 *
58 *
59 **** Alterations to Henry's code are...
60 ****
4bb101f2 61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
1129b882
NC
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
a687059c 64 ****
9ef589d8
LW
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
67
a687059c
LW
68 *
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
72 */
73#include "EXTERN.h"
864dbfa3 74#define PERL_IN_REGCOMP_C
a687059c 75#include "perl.h"
d06ea78c 76
acfe0abc 77#ifndef PERL_IN_XSUB_RE
d06ea78c
GS
78# include "INTERN.h"
79#endif
c277df42
IZ
80
81#define REG_COMP_C
54df2634
NC
82#ifdef PERL_IN_XSUB_RE
83# include "re_comp.h"
84#else
85# include "regcomp.h"
86#endif
a687059c 87
04e98a4d 88#include "dquote_static.c"
370b8f2f
TC
89#ifndef PERL_IN_XSUB_RE
90# include "charclass_invlists.h"
91#endif
04e98a4d 92
94dc5c2d
KW
93#define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
94
d4cce5f1 95#ifdef op
11343788 96#undef op
d4cce5f1 97#endif /* op */
11343788 98
fe14fcc3 99#ifdef MSDOS
7e4e8c89 100# if defined(BUGGY_MSC6)
fe14fcc3 101 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
7e4e8c89 102# pragma optimize("a",off)
fe14fcc3 103 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
7e4e8c89
NC
104# pragma optimize("w",on )
105# endif /* BUGGY_MSC6 */
fe14fcc3
LW
106#endif /* MSDOS */
107
a687059c
LW
108#ifndef STATIC
109#define STATIC static
110#endif
111
b1603ef8 112
830247a4 113typedef struct RExC_state_t {
514a91f1
DM
114 U32 flags; /* RXf_* are we folding, multilining? */
115 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 116 char *precomp; /* uncompiled string. */
288b8c02 117 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf
YO
118 regexp *rx; /* perl core regexp structure */
119 regexp_internal *rxi; /* internal data for regexp object pprivate field */
fac92740 120 char *start; /* Start of input for compile */
830247a4
IZ
121 char *end; /* End of input for compile */
122 char *parse; /* Input-scan pointer. */
123 I32 whilem_seen; /* number of WHILEM in this expr */
fac92740 124 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 125 regnode *emit_bound; /* First regnode outside of the allocated space */
ffc61ed2 126 regnode *emit; /* Code-emit pointer; &regdummy = don't = compiling */
830247a4
IZ
127 I32 naughty; /* How bad is this pattern? */
128 I32 sawback; /* Did we see \1, ...? */
129 U32 seen;
130 I32 size; /* Code size. */
c74340f9
YO
131 I32 npar; /* Capture buffer count, (OPEN). */
132 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 133 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
134 I32 extralen;
135 I32 seen_zerolen;
136 I32 seen_evals;
40d049e4
YO
137 regnode **open_parens; /* pointers to open parens */
138 regnode **close_parens; /* pointers to close parens */
139 regnode *opend; /* END node in program */
02daf0ab
YO
140 I32 utf8; /* whether the pattern is utf8 or not */
141 I32 orig_utf8; /* whether the pattern was originally in utf8 */
142 /* XXX use this for future optimisation of case
143 * where pattern must be upgraded to utf8. */
e40e74fe
KW
144 I32 uni_semantics; /* If a d charset modifier should use unicode
145 rules, even if the pattern is not in
146 utf8 */
81714fb9 147 HV *paren_names; /* Paren names */
1f1031fe 148
40d049e4
YO
149 regnode **recurse; /* Recurse regops */
150 I32 recurse_count; /* Number of recurse regops */
b57e4118 151 I32 in_lookbehind;
4624b182 152 I32 contains_locale;
bb3f3ed2 153 I32 override_recoding;
3d2bd50a 154 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 155 within pattern */
b1603ef8
DM
156 int num_code_blocks; /* size of code_blocks[] */
157 int code_index; /* next code_blocks[] slot */
830247a4
IZ
158#if ADD_TO_REGEXEC
159 char *starttry; /* -Dr: where regtry was called. */
160#define RExC_starttry (pRExC_state->starttry)
161#endif
d24ca0c5 162 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 163#ifdef DEBUGGING
be8e71aa 164 const char *lastparse;
3dab1dad 165 I32 lastnum;
1f1031fe 166 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
167#define RExC_lastparse (pRExC_state->lastparse)
168#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 169#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 170#endif
830247a4
IZ
171} RExC_state_t;
172
e2509266 173#define RExC_flags (pRExC_state->flags)
514a91f1 174#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 175#define RExC_precomp (pRExC_state->precomp)
288b8c02 176#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 177#define RExC_rx (pRExC_state->rx)
f8fc2ecf 178#define RExC_rxi (pRExC_state->rxi)
fac92740 179#define RExC_start (pRExC_state->start)
830247a4
IZ
180#define RExC_end (pRExC_state->end)
181#define RExC_parse (pRExC_state->parse)
182#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
183#ifdef RE_TRACK_PATTERN_OFFSETS
184#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
185#endif
830247a4 186#define RExC_emit (pRExC_state->emit)
fac92740 187#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 188#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
189#define RExC_naughty (pRExC_state->naughty)
190#define RExC_sawback (pRExC_state->sawback)
191#define RExC_seen (pRExC_state->seen)
192#define RExC_size (pRExC_state->size)
193#define RExC_npar (pRExC_state->npar)
e2e6a0f1 194#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
195#define RExC_extralen (pRExC_state->extralen)
196#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
197#define RExC_seen_evals (pRExC_state->seen_evals)
1aa99e6b 198#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 199#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 200#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
201#define RExC_open_parens (pRExC_state->open_parens)
202#define RExC_close_parens (pRExC_state->close_parens)
203#define RExC_opend (pRExC_state->opend)
81714fb9 204#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
205#define RExC_recurse (pRExC_state->recurse)
206#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 207#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 208#define RExC_contains_locale (pRExC_state->contains_locale)
bb3f3ed2 209#define RExC_override_recoding (pRExC_state->override_recoding)
830247a4 210
cde0cee5 211
a687059c
LW
212#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
213#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
214 ((*s) == '{' && regcurly(s)))
a687059c 215
35c8bce7
LW
216#ifdef SPSTART
217#undef SPSTART /* dratted cpp namespace... */
218#endif
a687059c
LW
219/*
220 * Flags to be passed up and down.
221 */
a687059c 222#define WORST 0 /* Worst case. */
a3b492c3 223#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee
KW
224
225/* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
cf8c372d
KW
226 * character, and if utf8, must be invariant. Note that this is not the same
227 * thing as REGNODE_SIMPLE */
fda99bee 228#define SIMPLE 0x02
a3b492c3
YO
229#define SPSTART 0x04 /* Starts with * or +. */
230#define TRYAGAIN 0x08 /* Weeded out a declaration. */
231#define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
a687059c 232
3dab1dad
YO
233#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
234
07be1b83
YO
235/* whether trie related optimizations are enabled */
236#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
237#define TRIE_STUDY_OPT
786e8c11 238#define FULL_TRIE_STUDY
07be1b83
YO
239#define TRIE_STCLASS
240#endif
1de06328
YO
241
242
40d049e4
YO
243
244#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
245#define PBITVAL(paren) (1 << ((paren) & 7))
246#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
247#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
248#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
249
bbd61b5f
KW
250/* If not already in utf8, do a longjmp back to the beginning */
251#define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
252#define REQUIRE_UTF8 STMT_START { \
253 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
254 } STMT_END
40d049e4 255
1de06328
YO
256/* About scan_data_t.
257
258 During optimisation we recurse through the regexp program performing
259 various inplace (keyhole style) optimisations. In addition study_chunk
260 and scan_commit populate this data structure with information about
261 what strings MUST appear in the pattern. We look for the longest
3b753521 262 string that must appear at a fixed location, and we look for the
1de06328
YO
263 longest string that may appear at a floating location. So for instance
264 in the pattern:
265
266 /FOO[xX]A.*B[xX]BAR/
267
268 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
269 strings (because they follow a .* construct). study_chunk will identify
270 both FOO and BAR as being the longest fixed and floating strings respectively.
271
272 The strings can be composites, for instance
273
274 /(f)(o)(o)/
275
276 will result in a composite fixed substring 'foo'.
277
278 For each string some basic information is maintained:
279
280 - offset or min_offset
281 This is the position the string must appear at, or not before.
282 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
283 characters must match before the string we are searching for.
284 Likewise when combined with minlenp and the length of the string it
1de06328
YO
285 tells us how many characters must appear after the string we have
286 found.
287
288 - max_offset
289 Only used for floating strings. This is the rightmost point that
3b753521 290 the string can appear at. If set to I32 max it indicates that the
1de06328
YO
291 string can occur infinitely far to the right.
292
293 - minlenp
294 A pointer to the minimum length of the pattern that the string
295 was found inside. This is important as in the case of positive
296 lookahead or positive lookbehind we can have multiple patterns
297 involved. Consider
298
299 /(?=FOO).*F/
300
301 The minimum length of the pattern overall is 3, the minimum length
302 of the lookahead part is 3, but the minimum length of the part that
303 will actually match is 1. So 'FOO's minimum length is 3, but the
304 minimum length for the F is 1. This is important as the minimum length
305 is used to determine offsets in front of and behind the string being
306 looked for. Since strings can be composites this is the length of the
486ec47a 307 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
308 the length is calculated by study_chunk, so that the minimum lengths
309 are not known until the full pattern has been compiled, thus the
310 pointer to the value.
311
312 - lookbehind
313
314 In the case of lookbehind the string being searched for can be
315 offset past the start point of the final matching string.
316 If this value was just blithely removed from the min_offset it would
317 invalidate some of the calculations for how many chars must match
318 before or after (as they are derived from min_offset and minlen and
319 the length of the string being searched for).
320 When the final pattern is compiled and the data is moved from the
321 scan_data_t structure into the regexp structure the information
322 about lookbehind is factored in, with the information that would
323 have been lost precalculated in the end_shift field for the
324 associated string.
325
326 The fields pos_min and pos_delta are used to store the minimum offset
327 and the delta to the maximum offset at the current point in the pattern.
328
329*/
2c2d71f5
JH
330
331typedef struct scan_data_t {
1de06328
YO
332 /*I32 len_min; unused */
333 /*I32 len_delta; unused */
2c2d71f5
JH
334 I32 pos_min;
335 I32 pos_delta;
336 SV *last_found;
1de06328 337 I32 last_end; /* min value, <0 unless valid. */
2c2d71f5
JH
338 I32 last_start_min;
339 I32 last_start_max;
1de06328
YO
340 SV **longest; /* Either &l_fixed, or &l_float. */
341 SV *longest_fixed; /* longest fixed string found in pattern */
342 I32 offset_fixed; /* offset where it starts */
486ec47a 343 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
344 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
345 SV *longest_float; /* longest floating string found in pattern */
346 I32 offset_float_min; /* earliest point in string it can appear */
347 I32 offset_float_max; /* latest point in string it can appear */
486ec47a 348 I32 *minlen_float; /* pointer to the minlen relevant to the string */
1de06328 349 I32 lookbehind_float; /* is the position of the string modified by LB */
2c2d71f5
JH
350 I32 flags;
351 I32 whilem_c;
cb434fcc 352 I32 *last_closep;
653099ff 353 struct regnode_charclass_class *start_class;
2c2d71f5
JH
354} scan_data_t;
355
a687059c 356/*
e50aee73 357 * Forward declarations for pregcomp()'s friends.
a687059c 358 */
a0d0e21e 359
27da23d5 360static const scan_data_t zero_scan_data =
1de06328 361 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
362
363#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
364#define SF_BEFORE_SEOL 0x0001
365#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
366#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
367#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
368
09b7f37c
CB
369#ifdef NO_UNARY_PLUS
370# define SF_FIX_SHIFT_EOL (0+2)
371# define SF_FL_SHIFT_EOL (0+4)
372#else
373# define SF_FIX_SHIFT_EOL (+2)
374# define SF_FL_SHIFT_EOL (+4)
375#endif
c277df42
IZ
376
377#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
378#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
379
380#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
381#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
382#define SF_IS_INF 0x0040
383#define SF_HAS_PAR 0x0080
384#define SF_IN_PAR 0x0100
385#define SF_HAS_EVAL 0x0200
386#define SCF_DO_SUBSTR 0x0400
653099ff
GS
387#define SCF_DO_STCLASS_AND 0x0800
388#define SCF_DO_STCLASS_OR 0x1000
389#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 390#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 391
786e8c11 392#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 393#define SCF_SEEN_ACCEPT 0x8000
07be1b83 394
43fead97 395#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
396
397/* The enums for all these are ordered so things work out correctly */
a62b1201 398#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
cfaf538b 399#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
00b27cfc 400#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
401#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
402#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1
KW
403#define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
404#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a62b1201 405
43fead97 406#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 407
ffc61ed2 408#define OOB_UNICODE 12345678
93733859 409#define OOB_NAMEDCLASS -1
b8c5462f 410
a0ed51b3
LW
411#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
412#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
413
8615cb43 414
b45f050a
JF
415/* length of regex to show in messages that don't mark a position within */
416#define RegexLengthToShowInErrorMessages 127
417
418/*
419 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
420 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
421 * op/pragma/warn/regcomp.
422 */
7253e4e3
RK
423#define MARKER1 "<-- HERE" /* marker as it appears in the description */
424#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 425
7253e4e3 426#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
427
428/*
429 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
430 * arg. Show regex, up to a maximum length. If it's too long, chop and add
431 * "...".
432 */
58e23c8d 433#define _FAIL(code) STMT_START { \
bfed75c6 434 const char *ellipses = ""; \
ccb2c380
MP
435 IV len = RExC_end - RExC_precomp; \
436 \
437 if (!SIZE_ONLY) \
288b8c02 438 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
439 if (len > RegexLengthToShowInErrorMessages) { \
440 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
441 len = RegexLengthToShowInErrorMessages - 10; \
442 ellipses = "..."; \
443 } \
58e23c8d 444 code; \
ccb2c380 445} STMT_END
8615cb43 446
58e23c8d
YO
447#define FAIL(msg) _FAIL( \
448 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
449 msg, (int)len, RExC_precomp, ellipses))
450
451#define FAIL2(msg,arg) _FAIL( \
452 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
453 arg, (int)len, RExC_precomp, ellipses))
454
b45f050a 455/*
b45f050a
JF
456 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
457 */
ccb2c380 458#define Simple_vFAIL(m) STMT_START { \
a28509cc 459 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
460 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
461 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
462} STMT_END
b45f050a
JF
463
464/*
465 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
466 */
ccb2c380
MP
467#define vFAIL(m) STMT_START { \
468 if (!SIZE_ONLY) \
288b8c02 469 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
470 Simple_vFAIL(m); \
471} STMT_END
b45f050a
JF
472
473/*
474 * Like Simple_vFAIL(), but accepts two arguments.
475 */
ccb2c380 476#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 477 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
478 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
479 (int)offset, RExC_precomp, RExC_precomp + offset); \
480} STMT_END
b45f050a
JF
481
482/*
483 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
484 */
ccb2c380
MP
485#define vFAIL2(m,a1) STMT_START { \
486 if (!SIZE_ONLY) \
288b8c02 487 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
488 Simple_vFAIL2(m, a1); \
489} STMT_END
b45f050a
JF
490
491
492/*
493 * Like Simple_vFAIL(), but accepts three arguments.
494 */
ccb2c380 495#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 496 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
497 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
498 (int)offset, RExC_precomp, RExC_precomp + offset); \
499} STMT_END
b45f050a
JF
500
501/*
502 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
503 */
ccb2c380
MP
504#define vFAIL3(m,a1,a2) STMT_START { \
505 if (!SIZE_ONLY) \
288b8c02 506 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
507 Simple_vFAIL3(m, a1, a2); \
508} STMT_END
b45f050a
JF
509
510/*
511 * Like Simple_vFAIL(), but accepts four arguments.
512 */
ccb2c380 513#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 514 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
515 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
516 (int)offset, RExC_precomp, RExC_precomp + offset); \
517} STMT_END
b45f050a 518
668c081a 519#define ckWARNreg(loc,m) STMT_START { \
a28509cc 520 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
521 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
522 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
523} STMT_END
524
668c081a 525#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 526 const IV offset = loc - RExC_precomp; \
d1d15184 527 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
528 m REPORT_LOCATION, \
529 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
530} STMT_END
531
2335b3d3
KW
532#define ckWARN2regdep(loc,m, a1) STMT_START { \
533 const IV offset = loc - RExC_precomp; \
534 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
535 m REPORT_LOCATION, \
536 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
537} STMT_END
538
668c081a 539#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 540 const IV offset = loc - RExC_precomp; \
668c081a 541 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
542 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
543} STMT_END
544
545#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 546 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
547 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
548 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
549} STMT_END
550
668c081a
NC
551#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
552 const IV offset = loc - RExC_precomp; \
553 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
554 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
555} STMT_END
556
ccb2c380 557#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 558 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
559 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
560 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
561} STMT_END
562
668c081a
NC
563#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
564 const IV offset = loc - RExC_precomp; \
565 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
566 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
567} STMT_END
568
ccb2c380 569#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 570 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
571 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
572 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
573} STMT_END
9d1d55b5 574
8615cb43 575
cd439c50 576/* Allow for side effects in s */
ccb2c380
MP
577#define REGC(c,s) STMT_START { \
578 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
579} STMT_END
cd439c50 580
fac92740
MJD
581/* Macros for recording node offsets. 20001227 mjd@plover.com
582 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
583 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
584 * Element 0 holds the number n.
07be1b83 585 * Position is 1 indexed.
fac92740 586 */
7122b237
YO
587#ifndef RE_TRACK_PATTERN_OFFSETS
588#define Set_Node_Offset_To_R(node,byte)
589#define Set_Node_Offset(node,byte)
590#define Set_Cur_Node_Offset
591#define Set_Node_Length_To_R(node,len)
592#define Set_Node_Length(node,len)
593#define Set_Node_Cur_Length(node)
594#define Node_Offset(n)
595#define Node_Length(n)
596#define Set_Node_Offset_Length(node,offset,len)
597#define ProgLen(ri) ri->u.proglen
598#define SetProgLen(ri,x) ri->u.proglen = x
599#else
600#define ProgLen(ri) ri->u.offsets[0]
601#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
602#define Set_Node_Offset_To_R(node,byte) STMT_START { \
603 if (! SIZE_ONLY) { \
604 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 605 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 606 if((node) < 0) { \
551405c4 607 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
608 } else { \
609 RExC_offsets[2*(node)-1] = (byte); \
610 } \
611 } \
612} STMT_END
613
614#define Set_Node_Offset(node,byte) \
615 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
616#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
617
618#define Set_Node_Length_To_R(node,len) STMT_START { \
619 if (! SIZE_ONLY) { \
620 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 621 __LINE__, (int)(node), (int)(len))); \
ccb2c380 622 if((node) < 0) { \
551405c4 623 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
624 } else { \
625 RExC_offsets[2*(node)] = (len); \
626 } \
627 } \
628} STMT_END
629
630#define Set_Node_Length(node,len) \
631 Set_Node_Length_To_R((node)-RExC_emit_start, len)
632#define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
633#define Set_Node_Cur_Length(node) \
634 Set_Node_Length(node, RExC_parse - parse_start)
fac92740
MJD
635
636/* Get offsets and lengths */
637#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
638#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
639
07be1b83
YO
640#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
641 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
642 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
643} STMT_END
7122b237 644#endif
07be1b83
YO
645
646#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
647#define EXPERIMENTAL_INPLACESCAN
f427392e 648#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 649
304ee84b
YO
650#define DEBUG_STUDYDATA(str,data,depth) \
651DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 652 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
653 "%*s" str "Pos:%"IVdf"/%"IVdf \
654 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
655 (int)(depth)*2, "", \
656 (IV)((data)->pos_min), \
657 (IV)((data)->pos_delta), \
304ee84b 658 (UV)((data)->flags), \
1de06328 659 (IV)((data)->whilem_c), \
304ee84b
YO
660 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
661 is_inf ? "INF " : "" \
1de06328
YO
662 ); \
663 if ((data)->last_found) \
664 PerlIO_printf(Perl_debug_log, \
665 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
666 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
667 SvPVX_const((data)->last_found), \
668 (IV)((data)->last_end), \
669 (IV)((data)->last_start_min), \
670 (IV)((data)->last_start_max), \
671 ((data)->longest && \
672 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
673 SvPVX_const((data)->longest_fixed), \
674 (IV)((data)->offset_fixed), \
675 ((data)->longest && \
676 (data)->longest==&((data)->longest_float)) ? "*" : "", \
677 SvPVX_const((data)->longest_float), \
678 (IV)((data)->offset_float_min), \
679 (IV)((data)->offset_float_max) \
680 ); \
681 PerlIO_printf(Perl_debug_log,"\n"); \
682});
683
acfe0abc 684static void clear_re(pTHX_ void *r);
4327152a 685
653099ff 686/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 687 Update the longest found anchored substring and the longest found
653099ff
GS
688 floating substrings if needed. */
689
4327152a 690STATIC void
304ee84b 691S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 692{
e1ec3a88
AL
693 const STRLEN l = CHR_SVLEN(data->last_found);
694 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 695 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 696
7918f24d
NC
697 PERL_ARGS_ASSERT_SCAN_COMMIT;
698
c277df42 699 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 700 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
701 if (*data->longest == data->longest_fixed) {
702 data->offset_fixed = l ? data->last_start_min : data->pos_min;
703 if (data->flags & SF_BEFORE_EOL)
b81d288d 704 data->flags
c277df42
IZ
705 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
706 else
707 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 708 data->minlen_fixed=minlenp;
1de06328 709 data->lookbehind_fixed=0;
a0ed51b3 710 }
304ee84b 711 else { /* *data->longest == data->longest_float */
c277df42 712 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
713 data->offset_float_max = (l
714 ? data->last_start_max
c277df42 715 : data->pos_min + data->pos_delta);
304ee84b 716 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 717 data->offset_float_max = I32_MAX;
c277df42 718 if (data->flags & SF_BEFORE_EOL)
b81d288d 719 data->flags
c277df42
IZ
720 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
721 else
722 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
723 data->minlen_float=minlenp;
724 data->lookbehind_float=0;
c277df42
IZ
725 }
726 }
727 SvCUR_set(data->last_found, 0);
0eda9292 728 {
a28509cc 729 SV * const sv = data->last_found;
097eb12c
AL
730 if (SvUTF8(sv) && SvMAGICAL(sv)) {
731 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
732 if (mg)
733 mg->mg_len = 0;
734 }
0eda9292 735 }
c277df42
IZ
736 data->last_end = -1;
737 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 738 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
739}
740
653099ff
GS
741/* Can match anything (initialization) */
742STATIC void
3fffb88a 743S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 744{
7918f24d
NC
745 PERL_ARGS_ASSERT_CL_ANYTHING;
746
f8bef550 747 ANYOF_BITMAP_SETALL(cl);
dd58aee1 748 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
3ad98780 749 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
3fffb88a
KW
750
751 /* If any portion of the regex is to operate under locale rules,
752 * initialization includes it. The reason this isn't done for all regexes
753 * is that the optimizer was written under the assumption that locale was
754 * all-or-nothing. Given the complexity and lack of documentation in the
755 * optimizer, and that there are inadequate test cases for locale, so many
756 * parts of it may not work properly, it is safest to avoid locale unless
757 * necessary. */
758 if (RExC_contains_locale) {
9d7a1e63 759 ANYOF_CLASS_SETALL(cl); /* /l uses class */
3fffb88a
KW
760 cl->flags |= ANYOF_LOCALE;
761 }
9d7a1e63
KW
762 else {
763 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
764 }
653099ff
GS
765}
766
767/* Can match anything (initialization) */
768STATIC int
5f66b61c 769S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
770{
771 int value;
772
7918f24d
NC
773 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
774
aaa51d5e 775 for (value = 0; value <= ANYOF_MAX; value += 2)
653099ff
GS
776 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
777 return 1;
1aa99e6b
IH
778 if (!(cl->flags & ANYOF_UNICODE_ALL))
779 return 0;
10edeb5d 780 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 781 return 0;
653099ff
GS
782 return 1;
783}
784
785/* Can match anything (initialization) */
786STATIC void
e755fd73 787S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 788{
7918f24d
NC
789 PERL_ARGS_ASSERT_CL_INIT;
790
8ecf7187 791 Zero(cl, 1, struct regnode_charclass_class);
653099ff 792 cl->type = ANYOF;
3fffb88a 793 cl_anything(pRExC_state, cl);
1411dba4 794 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
795}
796
1051e1c4
KW
797/* These two functions currently do the exact same thing */
798#define cl_init_zero S_cl_init
653099ff 799
dd58aee1
KW
800/* 'AND' a given class with another one. Can create false positives. 'cl'
801 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
802 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 803STATIC void
5f66b61c 804S_cl_and(struct regnode_charclass_class *cl,
a28509cc 805 const struct regnode_charclass_class *and_with)
653099ff 806{
7918f24d 807 PERL_ARGS_ASSERT_CL_AND;
40d049e4
YO
808
809 assert(and_with->type == ANYOF);
1e6ade67 810
c6b76537 811 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
812 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
813 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 814 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
815 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
816 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
653099ff
GS
817 int i;
818
819 if (and_with->flags & ANYOF_INVERT)
820 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
821 cl->bitmap[i] &= ~and_with->bitmap[i];
822 else
823 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
824 cl->bitmap[i] &= and_with->bitmap[i];
825 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 826
c6b76537 827 if (and_with->flags & ANYOF_INVERT) {
8951c461 828
c6b76537
KW
829 /* Here, the and'ed node is inverted. Get the AND of the flags that
830 * aren't affected by the inversion. Those that are affected are
831 * handled individually below */
832 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
833 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
834 cl->flags |= affected_flags;
835
836 /* We currently don't know how to deal with things that aren't in the
837 * bitmap, but we know that the intersection is no greater than what
838 * is already in cl, so let there be false positives that get sorted
839 * out after the synthetic start class succeeds, and the node is
840 * matched for real. */
841
842 /* The inversion of these two flags indicate that the resulting
843 * intersection doesn't have them */
844 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
845 cl->flags &= ~ANYOF_UNICODE_ALL;
846 }
c6b76537
KW
847 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
848 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 849 }
1aa99e6b 850 }
c6b76537 851 else { /* and'd node is not inverted */
3ad98780
KW
852 U8 outside_bitmap_but_not_utf8; /* Temp variable */
853
137165a6 854 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
855
856 /* Here 'and_with' doesn't match anything outside the bitmap
857 * (except possibly ANYOF_UNICODE_ALL), which means the
858 * intersection can't either, except for ANYOF_UNICODE_ALL, in
859 * which case we don't know what the intersection is, but it's no
860 * greater than what cl already has, so can just leave it alone,
861 * with possible false positives */
862 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
863 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 864 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 865 }
137165a6 866 }
c6b76537
KW
867 else if (! ANYOF_NONBITMAP(cl)) {
868
869 /* Here, 'and_with' does match something outside the bitmap, and cl
870 * doesn't have a list of things to match outside the bitmap. If
871 * cl can match all code points above 255, the intersection will
3ad98780
KW
872 * be those above-255 code points that 'and_with' matches. If cl
873 * can't match all Unicode code points, it means that it can't
874 * match anything outside the bitmap (since the 'if' that got us
875 * into this block tested for that), so we leave the bitmap empty.
876 */
c6b76537
KW
877 if (cl->flags & ANYOF_UNICODE_ALL) {
878 ARG_SET(cl, ARG(and_with));
3ad98780
KW
879
880 /* and_with's ARG may match things that don't require UTF8.
881 * And now cl's will too, in spite of this being an 'and'. See
882 * the comments below about the kludge */
883 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
884 }
885 }
886 else {
887 /* Here, both 'and_with' and cl match something outside the
888 * bitmap. Currently we do not do the intersection, so just match
889 * whatever cl had at the beginning. */
890 }
891
892
3ad98780
KW
893 /* Take the intersection of the two sets of flags. However, the
894 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
895 * kludge around the fact that this flag is not treated like the others
896 * which are initialized in cl_anything(). The way the optimizer works
897 * is that the synthetic start class (SSC) is initialized to match
898 * anything, and then the first time a real node is encountered, its
899 * values are AND'd with the SSC's with the result being the values of
900 * the real node. However, there are paths through the optimizer where
901 * the AND never gets called, so those initialized bits are set
902 * inappropriately, which is not usually a big deal, as they just cause
903 * false positives in the SSC, which will just mean a probably
904 * imperceptible slow down in execution. However this bit has a
905 * higher false positive consequence in that it can cause utf8.pm,
906 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
907 * bigger slowdown and also causes significant extra memory to be used.
908 * In order to prevent this, the code now takes a different tack. The
909 * bit isn't set unless some part of the regular expression needs it,
910 * but once set it won't get cleared. This means that these extra
911 * modules won't get loaded unless there was some path through the
912 * pattern that would have required them anyway, and so any false
913 * positives that occur by not ANDing them out when they could be
914 * aren't as severe as they would be if we treated this bit like all
915 * the others */
916 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
917 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 918 cl->flags &= and_with->flags;
3ad98780 919 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 920 }
653099ff
GS
921}
922
dd58aee1
KW
923/* 'OR' a given class with another one. Can create false positives. 'cl'
924 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
925 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 926STATIC void
3fffb88a 927S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 928{
7918f24d
NC
929 PERL_ARGS_ASSERT_CL_OR;
930
653099ff 931 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
932
933 /* Here, the or'd node is to be inverted. This means we take the
934 * complement of everything not in the bitmap, but currently we don't
935 * know what that is, so give up and match anything */
936 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 937 cl_anything(pRExC_state, cl);
c6b76537 938 }
653099ff
GS
939 /* We do not use
940 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
941 * <= (B1 | !B2) | (CL1 | !CL2)
942 * which is wasteful if CL2 is small, but we ignore CL2:
943 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
944 * XXXX Can we handle case-fold? Unclear:
945 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
946 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
947 */
c6b76537 948 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
949 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
950 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
653099ff
GS
951 int i;
952
953 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
954 cl->bitmap[i] |= ~or_with->bitmap[i];
955 } /* XXXX: logic is complicated otherwise */
956 else {
3fffb88a 957 cl_anything(pRExC_state, cl);
653099ff 958 }
c6b76537
KW
959
960 /* And, we can just take the union of the flags that aren't affected
961 * by the inversion */
962 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
963
964 /* For the remaining flags:
965 ANYOF_UNICODE_ALL and inverted means to not match anything above
966 255, which means that the union with cl should just be
967 what cl has in it, so can ignore this flag
968 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
969 is 127-255 to match them, but then invert that, so the
970 union with cl should just be what cl has in it, so can
971 ignore this flag
972 */
973 } else { /* 'or_with' is not inverted */
653099ff
GS
974 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
975 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
976 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
977 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
653099ff
GS
978 int i;
979
980 /* OR char bitmap and class bitmap separately */
981 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
982 cl->bitmap[i] |= or_with->bitmap[i];
1e6ade67 983 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
653099ff
GS
984 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
985 cl->classflags[i] |= or_with->classflags[i];
986 cl->flags |= ANYOF_CLASS;
987 }
988 }
989 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 990 cl_anything(pRExC_state, cl);
653099ff 991 }
9826f543 992
c6b76537
KW
993 if (ANYOF_NONBITMAP(or_with)) {
994
995 /* Use the added node's outside-the-bit-map match if there isn't a
996 * conflict. If there is a conflict (both nodes match something
997 * outside the bitmap, but what they match outside is not the same
998 * pointer, and hence not easily compared until XXX we extend
999 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
1000 * match everything outside the bitmap. If that stuff is all above
1001 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1002 if (! ANYOF_NONBITMAP(cl)) {
1003 ARG_SET(cl, ARG(or_with));
1004 }
1005 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1006
1007 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1008 cl_anything(pRExC_state, cl);
1009 }
1010 else {
1011 cl->flags |= ANYOF_UNICODE_ALL;
1012 }
c6b76537 1013 }
4c34a693 1014 }
0b9668ee
KW
1015
1016 /* Take the union */
1017 cl->flags |= or_with->flags;
1aa99e6b 1018 }
653099ff
GS
1019}
1020
a3621e74
YO
1021#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1022#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1023#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1024#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1025
3dab1dad
YO
1026
1027#ifdef DEBUGGING
07be1b83 1028/*
2b8b4781
NC
1029 dump_trie(trie,widecharmap,revcharmap)
1030 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1031 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1032
1033 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1034 The _interim_ variants are used for debugging the interim
1035 tables that are used to generate the final compressed
1036 representation which is what dump_trie expects.
1037
486ec47a 1038 Part of the reason for their existence is to provide a form
3dab1dad 1039 of documentation as to how the different representations function.
07be1b83
YO
1040
1041*/
3dab1dad
YO
1042
1043/*
3dab1dad
YO
1044 Dumps the final compressed table form of the trie to Perl_debug_log.
1045 Used for debugging make_trie().
1046*/
b9a59e08 1047
3dab1dad 1048STATIC void
2b8b4781
NC
1049S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1050 AV *revcharmap, U32 depth)
3dab1dad
YO
1051{
1052 U32 state;
ab3bbdeb 1053 SV *sv=sv_newmortal();
55eed653 1054 int colwidth= widecharmap ? 6 : 4;
2e64971a 1055 U16 word;
3dab1dad
YO
1056 GET_RE_DEBUG_FLAGS_DECL;
1057
7918f24d 1058 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1059
3dab1dad
YO
1060 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1061 (int)depth * 2 + 2,"",
1062 "Match","Base","Ofs" );
1063
1064 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1065 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1066 if ( tmp ) {
ab3bbdeb
YO
1067 PerlIO_printf( Perl_debug_log, "%*s",
1068 colwidth,
ddc5bc0f 1069 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1070 PL_colors[0], PL_colors[1],
1071 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1072 PERL_PV_ESCAPE_FIRSTCHAR
1073 )
1074 );
3dab1dad
YO
1075 }
1076 }
1077 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1078 (int)depth * 2 + 2,"");
1079
1080 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1081 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1082 PerlIO_printf( Perl_debug_log, "\n");
1083
1e2e3d02 1084 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1085 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1086
1087 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1088
1089 if ( trie->states[ state ].wordnum ) {
1090 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1091 } else {
1092 PerlIO_printf( Perl_debug_log, "%6s", "" );
1093 }
1094
1095 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1096
1097 if ( base ) {
1098 U32 ofs = 0;
1099
1100 while( ( base + ofs < trie->uniquecharcount ) ||
1101 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1102 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1103 ofs++;
1104
1105 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1106
1107 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1108 if ( ( base + ofs >= trie->uniquecharcount ) &&
1109 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1110 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1111 {
ab3bbdeb
YO
1112 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1113 colwidth,
3dab1dad
YO
1114 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1115 } else {
ab3bbdeb 1116 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1117 }
1118 }
1119
1120 PerlIO_printf( Perl_debug_log, "]");
1121
1122 }
1123 PerlIO_printf( Perl_debug_log, "\n" );
1124 }
2e64971a
DM
1125 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1126 for (word=1; word <= trie->wordcount; word++) {
1127 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1128 (int)word, (int)(trie->wordinfo[word].prev),
1129 (int)(trie->wordinfo[word].len));
1130 }
1131 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1132}
1133/*
3dab1dad
YO
1134 Dumps a fully constructed but uncompressed trie in list form.
1135 List tries normally only are used for construction when the number of
1136 possible chars (trie->uniquecharcount) is very high.
1137 Used for debugging make_trie().
1138*/
1139STATIC void
55eed653 1140S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1141 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1142 U32 depth)
3dab1dad
YO
1143{
1144 U32 state;
ab3bbdeb 1145 SV *sv=sv_newmortal();
55eed653 1146 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1147 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1148
1149 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1150
3dab1dad 1151 /* print out the table precompression. */
ab3bbdeb
YO
1152 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1153 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1154 "------:-----+-----------------\n" );
3dab1dad
YO
1155
1156 for( state=1 ; state < next_alloc ; state ++ ) {
1157 U16 charid;
1158
ab3bbdeb 1159 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1160 (int)depth * 2 + 2,"", (UV)state );
1161 if ( ! trie->states[ state ].wordnum ) {
1162 PerlIO_printf( Perl_debug_log, "%5s| ","");
1163 } else {
1164 PerlIO_printf( Perl_debug_log, "W%4x| ",
1165 trie->states[ state ].wordnum
1166 );
1167 }
1168 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1169 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1170 if ( tmp ) {
1171 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1172 colwidth,
ddc5bc0f 1173 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1174 PL_colors[0], PL_colors[1],
1175 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1176 PERL_PV_ESCAPE_FIRSTCHAR
1177 ) ,
1e2e3d02
YO
1178 TRIE_LIST_ITEM(state,charid).forid,
1179 (UV)TRIE_LIST_ITEM(state,charid).newstate
1180 );
1181 if (!(charid % 10))
664e119d
RGS
1182 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1183 (int)((depth * 2) + 14), "");
1e2e3d02 1184 }
ab3bbdeb
YO
1185 }
1186 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1187 }
1188}
1189
1190/*
3dab1dad
YO
1191 Dumps a fully constructed but uncompressed trie in table form.
1192 This is the normal DFA style state transition table, with a few
1193 twists to facilitate compression later.
1194 Used for debugging make_trie().
1195*/
1196STATIC void
55eed653 1197S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1198 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1199 U32 depth)
3dab1dad
YO
1200{
1201 U32 state;
1202 U16 charid;
ab3bbdeb 1203 SV *sv=sv_newmortal();
55eed653 1204 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1205 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1206
1207 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1208
1209 /*
1210 print out the table precompression so that we can do a visual check
1211 that they are identical.
1212 */
1213
1214 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1215
1216 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1217 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1218 if ( tmp ) {
ab3bbdeb
YO
1219 PerlIO_printf( Perl_debug_log, "%*s",
1220 colwidth,
ddc5bc0f 1221 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1222 PL_colors[0], PL_colors[1],
1223 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1224 PERL_PV_ESCAPE_FIRSTCHAR
1225 )
1226 );
3dab1dad
YO
1227 }
1228 }
1229
1230 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1231
1232 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1233 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1234 }
1235
1236 PerlIO_printf( Perl_debug_log, "\n" );
1237
1238 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1239
1240 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1241 (int)depth * 2 + 2,"",
1242 (UV)TRIE_NODENUM( state ) );
1243
1244 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1245 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1246 if (v)
1247 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1248 else
1249 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1250 }
1251 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1252 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1253 } else {
1254 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1255 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1256 }
1257 }
07be1b83 1258}
3dab1dad
YO
1259
1260#endif
1261
2e64971a 1262
786e8c11
YO
1263/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1264 startbranch: the first branch in the whole branch sequence
1265 first : start branch of sequence of branch-exact nodes.
1266 May be the same as startbranch
1267 last : Thing following the last branch.
1268 May be the same as tail.
1269 tail : item following the branch sequence
1270 count : words in the sequence
1271 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1272 depth : indent depth
3dab1dad 1273
786e8c11 1274Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1275
786e8c11
YO
1276A trie is an N'ary tree where the branches are determined by digital
1277decomposition of the key. IE, at the root node you look up the 1st character and
1278follow that branch repeat until you find the end of the branches. Nodes can be
1279marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1280
786e8c11 1281 /he|she|his|hers/
72f13be8 1282
786e8c11
YO
1283would convert into the following structure. Numbers represent states, letters
1284following numbers represent valid transitions on the letter from that state, if
1285the number is in square brackets it represents an accepting state, otherwise it
1286will be in parenthesis.
07be1b83 1287
786e8c11
YO
1288 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1289 | |
1290 | (2)
1291 | |
1292 (1) +-i->(6)-+-s->[7]
1293 |
1294 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1295
786e8c11
YO
1296 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1297
1298This shows that when matching against the string 'hers' we will begin at state 1
1299read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1300then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1301is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1302single traverse. We store a mapping from accepting to state to which word was
1303matched, and then when we have multiple possibilities we try to complete the
1304rest of the regex in the order in which they occured in the alternation.
1305
1306The only prior NFA like behaviour that would be changed by the TRIE support is
1307the silent ignoring of duplicate alternations which are of the form:
1308
1309 / (DUPE|DUPE) X? (?{ ... }) Y /x
1310
4b714af6 1311Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1312and without the optimisation. With the optimisations dupes will be silently
486ec47a 1313ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1314the following demonstrates:
1315
1316 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1317
1318which prints out 'word' three times, but
1319
1320 'words'=~/(word|word|word)(?{ print $1 })S/
1321
1322which doesnt print it out at all. This is due to other optimisations kicking in.
1323
1324Example of what happens on a structural level:
1325
486ec47a 1326The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1327
1328 1: CURLYM[1] {1,32767}(18)
1329 5: BRANCH(8)
1330 6: EXACT <ac>(16)
1331 8: BRANCH(11)
1332 9: EXACT <ad>(16)
1333 11: BRANCH(14)
1334 12: EXACT <ab>(16)
1335 16: SUCCEED(0)
1336 17: NOTHING(18)
1337 18: END(0)
1338
1339This would be optimizable with startbranch=5, first=5, last=16, tail=16
1340and should turn into:
1341
1342 1: CURLYM[1] {1,32767}(18)
1343 5: TRIE(16)
1344 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1345 <ac>
1346 <ad>
1347 <ab>
1348 16: SUCCEED(0)
1349 17: NOTHING(18)
1350 18: END(0)
1351
1352Cases where tail != last would be like /(?foo|bar)baz/:
1353
1354 1: BRANCH(4)
1355 2: EXACT <foo>(8)
1356 4: BRANCH(7)
1357 5: EXACT <bar>(8)
1358 7: TAIL(8)
1359 8: EXACT <baz>(10)
1360 10: END(0)
1361
1362which would be optimizable with startbranch=1, first=1, last=7, tail=8
1363and would end up looking like:
1364
1365 1: TRIE(8)
1366 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1367 <foo>
1368 <bar>
1369 7: TAIL(8)
1370 8: EXACT <baz>(10)
1371 10: END(0)
1372
1373 d = uvuni_to_utf8_flags(d, uv, 0);
1374
1375is the recommended Unicode-aware way of saying
1376
1377 *(d++) = uv;
1378*/
1379
fab2782b 1380#define TRIE_STORE_REVCHAR(val) \
786e8c11 1381 STMT_START { \
73031816 1382 if (UTF) { \
fab2782b 1383 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1384 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
fab2782b 1385 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
73031816
NC
1386 SvCUR_set(zlopp, kapow - flrbbbbb); \
1387 SvPOK_on(zlopp); \
1388 SvUTF8_on(zlopp); \
1389 av_push(revcharmap, zlopp); \
1390 } else { \
fab2782b 1391 char ooooff = (char)val; \
73031816
NC
1392 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1393 } \
1394 } STMT_END
786e8c11 1395
fab2782b
YO
1396#define TRIE_READ_CHAR STMT_START { \
1397 wordlen++; \
1398 if ( UTF ) { \
1399 /* if it is UTF then it is either already folded, or does not need folding */ \
1400 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1401 } \
1402 else if (folder == PL_fold_latin1) { \
1403 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1404 if ( foldlen > 0 ) { \
1405 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1406 foldlen -= len; \
1407 scan += len; \
1408 len = 0; \
1409 } else { \
1410 len = 1; \
1411 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1412 skiplen = UNISKIP(uvc); \
1413 foldlen -= skiplen; \
1414 scan = foldbuf + skiplen; \
1415 } \
1416 } else { \
1417 /* raw data, will be folded later if needed */ \
1418 uvc = (U32)*uc; \
1419 len = 1; \
1420 } \
786e8c11
YO
1421} STMT_END
1422
1423
1424
1425#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1426 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1427 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1428 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1429 } \
1430 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1431 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1432 TRIE_LIST_CUR( state )++; \
1433} STMT_END
07be1b83 1434
786e8c11
YO
1435#define TRIE_LIST_NEW(state) STMT_START { \
1436 Newxz( trie->states[ state ].trans.list, \
1437 4, reg_trie_trans_le ); \
1438 TRIE_LIST_CUR( state ) = 1; \
1439 TRIE_LIST_LEN( state ) = 4; \
1440} STMT_END
07be1b83 1441
786e8c11
YO
1442#define TRIE_HANDLE_WORD(state) STMT_START { \
1443 U16 dupe= trie->states[ state ].wordnum; \
1444 regnode * const noper_next = regnext( noper ); \
1445 \
786e8c11
YO
1446 DEBUG_r({ \
1447 /* store the word for dumping */ \
1448 SV* tmp; \
1449 if (OP(noper) != NOTHING) \
740cce10 1450 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1451 else \
740cce10 1452 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1453 av_push( trie_words, tmp ); \
786e8c11
YO
1454 }); \
1455 \
1456 curword++; \
2e64971a
DM
1457 trie->wordinfo[curword].prev = 0; \
1458 trie->wordinfo[curword].len = wordlen; \
1459 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1460 \
1461 if ( noper_next < tail ) { \
1462 if (!trie->jump) \
c944940b 1463 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1464 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1465 if (!jumper) \
1466 jumper = noper_next; \
1467 if (!nextbranch) \
1468 nextbranch= regnext(cur); \
1469 } \
1470 \
1471 if ( dupe ) { \
2e64971a
DM
1472 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1473 /* chain, so that when the bits of chain are later */\
1474 /* linked together, the dups appear in the chain */\
1475 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1476 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1477 } else { \
1478 /* we haven't inserted this word yet. */ \
1479 trie->states[ state ].wordnum = curword; \
1480 } \
1481} STMT_END
07be1b83 1482
3dab1dad 1483
786e8c11
YO
1484#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1485 ( ( base + charid >= ucharcount \
1486 && base + charid < ubound \
1487 && state == trie->trans[ base - ucharcount + charid ].check \
1488 && trie->trans[ base - ucharcount + charid ].next ) \
1489 ? trie->trans[ base - ucharcount + charid ].next \
1490 : ( state==1 ? special : 0 ) \
1491 )
3dab1dad 1492
786e8c11
YO
1493#define MADE_TRIE 1
1494#define MADE_JUMP_TRIE 2
1495#define MADE_EXACT_TRIE 4
3dab1dad 1496
a3621e74 1497STATIC I32
786e8c11 1498S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1499{
27da23d5 1500 dVAR;
a3621e74
YO
1501 /* first pass, loop through and scan words */
1502 reg_trie_data *trie;
55eed653 1503 HV *widecharmap = NULL;
2b8b4781 1504 AV *revcharmap = newAV();
a3621e74 1505 regnode *cur;
9f7f3913 1506 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1507 STRLEN len = 0;
1508 UV uvc = 0;
1509 U16 curword = 0;
1510 U32 next_alloc = 0;
786e8c11
YO
1511 regnode *jumper = NULL;
1512 regnode *nextbranch = NULL;
7f69552c 1513 regnode *convert = NULL;
2e64971a 1514 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1515 /* we just use folder as a flag in utf8 */
1e696034 1516 const U8 * folder = NULL;
a3621e74 1517
2b8b4781
NC
1518#ifdef DEBUGGING
1519 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1520 AV *trie_words = NULL;
1521 /* along with revcharmap, this only used during construction but both are
1522 * useful during debugging so we store them in the struct when debugging.
8e11feef 1523 */
2b8b4781
NC
1524#else
1525 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1526 STRLEN trie_charcount=0;
3dab1dad 1527#endif
2b8b4781 1528 SV *re_trie_maxbuff;
a3621e74 1529 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1530
1531 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1532#ifndef DEBUGGING
1533 PERL_UNUSED_ARG(depth);
1534#endif
a3621e74 1535
1e696034 1536 switch (flags) {
c46d03cf 1537 case EXACT: break;
2f7f8cb1 1538 case EXACTFA:
fab2782b
YO
1539 case EXACTFU_SS:
1540 case EXACTFU_TRICKYFOLD:
1e696034
KW
1541 case EXACTFU: folder = PL_fold_latin1; break;
1542 case EXACTF: folder = PL_fold; break;
1543 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1544 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1545 }
1546
c944940b 1547 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1548 trie->refcount = 1;
3dab1dad 1549 trie->startstate = 1;
786e8c11 1550 trie->wordcount = word_count;
f8fc2ecf 1551 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1552 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1553 if (flags == EXACT)
c944940b 1554 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1555 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1556 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1557
a3621e74 1558 DEBUG_r({
2b8b4781 1559 trie_words = newAV();
a3621e74 1560 });
a3621e74 1561
0111c4fd 1562 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1563 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1564 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1565 }
df826430 1566 DEBUG_TRIE_COMPILE_r({
3dab1dad 1567 PerlIO_printf( Perl_debug_log,
786e8c11 1568 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1569 (int)depth * 2 + 2, "",
1570 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1571 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1572 (int)depth);
3dab1dad 1573 });
7f69552c
YO
1574
1575 /* Find the node we are going to overwrite */
1576 if ( first == startbranch && OP( last ) != BRANCH ) {
1577 /* whole branch chain */
1578 convert = first;
1579 } else {
1580 /* branch sub-chain */
1581 convert = NEXTOPER( first );
1582 }
1583
a3621e74
YO
1584 /* -- First loop and Setup --
1585
1586 We first traverse the branches and scan each word to determine if it
1587 contains widechars, and how many unique chars there are, this is
1588 important as we have to build a table with at least as many columns as we
1589 have unique chars.
1590
1591 We use an array of integers to represent the character codes 0..255
38a44b82 1592 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1593 native representation of the character value as the key and IV's for the
1594 coded index.
1595
1596 *TODO* If we keep track of how many times each character is used we can
1597 remap the columns so that the table compression later on is more
3b753521 1598 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1599 middle and the least common are on the outside. IMO this would be better
1600 than a most to least common mapping as theres a decent chance the most
1601 common letter will share a node with the least common, meaning the node
486ec47a 1602 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1603 case is when we have the least common nodes twice.
1604
1605 */
1606
a3621e74 1607 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1608 regnode *noper = NEXTOPER( cur );
e1ec3a88 1609 const U8 *uc = (U8*)STRING( noper );
df826430 1610 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1611 STRLEN foldlen = 0;
1612 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1613 STRLEN skiplen = 0;
2af232bd 1614 const U8 *scan = (U8*)NULL;
07be1b83 1615 U32 wordlen = 0; /* required init */
02daf0ab
YO
1616 STRLEN chars = 0;
1617 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1618
3dab1dad 1619 if (OP(noper) == NOTHING) {
df826430
YO
1620 regnode *noper_next= regnext(noper);
1621 if (noper_next != tail && OP(noper_next) == flags) {
1622 noper = noper_next;
1623 uc= (U8*)STRING(noper);
1624 e= uc + STR_LEN(noper);
1625 trie->minlen= STR_LEN(noper);
1626 } else {
1627 trie->minlen= 0;
1628 continue;
1629 }
3dab1dad 1630 }
df826430 1631
fab2782b 1632 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1633 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1634 regardless of encoding */
fab2782b
YO
1635 if (OP( noper ) == EXACTFU_SS) {
1636 /* false positives are ok, so just set this */
1637 TRIE_BITMAP_SET(trie,0xDF);
1638 }
1639 }
a3621e74 1640 for ( ; uc < e ; uc += len ) {
3dab1dad 1641 TRIE_CHARCOUNT(trie)++;
a3621e74 1642 TRIE_READ_CHAR;
3dab1dad 1643 chars++;
a3621e74 1644 if ( uvc < 256 ) {
fab2782b
YO
1645 if ( folder ) {
1646 U8 folded= folder[ (U8) uvc ];
1647 if ( !trie->charmap[ folded ] ) {
1648 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1649 TRIE_STORE_REVCHAR( folded );
1650 }
1651 }
a3621e74
YO
1652 if ( !trie->charmap[ uvc ] ) {
1653 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1654 TRIE_STORE_REVCHAR( uvc );
a3621e74 1655 }
02daf0ab 1656 if ( set_bit ) {
62012aee
KW
1657 /* store the codepoint in the bitmap, and its folded
1658 * equivalent. */
fab2782b 1659 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1660
1661 /* store the folded codepoint */
fab2782b 1662 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1663
1664 if ( !UTF ) {
1665 /* store first byte of utf8 representation of
acdf4139
KW
1666 variant codepoints */
1667 if (! UNI_IS_INVARIANT(uvc)) {
1668 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1669 }
1670 }
02daf0ab
YO
1671 set_bit = 0; /* We've done our bit :-) */
1672 }
a3621e74
YO
1673 } else {
1674 SV** svpp;
55eed653
NC
1675 if ( !widecharmap )
1676 widecharmap = newHV();
a3621e74 1677
55eed653 1678 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1679
1680 if ( !svpp )
e4584336 1681 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1682
1683 if ( !SvTRUE( *svpp ) ) {
1684 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1685 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1686 }
1687 }
1688 }
3dab1dad 1689 if( cur == first ) {
fab2782b
YO
1690 trie->minlen = chars;
1691 trie->maxlen = chars;
3dab1dad 1692 } else if (chars < trie->minlen) {
fab2782b 1693 trie->minlen = chars;
3dab1dad 1694 } else if (chars > trie->maxlen) {
fab2782b
YO
1695 trie->maxlen = chars;
1696 }
1697 if (OP( noper ) == EXACTFU_SS) {
1698 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1699 if (trie->minlen > 1)
1700 trie->minlen= 1;
1701 }
1702 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1703 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1704 * - We assume that any such sequence might match a 2 byte string */
1705 if (trie->minlen > 2 )
1706 trie->minlen= 2;
3dab1dad
YO
1707 }
1708
a3621e74
YO
1709 } /* end first pass */
1710 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1711 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1712 (int)depth * 2 + 2,"",
55eed653 1713 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1714 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1715 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1716 );
a3621e74
YO
1717
1718 /*
1719 We now know what we are dealing with in terms of unique chars and
1720 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1721 representation using a flat table will take. If it's over a reasonable
1722 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1723 conservative but potentially much slower representation using an array
1724 of lists.
1725
1726 At the end we convert both representations into the same compressed
1727 form that will be used in regexec.c for matching with. The latter
1728 is a form that cannot be used to construct with but has memory
1729 properties similar to the list form and access properties similar
1730 to the table form making it both suitable for fast searches and
1731 small enough that its feasable to store for the duration of a program.
1732
1733 See the comment in the code where the compressed table is produced
1734 inplace from the flat tabe representation for an explanation of how
1735 the compression works.
1736
1737 */
1738
1739
2e64971a
DM
1740 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1741 prev_states[1] = 0;
1742
3dab1dad 1743 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1744 /*
1745 Second Pass -- Array Of Lists Representation
1746
1747 Each state will be represented by a list of charid:state records
1748 (reg_trie_trans_le) the first such element holds the CUR and LEN
1749 points of the allocated array. (See defines above).
1750
1751 We build the initial structure using the lists, and then convert
1752 it into the compressed table form which allows faster lookups
1753 (but cant be modified once converted).
a3621e74
YO
1754 */
1755
a3621e74
YO
1756 STRLEN transcount = 1;
1757
1e2e3d02
YO
1758 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1759 "%*sCompiling trie using list compiler\n",
1760 (int)depth * 2 + 2, ""));
686b73d4 1761
c944940b
JH
1762 trie->states = (reg_trie_state *)
1763 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1764 sizeof(reg_trie_state) );
a3621e74
YO
1765 TRIE_LIST_NEW(1);
1766 next_alloc = 2;
1767
1768 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1769
df826430 1770 regnode *noper = NEXTOPER( cur );
c445ea15 1771 U8 *uc = (U8*)STRING( noper );
df826430 1772 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1773 U32 state = 1; /* required init */
1774 U16 charid = 0; /* sanity init */
1775 U8 *scan = (U8*)NULL; /* sanity init */
1776 STRLEN foldlen = 0; /* required init */
07be1b83 1777 U32 wordlen = 0; /* required init */
c445ea15 1778 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1779 STRLEN skiplen = 0;
c445ea15 1780
df826430
YO
1781 if (OP(noper) == NOTHING) {
1782 regnode *noper_next= regnext(noper);
1783 if (noper_next != tail && OP(noper_next) == flags) {
1784 noper = noper_next;
1785 uc= (U8*)STRING(noper);
1786 e= uc + STR_LEN(noper);
1787 }
1788 }
1789
3dab1dad 1790 if (OP(noper) != NOTHING) {
786e8c11 1791 for ( ; uc < e ; uc += len ) {
c445ea15 1792
786e8c11 1793 TRIE_READ_CHAR;
c445ea15 1794
786e8c11
YO
1795 if ( uvc < 256 ) {
1796 charid = trie->charmap[ uvc ];
c445ea15 1797 } else {
55eed653 1798 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1799 if ( !svpp ) {
1800 charid = 0;
1801 } else {
1802 charid=(U16)SvIV( *svpp );
1803 }
c445ea15 1804 }
786e8c11
YO
1805 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1806 if ( charid ) {
a3621e74 1807
786e8c11
YO
1808 U16 check;
1809 U32 newstate = 0;
a3621e74 1810
786e8c11
YO
1811 charid--;
1812 if ( !trie->states[ state ].trans.list ) {
1813 TRIE_LIST_NEW( state );
c445ea15 1814 }
786e8c11
YO
1815 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1816 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1817 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1818 break;
1819 }
1820 }
1821 if ( ! newstate ) {
1822 newstate = next_alloc++;
2e64971a 1823 prev_states[newstate] = state;
786e8c11
YO
1824 TRIE_LIST_PUSH( state, charid, newstate );
1825 transcount++;
1826 }
1827 state = newstate;
1828 } else {
1829 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1830 }
a28509cc 1831 }
c445ea15 1832 }
3dab1dad 1833 TRIE_HANDLE_WORD(state);
a3621e74
YO
1834
1835 } /* end second pass */
1836
1e2e3d02
YO
1837 /* next alloc is the NEXT state to be allocated */
1838 trie->statecount = next_alloc;
c944940b
JH
1839 trie->states = (reg_trie_state *)
1840 PerlMemShared_realloc( trie->states,
1841 next_alloc
1842 * sizeof(reg_trie_state) );
a3621e74 1843
3dab1dad 1844 /* and now dump it out before we compress it */
2b8b4781
NC
1845 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1846 revcharmap, next_alloc,
1847 depth+1)
1e2e3d02 1848 );
a3621e74 1849
c944940b
JH
1850 trie->trans = (reg_trie_trans *)
1851 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1852 {
1853 U32 state;
a3621e74
YO
1854 U32 tp = 0;
1855 U32 zp = 0;
1856
1857
1858 for( state=1 ; state < next_alloc ; state ++ ) {
1859 U32 base=0;
1860
1861 /*
1862 DEBUG_TRIE_COMPILE_MORE_r(
1863 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1864 );
1865 */
1866
1867 if (trie->states[state].trans.list) {
1868 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1869 U16 maxid=minid;
a28509cc 1870 U16 idx;
a3621e74
YO
1871
1872 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1873 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1874 if ( forid < minid ) {
1875 minid=forid;
1876 } else if ( forid > maxid ) {
1877 maxid=forid;
1878 }
a3621e74
YO
1879 }
1880 if ( transcount < tp + maxid - minid + 1) {
1881 transcount *= 2;
c944940b
JH
1882 trie->trans = (reg_trie_trans *)
1883 PerlMemShared_realloc( trie->trans,
446bd890
NC
1884 transcount
1885 * sizeof(reg_trie_trans) );
a3621e74
YO
1886 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1887 }
1888 base = trie->uniquecharcount + tp - minid;
1889 if ( maxid == minid ) {
1890 U32 set = 0;
1891 for ( ; zp < tp ; zp++ ) {
1892 if ( ! trie->trans[ zp ].next ) {
1893 base = trie->uniquecharcount + zp - minid;
1894 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1895 trie->trans[ zp ].check = state;
1896 set = 1;
1897 break;
1898 }
1899 }
1900 if ( !set ) {
1901 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1902 trie->trans[ tp ].check = state;
1903 tp++;
1904 zp = tp;
1905 }
1906 } else {
1907 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1908 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1909 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1910 trie->trans[ tid ].check = state;
1911 }
1912 tp += ( maxid - minid + 1 );
1913 }
1914 Safefree(trie->states[ state ].trans.list);
1915 }
1916 /*
1917 DEBUG_TRIE_COMPILE_MORE_r(
1918 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1919 );
1920 */
1921 trie->states[ state ].trans.base=base;
1922 }
cc601c31 1923 trie->lasttrans = tp + 1;
a3621e74
YO
1924 }
1925 } else {
1926 /*
1927 Second Pass -- Flat Table Representation.
1928
1929 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1930 We know that we will need Charcount+1 trans at most to store the data
1931 (one row per char at worst case) So we preallocate both structures
1932 assuming worst case.
1933
1934 We then construct the trie using only the .next slots of the entry
1935 structs.
1936
3b753521 1937 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1938 make compression both faster and easier by keeping track of how many non
1939 zero fields are in the node.
1940
1941 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1942 transition.
1943
1944 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1945 number representing the first entry of the node, and state as a
1946 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1947 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1948 are 2 entrys per node. eg:
1949
1950 A B A B
1951 1. 2 4 1. 3 7
1952 2. 0 3 3. 0 5
1953 3. 0 0 5. 0 0
1954 4. 0 0 7. 0 0
1955
1956 The table is internally in the right hand, idx form. However as we also
1957 have to deal with the states array which is indexed by nodenum we have to
1958 use TRIE_NODENUM() to convert.
1959
1960 */
1e2e3d02
YO
1961 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1962 "%*sCompiling trie using table compiler\n",
1963 (int)depth * 2 + 2, ""));
3dab1dad 1964
c944940b
JH
1965 trie->trans = (reg_trie_trans *)
1966 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1967 * trie->uniquecharcount + 1,
1968 sizeof(reg_trie_trans) );
1969 trie->states = (reg_trie_state *)
1970 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1971 sizeof(reg_trie_state) );
a3621e74
YO
1972 next_alloc = trie->uniquecharcount + 1;
1973
3dab1dad 1974
a3621e74
YO
1975 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1976
df826430 1977 regnode *noper = NEXTOPER( cur );
a28509cc 1978 const U8 *uc = (U8*)STRING( noper );
df826430 1979 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1980
1981 U32 state = 1; /* required init */
1982
1983 U16 charid = 0; /* sanity init */
1984 U32 accept_state = 0; /* sanity init */
1985 U8 *scan = (U8*)NULL; /* sanity init */
1986
1987 STRLEN foldlen = 0; /* required init */
07be1b83 1988 U32 wordlen = 0; /* required init */
fab2782b 1989 STRLEN skiplen = 0;
a3621e74
YO
1990 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1991
df826430
YO
1992 if (OP(noper) == NOTHING) {
1993 regnode *noper_next= regnext(noper);
1994 if (noper_next != tail && OP(noper_next) == flags) {
1995 noper = noper_next;
1996 uc= (U8*)STRING(noper);
1997 e= uc + STR_LEN(noper);
1998 }
1999 }
fab2782b 2000
3dab1dad 2001 if ( OP(noper) != NOTHING ) {
786e8c11 2002 for ( ; uc < e ; uc += len ) {
a3621e74 2003
786e8c11 2004 TRIE_READ_CHAR;
a3621e74 2005
786e8c11
YO
2006 if ( uvc < 256 ) {
2007 charid = trie->charmap[ uvc ];
2008 } else {
55eed653 2009 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2010 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2011 }
786e8c11
YO
2012 if ( charid ) {
2013 charid--;
2014 if ( !trie->trans[ state + charid ].next ) {
2015 trie->trans[ state + charid ].next = next_alloc;
2016 trie->trans[ state ].check++;
2e64971a
DM
2017 prev_states[TRIE_NODENUM(next_alloc)]
2018 = TRIE_NODENUM(state);
786e8c11
YO
2019 next_alloc += trie->uniquecharcount;
2020 }
2021 state = trie->trans[ state + charid ].next;
2022 } else {
2023 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2024 }
2025 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2026 }
a3621e74 2027 }
3dab1dad
YO
2028 accept_state = TRIE_NODENUM( state );
2029 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2030
2031 } /* end second pass */
2032
3dab1dad 2033 /* and now dump it out before we compress it */
2b8b4781
NC
2034 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2035 revcharmap,
2036 next_alloc, depth+1));
a3621e74 2037
a3621e74
YO
2038 {
2039 /*
2040 * Inplace compress the table.*
2041
2042 For sparse data sets the table constructed by the trie algorithm will
2043 be mostly 0/FAIL transitions or to put it another way mostly empty.
2044 (Note that leaf nodes will not contain any transitions.)
2045
2046 This algorithm compresses the tables by eliminating most such
2047 transitions, at the cost of a modest bit of extra work during lookup:
2048
2049 - Each states[] entry contains a .base field which indicates the
2050 index in the state[] array wheres its transition data is stored.
2051
3b753521 2052 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2053
2054 - If .base is nonzero then charid is added to it to find an entry in
2055 the trans array.
2056
2057 -If trans[states[state].base+charid].check!=state then the
2058 transition is taken to be a 0/Fail transition. Thus if there are fail
2059 transitions at the front of the node then the .base offset will point
2060 somewhere inside the previous nodes data (or maybe even into a node
2061 even earlier), but the .check field determines if the transition is
2062 valid.
2063
786e8c11 2064 XXX - wrong maybe?
a3621e74 2065 The following process inplace converts the table to the compressed
3b753521 2066 table: We first do not compress the root node 1,and mark all its
a3621e74 2067 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2068 allows us to do a DFA construction from the compressed table later,
2069 and ensures that any .base pointers we calculate later are greater
2070 than 0.
a3621e74
YO
2071
2072 - We set 'pos' to indicate the first entry of the second node.
2073
2074 - We then iterate over the columns of the node, finding the first and
2075 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2076 and set the .check pointers accordingly, and advance pos
2077 appropriately and repreat for the next node. Note that when we copy
2078 the next pointers we have to convert them from the original
2079 NODEIDX form to NODENUM form as the former is not valid post
2080 compression.
2081
2082 - If a node has no transitions used we mark its base as 0 and do not
2083 advance the pos pointer.
2084
2085 - If a node only has one transition we use a second pointer into the
2086 structure to fill in allocated fail transitions from other states.
2087 This pointer is independent of the main pointer and scans forward
2088 looking for null transitions that are allocated to a state. When it
2089 finds one it writes the single transition into the "hole". If the
786e8c11 2090 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2091
2092 - Once compressed we can Renew/realloc the structures to release the
2093 excess space.
2094
2095 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2096 specifically Fig 3.47 and the associated pseudocode.
2097
2098 demq
2099 */
a3b680e6 2100 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2101 U32 state, charid;
a3621e74 2102 U32 pos = 0, zp=0;
1e2e3d02 2103 trie->statecount = laststate;
a3621e74
YO
2104
2105 for ( state = 1 ; state < laststate ; state++ ) {
2106 U8 flag = 0;
a28509cc
AL
2107 const U32 stateidx = TRIE_NODEIDX( state );
2108 const U32 o_used = trie->trans[ stateidx ].check;
2109 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2110 trie->trans[ stateidx ].check = 0;
2111
2112 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2113 if ( flag || trie->trans[ stateidx + charid ].next ) {
2114 if ( trie->trans[ stateidx + charid ].next ) {
2115 if (o_used == 1) {
2116 for ( ; zp < pos ; zp++ ) {
2117 if ( ! trie->trans[ zp ].next ) {
2118 break;
2119 }
2120 }
2121 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2122 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2123 trie->trans[ zp ].check = state;
2124 if ( ++zp > pos ) pos = zp;
2125 break;
2126 }
2127 used--;
2128 }
2129 if ( !flag ) {
2130 flag = 1;
2131 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2132 }
2133 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2134 trie->trans[ pos ].check = state;
2135 pos++;
2136 }
2137 }
2138 }
cc601c31 2139 trie->lasttrans = pos + 1;
c944940b
JH
2140 trie->states = (reg_trie_state *)
2141 PerlMemShared_realloc( trie->states, laststate
2142 * sizeof(reg_trie_state) );
a3621e74 2143 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2144 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2145 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2146 (int)depth * 2 + 2,"",
2147 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2148 (IV)next_alloc,
2149 (IV)pos,
a3621e74
YO
2150 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2151 );
2152
2153 } /* end table compress */
2154 }
1e2e3d02
YO
2155 DEBUG_TRIE_COMPILE_MORE_r(
2156 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2157 (int)depth * 2 + 2, "",
2158 (UV)trie->statecount,
2159 (UV)trie->lasttrans)
2160 );
cc601c31 2161 /* resize the trans array to remove unused space */
c944940b
JH
2162 trie->trans = (reg_trie_trans *)
2163 PerlMemShared_realloc( trie->trans, trie->lasttrans
2164 * sizeof(reg_trie_trans) );
a3621e74 2165
3b753521 2166 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2167 U8 nodetype =(U8)(flags & 0xFF);
2168 char *str=NULL;
786e8c11 2169
07be1b83 2170#ifdef DEBUGGING
e62cc96a 2171 regnode *optimize = NULL;
7122b237
YO
2172#ifdef RE_TRACK_PATTERN_OFFSETS
2173
b57a0404
JH
2174 U32 mjd_offset = 0;
2175 U32 mjd_nodelen = 0;
7122b237
YO
2176#endif /* RE_TRACK_PATTERN_OFFSETS */
2177#endif /* DEBUGGING */
a3621e74 2178 /*
3dab1dad
YO
2179 This means we convert either the first branch or the first Exact,
2180 depending on whether the thing following (in 'last') is a branch
2181 or not and whther first is the startbranch (ie is it a sub part of
2182 the alternation or is it the whole thing.)
3b753521 2183 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2184 the whole branch sequence, including the first.
a3621e74 2185 */
3dab1dad 2186 /* Find the node we are going to overwrite */
7f69552c 2187 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2188 /* branch sub-chain */
3dab1dad 2189 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2190#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2191 DEBUG_r({
2192 mjd_offset= Node_Offset((convert));
2193 mjd_nodelen= Node_Length((convert));
2194 });
7122b237 2195#endif
7f69552c 2196 /* whole branch chain */
7122b237
YO
2197 }
2198#ifdef RE_TRACK_PATTERN_OFFSETS
2199 else {
7f69552c
YO
2200 DEBUG_r({
2201 const regnode *nop = NEXTOPER( convert );
2202 mjd_offset= Node_Offset((nop));
2203 mjd_nodelen= Node_Length((nop));
2204 });
07be1b83
YO
2205 }
2206 DEBUG_OPTIMISE_r(
2207 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2208 (int)depth * 2 + 2, "",
786e8c11 2209 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2210 );
7122b237 2211#endif
3dab1dad
YO
2212 /* But first we check to see if there is a common prefix we can
2213 split out as an EXACT and put in front of the TRIE node. */
2214 trie->startstate= 1;
55eed653 2215 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2216 U32 state;
1e2e3d02 2217 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2218 U32 ofs = 0;
8e11feef
RGS
2219 I32 idx = -1;
2220 U32 count = 0;
2221 const U32 base = trie->states[ state ].trans.base;
a3621e74 2222
3dab1dad 2223 if ( trie->states[state].wordnum )
8e11feef 2224 count = 1;
a3621e74 2225
8e11feef 2226 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2227 if ( ( base + ofs >= trie->uniquecharcount ) &&
2228 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2229 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2230 {
3dab1dad 2231 if ( ++count > 1 ) {
2b8b4781 2232 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2233 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2234 if ( state == 1 ) break;
3dab1dad
YO
2235 if ( count == 2 ) {
2236 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2237 DEBUG_OPTIMISE_r(
8e11feef
RGS
2238 PerlIO_printf(Perl_debug_log,
2239 "%*sNew Start State=%"UVuf" Class: [",
2240 (int)depth * 2 + 2, "",
786e8c11 2241 (UV)state));
be8e71aa 2242 if (idx >= 0) {
2b8b4781 2243 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2244 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2245
3dab1dad 2246 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2247 if ( folder )
2248 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2249 DEBUG_OPTIMISE_r(
f1f66076 2250 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2251 );
8e11feef
RGS
2252 }
2253 }
2254 TRIE_BITMAP_SET(trie,*ch);
2255 if ( folder )
2256 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2257 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2258 }
2259 idx = ofs;
2260 }
3dab1dad
YO
2261 }
2262 if ( count == 1 ) {
2b8b4781 2263 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2264 STRLEN len;
2265 char *ch = SvPV( *tmp, len );
de734bd5
A
2266 DEBUG_OPTIMISE_r({
2267 SV *sv=sv_newmortal();
8e11feef
RGS
2268 PerlIO_printf( Perl_debug_log,
2269 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2270 (int)depth * 2 + 2, "",
de734bd5
A
2271 (UV)state, (UV)idx,
2272 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2273 PL_colors[0], PL_colors[1],
2274 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2275 PERL_PV_ESCAPE_FIRSTCHAR
2276 )
2277 );
2278 });
3dab1dad
YO
2279 if ( state==1 ) {
2280 OP( convert ) = nodetype;
2281 str=STRING(convert);
2282 STR_LEN(convert)=0;
2283 }
c490c714
YO
2284 STR_LEN(convert) += len;
2285 while (len--)
de734bd5 2286 *str++ = *ch++;
8e11feef 2287 } else {
f9049ba1 2288#ifdef DEBUGGING
8e11feef
RGS
2289 if (state>1)
2290 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2291#endif
8e11feef
RGS
2292 break;
2293 }
2294 }
2e64971a 2295 trie->prefixlen = (state-1);
3dab1dad 2296 if (str) {
8e11feef 2297 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2298 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2299 trie->startstate = state;
07be1b83
YO
2300 trie->minlen -= (state - 1);
2301 trie->maxlen -= (state - 1);
33809eae
JH
2302#ifdef DEBUGGING
2303 /* At least the UNICOS C compiler choked on this
2304 * being argument to DEBUG_r(), so let's just have
2305 * it right here. */
2306 if (
2307#ifdef PERL_EXT_RE_BUILD
2308 1
2309#else
2310 DEBUG_r_TEST
2311#endif
2312 ) {
2313 regnode *fix = convert;
2314 U32 word = trie->wordcount;
2315 mjd_nodelen++;
2316 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2317 while( ++fix < n ) {
2318 Set_Node_Offset_Length(fix, 0, 0);
2319 }
2320 while (word--) {
2321 SV ** const tmp = av_fetch( trie_words, word, 0 );
2322 if (tmp) {
2323 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2324 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2325 else
2326 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2327 }
2328 }
2329 }
2330#endif
8e11feef
RGS
2331 if (trie->maxlen) {
2332 convert = n;
2333 } else {
3dab1dad 2334 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2335 DEBUG_r(optimize= n);
3dab1dad
YO
2336 }
2337 }
2338 }
a5ca303d
YO
2339 if (!jumper)
2340 jumper = last;
3dab1dad 2341 if ( trie->maxlen ) {
8e11feef
RGS
2342 NEXT_OFF( convert ) = (U16)(tail - convert);
2343 ARG_SET( convert, data_slot );
786e8c11
YO
2344 /* Store the offset to the first unabsorbed branch in
2345 jump[0], which is otherwise unused by the jump logic.
2346 We use this when dumping a trie and during optimisation. */
2347 if (trie->jump)
7f69552c 2348 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2349
6c48061a
YO
2350 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2351 * and there is a bitmap
2352 * and the first "jump target" node we found leaves enough room
2353 * then convert the TRIE node into a TRIEC node, with the bitmap
2354 * embedded inline in the opcode - this is hypothetically faster.
2355 */
2356 if ( !trie->states[trie->startstate].wordnum
2357 && trie->bitmap
2358 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2359 {
2360 OP( convert ) = TRIEC;
2361 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2362 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2363 trie->bitmap= NULL;
2364 } else
2365 OP( convert ) = TRIE;
a3621e74 2366
3dab1dad
YO
2367 /* store the type in the flags */
2368 convert->flags = nodetype;
a5ca303d
YO
2369 DEBUG_r({
2370 optimize = convert
2371 + NODE_STEP_REGNODE
2372 + regarglen[ OP( convert ) ];
2373 });
2374 /* XXX We really should free up the resource in trie now,
2375 as we won't use them - (which resources?) dmq */
3dab1dad 2376 }
a3621e74 2377 /* needed for dumping*/
e62cc96a 2378 DEBUG_r(if (optimize) {
07be1b83 2379 regnode *opt = convert;
bcdf7404 2380
e62cc96a 2381 while ( ++opt < optimize) {
07be1b83
YO
2382 Set_Node_Offset_Length(opt,0,0);
2383 }
786e8c11
YO
2384 /*
2385 Try to clean up some of the debris left after the
2386 optimisation.
a3621e74 2387 */
786e8c11 2388 while( optimize < jumper ) {
07be1b83 2389 mjd_nodelen += Node_Length((optimize));
a3621e74 2390 OP( optimize ) = OPTIMIZED;
07be1b83 2391 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2392 optimize++;
2393 }
07be1b83 2394 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2395 });
2396 } /* end node insert */
2e64971a
DM
2397
2398 /* Finish populating the prev field of the wordinfo array. Walk back
2399 * from each accept state until we find another accept state, and if
2400 * so, point the first word's .prev field at the second word. If the
2401 * second already has a .prev field set, stop now. This will be the
2402 * case either if we've already processed that word's accept state,
3b753521
FN
2403 * or that state had multiple words, and the overspill words were
2404 * already linked up earlier.
2e64971a
DM
2405 */
2406 {
2407 U16 word;
2408 U32 state;
2409 U16 prev;
2410
2411 for (word=1; word <= trie->wordcount; word++) {
2412 prev = 0;
2413 if (trie->wordinfo[word].prev)
2414 continue;
2415 state = trie->wordinfo[word].accept;
2416 while (state) {
2417 state = prev_states[state];
2418 if (!state)
2419 break;
2420 prev = trie->states[state].wordnum;
2421 if (prev)
2422 break;
2423 }
2424 trie->wordinfo[word].prev = prev;
2425 }
2426 Safefree(prev_states);
2427 }
2428
2429
2430 /* and now dump out the compressed format */
2431 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2432
55eed653 2433 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2434#ifdef DEBUGGING
2435 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2436 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2437#else
2438 SvREFCNT_dec(revcharmap);
07be1b83 2439#endif
786e8c11
YO
2440 return trie->jump
2441 ? MADE_JUMP_TRIE
2442 : trie->startstate>1
2443 ? MADE_EXACT_TRIE
2444 : MADE_TRIE;
2445}
2446
2447STATIC void
2448S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2449{
3b753521 2450/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2451
2452 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2453 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2454 ISBN 0-201-10088-6
2455
2456 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2457 suffix of the current state's 'word' that is also a proper prefix of another word in our
2458 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2459 the DFA not to have to restart after its tried and failed a word at a given point, it
2460 simply continues as though it had been matching the other word in the first place.
2461 Consider
2462 'abcdgu'=~/abcdefg|cdgu/
2463 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2464 fail, which would bring us to the state representing 'd' in the second word where we would
2465 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2466 */
2467 /* add a fail transition */
3251b653
NC
2468 const U32 trie_offset = ARG(source);
2469 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2470 U32 *q;
2471 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2472 const U32 numstates = trie->statecount;
786e8c11
YO
2473 const U32 ubound = trie->lasttrans + ucharcount;
2474 U32 q_read = 0;
2475 U32 q_write = 0;
2476 U32 charid;
2477 U32 base = trie->states[ 1 ].trans.base;
2478 U32 *fail;
2479 reg_ac_data *aho;
2480 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2481 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2482
2483 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2484#ifndef DEBUGGING
2485 PERL_UNUSED_ARG(depth);
2486#endif
2487
2488
2489 ARG_SET( stclass, data_slot );
c944940b 2490 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2491 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2492 aho->trie=trie_offset;
446bd890
NC
2493 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2494 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2495 Newxz( q, numstates, U32);
c944940b 2496 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2497 aho->refcount = 1;
2498 fail = aho->fail;
2499 /* initialize fail[0..1] to be 1 so that we always have
2500 a valid final fail state */
2501 fail[ 0 ] = fail[ 1 ] = 1;
2502
2503 for ( charid = 0; charid < ucharcount ; charid++ ) {
2504 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2505 if ( newstate ) {
2506 q[ q_write ] = newstate;
2507 /* set to point at the root */
2508 fail[ q[ q_write++ ] ]=1;
2509 }
2510 }
2511 while ( q_read < q_write) {
2512 const U32 cur = q[ q_read++ % numstates ];
2513 base = trie->states[ cur ].trans.base;
2514
2515 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2516 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2517 if (ch_state) {
2518 U32 fail_state = cur;
2519 U32 fail_base;
2520 do {
2521 fail_state = fail[ fail_state ];
2522 fail_base = aho->states[ fail_state ].trans.base;
2523 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2524
2525 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2526 fail[ ch_state ] = fail_state;
2527 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2528 {
2529 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2530 }
2531 q[ q_write++ % numstates] = ch_state;
2532 }
2533 }
2534 }
2535 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2536 when we fail in state 1, this allows us to use the
2537 charclass scan to find a valid start char. This is based on the principle
2538 that theres a good chance the string being searched contains lots of stuff
2539 that cant be a start char.
2540 */
2541 fail[ 0 ] = fail[ 1 ] = 0;
2542 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2543 PerlIO_printf(Perl_debug_log,
2544 "%*sStclass Failtable (%"UVuf" states): 0",
2545 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2546 );
786e8c11
YO
2547 for( q_read=1; q_read<numstates; q_read++ ) {
2548 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2549 }
2550 PerlIO_printf(Perl_debug_log, "\n");
2551 });
2552 Safefree(q);
2553 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2554}
2555
786e8c11 2556
a3621e74 2557/*
5d1c421c
JH
2558 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2559 * These need to be revisited when a newer toolchain becomes available.
2560 */
2561#if defined(__sparc64__) && defined(__GNUC__)
2562# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2563# undef SPARC64_GCC_WORKAROUND
2564# define SPARC64_GCC_WORKAROUND 1
2565# endif
2566#endif
2567
07be1b83 2568#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2569 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2570 SV * const mysv=sv_newmortal(); \
2571 regnode *Next = regnext(scan); \
2572 regprop(RExC_rx, mysv, scan); \
7f69552c 2573 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2574 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2575 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2576 }});
07be1b83 2577
1de06328 2578
bb914485
KW
2579/* The below joins as many adjacent EXACTish nodes as possible into a single
2580 * one, and looks for problematic sequences of characters whose folds vs.
2581 * non-folds have sufficiently different lengths, that the optimizer would be
2582 * fooled into rejecting legitimate matches of them, and the trie construction
2583 * code can't cope with them. The joining is only done if:
2584 * 1) there is room in the current conglomerated node to entirely contain the
2585 * next one.
2586 * 2) they are the exact same node type
2587 *
2588 * The adjacent nodes actually may be separated by NOTHING kind nodes, and
2589 * these get optimized out
2590 *
9d071ca8
KW
2591 * If there are problematic code sequences, *min_subtract is set to the delta
2592 * that the minimum size of the node can be less than its actual size. And,
2593 * the node type of the result is changed to reflect that it contains these
bb914485
KW
2594 * sequences.
2595 *
a0c4c608
KW
2596 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2597 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2598 *
bb914485
KW
2599 * This is as good a place as any to discuss the design of handling these
2600 * problematic sequences. It's been wrong in Perl for a very long time. There
2601 * are three code points in Unicode whose folded lengths differ so much from
2602 * the un-folded lengths that it causes problems for the optimizer and trie
2603 * construction. Why only these are problematic, and not others where lengths
2604 * also differ is something I (khw) do not understand. New versions of Unicode
2605 * might add more such code points. Hopefully the logic in fold_grind.t that
287722f3 2606 * figures out what to test (in part by verifying that each size-combination
bb914485 2607 * gets tested) will catch any that do come along, so they can be added to the
287722f3
KW
2608 * special handling below. The chances of new ones are actually rather small,
2609 * as most, if not all, of the world's scripts that have casefolding have
2610 * already been encoded by Unicode. Also, a number of Unicode's decisions were
2611 * made to allow compatibility with pre-existing standards, and almost all of
2612 * those have already been dealt with. These would otherwise be the most
2613 * likely candidates for generating further tricky sequences. In other words,
2614 * Unicode by itself is unlikely to add new ones unless it is for compatibility
a0c4c608 2615 * with pre-existing standards, and there aren't many of those left.
bb914485
KW
2616 *
2617 * The previous designs for dealing with these involved assigning a special
2618 * node for them. This approach doesn't work, as evidenced by this example:
a0c4c608 2619 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
bb914485
KW
2620 * Both these fold to "sss", but if the pattern is parsed to create a node of
2621 * that would match just the \xDF, it won't be able to handle the case where a
2622 * successful match would have to cross the node's boundary. The new approach
2623 * that hopefully generally solves the problem generates an EXACTFU_SS node
2624 * that is "sss".
2625 *
2626 * There are a number of components to the approach (a lot of work for just
2627 * three code points!):
2628 * 1) This routine examines each EXACTFish node that could contain the
9d071ca8
KW
2629 * problematic sequences. It returns in *min_subtract how much to
2630 * subtract from the the actual length of the string to get a real minimum
2631 * for one that could match it. This number is usually 0 except for the
2632 * problematic sequences. This delta is used by the caller to adjust the
2633 * min length of the match, and the delta between min and max, so that the
2634 * optimizer doesn't reject these possibilities based on size constraints.
bb914485
KW
2635 * 2) These sequences are not currently correctly handled by the trie code
2636 * either, so it changes the joined node type to ops that are not handled
fab2782b 2637 * by trie's, those new ops being EXACTFU_SS and EXACTFU_TRICKYFOLD.
bb914485
KW
2638 * 3) This is sufficient for the two Greek sequences (described below), but
2639 * the one involving the Sharp s (\xDF) needs more. The node type
2640 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2641 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2642 * case where there is a possible fold length change. That means that a
2643 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2644 * itself with length changes, and so can be processed faster. regexec.c
2645 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2646 * is pre-folded by regcomp.c. This saves effort in regex matching.
2647 * However, probably mostly for historical reasons, the pre-folding isn't
a0c4c608
KW
2648 * done for non-UTF8 patterns (and it can't be for EXACTF and EXACTFL
2649 * nodes, as what they fold to isn't known until runtime.) The fold
2650 * possibilities for the non-UTF8 patterns are quite simple, except for
2651 * the sharp s. All the ones that don't involve a UTF-8 target string
2652 * are members of a fold-pair, and arrays are set up for all of them
2653 * that quickly find the other member of the pair. It might actually
2654 * be faster to pre-fold these, but it isn't currently done, except for
2655 * the sharp s. Code elsewhere in this file makes sure that it gets
2656 * folded to 'ss', even if the pattern isn't UTF-8. This avoids the
2657 * issues described in the next item.
bb914485
KW
2658 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2659 * 'ss' or not is not knowable at compile time. It will match iff the
2660 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2661 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2662 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2663 * described in item 3). An assumption that the optimizer part of
2664 * regexec.c (probably unwittingly) makes is that a character in the
2665 * pattern corresponds to at most a single character in the target string.
2666 * (And I do mean character, and not byte here, unlike other parts of the
2667 * documentation that have never been updated to account for multibyte
2668 * Unicode.) This assumption is wrong only in this case, as all other
2669 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2670 * virtue of having this file pre-fold UTF-8 patterns. I'm
2671 * reluctant to try to change this assumption, so instead the code punts.
9d071ca8
KW
2672 * This routine examines EXACTF nodes for the sharp s, and returns a
2673 * boolean indicating whether or not the node is an EXACTF node that
2674 * contains a sharp s. When it is true, the caller sets a flag that later
2675 * causes the optimizer in this file to not set values for the floating
2676 * and fixed string lengths, and thus avoids the optimizer code in
2677 * regexec.c that makes the invalid assumption. Thus, there is no
2678 * optimization based on string lengths for EXACTF nodes that contain the
2679 * sharp s. This only happens for /id rules (which means the pattern
2680 * isn't in UTF-8).
bb914485 2681 */
1de06328 2682
9d071ca8 2683#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2684 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2685 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2686
be8e71aa 2687STATIC U32
9d071ca8 2688S_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
2689 /* Merge several consecutive EXACTish nodes into one. */
2690 regnode *n = regnext(scan);
2691 U32 stringok = 1;
2692 regnode *next = scan + NODE_SZ_STR(scan);
2693 U32 merged = 0;
2694 U32 stopnow = 0;
2695#ifdef DEBUGGING
2696 regnode *stop = scan;
72f13be8 2697 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2698#else
d47053eb
RGS
2699 PERL_UNUSED_ARG(depth);
2700#endif
7918f24d
NC
2701
2702 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2703#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2704 PERL_UNUSED_ARG(flags);
2705 PERL_UNUSED_ARG(val);
07be1b83 2706#endif
07be1b83 2707 DEBUG_PEEP("join",scan,depth);
bb914485 2708
3f410cf6
KW
2709 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2710 * EXACT ones that are mergeable to the current one. */
2711 while (n
2712 && (PL_regkind[OP(n)] == NOTHING
2713 || (stringok && OP(n) == OP(scan)))
07be1b83 2714 && NEXT_OFF(n)
3f410cf6
KW
2715 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2716 {
07be1b83
YO
2717
2718 if (OP(n) == TAIL || n > next)
2719 stringok = 0;
2720 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2721 DEBUG_PEEP("skip:",n,depth);
2722 NEXT_OFF(scan) += NEXT_OFF(n);
2723 next = n + NODE_STEP_REGNODE;
2724#ifdef DEBUGGING
2725 if (stringok)
2726 stop = n;
2727#endif
2728 n = regnext(n);
2729 }
2730 else if (stringok) {
786e8c11 2731 const unsigned int oldl = STR_LEN(scan);
07be1b83 2732 regnode * const nnext = regnext(n);
b2230d39
KW
2733
2734 if (oldl + STR_LEN(n) > U8_MAX)
2735 break;
07be1b83
YO
2736
2737 DEBUG_PEEP("merg",n,depth);
07be1b83 2738 merged++;
b2230d39 2739
07be1b83
YO
2740 NEXT_OFF(scan) += NEXT_OFF(n);
2741 STR_LEN(scan) += STR_LEN(n);
2742 next = n + NODE_SZ_STR(n);
2743 /* Now we can overwrite *n : */
2744 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2745#ifdef DEBUGGING
2746 stop = next - 1;
2747#endif
2748 n = nnext;
2749 if (stopnow) break;
2750 }
2751
d47053eb
RGS
2752#ifdef EXPERIMENTAL_INPLACESCAN
2753 if (flags && !NEXT_OFF(n)) {
2754 DEBUG_PEEP("atch", val, depth);
2755 if (reg_off_by_arg[OP(n)]) {
2756 ARG_SET(n, val - n);
2757 }
2758 else {
2759 NEXT_OFF(n) = val - n;
2760 }
2761 stopnow = 1;
2762 }
07be1b83
YO
2763#endif
2764 }
2c2b7f86 2765
9d071ca8 2766 *min_subtract = 0;
f758bddf 2767 *has_exactf_sharp_s = FALSE;
f646642f 2768
3f410cf6
KW
2769 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2770 * can now analyze for sequences of problematic code points. (Prior to
2771 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2772 * hence missed). The sequences only happen in folding, hence for any
2773 * non-EXACT EXACTish node */
86d6fcad 2774 if (OP(scan) != EXACT) {
f758bddf
KW
2775 U8 *s;
2776 U8 * s0 = (U8*) STRING(scan);
2777 U8 * const s_end = s0 + STR_LEN(scan);
2778
2779 /* The below is perhaps overboard, but this allows us to save a test
2780 * each time through the loop at the expense of a mask. This is
2781 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2782 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2783 * This uses an exclusive 'or' to find that bit and then inverts it to
2784 * form a mask, with just a single 0, in the bit position where 'S' and
2785 * 's' differ. */
dbeb8947 2786 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
f758bddf
KW
2787 const U8 s_masked = 's' & S_or_s_mask;
2788
2789 /* One pass is made over the node's string looking for all the
2790 * possibilities. to avoid some tests in the loop, there are two main
2791 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2792 * non-UTF-8 */
2793 if (UTF) {
86d6fcad 2794
f758bddf
KW
2795 /* There are two problematic Greek code points in Unicode
2796 * casefolding
86d6fcad
KW
2797 *
2798 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2799 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2800 *
2801 * which casefold to
2802 *
2803 * Unicode UTF-8
2804 *
2805 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2806 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2807 *
2808 * This means that in case-insensitive matching (or "loose
2809 * matching", as Unicode calls it), an EXACTF of length six (the
2810 * UTF-8 encoded byte length of the above casefolded versions) can
2811 * match a target string of length two (the byte length of UTF-8
2812 * encoded U+0390 or U+03B0). This would rather mess up the
2813 * minimum length computation. (there are other code points that
2814 * also fold to these two sequences, but the delta is smaller)
2815 *
f758bddf
KW
2816 * If these sequences are found, the minimum length is decreased by
2817 * four (six minus two).
86d6fcad 2818 *
f758bddf
KW
2819 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2820 * LETTER SHARP S. We decrease the min length by 1 for each
2821 * occurrence of 'ss' found */
3f410cf6 2822
e294cc5d 2823#ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
f758bddf
KW
2824# define U390_first_byte 0xb4
2825 const U8 U390_tail[] = "\x68\xaf\x49\xaf\x42";
2826# define U3B0_first_byte 0xb5
2827 const U8 U3B0_tail[] = "\x46\xaf\x49\xaf\x42";
e294cc5d 2828#else
f758bddf
KW
2829# define U390_first_byte 0xce
2830 const U8 U390_tail[] = "\xb9\xcc\x88\xcc\x81";
2831# define U3B0_first_byte 0xcf
2832 const U8 U3B0_tail[] = "\x85\xcc\x88\xcc\x81";
e294cc5d 2833#endif
f758bddf
KW
2834 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2835 yields a net of 0 */
2836 /* Examine the string for one of the problematic sequences */
2837 for (s = s0;
2838 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2839 * sequence we are looking for is 2 */
2840 s += UTF8SKIP(s))
86d6fcad 2841 {
bb914485 2842
f758bddf
KW
2843 /* Look for the first byte in each problematic sequence */
2844 switch (*s) {
2845 /* We don't have to worry about other things that fold to
2846 * 's' (such as the long s, U+017F), as all above-latin1
2847 * code points have been pre-folded */
2848 case 's':
2849 case 'S':
2850
a0c4c608
KW
2851 /* Current character is an 's' or 'S'. If next one is
2852 * as well, we have the dreaded sequence */
f758bddf
KW
2853 if (((*(s+1) & S_or_s_mask) == s_masked)
2854 /* These two node types don't have special handling
2855 * for 'ss' */
2856 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2857 {
9d071ca8 2858 *min_subtract += 1;
f758bddf
KW
2859 OP(scan) = EXACTFU_SS;
2860 s++; /* No need to look at this character again */
2861 }
2862 break;
2863
2864 case U390_first_byte:
2865 if (s_end - s >= len
2866
2867 /* The 1's are because are skipping comparing the
2868 * first byte */
2869 && memEQ(s + 1, U390_tail, len - 1))
2870 {
2871 goto greek_sequence;
2872 }
2873 break;
2874
2875 case U3B0_first_byte:
2876 if (! (s_end - s >= len
2877 && memEQ(s + 1, U3B0_tail, len - 1)))
2878 {
2879 break;
2880 }
2881 greek_sequence:
9d071ca8 2882 *min_subtract += 4;
f758bddf
KW
2883
2884 /* This can't currently be handled by trie's, so change
2885 * the node type to indicate this. If EXACTFA and
2886 * EXACTFL were ever to be handled by trie's, this
2887 * would have to be changed. If this node has already
2888 * been changed to EXACTFU_SS in this loop, leave it as
2889 * is. (I (khw) think it doesn't matter in regexec.c
2890 * for UTF patterns, but no need to change it */
2891 if (OP(scan) == EXACTFU) {
fab2782b 2892 OP(scan) = EXACTFU_TRICKYFOLD;
f758bddf
KW
2893 }
2894 s += 6; /* We already know what this sequence is. Skip
2895 the rest of it */
2896 break;
bb914485
KW
2897 }
2898 }
2899 }
f758bddf 2900 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
bb914485 2901
f758bddf
KW
2902 /* Here, the pattern is not UTF-8. We need to look only for the
2903 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2904 * in the final position. Otherwise we can stop looking 1 byte
2905 * earlier because have to find both the first and second 's' */
2906 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2907
2908 for (s = s0; s < upper; s++) {
2909 switch (*s) {
2910 case 'S':
2911 case 's':
2912 if (s_end - s > 1
2913 && ((*(s+1) & S_or_s_mask) == s_masked))
2914 {
9d071ca8 2915 *min_subtract += 1;
f758bddf
KW
2916
2917 /* EXACTF nodes need to know that the minimum
2918 * length changed so that a sharp s in the string
2919 * can match this ss in the pattern, but they
2920 * remain EXACTF nodes, as they are not trie'able,
2921 * so don't have to invent a new node type to
2922 * exclude them from the trie code */
2923 if (OP(scan) != EXACTF) {
2924 OP(scan) = EXACTFU_SS;
2925 }
2926 s++;
2927 }
2928 break;
2929 case LATIN_SMALL_LETTER_SHARP_S:
2930 if (OP(scan) == EXACTF) {
2931 *has_exactf_sharp_s = TRUE;
2932 }
2933 break;
86d6fcad
KW
2934 }
2935 }
2936 }
07be1b83 2937 }
3f410cf6 2938
07be1b83 2939#ifdef DEBUGGING
bb789b09
DM
2940 /* Allow dumping but overwriting the collection of skipped
2941 * ops and/or strings with fake optimized ops */
07be1b83
YO
2942 n = scan + NODE_SZ_STR(scan);
2943 while (n <= stop) {
bb789b09
DM
2944 OP(n) = OPTIMIZED;
2945 FLAGS(n) = 0;
2946 NEXT_OFF(n) = 0;
07be1b83
YO
2947 n++;
2948 }
2949#endif
2950 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2951 return stopnow;
2952}
2953
486ec47a 2954/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2955 Finds fixed substrings. */
2956
a0288114 2957/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2958 to the position after last scanned or to NULL. */
2959
40d049e4
YO
2960#define INIT_AND_WITHP \
2961 assert(!and_withp); \
2962 Newx(and_withp,1,struct regnode_charclass_class); \
2963 SAVEFREEPV(and_withp)
07be1b83 2964
b515a41d 2965/* this is a chain of data about sub patterns we are processing that
486ec47a 2966 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2967 we can simulate recursion without losing state. */
2968struct scan_frame;
2969typedef struct scan_frame {
2970 regnode *last; /* last node to process in this frame */
2971 regnode *next; /* next node to process when last is reached */
2972 struct scan_frame *prev; /*previous frame*/
2973 I32 stop; /* what stopparen do we use */
2974} scan_frame;
2975
304ee84b
YO
2976
2977#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2978
e1d1eefb
YO
2979#define CASE_SYNST_FNC(nAmE) \
2980case nAmE: \
2981 if (flags & SCF_DO_STCLASS_AND) { \
2982 for (value = 0; value < 256; value++) \
2983 if (!is_ ## nAmE ## _cp(value)) \
2984 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2985 } \
2986 else { \
2987 for (value = 0; value < 256; value++) \
2988 if (is_ ## nAmE ## _cp(value)) \
2989 ANYOF_BITMAP_SET(data->start_class, value); \
2990 } \
2991 break; \
2992case N ## nAmE: \
2993 if (flags & SCF_DO_STCLASS_AND) { \
2994 for (value = 0; value < 256; value++) \
2995 if (is_ ## nAmE ## _cp(value)) \
2996 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2997 } \
2998 else { \
2999 for (value = 0; value < 256; value++) \
3000 if (!is_ ## nAmE ## _cp(value)) \
3001 ANYOF_BITMAP_SET(data->start_class, value); \
3002 } \
3003 break
3004
3005
3006
76e3520e 3007STATIC I32
40d049e4 3008S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 3009 I32 *minlenp, I32 *deltap,
40d049e4
YO
3010 regnode *last,
3011 scan_data_t *data,
3012 I32 stopparen,
3013 U8* recursed,
3014 struct regnode_charclass_class *and_withp,
3015 U32 flags, U32 depth)
c277df42
IZ
3016 /* scanp: Start here (read-write). */
3017 /* deltap: Write maxlen-minlen here. */
3018 /* last: Stop before this one. */
40d049e4
YO
3019 /* data: string data about the pattern */
3020 /* stopparen: treat close N as END */
3021 /* recursed: which subroutines have we recursed into */
3022 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 3023{
97aff369 3024 dVAR;
c277df42
IZ
3025 I32 min = 0, pars = 0, code;
3026 regnode *scan = *scanp, *next;
3027 I32 delta = 0;
3028 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 3029 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
3030 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3031 scan_data_t data_fake;
a3621e74 3032 SV *re_trie_maxbuff = NULL;
786e8c11 3033 regnode *first_non_open = scan;
e2e6a0f1 3034 I32 stopmin = I32_MAX;
8aa23a47 3035 scan_frame *frame = NULL;
a3621e74 3036 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 3037
7918f24d
NC
3038 PERL_ARGS_ASSERT_STUDY_CHUNK;
3039
13a24bad 3040#ifdef DEBUGGING
40d049e4 3041 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 3042#endif
40d049e4 3043
786e8c11 3044 if ( depth == 0 ) {
40d049e4 3045 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
3046 first_non_open=regnext(first_non_open);
3047 }
3048
b81d288d 3049
8aa23a47
YO
3050 fake_study_recurse:
3051 while ( scan && OP(scan) != END && scan < last ){
9d071ca8
KW
3052 UV min_subtract = 0; /* How much to subtract from the minimum node
3053 length to get a real minimum (because the
3054 folded version may be shorter) */
f758bddf 3055 bool has_exactf_sharp_s = FALSE;
8aa23a47 3056 /* Peephole optimizer: */
304ee84b 3057 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3058 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3059
3060 /* Its not clear to khw or hv why this is done here, and not in the
3061 * clauses that deal with EXACT nodes. khw's guess is that it's
3062 * because of a previous design */
9d071ca8 3063 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3064
3065 /* Follow the next-chain of the current node and optimize
3066 away all the NOTHINGs from it. */
3067 if (OP(scan) != CURLYX) {
3068 const int max = (reg_off_by_arg[OP(scan)]
3069 ? I32_MAX
3070 /* I32 may be smaller than U16 on CRAYs! */
3071 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3072 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3073 int noff;
3074 regnode *n = scan;
686b73d4 3075
8aa23a47
YO
3076 /* Skip NOTHING and LONGJMP. */
3077 while ((n = regnext(n))
3078 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3079 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3080 && off + noff < max)
3081 off += noff;
3082 if (reg_off_by_arg[OP(scan)])
3083 ARG(scan) = off;
3084 else
3085 NEXT_OFF(scan) = off;
3086 }
a3621e74 3087
c277df42 3088
8aa23a47
YO
3089
3090 /* The principal pseudo-switch. Cannot be a switch, since we
3091 look into several different things. */
3092 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3093 || OP(scan) == IFTHEN) {
3094 next = regnext(scan);
3095 code = OP(scan);
3096 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3097
8aa23a47
YO
3098 if (OP(next) == code || code == IFTHEN) {
3099 /* NOTE - There is similar code to this block below for handling
3100 TRIE nodes on a re-study. If you change stuff here check there
3101 too. */
3102 I32 max1 = 0, min1 = I32_MAX, num = 0;
3103 struct regnode_charclass_class accum;
3104 regnode * const startbranch=scan;
686b73d4 3105
8aa23a47 3106 if (flags & SCF_DO_SUBSTR)
304ee84b 3107 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3108 if (flags & SCF_DO_STCLASS)
e755fd73 3109 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3110
3111 while (OP(scan) == code) {
3112 I32 deltanext, minnext, f = 0, fake;
3113 struct regnode_charclass_class this_class;
3114
3115 num++;
3116 data_fake.flags = 0;
3117 if (data) {
3118 data_fake.whilem_c = data->whilem_c;
3119 data_fake.last_closep = data->last_closep;
3120 }
3121 else
3122 data_fake.last_closep = &fake;
58e23c8d
YO
3123
3124 data_fake.pos_delta = delta;
8aa23a47
YO
3125 next = regnext(scan);
3126 scan = NEXTOPER(scan);
3127 if (code != BRANCH)
c277df42 3128 scan = NEXTOPER(scan);
8aa23a47 3129 if (flags & SCF_DO_STCLASS) {
e755fd73 3130 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3131 data_fake.start_class = &this_class;
3132 f = SCF_DO_STCLASS_AND;
58e23c8d 3133 }
8aa23a47
YO
3134 if (flags & SCF_WHILEM_VISITED_POS)
3135 f |= SCF_WHILEM_VISITED_POS;
3136
3137 /* we suppose the run is continuous, last=next...*/
3138 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3139 next, &data_fake,
3140 stopparen, recursed, NULL, f,depth+1);
3141 if (min1 > minnext)
3142 min1 = minnext;
3143 if (max1 < minnext + deltanext)
3144 max1 = minnext + deltanext;
3145 if (deltanext == I32_MAX)
3146 is_inf = is_inf_internal = 1;
3147 scan = next;
3148 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3149 pars++;
3150 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3151 if ( stopmin > minnext)
3152 stopmin = min + min1;
3153 flags &= ~SCF_DO_SUBSTR;
3154 if (data)
3155 data->flags |= SCF_SEEN_ACCEPT;
3156 }
3157 if (data) {
3158 if (data_fake.flags & SF_HAS_EVAL)
3159 data->flags |= SF_HAS_EVAL;
3160 data->whilem_c = data_fake.whilem_c;
3dab1dad 3161 }
8aa23a47 3162 if (flags & SCF_DO_STCLASS)
3fffb88a 3163 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3164 }
3165 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3166 min1 = 0;
3167 if (flags & SCF_DO_SUBSTR) {
3168 data->pos_min += min1;
3169 data->pos_delta += max1 - min1;
3170 if (max1 != min1 || is_inf)
3171 data->longest = &(data->longest_float);
3172 }
3173 min += min1;
3174 delta += max1 - min1;
3175 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3176 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3177 if (min1) {
3178 cl_and(data->start_class, and_withp);
3179 flags &= ~SCF_DO_STCLASS;
653099ff 3180 }
8aa23a47
YO
3181 }
3182 else if (flags & SCF_DO_STCLASS_AND) {
3183 if (min1) {
3184 cl_and(data->start_class, &accum);
3185 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3186 }
8aa23a47
YO
3187 else {
3188 /* Switch to OR mode: cache the old value of
3189 * data->start_class */
3190 INIT_AND_WITHP;
3191 StructCopy(data->start_class, and_withp,
3192 struct regnode_charclass_class);
3193 flags &= ~SCF_DO_STCLASS_AND;
3194 StructCopy(&accum, data->start_class,
3195 struct regnode_charclass_class);
3196 flags |= SCF_DO_STCLASS_OR;
3197 data->start_class->flags |= ANYOF_EOS;
de0c8cb8 3198 }
8aa23a47 3199 }
a3621e74 3200
8aa23a47
YO
3201 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3202 /* demq.
a3621e74 3203
8aa23a47
YO
3204 Assuming this was/is a branch we are dealing with: 'scan' now
3205 points at the item that follows the branch sequence, whatever
3206 it is. We now start at the beginning of the sequence and look
3207 for subsequences of
a3621e74 3208
8aa23a47
YO
3209 BRANCH->EXACT=>x1
3210 BRANCH->EXACT=>x2
3211 tail
a3621e74 3212
8aa23a47 3213 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3214
486ec47a 3215 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3216 element into a trie and then add the subsequent branch exact
3217 strings to the trie.
a3621e74 3218
8aa23a47 3219 We have two cases
a3621e74 3220
3b753521 3221 1. patterns where the whole set of branches can be converted.
a3621e74 3222
8aa23a47 3223 2. patterns where only a subset can be converted.
a3621e74 3224
8aa23a47
YO
3225 In case 1 we can replace the whole set with a single regop
3226 for the trie. In case 2 we need to keep the start and end
3b753521 3227 branches so
a3621e74 3228
8aa23a47
YO
3229 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3230 becomes BRANCH TRIE; BRANCH X;
786e8c11 3231
8aa23a47
YO
3232 There is an additional case, that being where there is a
3233 common prefix, which gets split out into an EXACT like node
3234 preceding the TRIE node.
a3621e74 3235
8aa23a47
YO
3236 If x(1..n)==tail then we can do a simple trie, if not we make
3237 a "jump" trie, such that when we match the appropriate word
486ec47a 3238 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3239 a nested if into a case structure of sorts.
b515a41d 3240
8aa23a47 3241 */
686b73d4 3242
8aa23a47
YO
3243 int made=0;
3244 if (!re_trie_maxbuff) {
3245 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3246 if (!SvIOK(re_trie_maxbuff))
3247 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3248 }
3249 if ( SvIV(re_trie_maxbuff)>=0 ) {
3250 regnode *cur;
3251 regnode *first = (regnode *)NULL;
3252 regnode *last = (regnode *)NULL;
3253 regnode *tail = scan;
fab2782b 3254 U8 trietype = 0;
8aa23a47 3255 U32 count=0;
a3621e74
YO
3256
3257#ifdef DEBUGGING
8aa23a47 3258 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3259#endif
8aa23a47
YO
3260 /* var tail is used because there may be a TAIL
3261 regop in the way. Ie, the exacts will point to the
3262 thing following the TAIL, but the last branch will
3263 point at the TAIL. So we advance tail. If we
3264 have nested (?:) we may have to move through several
3265 tails.
3266 */
3267
3268 while ( OP( tail ) == TAIL ) {
3269 /* this is the TAIL generated by (?:) */
3270 tail = regnext( tail );
3271 }
a3621e74 3272
8aa23a47 3273
df826430 3274 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3275 regprop(RExC_rx, mysv, tail );
3276 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3277 (int)depth * 2 + 2, "",
3278 "Looking for TRIE'able sequences. Tail node is: ",
3279 SvPV_nolen_const( mysv )
3280 );
3281 });
3282
3283 /*
3284
fab2782b
YO
3285 Step through the branches
3286 cur represents each branch,
3287 noper is the first thing to be matched as part of that branch
3288 noper_next is the regnext() of that node.
3289
3290 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3291 via a "jump trie" but we also support building with NOJUMPTRIE,
3292 which restricts the trie logic to structures like /FOO|BAR/.
3293
3294 If noper is a trieable nodetype then the branch is a possible optimization
3295 target. If we are building under NOJUMPTRIE then we require that noper_next
3296 is the same as scan (our current position in the regex program).
3297
3298 Once we have two or more consecutive such branches we can create a
3299 trie of the EXACT's contents and stitch it in place into the program.
3300
3301 If the sequence represents all of the branches in the alternation we
3302 replace the entire thing with a single TRIE node.
3303
3304 Otherwise when it is a subsequence we need to stitch it in place and
3305 replace only the relevant branches. This means the first branch has
3306 to remain as it is used by the alternation logic, and its next pointer,
3307 and needs to be repointed at the item on the branch chain following
3308 the last branch we have optimized away.
3309
3310 This could be either a BRANCH, in which case the subsequence is internal,
3311 or it could be the item following the branch sequence in which case the
3312 subsequence is at the end (which does not necessarily mean the first node
3313 is the start of the alternation).
3314
3315 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3316
3317 optype | trietype
3318 ----------------+-----------
3319 NOTHING | NOTHING
3320 EXACT | EXACT
3321 EXACTFU | EXACTFU
3322 EXACTFU_SS | EXACTFU
3323 EXACTFU_TRICKYFOLD | EXACTFU
3324 EXACTFA | 0
3325
8aa23a47
YO
3326
3327 */
fab2782b
YO
3328#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3329 ( EXACT == (X) ) ? EXACT : \
3330 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3331 0 )
8aa23a47
YO
3332
3333 /* dont use tail as the end marker for this traverse */
3334 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3335 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3336 U8 noper_type = OP( noper );
3337 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3338#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3339 regnode * const noper_next = regnext( noper );
df826430
YO
3340 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3341 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3342#endif
3343
df826430 3344 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3345 regprop(RExC_rx, mysv, cur);
3346 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3347 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3348
3349 regprop(RExC_rx, mysv, noper);
3350 PerlIO_printf( Perl_debug_log, " -> %s",
3351 SvPV_nolen_const(mysv));
3352
3353 if ( noper_next ) {
3354 regprop(RExC_rx, mysv, noper_next );
3355 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3356 SvPV_nolen_const(mysv));
3357 }
df826430
YO
3358 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3359 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3360 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3361 );
8aa23a47 3362 });
fab2782b
YO
3363
3364 /* Is noper a trieable nodetype that can be merged with the
3365 * current trie (if there is one)? */
3366 if ( noper_trietype
3367 &&
3368 (
df826430
YO
3369 ( noper_trietype == NOTHING)
3370 || ( trietype == NOTHING )
a40630bf 3371 || ( trietype == noper_trietype )
fab2782b 3372 )
786e8c11 3373#ifdef NOJUMPTRIE
8aa23a47 3374 && noper_next == tail
786e8c11 3375#endif
8aa23a47
YO
3376 && count < U16_MAX)
3377 {
fab2782b
YO
3378 /* Handle mergable triable node
3379 * Either we are the first node in a new trieable sequence,
3380 * in which case we do some bookkeeping, otherwise we update
3381 * the end pointer. */
fab2782b 3382 if ( !first ) {
3b6759a6
YO
3383 first = cur;
3384 trietype = noper_trietype;
df826430
YO
3385 if ( noper_trietype == NOTHING ) {
3386#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3387 regnode * const noper_next = regnext( noper );
3b6759a6 3388 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3389 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3390#endif
3391
3b6759a6 3392 if ( noper_next_trietype )
df826430 3393 trietype = noper_next_trietype;
3b6759a6 3394 }
8aa23a47 3395 } else {
fab2782b
YO
3396 if ( trietype == NOTHING )
3397 trietype = noper_trietype;
8aa23a47
YO
3398 last = cur;
3399 }
df826430
YO
3400 if (first)
3401 count++;
fab2782b
YO
3402 } /* end handle mergable triable node */
3403 else {
3404 /* handle unmergable node -
3405 * noper may either be a triable node which can not be tried
3406 * together with the current trie, or a non triable node */
729aaeb5
YO
3407 if ( last ) {
3408 /* If last is set and trietype is not NOTHING then we have found
3409 * at least two triable branch sequences in a row of a similar
3410 * trietype so we can turn them into a trie. If/when we
3411 * allow NOTHING to start a trie sequence this condition will be
3412 * required, and it isn't expensive so we leave it in for now. */
3413 if ( trietype != NOTHING )
3414 make_trie( pRExC_state,
3415 startbranch, first, cur, tail, count,
3416 trietype, depth+1 );
fab2782b 3417 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3418 }
fab2782b 3419 if ( noper_trietype
786e8c11 3420#ifdef NOJUMPTRIE
8aa23a47 3421 && noper_next == tail
786e8c11 3422#endif
8aa23a47 3423 ){
fab2782b 3424 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3425 count = 1;
3426 first = cur;
fab2782b
YO
3427 trietype = noper_trietype;
3428 } else if (first) {
3429 /* if we already saw a first but the current node is not triable then we have
3430 * to reset the first information. */
8aa23a47
YO
3431 count = 0;
3432 first = NULL;
fab2782b 3433 trietype = 0;
8aa23a47 3434 }
fab2782b
YO
3435 } /* end handle unmergable node */
3436 } /* loop over branches */
df826430 3437 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3438 regprop(RExC_rx, mysv, cur);
3439 PerlIO_printf( Perl_debug_log,
3440 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3441 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3442
3443 });
3b6759a6
YO
3444 if ( last ) {
3445 if ( trietype != NOTHING ) {
3446 /* the last branch of the sequence was part of a trie,
3447 * so we have to construct it here outside of the loop
3448 */
3449 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3450#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3451 if ( ((made == MADE_EXACT_TRIE &&
3452 startbranch == first)
3453 || ( first_non_open == first )) &&
3454 depth==0 ) {
3455 flags |= SCF_TRIE_RESTUDY;
3456 if ( startbranch == first
3457 && scan == tail )
3458 {
3459 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3460 }
8aa23a47 3461 }
3dab1dad 3462#endif
3b6759a6
YO
3463 } else {
3464 /* at this point we know whatever we have is a NOTHING sequence/branch
3465 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3466 */
3467 if ( startbranch == first ) {
3468 regnode *opt;
3469 /* the entire thing is a NOTHING sequence, something like this:
3470 * (?:|) So we can turn it into a plain NOTHING op. */
3471 DEBUG_TRIE_COMPILE_r({
3472 regprop(RExC_rx, mysv, cur);
3473 PerlIO_printf( Perl_debug_log,
3474 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3475 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3476
3477 });
3478 OP(startbranch)= NOTHING;
3479 NEXT_OFF(startbranch)= tail - startbranch;
3480 for ( opt= startbranch + 1; opt < tail ; opt++ )
3481 OP(opt)= OPTIMIZED;
3482 }
3483 }
fab2782b
YO
3484 } /* end if ( last) */
3485 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3486
3487 } /* do trie */
3488
653099ff 3489 }
8aa23a47
YO
3490 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3491 scan = NEXTOPER(NEXTOPER(scan));
3492 } else /* single branch is optimized. */
3493 scan = NEXTOPER(scan);
3494 continue;
3495 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3496 scan_frame *newframe = NULL;
3497 I32 paren;
3498 regnode *start;
3499 regnode *end;
3500
3501 if (OP(scan) != SUSPEND) {
3502 /* set the pointer */
3503 if (OP(scan) == GOSUB) {
3504 paren = ARG(scan);
3505 RExC_recurse[ARG2L(scan)] = scan;
3506 start = RExC_open_parens[paren-1];
3507 end = RExC_close_parens[paren-1];
3508 } else {
3509 paren = 0;
f8fc2ecf 3510 start = RExC_rxi->program + 1;
8aa23a47
YO
3511 end = RExC_opend;
3512 }
3513 if (!recursed) {
3514 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3515 SAVEFREEPV(recursed);
3516 }
3517 if (!PAREN_TEST(recursed,paren+1)) {
3518 PAREN_SET(recursed,paren+1);
3519 Newx(newframe,1,scan_frame);
3520 } else {
3521 if (flags & SCF_DO_SUBSTR) {
304ee84b 3522 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3523 data->longest = &(data->longest_float);
3524 }
3525 is_inf = is_inf_internal = 1;
3526 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3527 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3528 flags &= ~SCF_DO_STCLASS;
3529 }
3530 } else {
3531 Newx(newframe,1,scan_frame);
3532 paren = stopparen;
3533 start = scan+2;
3534 end = regnext(scan);
3535 }
3536 if (newframe) {
3537 assert(start);
3538 assert(end);
3539 SAVEFREEPV(newframe);
3540 newframe->next = regnext(scan);
3541 newframe->last = last;
3542 newframe->stop = stopparen;
3543 newframe->prev = frame;
3544
3545 frame = newframe;
3546 scan = start;
3547 stopparen = paren;
3548 last = end;
3549
3550 continue;
3551 }
3552 }
3553 else if (OP(scan) == EXACT) {
3554 I32 l = STR_LEN(scan);
3555 UV uc;
3556 if (UTF) {
3557 const U8 * const s = (U8*)STRING(scan);
4b88fb76 3558 uc = utf8_to_uvchr_buf(s, s + l, NULL);
8aa23a47 3559 l = utf8_length(s, s + l);
8aa23a47
YO
3560 } else {
3561 uc = *((U8*)STRING(scan));
3562 }
3563 min += l;
3564 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3565 /* The code below prefers earlier match for fixed
3566 offset, later match for variable offset. */
3567 if (data->last_end == -1) { /* Update the start info. */
3568 data->last_start_min = data->pos_min;
3569 data->last_start_max = is_inf
3570 ? I32_MAX : data->pos_min + data->pos_delta;
b515a41d 3571 }
8aa23a47
YO
3572 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3573 if (UTF)
3574 SvUTF8_on(data->last_found);
3575 {
3576 SV * const sv = data->last_found;
3577 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3578 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3579 if (mg && mg->mg_len >= 0)
3580 mg->mg_len += utf8_length((U8*)STRING(scan),
3581 (U8*)STRING(scan)+STR_LEN(scan));
b515a41d 3582 }
8aa23a47
YO
3583 data->last_end = data->pos_min + l;
3584 data->pos_min += l; /* As in the first entry. */
3585 data->flags &= ~SF_BEFORE_EOL;
3586 }
3587 if (flags & SCF_DO_STCLASS_AND) {
3588 /* Check whether it is compatible with what we know already! */
3589 int compat = 1;
3590
54251c2e 3591
486ec47a 3592 /* If compatible, we or it in below. It is compatible if is
54251c2e
KW
3593 * in the bitmp and either 1) its bit or its fold is set, or 2)
3594 * it's for a locale. Even if there isn't unicode semantics
3595 * here, at runtime there may be because of matching against a
3596 * utf8 string, so accept a possible false positive for
3597 * latin1-range folds */
8aa23a47
YO
3598 if (uc >= 0x100 ||
3599 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))